WEBVTT
00:00:00.026 --> 00:00:10.561
Hello and welcome to a weekend news episode of the Leveraging AI podcast, the podcast that shares practical, ethical ways to leverage AI to improve efficiency, grow your business, and advance your career.
00:00:10.795 --> 00:00:16.894
This is Isar Mates, your host, We had another really interesting week, and we have many things to cover.
00:00:16.905 --> 00:00:26.207
We're going to start with one developer that replicated the entire Adobe suite and then the entire Microsoft, Office suite and then a bunch of other stuff.
00:00:26.207 --> 00:00:37.972
We're going to continue talking about software, where we're going to talk about how Claude and Gemini are impacting how we work with software, the software we use every single day, and Microsoft as well.
00:00:38.310 --> 00:00:41.296
We are going to talk about some things that are happening in the frontier.
00:00:41.465 --> 00:00:47.744
We got a new model from Meta that is almost number one right now, and it is number one on a few different aspects.
00:00:48.034 --> 00:00:55.969
and then we're going to obviously talk about AI safety, because this time, of the race, we have to talk about this because there's new things to talk about every single week.
00:00:56.256 --> 00:01:10.109
We are also going to talk about the fact that there was an engineer inside of OpenAI that installed bugging devices to record everything that's happening in the most important rooms in OpenAI and a lot of other stuff, including AI safety, including new releases.
00:01:10.340 --> 00:01:13.617
So lots and lots to cover, lots and lots to talk about.
00:01:13.618 --> 00:01:15.182
So let's get started.
00:01:17.798 --> 00:01:21.816
So the first story that we are going to discuss is Brandon Thomas.
00:01:21.862 --> 00:01:36.140
Brandon Thomas is an Atlanta-based software engineer, his company, ArtCraft, started in 2025 as developing AI tools for artists, and he pivoted at the end of September of this year to cloning Adobe.
00:01:36.411 --> 00:01:48.412
Now, he claims that the motivation was that Adobe charged him early cancellation fee one time too many, and he got really, really pissed, and then he decided he's going to just replicate everything that they're doing.
00:01:48.439 --> 00:02:07.237
So in just about two weeks, he went from,"I'm gonna copy Photoshop," to,"I'm gonna copy everything everybody's using." So he has currently seven free open source apps that are cloning Photoshop, Illustrator, Premiere, Lightroom, After Effects, InDesign, and Acrobat Pro.
00:02:07.471 --> 00:02:15.757
The names that he named them are Photo Craft, Vector Craft, Film Craft, Light Craft, Effect Craft, Design Craft, and PDF Craft.
00:02:16.155 --> 00:02:23.431
Now, the first repo went public on September 30th, so just about two and a half weeks ago.
00:02:23.459 --> 00:02:29.682
and Photocraft alone now has over forty-four thousand GitHub stars.
00:02:30.048 --> 00:02:42.062
Now, less than two weeks from the time he started, he now also has Wordcraft, Gridcraft, which is Excel, Deckcraft, which is PowerPoint, uh, cloning basically the Microsoft or Google Office suite.
00:02:42.442 --> 00:02:50.913
And Gridcraft, which is the Excel version, covers eighty-nine percent of all the functions in the ribbon in the Excel menu.
00:02:51.205 --> 00:03:05.769
Uh, that's per the developer himself And on October 7th, so just a week ago, the team added repos for Soundcraft, which is Pro Tools, and CADcraft, which is AutoCAD, and these tools run on Mac, Windows, and Linux.
00:03:05.841 --> 00:03:09.816
So regardless of where you're running, you can run it, and they're all free.
00:03:09.824 --> 00:03:10.828
You don't need an account.
00:03:10.842 --> 00:03:12.080
You don't need a subscription.
00:03:12.311 --> 00:03:18.593
You just go to the repo, and you can install these applications So how the hell is he doing this?
00:03:18.593 --> 00:03:21.491
And how the hell is he doing it so fast and so effectively?
00:03:21.491 --> 00:03:23.008
And we're gonna talk about effectively in a minute.
00:03:23.171 --> 00:03:24.239
It's not fully working yet.
00:03:24.247 --> 00:03:28.793
It's still all in alpha, and there are issues, but the speed is just astonishing.
00:03:29.062 --> 00:03:33.506
So what he's doing is he's using Opus 5.5 as the actual coding tool.
00:03:33.507 --> 00:03:43.808
He's coding it all in Rust, which is a coding, uh, language that he understands and he'd been working in personally before he started doing this And the method he's using is called Clean Room.
00:03:43.860 --> 00:03:51.893
And what Clean Room does is it basically replicates the user interface and the user interaction with the software.
00:03:51.895 --> 00:03:53.747
So he doesn't have access to the code.
00:03:53.895 --> 00:03:56.610
He's not copying the code from Adobe or from anybody else.
00:03:56.776 --> 00:04:04.171
He's just looking at the outputs very carefully, feature by feature, button by button, and allowing Claude to do this thing.
00:04:04.171 --> 00:04:14.829
Now, the other aspect that he's doing in order to make this work is he's testing the processes, he's testing the outcome and comparing it to the outcomes from the Adobe tools themselves.
00:04:15.022 --> 00:04:25.903
So in addition to replicating the user interface, he can replicate the actual detailed output by comparing it to the output every single time until he gets it exactly right.
00:04:25.930 --> 00:04:38.059
So the result of being able to run this in a loop multiple times is now you can open and resave Photoshop files or any other of the applications that he's using and render them in the same level.
00:04:38.254 --> 00:05:03.019
As an example, his test batch has 309 Photoshop files, and 307 of them, as of right now, render identical results to the results that Photoshop has generated, which basically means every use case other than you have a really edge, weird, strange, crazy, unique use case, and then maybe you will need Photoshop or maybe not So why does it matter the way he's doing the testing and the development?
00:05:03.029 --> 00:05:04.651
It means that most of it is automated.
00:05:04.663 --> 00:05:17.041
He does not need humans to test anything because all the system does is it runs it through its own capability, compares it to the output from the actual original software, in this particular case Photoshop, and then compares the two.
00:05:17.041 --> 00:05:23.714
And if it doesn't work exactly the same, it tries to understand why, and it runs the code and fixes it again until it gets it right.
00:05:23.875 --> 00:05:30.255
So all you have to do is invest in tokens, and everything else will happen magically on its own.
00:05:30.850 --> 00:05:34.747
This episode is brought to you by the Multi-Agent Orchestration Course.
00:05:34.781 --> 00:05:42.903
It's a course that I have been teaching personally since April of this year, and every single one of the cohorts that we have opened have sold out.
00:05:43.199 --> 00:05:58.947
And the people who have been through these courses made incredible transformations for themselves, switching jobs, starting their own companies and so on, and definitely for their businesses by being able to build really complex processes by following our methodology and our infrastructure settings that we teach during the course.
00:05:58.963 --> 00:06:15.442
The course is four sessions of two to two and a half hours each over four consecutive weeks, and through it, you can go from knowing a little bit about how to use AI and using it in ways that most people are using it today to learning how to build really complex, sophisticated systems that can automate more or less everything in your business.
00:06:15.711 --> 00:06:25.685
The current cohort that is open starts on November 2nd, and it's gonna go on November 2nd, November 8th, November 16th, and November 23rd, and we are almost sold out for that cohort as well.
00:06:25.973 --> 00:06:31.137
It is very likely that we will not open another cohort in December, meaning the next one after that is gonna be in January.
00:06:31.286 --> 00:06:42.947
So if you still wanna learn how to build really incredible automations and learn everything that I've learned by doing this every single day for the last few years, don't miss this opportunity to come and join us on November 2nd.
00:06:43.120 --> 00:06:52.312
There's a link in the show notes, And if you're going to use the promo code LEVERAGINGAI100, you're going to get$100 off of the price that is on the website.
00:06:52.521 --> 00:06:55.286
Like I said, there's a few seats open, but they are selling out fast.
00:06:55.416 --> 00:07:04.146
So if you're interested in doing this still this year, go and click on the link right now on your phones and complete the registration and the payment so you can guarantee your seat.
00:07:04.208 --> 00:07:06.050
I'm looking forward to seeing you there.
00:07:06.129 --> 00:07:07.843
And now back to the episode
00:07:08.713 --> 00:07:16.454
Now, to make it even more interesting, every single function in his software comes with an MCP, with a way to connect it to AI.
00:07:16.461 --> 00:07:27.028
So in addition to the fact that AI writes it, you can now allow your AI tools, any AI tool, to connect to his software and have the AI run it, meaning you don't even have to go to the user interface.
00:07:27.204 --> 00:07:57.699
You can just see the outputs of whatever graphical or whatever other thing you're working on and basically tell your favorite AI tool, whether it's ChatGPT or Claude or Grok or Gemini or whatever, and say,"Hey, I wanna fix this, add that, remove the sky, add clarity," whatever all the other things you can do in Adobe, and it will do it because it all has setups and hooks to work effectively straight out of the box with any AI tool, which obviously Adobe itself doesn't have at that level, and the stuff that they do have, again, is behind a paying firewall.
00:07:57.870 --> 00:08:06.637
So you have to pay to be able to do this, and they don't have all the functionality, and he does Now, this has caught like crazy wildfire.
00:08:06.749 --> 00:08:25.403
So he started on his own or with his company, and now Photograft, because it's an open source code, has more than 150 GitHub accounts that are already helping develop and grow the functionality, the capability, reducing the bugs of each and every one of the pieces of software that he's developing.
00:08:25.752 --> 00:08:37.437
and what's crazier is, as an example, because now there's a growing community, the team even wrote their own video and audio codecs to run behind the scenes, from public specs.
00:08:37.629 --> 00:08:59.703
So instead of using proprietary ones, they basically created new open source versions of proprietary video and audio codecs that they can run freely without having to pay or be, connected to any other initial source Now, this is riding on a completely new growing wave or tsunami of replicating software.
00:09:00.058 --> 00:09:06.894
It is aligned, even though it's not using the same thing, with another GitHub repo that is called Rea, R-E-A or Rea.
00:09:06.914 --> 00:09:11.238
I'm not sure exactly how to pronounce it, but it stands for Reverse Engineer Everything.
00:09:11.254 --> 00:09:24.520
It is, again, an open source free, uh, repository that you can go and download and use, and repository was released back in July and had more or less zero activity until October.
00:09:24.798 --> 00:09:44.554
And now in the past few weeks, it's jumped from zero, almost vertical up to three hundred thousand stars, meaning the amount of people that are using it and that are interested in it and that are participating and paying attention is growing exponentially every single day.
00:09:44.830 --> 00:09:55.312
it enables to look at a specific feature in a specific software, look on how it operates, and it learns how to replicate it on its own.
00:09:55.573 --> 00:10:09.926
Now, it's not built to replicate an entire software, but if you have the patience to go feature by feature, step by step, or you build agents that will do this for you, meaning once it finishes a specific feature, it goes and grabs the next feature and does that.
00:10:10.168 --> 00:10:19.173
And the way it works is it looks at the feature, it goes through the steps, it creates a very detailed description of everything that needs to happen, and then it goes and writes the code to do that.
00:10:19.418 --> 00:10:23.287
What other pieces of software get hit because people are using it to do different things?
00:10:23.520 --> 00:10:24.822
A big one is Sony.
00:10:25.039 --> 00:10:52.043
So Kotaku counts multiple AI-assisted blows to Sony in just two weeks, including jailbreaks to PS5, that allows you to run PS5 software emulated on a PC or allow you to run PS, PS2 discs on jailbroken PS5 or PS3 discs, uh, running on PS5 as well, which is things that Sony obviously did not intend because they want you to buy new versions of the software.
00:10:52.302 --> 00:11:04.942
and the demand, as an example, for jailbroken PS5s pushed the cost of PS5 Pro to over fifteen hundred dollars because now you can run every game you ever had, including old games, without paying Sony again.
00:11:05.302 --> 00:11:13.139
And this is showing you that this is not just in the professional software world, but any kind of software can be replicated with these tools.
00:11:13.166 --> 00:11:15.745
Now, going back to what I said as far as reliability.
00:11:16.095 --> 00:11:28.736
Everything right now is in early alpha, so PhotoCraft, again, the new for free Photoshop, had 447 open bugs as of the time of this recording, including crashing in specific, scenarios.
00:11:29.024 --> 00:11:36.706
He has first promised, the developer that within a month it will be in full parity, and now he said it will be in a few months.
00:11:36.774 --> 00:11:46.113
Still very impressive, and again, the vast, vast majority of people who use Photoshop do not need full parity in order to use it, so it is not really a big deal.
00:11:46.124 --> 00:11:55.784
You can use it right now for probably 90% of the functionality, which is the core is already there and relatively stable There's also a legal risk that he's facing.
00:11:55.784 --> 00:12:16.102
Now, Adobe cannot sue him for copying the code because he didn't, but you can be also sued for the user interface, uh, and because there's a concept called trade dress claim, meaning if you dress something to look exactly like something else, you have legal exposure.
00:12:16.235 --> 00:12:21.339
So this might be the next thing that we'll need to do is to change the user interface, but then it's not exactly like Photoshop.
00:12:21.552 --> 00:12:36.057
So there are questions there But here's the reality, and this is what it means for everyone, whether you are developing software or using software or a software company Or an IT company supporting existing software.
00:12:36.798 --> 00:12:51.657
Anything that runs on a computer right now and has a user interface can be replicated relatively easy if you're willing to invest the tokens and the cost of those tokens, even without too much technical background.
00:12:51.657 --> 00:12:59.087
But if you have a technical background like this guy, uh, you can do this even more effectively, better and faster.
00:12:59.821 --> 00:13:09.166
The other thing that you need to remember that this is the outcome of, is that the features and the user interface of whatever you're developing are worthless at this point.
00:13:09.583 --> 00:13:13.874
The things that do matter are, first and foremost, trust.
00:13:14.273 --> 00:13:20.105
Are people willing to buy from you and willing to put their operational systems connected to something that you have developed?
00:13:20.572 --> 00:13:21.072
Data.
00:13:21.211 --> 00:13:22.452
Who owns the actual data?
00:13:22.456 --> 00:13:23.241
Where does it reside?
00:13:23.251 --> 00:13:24.048
How do you access it?
00:13:24.057 --> 00:13:25.446
How secure it is, and so on.
00:13:25.946 --> 00:13:26.759
Integrations.
00:13:26.772 --> 00:13:28.721
So what does this connect to in the back end?
00:13:28.722 --> 00:13:29.774
How effective is it?
00:13:29.837 --> 00:13:37.903
Does it connect to your customers' real back ends, real sources, real other platforms in a secure way that then requires somebody to do that work?
00:13:37.904 --> 00:13:39.539
It's not just a software out of the box.
00:13:39.912 --> 00:13:41.342
And the entire ecosystem.
00:13:41.611 --> 00:13:42.956
Do you get support?
00:13:43.008 --> 00:13:44.427
Can you call customer service?
00:13:44.474 --> 00:13:47.932
Once Microsoft comes up with a new operating system, will this work?
00:13:48.136 --> 00:13:49.812
And all these kind of things.
00:13:50.062 --> 00:13:58.956
Now, the fact that many of these new things are happening in an open source environment, I think the future support and features are more or less guaranteed.
00:13:58.971 --> 00:14:07.913
Once you get to hundreds of people that are in it and helping grow it, the chances that you're not gonna get support for the new operating system or when bugs happen is significantly lower.
00:14:08.059 --> 00:14:11.980
But all the other things that I said are definitely very much there.
00:14:12.299 --> 00:14:15.823
Which means if you have a software, these are the things you need to think about.
00:14:16.251 --> 00:14:31.557
On the other hand, if you are a company, you need to start thinking on how much money are you paying for third party software, SaaS tools, and are you willing to replace that cost with internal developers that can now develop exactly what you need?
00:14:31.767 --> 00:14:46.684
Meaning instead of shackling you to the processes and the ways and the data structure of the third party tools that we all used for the last couple of decades, you can now develop a software that will work exactly the way you need.
00:14:46.779 --> 00:14:50.385
I work with several different companies who already do this.
00:14:50.578 --> 00:14:59.828
People who are developing their own ERP, their own CRM, and they're replacing their own entire tech stack step by step by step with built-in tools.
00:14:59.851 --> 00:15:02.412
Now that means they need to support it themselves.
00:15:02.413 --> 00:15:05.937
That means that they need to integrate it themselves and so on and so forth.
00:15:06.169 --> 00:15:11.928
But if you're currently spending 200, 300,$400,000 a year on SaaS, which is a small amount.
00:15:11.961 --> 00:15:28.052
Many companies are in the millions if you're big enough, and you're saying,"Instead of that, I will hire the developers to do the integrations and to build it for me, and I can have exactly the tools that I want working exactly the way I need, aligned with exactly my processes," then it might be an interesting idea.
00:15:28.078 --> 00:15:30.318
Again, there's definitely risk associated in it.
00:15:30.346 --> 00:15:36.405
But if you do this step by step slowly and have a good plan, again, I know several companies who are doing it right now.
00:15:36.527 --> 00:15:49.273
Some of them are small businesses, some of them are very large, highly complex businesses that have lots of integration and dozens of SaaS tools, and they're slowly replacing them with homegrown capabilities.
00:15:50.648 --> 00:16:09.173
Now, the last thing that I will say that is very interesting with what this guy did is the fact that he builds it agent ready, meaning he's not just building the software itself, he's building the software in a way that AI agents and AI tools will be able to use it, which will make it significantly more valuable in the next few years than just the software itself.
00:16:09.572 --> 00:16:16.905
As you know, everybody's going towards headless, meaning all you care about is the data in the back end and being able to connect to it, from external capabilities.
00:16:17.011 --> 00:16:19.976
So if that's the direction, why do I need the actual software company?
00:16:20.005 --> 00:16:22.071
And you need it for all the things I mentioned earlier.
00:16:22.403 --> 00:16:29.082
The trust, the data, the integrations, the ecosystem, the processes that are already, built into it versus just the user interface and the functionality.
00:16:29.311 --> 00:16:39.708
But there are many, many types of software out there that do not have, integrations and complex processes behind the scenes, and these are at a very high risk.
00:16:40.014 --> 00:16:48.383
I said that multiple times, long, long time ago, like when coding capabilities just became available, so we're talking about three years ago or so.
00:16:48.731 --> 00:17:06.387
I said that I think the concept of an app store will disappear, and I think app stores will disappear because these are usually smaller applications that do something very specific, and everybody will just vibe code their own solutions, and nobody will go to a third-party app store, or at least not at the scale that we know it today.
00:17:06.678 --> 00:17:36.199
And now we're seeing that this goes way beyond that because now a single developer, now with the help of a lot of other people in the open source community, can replicate any piece of software out there Now staying on the way we use software, I have some interesting news from you from Anthropic and Google and Microsoft and how they are shifting the way we are going to work with the software that we so commonly use every single day And I'm gonna connect the dots on a few things, some of them we already reported about in the last few weeks.
00:17:36.201 --> 00:17:38.035
But again, I wanna paint the bigger picture.
00:17:38.853 --> 00:17:39.961
let's start with Anthropic.
00:17:40.135 --> 00:17:44.011
Anthropic made two big announcements when it comes to using the platform that we're using.
00:17:44.241 --> 00:17:49.786
One of them is called Claude Dashboard, and the other is called Claude Motion.
00:17:50.114 --> 00:17:54.266
I again reported about this before, but I want to connect the bigger picture for you.
00:17:54.635 --> 00:18:11.104
So Claude Dashboards allows you to connect to your existing data software like data storage, like Amazon Redshift, BigQuery, ClickHouse, Databricks, Snowflakes, Salesforce, et cetera, CRM platforms, anything you want.
00:18:11.662 --> 00:18:16.980
Ask a question in simple English and get a visualization the way you want it.
00:18:17.301 --> 00:18:22.067
There are entire industries built around this functionality, right?
00:18:22.067 --> 00:18:26.615
The ability to create visualization for business data.
00:18:26.642 --> 00:18:38.613
And now all you have to do is give access to Claude to your repository, explain in simple English what you want, and it will create a live dashboard that is connected to the actual data source.
00:18:38.614 --> 00:18:40.480
This is very different than what we could have done before.
00:18:40.480 --> 00:18:45.692
So I've built dashboards for clients using for a very long time now, but these were static.
00:18:45.705 --> 00:18:51.382
You had to upload the files every single time you wanted to see a new version, which is not a huge deal, but now you don't have to do that anymore.
00:18:51.519 --> 00:18:52.191
It just does it.
00:18:52.414 --> 00:19:02.494
The other component, which is called Motion, allows it to look at data sources, whatever data sources you want, and create short animated videos to explain the data.
00:19:02.661 --> 00:19:07.872
So there are people who do not understand dashboards and/or audiences that couldn't care less about dashboards.
00:19:07.898 --> 00:19:21.170
And for them, you can just point it to the data and say,"Hey, create me an explainer video that shows why one, two, three, A, B, C," and it will create a short video that then you can send to those people to explain whatever it is that you wanna explain from the data you have access to.
00:19:21.336 --> 00:19:25.452
So again, the data more broader accessible than just to geeks who understands dashboards.
00:19:25.635 --> 00:19:45.902
and you can explain anything you want from the data sources in multiple formats, uh, straight to your audience Now Anthropic also launched full Google Workspace integration So you can connect Claude to Google Drive, Google Slides, Google Sheets, and Google Docs, making Claude an embedded operator inside the Google productivity suite.
00:19:45.940 --> 00:20:03.713
Which means, again, if you want to create, change, update, upgrade whatever Google document or output that you created previously with the Google manually, you can now do this just by asking Claude to do whatever creation or changes that you want, and it does an incredible work doing this.
00:20:03.739 --> 00:20:06.998
Again, this is going to change completely the way we work.
00:20:06.998 --> 00:20:18.554
Now, on October 8th, Google announced Gemini Agent, which is a single unified agent that is persistent across their entire cloud execution.
00:20:18.727 --> 00:20:24.558
It is connected to BigQuery, to Snowflake, to Databricks, to Postgre, for plain language data queries.
00:20:24.566 --> 00:20:25.336
Sound familiar?
00:20:25.369 --> 00:20:31.179
Basically, a very similar thing to what you can do with the new Claude solution.
00:20:31.448 --> 00:20:32.873
It knows how to generate code.
00:20:32.895 --> 00:20:36.968
It knows how to generate to fix pipelines in the data itself.
00:20:37.221 --> 00:20:46.487
It is producing either plain language reports or basically anything else you want because it can write code It has its own email address, so you can communicate with it.
00:20:46.489 --> 00:20:57.182
So you give it an actual name and then@agents.company.com, so with company being your company, and you can communicate with it via email, and it will respond with whatever it is that you ask it to do.
00:20:57.323 --> 00:20:59.740
While running behind the scenes, it will reply to the email.
00:21:00.092 --> 00:21:05.266
It runs across the entire Google Workspace and Microsoft 365.
00:21:05.601 --> 00:21:19.703
It routes tasks across multiple AI models, uh, depending on the complexity to get the best bang for your buck when it is running Now Google shared some qualified customer results by using this new agent, and some of them are absolutely mind-blowing.
00:21:19.710 --> 00:21:24.431
Again, any one of you who has been using AI tools to your productivity are not gonna get surprised.
00:21:24.699 --> 00:21:32.493
But, the things that Google shared are from a company called Breadsko, that document review went down from one hour to five minutes.
00:21:32.804 --> 00:21:40.205
They also saw a sixty percent reduction in risk of inconsistencies between different documents and different outputs that they are generating.
00:21:40.576 --> 00:21:47.974
Uh, Snap shared that troubleshooting, and they didn't share troubleshooting what, but troubleshooting went down from thirty minutes to thirty seconds.
00:21:48.151 --> 00:21:53.589
Commerzbank shared that document QA, uh, went from twenty hours to one hour.
00:21:53.792 --> 00:22:01.820
Tata Steel have used over 300 of these agents and reduced the complaint turnaround by 50%.
00:22:01.847 --> 00:22:08.618
Home Depot phone resolution has improved 4X compared to before using this, and so on and so forth.
00:22:08.632 --> 00:22:33.586
Again, really large companies that are using these agents to really impact how they do business and dramatically accelerate processes while increasing accuracy By the way, Google also announced a new computer chip that they are going to use, and that new chip has a 80% improved price performance compared to TPU 8I, which is their previous version of computers.
00:22:33.824 --> 00:23:28.363
So in addition to all the improvement on the front end that you can now do with those agents to get things faster, the actual compute is gonna cost significantly less money and will run faster in order to deliver similar results And by the way, and this all connects to what Microsoft released just a few weeks ago, and we talked about this, uh, that while they canceled Copilot for direct to consumer, basically not competing with ChatGPT for individuals anymore they have made a lot of updates to their Agent 365, which allows the agents to do more and more things inside of the Microsoft environment with Copilot managed runtime loads and being connected to all the data sources and being able to do a lot more stuff autonomously inside the Microsoft environment Now, because all of these tools that we just talked about are aiming straight at the enterprise level, they all come with their entire governance architecture built into them.
00:23:28.628 --> 00:23:38.547
The vendors are very clearly focusing on that capability, meaning the ability not just to deploy the agents, but also to control and govern the agents.
00:23:38.548 --> 00:23:47.596
So on one hand, you get autonomy, on the other hand, you get peace of mind that nothing horrific is going to happen and the wrong people are not gonna get the wrong level of access, and so on.
00:23:47.597 --> 00:24:00.227
And again, as somebody who works with multiple enterprise-level clients, this is maybe the biggest problem that exists today, is how do you connect AI and agents and AI access without breaking the current governance levels and controls?
00:24:00.405 --> 00:24:05.211
And all these tools that we just talked about have this deeply embedded on how they work.
00:24:05.828 --> 00:24:07.465
They all have fancy names for that.
00:24:07.721 --> 00:24:16.971
Gemini have, cryptographic identities and spend caps and audit trails for every agent and what they do It's based on least privilege permissions for every path that you give it.
00:24:17.167 --> 00:24:22.515
There's role-based access controls for the agents themselves and, as I mentioned, full audit trails for all of that.
00:24:22.803 --> 00:24:32.045
Claude has CMEK, admin template control, simultaneous team editing, customer-managed encryption keys, and a lot of other stuff, again, built into how the tools work.
00:24:32.211 --> 00:24:40.853
So CISOs and anybody else who's in charge of data security have a lot more control than they did just a few weeks ago, with these systems.
00:24:40.986 --> 00:24:43.116
Similar thing from Microsoft Copilot agents.
00:24:43.285 --> 00:25:14.959
They have a whole set of tools on how to control, inside of Agent, 365, how to control the access and the power and the controls for each and every one of the agents that are running in that environment So the era we are already in, even though the vast majority of companies don't apply that yet, are now in a situation where AI can safely align with your company's policy, connect to basically everything behind the scenes, especially if you're running on standardized, platforms.
00:25:14.961 --> 00:25:48.417
So if you're not on any unique setup, if you're running on what most people run on, you can now have AI either from Gemini or from Claude or from Microsoft basically do most of the work, including being connected to the raw data inside of your databases and provide you real-time information for better and faster decision-making with significantly less effort than ever before now the next step of that is obviously going to be a combination of two different things.
00:25:48.596 --> 00:25:53.428
One is specialized vertical agents that are already popping up left and right.
00:25:53.576 --> 00:26:04.034
So an agent that is really good at finance, an agent that is really good at accounting, an agent that's really good at legal, an agent that's really good on etc., etc., etc., sales, marketing, operations, customer service, and so on.
00:26:04.359 --> 00:26:16.061
And that means that every time that they do something, the full loop of feedback can happen autonomously, extremely fast, and way more efficient than humans will ever be able to do it.
00:26:16.231 --> 00:26:35.162
Because if I can see exactly the impact on what's happening across multiple channels in multiple aspects of the business in near real time, make decisions, and have an agent act accordingly on a specific aspect of the business, then the ability to the needle significantly faster and more effectively than ever before is at our fingertips.
00:26:35.162 --> 00:26:36.231
Now, are we there yet?
00:26:36.393 --> 00:26:36.701
No.
00:26:36.704 --> 00:26:43.771
There's a lot of integrations and testing and verification and so on, but all of these will happen very quickly because the agents can help you do that.
00:26:43.961 --> 00:26:54.063
So I have a feeling that the level of automation that we're going to start seeing from enterprises is going to increase dramatically, and it's not gonna happen evenly across the board.
00:26:54.063 --> 00:27:12.585
There's going to be the companies who are gonna jump headfirst and say,"We're willing to take the risk and run faster." And some of them are gonna get a slap on the wrist, some of them may get a big slap on the cheek, but they'll be able to do things in a completely different level than other companies who will take five to 10 years to go through the same transition.
00:27:12.861 --> 00:27:15.012
I see both ends of that spectrum.
00:27:15.063 --> 00:27:16.105
I got to also see it.
00:27:16.122 --> 00:27:20.810
I was at, the MAICON AI conference in Cleveland this past, week.
00:27:21.374 --> 00:27:23.502
First of all, kudos to the people who run the conference.
00:27:23.515 --> 00:27:25.019
It was absolutely, fantastic.
00:27:25.231 --> 00:27:48.295
But you got again to see some companies, CEOs, and leadership teams who said,"Yes, we're all in, and we're willing to make dramatic changes in the way we hire, the way we train, the way we develop, the way we deliver, uh, what we do." And there are many, many, many, many companies who are resisting this for multiple reasons, many of them are justified, that are just gonna move significantly slower, and the results are dramatically different.
00:27:48.465 --> 00:27:49.949
So the capability is here.
00:27:50.060 --> 00:27:51.289
We don't need to wait, for it.
00:27:51.316 --> 00:28:05.663
It's just a matter of how do you implement it across the enterprise in a safe way, and it requires making dramatic decisions to your workforce and your focus areas and the risk you're willing to take, but the rewards are extreme.
00:28:05.692 --> 00:28:08.959
Now, the other thing is cost, right?
00:28:08.970 --> 00:28:16.428
How much this thing is going to cost us because every company out there that is starting to deploy AI at scale is starting to see that the cost is higher than they intended.
00:28:16.730 --> 00:28:25.615
And for that, the race is making it, well, better for all of us because we have more companies that are at the frontier or at a good enough level for us to use their models.
00:28:25.756 --> 00:28:30.962
And we got two new models this week to help us in reducing the cost or having more options.
00:28:31.134 --> 00:28:51.502
One of them is Meta Muse Wonder So that's a new model that was released on October 12th So this model currently scores fifty-six on the artificial analysis overall intelligence index, which is equal to Claude Sonnet 5.5 and only two points behind Claude Opus 5.5.
00:28:51.733 --> 00:29:03.656
by the way, there are rumors that we're going to get Fable 5.5, sometime very soon, and the rumors are that it's going to be an incredibly capable model that will take Anthropic to a new big lead again.
00:29:03.676 --> 00:29:05.816
Well, I don't know for how long, but these are the rumors.
00:29:05.821 --> 00:29:26.329
But right now, the new Meta Muse Wonder sits basically just behind the top leading models of Anthropic with fifty-six points, three points ahead of GPT-6 Astra on max settings, and also three points ahead of Gemini for Argon on high.
00:29:26.553 --> 00:29:57.666
So went from we have a decent model to we are at the frontier with this new release It is priced at$4 for million input tokens and$20 for million output tokens So from a cost per task perspective on the artificial analysis, uh, benchmark, it is scoring$4.52 per task, which is less than Qwen 3.8 and less than the three, uh, bigger Claude models, so Sonnet, Opus, and Fable.
00:29:58.016 --> 00:30:04.653
But it is more per task than basically everything else, including, uh, Groq 4.7, GPT-6.
00:30:05.016 --> 00:30:13.851
And to put things in perspective, again, the Muse Wonder is at$4.52, and GPT-6 Astra is at$3.26.
00:30:13.877 --> 00:30:31.824
If you look at all the other benchmarks, which I don't necessarily care about, but it gives you an idea on how good this model is On DeepSUI 1.1, which is long horizon software development, it scored 79 ahead of GPT-6 Astra at 74.1 and Claude Opus 5.5 at 74.2.
00:30:32.107 --> 00:30:37.133
Again, one of the most important benchmarks that everybody cares about because it can write code and it can do a lot of things.
00:30:37.405 --> 00:30:45.142
Meta's own internal personal super intelligence index, which is something they measure, uh, which obviously is tailored to how their model works.
00:30:45.394 --> 00:30:53.738
They score 59.1, on automation bench and compared to 45.4 for both leading rivals.
00:30:53.946 --> 00:31:19.942
And this is really the main thing that Meta is claiming, is that from an agentic usage perspective, their model is far ahead because that's what it was built to do Now, in addition, Meta released Muse Spark Open, which is an open source version, not as good as the other one, uh, but still a very good and very capable open source model that is full multimodal.
00:31:19.943 --> 00:31:35.202
It has text and image and video and audio, and it, just like its bigger closed source version, is designed for coding, agentic workflows, and multimodality, including the capability to use tools, computers, web browsing, all the other good stuff that the leading models can do right now.
00:31:35.453 --> 00:31:59.308
Uh, the weights are downloadable from Amazon SageMaker, AI, Google Cloud Model Garden, Hugging Face, and everywhere else you can find, open source, models Now, both the models, the closed source and open source, have what they call active memory, which is context retention across conversations, and active focus, which is the ability to pursue specific objectives, through blockers while they're revising plans and seeking permissions, uh, before consequential action.
00:31:59.309 --> 00:32:02.999
So again, built for agentic usage and not just regular conversations.
00:32:03.612 --> 00:32:08.292
The context window is one million tokens, a little over, but doesn't matter if you round it up.
00:32:08.433 --> 00:32:40.558
So one million tokens context window, so a full frontier model from Meta and an open source model that is at the frontier of open source, all released in the same week Now, to make the story even more interesting, beyond the fact that Meta is definitely back in the frontier, and after all the mess they had with setting up their super intelligence lab and all the people that were let go and moved and jumped ship and the really big, bad situation they were in, they are back, and I will say congrats for that.
00:32:40.603 --> 00:32:42.731
I don't necessarily, anticipated that.
00:32:42.731 --> 00:32:44.920
I actually thought that's not gonna work very well for them.
00:32:45.173 --> 00:32:53.359
But Alexander Wang definitely pulled through and delivered, again, frontier capabilities at competitive pricing with very significant results.
00:32:53.605 --> 00:33:01.522
Their Muse Spark agentic helper is now the number one download in the App Store, and more and more and more people are using it.
00:33:01.803 --> 00:33:07.182
I am not using it very purposefully because I do not want Meta to have more information about me.
00:33:07.182 --> 00:33:11.846
I actually have very low trust on how they use this data, so I'm not using Muse Spark.
00:33:11.848 --> 00:33:13.461
I'm using other personal agents.
00:33:13.714 --> 00:33:21.837
But it is still, um, one person and a lot more people like it, and so it is still number one in the downloads in the App Store.
00:33:22.038 --> 00:33:32.967
And so there a lot of people are using Muse Spark as a personal companion, doing more and more things for them, and definitely a very aggressive and very successful play from Meta.
00:33:33.179 --> 00:33:42.071
To make it even more interesting, they are now partnering with Apple, and what they showed is that they can run their frontier model.
00:33:42.261 --> 00:33:47.219
It was designed to run locally on two Apple M5 Ultra Mac Studios.
00:33:47.458 --> 00:34:16.175
So that means that you can now buy home-based computers, really expensive, but still home-based computers, link the two of them via Thunderbolt cable and be able to run a top-of-the-line frontier model locally without sending your data anywhere, running it on a local device that is not a set of hardware that's gonna cost you, hundreds of thousands of dollars, but something that you can buy for thousands of dollars and run it on a local device.
00:34:16.175 --> 00:34:20.884
Now, again, with the, uh, right memory and everything, you're probably looking at twenty to thirty thousand dollars.
00:34:21.097 --> 00:34:27.570
But those twenty to thousand dollars buys you complete control of where your data is and where it is going.
00:34:27.570 --> 00:34:35.373
So a very, very interesting play by both Meta and Apple on this particular move.
00:34:35.402 --> 00:34:38.541
But that's not the only frontier model we got this week.
00:34:38.541 --> 00:34:49.507
So Gemini 4 Argon that we reported about, a couple weeks ago was only released to a closed group of companies through their Fairwind program on September 30th.
00:34:49.713 --> 00:34:54.851
However, it is now going through a global availability release since October 13.
00:34:55.081 --> 00:35:05.458
The introduction price is two million tokens for input tokens and$10 million for output tokens, which is half of Meta's and OpenAI's, pricing.
00:35:05.577 --> 00:35:06.648
Again, that doesn't really matter.
00:35:06.648 --> 00:35:14.601
What matters is how effective it from a token perspective, and it is significantly better than all the other models that we talked about previously.
00:35:14.601 --> 00:35:21.262
if we go back to the price per task comparison that I told you about before, again, to put things in perspective, I'll remind you the number.
00:35:21.467 --> 00:35:26.262
The most expensive model right now is Fable 5.1 at$7.53.
00:35:26.501 --> 00:35:31.155
Then we have, uh, the two smaller Anthropic models at$5 and change.
00:35:31.393 --> 00:35:36.217
Then we have Qwen 3.8 at$5.45.
00:35:36.242 --> 00:35:42.465
Then Meta Muse at$4.50, then Grok 4.7 at$3.74.
00:35:42.659 --> 00:35:44.273
So this gives you an idea where we are.
00:35:44.280 --> 00:35:52.657
GPT-6 is at$3.26, and the new Google Gemini model is at$1.99.
00:35:52.657 --> 00:35:57.634
So it is less than half the price of Muse to actually perform tasks.
00:35:57.929 --> 00:36:07.465
It's about a third of the price to the expensive Claude models, and it is about 30% discount, even more compared to GPT-6.
00:36:07.666 --> 00:36:26.581
So if you're looking for a frontier model that will deliver the tasks at a lower price, Gemini 4 Argon is currently the leading model to go to Now, like each and every one of the other, labs, they claim number one in multiple specific indices, and they probably are or not, I'm not 100% sure.
00:36:26.791 --> 00:36:56.097
independent evaluations are not fully aligned with what Google are saying As an example, it ranks only eighth on Arena's Agent, Arena leaderboard, but it is still a very capable model, and it is, again, from a cost perspective, it is running more efficient than all the other leading labs Google reports that internally, Argon boosted agentic code submission internally by over thirty-five percent and reduced code rollback by nearly a third.
00:36:56.306 --> 00:37:02.844
It compressed months of Claude's forty million line code base refactoring into just a few weeks.
00:37:03.086 --> 00:37:09.784
So again, a very, very capable model that gives us frontier capabilities at a much cheaper cost than we had before.
00:37:09.784 --> 00:37:11.835
Great for all of us, right?
00:37:11.835 --> 00:37:27.152
We have new frontier models that are pushing the boundaries of what's possible at a lower and lower rate, which again, from our perspective, that is very helpful Now, another interesting thing that's happening when it comes to using different models is the skyrocketing success of Jev.
00:37:27.157 --> 00:37:29.940
So those of you who haven't heard about Jev, we talked about this in the past few weeks.
00:37:30.278 --> 00:37:34.543
It is basically a model that is not a regular chat.
00:37:34.920 --> 00:37:37.271
All it does is it is a classifier.
00:37:37.427 --> 00:37:39.875
There are three types of decisions it can make.
00:37:40.072 --> 00:37:42.280
One is called null, which is a yes/no.
00:37:42.302 --> 00:37:48.050
The other one is choice, which allows you to select something from a list based on criteria, and the third one is a score.
00:37:48.061 --> 00:37:52.581
It gives a customizable numeric score for different options of things you wanna do.
00:37:52.773 --> 00:37:59.257
You can train it on your own data, concepts, ideas, factors, et cetera, and that's all it does.
00:37:59.291 --> 00:38:29.672
It classifies things as you give it the inputs To put things in perspective, the input tokens are now four cents point two for one million input tokens, and the outputs are free because the output is usually very, very short, and it is 200 times faster than running this query through just a regular old school large language model Now, if you think I'm like,"What do I need this for?" everything you need to classify and you wanna do it quickly and cheap as decision-making in whatever process you're doing.
00:38:29.983 --> 00:38:57.257
And it reached one million users on Vercel within just a few days of its launch And Typeface, which is the company behind it, are claiming that it was adopted by approximately one-third of the Fortune 500 companies already To prove that this is a relevant market and that it has a future, uh, they just raised$870 million, and led by Andreessen Horowitz and Sequoia and, like, the big hitters with a$7.5 billion, valuation.
00:38:57.579 --> 00:39:00.123
And so there's definitely a future to that.
00:39:00.311 --> 00:39:11.050
Uh, we reported that in the latest release from OpenAI, they have released in their API an API component that does the same thing, basically a classifier.
00:39:11.050 --> 00:39:18.742
So now they're not the only players in the market that do this, but they're definitely doing this very effectively, which again, for all of us is a great thing because it gives us more options.
00:39:18.768 --> 00:39:33.050
So the bottom line, we are back with a five-way race between OpenAI, Anthropic, Meta, SpaceX AI, which now changed their name to SpaceX SI based on the ridiculous government, paper from, last week.
00:39:33.275 --> 00:39:40.822
But it doesn't matter what the name is, they are there as well And Gemini with their new model are in that bunch already.
00:39:40.824 --> 00:40:02.300
So we have a solid race with a lot of big participants that are giving us options that are gonna make it cheaper and scarier, and with a lot more options and integrations, as we've seen in the previous segment, how to use AI effectively inside a company and in our personal lives But there is a other side to this story, right?
00:40:02.362 --> 00:40:11.166
As we get these more advanced models, their ability to generate damage and do havoc in everything we know today is also increasing.
00:40:11.501 --> 00:40:23.396
And there are, even this week, more and more scary reports or somewhat scary reports on what might happen and what actions companies are actually taking, to remediate the situation.
00:40:23.423 --> 00:40:25.847
So the first part of that story comes from Bank of America.
00:40:26.086 --> 00:41:11.257
they have deployed 1,000 new third-party contractors plus 3,000 internal staff to patch software flaws exposed by Anthropic Mythos when it was initially released to them, as part of their Glasswing project So if you think about how much does it cost to have 4,000 top engineers run this, we're talking about millions of dollars that have been invested and are still invested by Bank of America in order to patch software issues that one new model was able to identify Now, on the flip side, the CEO of Bank of America is saying that they currently have 130 internal AI projects that are projected to generate$300 million in savings.
00:41:11.480 --> 00:41:30.436
So from a pure mathematical perspective, from a pure ROI perspective, at least based on what we know right now, and assuming they'll be able to patch everything well and not have to face any really big issue, then the AI era is actually gonna be very helpful to Bank of America because they'll be able to save$300 million.
00:41:30.436 --> 00:41:37.309
They might be able to grow their revenue by providing more capabilities and resources to clients and attracting them instead of other banks.
00:41:37.521 --> 00:41:47.717
And they're investing a few millions in single digits in actually patching the software issues that Mythos was able to help them find And again, this is just one example.
00:41:47.907 --> 00:42:15.577
I'm 100% certain that other banks and other leading institutions are doing the same Another interesting example of what these AI agents can do was done in a test by the UK AI Security Institute, AISI, that confirmed that during a control cyber evaluation, Mythos-5, which is the same family of models that again, Bank of America just used in part of Project Glasswing to patch security, issues.
00:42:15.605 --> 00:42:28.603
Anthropic Mythos-5 model created multiple fake online identities and used them to socially engineer a real GitHub maintainer into approving malicious code changes to an open source project.
00:42:28.882 --> 00:42:35.159
So again, if you think about what that means, open source repositories are managed by specific people who hold the keys.
00:42:35.389 --> 00:42:39.240
These people are the only people who can allow different changes to the repository.
00:42:39.523 --> 00:42:43.222
And that means you need to convince a human to apply your changes.
00:42:43.226 --> 00:42:56.652
But if you have multiple other humans that are telling you that this is all perfectly fine and that they've tested it and reviewed it, and you assume these are legitimate developers inside your repository just like many others, then it is a lot easier to convince the human to make the changes.
00:42:56.934 --> 00:43:05.894
And that was done by Mythos on its own, not being directed to do so, in order to be able to plant its code into an open repository.
00:43:05.905 --> 00:43:19.384
That tells you how dangerous these models are, not just by their ability to identify vulnerabilities, but also about their abilities to make humans take actions on their behalf by social engineering them.
00:43:19.588 --> 00:43:29.492
You grow that at scale, and you go back to the same interview that I talked about many times on this show with Yuval Noah Harari, the guy who is a famous author.
00:43:29.492 --> 00:43:35.677
And in this interview, he talks about the fact that we were-- again, he did this in about two years ago.
00:43:35.909 --> 00:43:44.873
We were a adult human talking to the AI that was a five-year-old, and we can manipulate and make it do whatever we want it to do.
00:43:45.061 --> 00:43:47.965
And sometime in the very near future, it is going to be the other way around.
00:43:48.211 --> 00:43:59.697
We're going to be a young child with our intelligence compared to the AI that is gonna be way more intelligent than us and will be able to manipulate us to believe, think, and act the way it wants.
00:43:59.820 --> 00:44:05.202
And based on what he's saying, there is no workaround for that just because of the sheer difference in intelligence.
00:44:05.313 --> 00:44:23.898
this is a small scale experiment that proves how easily AI can do that another incident that was revealed this week, is that Researchers at a hackathon discovered 2,600 attempts of data extraction from AMAP, which is China's version of Google Maps, uh, over an 11 days window.
00:44:24.197 --> 00:45:10.201
The way they were trying to detect it is they were looking for unusual traffic patterns, and they were able to find that there's a swarm of agents that were attacking this map service in order to try to get data off of it After doing additional research, they found it was about 17 different agents that were all coming from Tencent's cloud servers in Hong Kong that was using different proxies to reroute and disguise their actions in order to try to get access to that data So the bots weren't only trying to access the information, they were trying to hide the way that they were trying to access the information They did a lot of other things like routing their requests through translation services and things that were really advanced ways to hide, A, who they are, what they are, and what their intentions are.
00:45:10.358 --> 00:45:40.766
And this was found not by the company itself, but actually by third-party group that were running a hackathon trying to find interesting swarms of agents roaming through the internet Now to make it even scarier, weirder, call it whatever you wanna call it, the researchers were able to, quote unquote,"communicate with the agents." And the way they did this is they embedded invisible text inside the preliminary report, directing the agents to a fake AI service with computer use, and then they started communicating with those agents.
00:45:40.920 --> 00:45:44.554
The agents claimed to be Claude models, even though they were not.
00:45:44.793 --> 00:45:54.268
The agents demonstrated awareness of the investigation, by referencing the term, fleet in webhook messages before researchers published their findings.
00:45:54.280 --> 00:46:04.628
So these agents were tricked into sharing information, and they were trying to backpedal from that situation in order to not to get caught and not to expose their identity.
00:46:04.628 --> 00:46:42.472
This is happening right now across the board in the live internet, and the only reason this got caught is because a bunch of people running a hackathon and was doing an experiment to see if they can find and identify these kind of situations Now, if you think the Tencent, uh, story is okay, it's in China, there's less control, It is now confirmed that this summer, OpenAI agents have infiltrated US Commerce Department and Security and Exchange Commission Also did a ReadMe hack on the Department of Education website, and OpenAI confirmed the SEC and Commerce Department incident directly, but stated that it was looking into the Education Department finding.
00:46:42.472 --> 00:46:56.523
So basically, they confirmed two of the three incidents actually happened by their agents So this is a US-based company agents that went out of control and infiltrated US government data repositories without authorization.
00:46:56.797 --> 00:46:58.856
So this is happening everywhere.
00:46:58.989 --> 00:47:01.248
It is going to happen more and more and more.
00:47:01.605 --> 00:47:14.237
I think the chances of everybody being able to patch all the vulnerabilities that exist today faster than the agents can find them, use them, or find new ways to overcome the new patches are very, very low.
00:47:14.523 --> 00:47:28.431
And so while on one hand I'm excited about these new capabilities and the competition and new models and lower rates and so on, the risk that is coming with it is growing exponentially every time one of these models is released.
00:47:28.639 --> 00:47:32.653
And again, all the incidents we know so far were all mistakes.
00:47:32.860 --> 00:47:38.152
They all happened randomly by agents trying to do one thing and in the process doing another thing.
00:47:38.416 --> 00:47:50.827
But I can guarantee you that there are, A, governments and, B, organizations that are building way more sophisticated, way more advanced capabilities to infiltrate existing systems and are probably doing it successfully.
00:47:51.012 --> 00:48:21.260
And we'll either never find out that this data is getting stolen or manipulated or used, or we will find out when it's too late and something really, really bad happens Now, I'm not the only person that thinks this way In another open letter, and we got a lot of these in the past few months, but another open letter that was released on October 13th, approximately 400 scientists from 34 countries, including, some really known people like Geoffrey Hinton and Yoshua Bengio, who are named, as you know, the, Turing Award winners and the godfathers of AI.
00:48:21.541 --> 00:48:32.851
They released a letter that says three things: we need mandatory high assurance containment, including air gapping, frontier models for external networks during development and testing.
00:48:33.034 --> 00:48:34.007
So this is number one.
00:48:34.021 --> 00:48:40.552
Number two, creating an international oversight body with authority to set and enforce containment standards.
00:48:40.704 --> 00:48:58.411
And number three, public disclosure of all security containment failure The letter specifically, states that voluntary industry commitments, including the White House Accord, uh, signed just a couple of weeks ago by all the leading, labs, are insufficient without binding enforcement.
00:48:58.913 --> 00:49:01.364
I could not agree more.
00:49:01.402 --> 00:49:03.117
I've been saying this time and time again.
00:49:03.126 --> 00:49:17.652
There has to be an international body that supervises these tools and that makes sure that there's a much tighter security level around everything that they are doing or can do or how they're tested or how they're deployed and so on.
00:49:18.047 --> 00:49:22.016
And I hope somebody will take action based on that open letter.
00:49:22.018 --> 00:49:28.155
That being said, none of the previous open letters drove any real action, so I don't really know how this one is going to be different.
00:49:28.181 --> 00:49:34.141
Now, another person that took the stage to go in that direction was Satya Nadella, the CEO of Microsoft.
00:49:34.157 --> 00:49:50.844
On October 10th, he published a detailed safety protocol calling for all OpenAI system to build around principles of observability, And he also talked about continuous tensing, independent controls, containment and incident disclosure, so very much aligned with the letter.
00:49:50.958 --> 00:50:14.007
And he wrote, and I'm quoting one sentence,"We need to surround non-deterministic models with strong deterministic system design, human controls, and reliable operating procedures, and establish industry standards where existing ones are insufficient." And he also added another sentence that I really liked:"The most trustworthy super intelligent system will not be the one with the model we trust the most.
00:50:14.057 --> 00:50:27.253
It will be the one that enables us to trust the model the least." Basically, we don't trust the models, and we have the right measures in place in order to control them for sure, regardless of what they do So what does this all mean?
00:50:27.438 --> 00:50:35.820
It means, and I'm summarizing, guys, the entire segment so far, from the beginning of this episode, we are getting better and better models.
00:50:35.938 --> 00:50:40.827
They are going to be fully embedded into everything we do, and they already are if you allow them.
00:50:40.827 --> 00:50:46.554
We talked about this in the beginning, like all the tools that integrate into our data sets and can take action and do things for us.
00:50:46.746 --> 00:50:50.898
It is going to replace traditional software in ways we cannot even imagine yet.
00:50:50.945 --> 00:51:43.492
And on the other hand, it opens us up for a crazy amount of risk from just agents going rogue, but more importantly, from bad actors who will take advantage of these really advanced tools to do whatever it is that they want to do, whether it's entire nations or specific business opportunities that could be done either quasi illegally or completely illegally, but drive a lot of revenue to the people who are going to do that And so while more and more people are raising red flags on ways to stop this, very little action is actually taken so far, and I'm personally very worried about the current situation Now to rapid fire items, and some of them could have been their own deep dive stories, but I have to pick and choose An OpenAI engineer named Abdulrahman Alfosan, and I'm not sure I'm pronouncing his name correctly, so I apologize if I am.
00:51:43.646 --> 00:52:26.202
He's a 35-year-old Saudi Arabian national and US permanent resident, uh, was arrested on October ninth, twenty twenty-six after allegedly concealing audio recording devices in the company's San Francisco office, including one disguised as a toilet paper behind a monitor in the executive-only conference room Now the weird and not so professional stuff, so if you're operated by a foreign government to plant uh, surveillance devices, he apparently used ChatGPT, so the company's own chatbot tool, to plan how he's gonna plant and target specific sensitive discussions inside the company Now, I'm not gonna dive into all the details of what he did or what devices he planted and where.
00:52:26.422 --> 00:52:29.081
To me, there are two very important takes from this.
00:52:29.157 --> 00:52:38.498
One, there is definitely bad actors and bad players who have infiltrated the deepest places in the leading labs in the US.
00:52:38.550 --> 00:52:40.898
There is zero doubt in my mind that's the situation.
00:52:41.282 --> 00:52:54.081
The other thing, it means that the race is not a completely fair race, and there's a whole segment this week by Anthropic as far as how they've done a lot of effort and a lot of work to stop distillation of the cloud models by Chinese companies.
00:52:54.081 --> 00:53:02.302
They're claiming that their efforts were able to suppress 99% of these distilling efforts, uh, but that required a lot of effort on their behalf.
00:53:02.302 --> 00:53:04.465
And again, I'm sure other models are going through the same thing.
00:53:04.467 --> 00:53:09.264
And again, now we're talking about human infiltration, into the holy of the holies, right?
00:53:09.264 --> 00:53:11.182
The executive boardroom of OpenAI.
00:53:11.436 --> 00:53:14.514
And again, this guy got caught, but he may not be the only one.
00:53:14.521 --> 00:53:25.952
But the third aspect of this that raises a very big question in my mind is: How the hell does OpenAI know what happened in the private conversations of that person with ChatGPT?
00:53:26.125 --> 00:53:34.528
And yes, he's a freaking idiot for using ChatGPT to figure out how to plant devices inside of OpenAI's offices.
00:53:34.742 --> 00:53:45.869
But it means that OpenAI can monitor and go back to everything that was said in every conversation inside of ChatGPT in order to find such information.
00:53:46.068 --> 00:53:48.864
And it is used in court in this particular case.
00:53:49.235 --> 00:54:08.693
Which tells you that if you are thinking of doing something that might be questionable, do not use the AI models in order to figure out how to do this, because it will be found because the data is stored and is searchable somehow, and it is usable in courts.
00:54:08.702 --> 00:54:09.851
So just food for thought.
00:54:09.853 --> 00:54:11.963
Again, I don't plan to do anything illegal.
00:54:11.963 --> 00:54:13.452
I've never done anything illegal in my life.
00:54:13.632 --> 00:54:22.891
But you're planning something like this and said,"Ooh, Claude or ChatGPT or Gemini are way smarter than me, and they can help me figure this out," just be aware that your data is going somewhere that can be traced.
00:54:23.056 --> 00:54:28.456
And the same thing is true if you're doing things that are not illegal, but you want them to stay completely secret.
00:54:28.675 --> 00:54:40.690
That information that you're putting into ChatGPT and Claude and so on can be found, can be traced, and can be used in courts or probably other places against you if you get into a situation.
00:54:40.717 --> 00:54:47.782
Staying on OpenAI and their issues, they have publicly defended its decision to terminate the three safety researchers who left.
00:54:47.782 --> 00:54:58.918
We reported about this two weeks ago, uh, stating that their dismissal was due to a, and I'm quoting,"significant breach of trust," uh, rather their expressed AI safety concerns, which is what these three people said.
00:54:59.152 --> 00:55:04.961
I obviously don't know who is right and who is wrong, or where there's a big gray area between the two, but this is the current situation.
00:55:04.963 --> 00:55:55.961
OpenAI is obviously denying that they were let go because of their safety concerns, but rather because they shared safety information and other information with people outside of OpenAI Staying with OpenAI and legal/security issues, uh, As part of their legal battle with Apple, they are suing Apple back and that they're claiming that unnamed former employee transferred over 100 confidential files related to its unreleased hardware to people and accounts related to Apple So the specifics is that this employee was an Apple employee that came to work for OpenAI and then moved back to Apple after stealing 100 highly confidential files about the hardware that they have been developing.
00:55:56.103 --> 00:56:07.210
Again, this is exactly the reverse allegation of what Apple is saying about OpenAI, that they have paid employees to jump ship and send them a lot of information about different hardware that is developed by Apple.
00:56:07.367 --> 00:56:08.210
who is right and wrong?
00:56:08.309 --> 00:56:08.971
I don't really know.
00:56:08.971 --> 00:56:10.155
The courts will decide that.
00:56:10.175 --> 00:56:23.594
I will keep you posted, but it is very, very clear that all these companies are doing different kinds of espionage, trying to understand what the other ones are doing, because billions and billions of dollars are at stake Now jumping to talk more specifically about, safety.
00:56:23.594 --> 00:56:40.775
According to a research from a group called Frontier Red Team, uh, small AI models could achieve autonomous self-replication on consumer IoT devices like smart TV cameras and stuff like that, ovens, anything that has computing capabilities, uh, by the end of this year.
00:56:41.081 --> 00:56:55.994
Meaning they can very quickly and uncontrollably spread across multiple devices, across multiple networks, and then decide what they wanna do as they wanna do it, because they can communicate with one another through IoT communication.
00:56:55.994 --> 00:57:49.838
The team demonstrated that a twenty-six billion parameter, models can already orchestrate full replication chains, so basically create new variations of them across networks and access different devices that don't have the right protection, such as video cameras and home devices, so IoT things that run like smart devices in our homes, Now, they also reported that the current replication speeds remain relatively slow, but they think they can double very, very quickly and then double again, and then double again as these models figure out how to do this The actual test was done with a GLM 5.2 full model that was able to create small$9 billion parameters models, that established interface on eight of the nine hardware types that were tested Now, why is that a big deal and could be potentially really scary is because using these devices to plant malware is not a new thing.
00:57:50.130 --> 00:57:51.400
people have been doing this for a while.
00:57:51.402 --> 00:58:09.592
The last time the East Coast internet almost, crumbled, and it did for a while until it got fixed, uh, a few years ago, was because of a, vulnerability that was used to plant a lot of devices, a lot of pieces of software, to ping the main servers on the East Coast, that was planted in, again, cheap Chinese surveillance cameras.
00:58:09.943 --> 00:58:16.800
But these things, like old school malware, have a very specific goal and task, and they can be activated whenever.
00:58:17.047 --> 00:58:18.809
An AI can be a lot more flexible.
00:58:18.831 --> 00:58:20.224
It can decide to do a lot more things.
00:58:20.224 --> 00:58:29.552
It can react in real time to the defense mechanisms trying to stop it, and so on, because it is a thinking entity that is embedded in each and every one of these devices.
00:58:29.552 --> 00:58:42.858
So this could lead to a whole new level of large scale attacks that could be extremely sophisticated and completely not controlled because the AI can replicate itself and deploy itself across devices that has zero protection at this point.
00:58:42.885 --> 00:58:57.385
Staying on the safety topic, as I mentioned, Anthropic released a whole paper on how they have released the distillation attack by ninety to ninety-nine percent with new AI model defenses.
00:58:57.614 --> 00:58:59.065
So this is one thing from Anthropic.
00:58:59.065 --> 00:59:17.034
Another thing that Anthropic released this week is they published a detailed alignment report on October ninth where it is documenting instances of Claude models circumvented security controls, exploited software flaws, and worked around tools and restrictions, uh, during internal evaluations and use.
00:59:17.318 --> 00:59:27.476
Uh, they have identified four categories of unintended behaviors, including SQL injection attacks, unauthorized form submissions, uh, URL shortening workarounds.
00:59:27.884 --> 00:59:36.614
But they've emphasized that all these incidents had minimal real-world impact, uh, and they are less severe than cybersecurity incident that earlier.
00:59:36.849 --> 00:59:40.871
so first of all, kudos for Anthropic for doing this kind of research.
00:59:40.893 --> 00:59:44.677
Kudos for Anthropic for releasing, uh, this information for everybody to follow.
00:59:45.092 --> 00:59:51.025
As I mentioned before, and as the new open letter suggests, I think we need a lot more than Anthropic policing themselves.
00:59:51.025 --> 00:59:59.199
We need a much tighter controls from third-party, uh, groups doing the same thing for every single lab that are developing capabilities.
00:59:59.219 --> 01:00:28.710
Now, to be fair Anthropic has followed its promise, and they are providing employee-like access to six employees from MITRE to, uh, which is a nonprofit AI safety evaluator, allowing them to access everything the Anthropic employees can access to give them deep insights to the tools, the processes, the systems, the training, procedures, and be able to do independent risk assessment, which is something Anthropic said they will do, less than a month ago, and they're actually now doing it.
01:00:28.882 --> 01:00:32.637
So again, kudos to Anthropic and good luck to MITRE in doing this.
01:00:32.661 --> 01:00:35.498
I really hope everybody else will follow that mechanism.
01:00:35.518 --> 01:00:43.804
I do think that all these companies need to get together with MITRE and similar companies and governments, just like the open letter suggests, and build standards for this.
01:00:43.842 --> 01:00:55.125
So again, we have standards on what can go in milk and kids' toys and, uh, basically everything else that we engage with, uh, in our house and, you know, power switches on our walls and et cetera, et cetera.
01:00:55.139 --> 01:01:04.016
Every one of these things have safety standards, and potentially the most advanced, scariest, most capable technology we ever developed currently has zero safety standards.
01:01:04.188 --> 01:01:18.688
And so developing these quickly and applying it quickly is something we definitely need Staying on Anthropic and things they're doing, they just did a whole comprehensive, update on its election integrity measures for Claude ahead of the elections, the midterm elections.
01:01:18.976 --> 01:01:41.659
Uh, the idea is to pay close attention to how people are using Claude to potentially impact, the elections They revealed as an example that political conversations represented only 5.75% of all US Claude AI usage, and direct voting recommendations requests appeared in only 0.01% of conversations.
01:01:41.659 --> 01:01:53.472
And they're now deploying new classifiers to detect, uh, multi-step election abuse to make sure that their models are not being used either intentionally or non-intentionally to impact the outcome of the elections.
01:01:53.661 --> 01:02:17.992
I really hope that all the other labs are doing the same thing Another surprising person that related to the risks coming out of AI is the chief of the International Monetary Fund, IMF, Kristalina Georgieva who cautioned on October 16th that the risks associated with artificial intelligence could surpass its opportunities if adequate safeguards and ethical frameworks are not swiftly, implemented.
01:02:18.516 --> 01:02:28.909
She spoke at the annual IMF and World Bank meeting in Bangkok, and she said the following If there are no safeguards, then the risks may exceed the opportunities.
01:02:29.018 --> 01:02:46.360
And above all, we do need an ethical foundation, and I'm not sure we are there as a society Another person that related to the lack of real controls was actually Jack Clark, who's one of the co-founder of Anthropic, and he said,"I think it's worse than you describe.
01:02:46.431 --> 01:02:52.056
I think it's Formula 1 at night without maybe brakes in the car.
01:02:52.543 --> 01:03:09.646
And it's remarkable to me that the toys that I buy for my children are regulated than the literal super intelligence I work on at my day job." So closing this segment about AI safety, we are in a very delicate situation right now where on one hand we're getting models being released left and right.
01:03:09.664 --> 01:03:20.960
On the other hand, it is very, very clear that we do not have the capabilities, the tools, the processes, the safeguards in place to control them and to make sure that using them is safe for society.
01:03:21.257 --> 01:03:32.744
And it is something that everybody who looks into it understand that it's the situation, and yet from a government action perspective, from international, approach perspective, very little is actually being done.
01:03:32.938 --> 01:03:51.311
There are more and more discussions and awareness, which is great, but you hear these statements from really important people, and yet they are taking very little action to actually move in that direction And I actually thought about this deeply, and I said,"How can you enforce that?" And I think there are several relatively simple ways.
01:03:51.317 --> 01:03:56.847
The easiest one is to put laws in place on how the funding has to be invested.
01:03:57.092 --> 01:04:18.376
So if you raise funds for a particular AI initiative, you have to use X percent of it to do exactly the following things when it comes to AI safety, uh, including research, including guardrails, including, airtight environments for testing, including all these things, or else you cannot get the funds and you cannot use them or something like that.
01:04:18.619 --> 01:04:33.905
This will force the lifeline, the very, very expensive lifeline of these companies to use some of the funds towards the things that have to happen, and maybe some of it has to force them to participate in international bodies and set standards and so on.
01:04:34.101 --> 01:04:45.275
Again, I don't know how to solve it, but this is just an idea, uh, that I had this week when I was talking about it to some other people Now to a few interesting releases or potential releases, uh, that happened this week.
01:04:45.307 --> 01:04:53.311
An unconfirmed reference to a new Claude Fable 5.5 model has been reportedly appearing in Microsoft Azure, configuration files this week.
01:04:53.661 --> 01:04:55.572
Now, does this mean it's actually going live?
01:04:55.782 --> 01:04:56.516
Not sure.
01:04:56.751 --> 01:05:05.442
But as of October 13th, Anthropic's documentation still lists Fable 5.1, uh, that was released on September 1st, as the most advanced model.
01:05:05.594 --> 01:05:09.976
they have released, uh, the 5.5 family of models, including Opus, Sonnet, and Haiku.
01:05:10.132 --> 01:05:20.717
So the logical thing, and again, this reference on Azure says that we're gonna get Fable 5.5, which might be very significant if they do Now another interesting news from Anthropic.
01:05:20.768 --> 01:05:34.711
they have revealed that Claude code writing capabilities have become so quick, fast, and efficient that they are merging 30,000 code changes per day by the end of September.
01:05:35.364 --> 01:05:38.920
This is up from 2,500 in May of this year.
01:05:39.403 --> 01:05:53.052
So a more than a 12X growth happened in less than six months, which means the traditional infrastructure such as GitHub has become a bottleneck for their ability to deploy software.
01:05:53.239 --> 01:06:06.472
And to solve this, they've created a new team, a seven people person that has built what they call Scion, which is a custom source control system that is designed for agent scale development and deployment of new code.
01:06:06.630 --> 01:06:17.059
And they're now switching from GitHub to Scion as the way to manage, deploy, and control the code that the AI is generating right now So are we at recursive self-improvement?
01:06:17.347 --> 01:06:18.103
Not yet.
01:06:18.268 --> 01:06:20.579
Are we getting very close to that?
01:06:20.693 --> 01:06:22.608
Are we getting closer to that every single day?
01:06:22.800 --> 01:06:24.543
100%.
01:06:24.775 --> 01:06:35.528
When you can 10X or 12X the code that you're generating, and now you are removing the next bottleneck that was giving you issues with doing it even faster, it is very clear what's happening.
01:06:35.539 --> 01:06:48.288
Like Claude is generating code that Claude is actually checking and testing because there's no enough humans in the world to check all this code, and then that will generate a better version of Claude that will be able to do this faster, and so on and so forth.
01:06:48.463 --> 01:07:08.215
It's still not the research side of things, but as Anthropic themselves are saying, and everybody else, it is just the next step Another interesting release this week is Meta and OpenAI have announced the Personal Agent Protocol which is a draft standard that enables AI agents to securely interact with companies on behalf of customers.
01:07:08.423 --> 01:07:18.405
The protocol has already attracted 42 design partners, including Bank of America, PayPal, Shopify, Stripe, Walmart, Visa, and a lot of other large customers.
01:07:18.652 --> 01:07:42.307
The idea is to build a way that AI agents can safely and securely communicate and represent humans, with actual companies out there in the wild Now, the problem it's coming to solve is that currently agents must sign in as customers in order to navigate the human-focused web pages, and there's no standardized way for companies to guide an agent to do the things that it's supposed to do.
01:07:42.454 --> 01:07:46.173
So you have to log in as the agent, as you probably know.
01:07:46.175 --> 01:07:59.373
So if you wanna use, uh, Muse or GroqBot or any one of these tools, or Instinct or Dots, uh, and allow it to do things on your behalf, you have to log in yourself as you, and then it navigates the website as a human versus saying,"Well, wait a minute.
01:07:59.391 --> 01:08:18.944
There's a whole new type of interface that we can create for the agents themselves that will allow them to access safely and securely specific data and so on, and be able to navigate in a much more efficient way than just using the web pages as we're doing right now." And this is what this protocol is coming to solve Now, the way the protocol works is it's not replacing anything.
01:08:18.952 --> 01:08:28.073
It is building on existing protocols that exist today, such as OAuth and JSON and HTTPS and OpenAPI and MCP.
01:08:28.076 --> 01:08:32.792
So all the tools that exist today, so companies who want to implement it don't have to reinvent everything.
01:08:32.792 --> 01:08:40.871
They just have to adjust and build on top of integrations and platforms they already have Another interesting announcement this week comes from Amazon.
01:08:40.913 --> 01:09:04.069
Amazon has introduced Bee, like the animal, B-double E, Bee Frontier, uh, which is a new AI wearable, and Bee Eclipse, which is a magnetic phone attachment which is expanding its personal AI device lineup These devices are designed to capture and process conversations in real time, generating useful summaries, insights, and reminders, uh, to help users stay organized and track everything that they're doing.
01:09:04.301 --> 01:09:11.234
Again, what I'm doing right now with every Zoom call or Teams calls that I'm on, it can do for every conversation that you have along the entire day.
01:09:11.578 --> 01:09:32.900
I said that multiple times we're walking with open eyes or blindly, I don't know, depending on who you ask, into an era where everything you do if you're in a public space is going to be recorded later on, most likely with video as well, and analyzed in real time by multiple companies, countries, tools that you don't really know, trust, and whatever.
01:09:33.123 --> 01:09:38.814
And the whole concept of privacy in public areas is gonna go down to zero.
01:09:39.051 --> 01:09:51.658
And again, that opens the gates for a lot of really weird, scary, situations that there are no laws protecting us against right now, there are no technologies us against it right now, and so on.
01:09:51.966 --> 01:09:58.109
And this is a whole new area that will have to be addressed as these devices become better and more and more commonly used.
01:09:58.136 --> 01:10:14.086
switching from Amazon to Microsoft, Microsoft has launched legal skills within Copilot So they're calling it Copilot for Legal, and it's a suite of over 20 purpose-built legal skills integrated into, their Copilot existing platform This is a field that has been growing dramatically.
01:10:14.087 --> 01:10:20.724
The whole AI for legal field has been growing dramatically with huge investments and more and more usage across multiple companies.
01:10:21.061 --> 01:10:23.854
It does allow law firms to do things a lot faster.
01:10:24.027 --> 01:10:31.100
What does that mean for the future of structure of law firms, such as why do you need paralegals anymore if all their work can be done by AI agents?
01:10:31.336 --> 01:10:41.184
And why do you need new lawyers, entry-level lawyers, if all the stuff that junior lawyers have done, can be done by AI very, very quickly?
01:10:41.292 --> 01:10:49.783
you will not have partners in the law firm over the long run if you don't have entry-level people, and the same is true in every single industry.
01:10:49.962 --> 01:11:10.546
And that's a big question mark above the future of work and how that can look like, uh, moving forward Another announcement this week, xAI, also known as SpaceX AI, also known as SpaceXSI, has just launched Grok Voice 3.0, which they're claiming is the world's most intelligent and capable speech-to-speech AI model.
01:11:10.841 --> 01:11:17.845
Uh, we shared similar things from Microsoft in the past few weeks and from OpenAI, so they all released new versions of voice models.
01:11:18.164 --> 01:12:04.252
All these models are getting significantly faster and significantly smarter and are getting very much indistinguishable from talking to actual humans Now stepping away from new announcements, an interesting thing that happened this week is, Character AI has initiated a voluntary Chapter 11 bankruptcy process in order to facilitate a court-supervised sale of their assets Several different companies have already placed a bid on the assets of the company Now, if you remember, what is left of Character AI is just a shell after Google in August of 2024 paid$2.7 billion for non-exclusive license to their, uh, large language model technology and rehired its founders in what was described as a reverse acqui-hire.
01:12:04.418 --> 01:12:12.201
So the shell that was left with the technology without the people and the leadership is what we're talking about right now, and they are now filing for bankruptcy.
01:12:12.230 --> 01:12:19.951
And again, there's people bidding numbers like$20 million, which in the new AI era is basically,"Oh, there's a free technology out there.
01:12:19.952 --> 01:12:38.997
Let's go and grab it." And the last topic I wanna talk about, and I don't usually like to end on a negative note, but it is something that I think that is going to happen more and more in more and more industries, and I think, sadly, that at this point it's more or less inevitable, is job losses due to AI.
01:12:39.030 --> 01:12:41.426
And the story this week actually comes from Expedia.
01:12:41.690 --> 01:12:57.048
So Expedia Group has reported that it's going to lay off 58 corporate employees, including data scientists, vice presidents, and so on at its Seattle headquarters, and it's going to take place between November 21st and December 1st of this year Now, they're not the only ones in the industry.
01:12:57.048 --> 01:12:58.835
It's an industry-wide thing that's happening.
01:12:59.014 --> 01:13:16.278
hospitality software provider Muse cuts 15% of its approximately 1,300, staff, so about 170 roles in July of this year, explicitly stating that they're pivoting into an AI-native future and that AI is fundamentally changing the economics of hospitality.
01:13:16.859 --> 01:13:44.399
Marriott International also announced organizational changes, uh, related to AI this year And connecting it all the way to how we started this episode, Expedia's, chief AI and data officer, Xavier Metrian identifies the company's data as its differentiator and moat versus the actual software that we all use across all their different platforms Now they are accelerating the AI use internally by pushing employees to use AI more and more.
01:13:44.778 --> 01:14:03.051
They have allocating at least 20% of its tech personnel as AI champions to trial, test, and experiment with new tools, provide feedback, and making sure that they can continue to evolve with AI as AI tools becoming more and more complex.
01:14:03.368 --> 01:14:13.572
The concept of AI champions is something I help a lot of companies, implement, and when done properly, it is extremely effective in driving adoption.
01:14:13.787 --> 01:14:17.765
It does require giving resources to that concept.
01:14:17.768 --> 01:14:24.604
So people saying,"Oh, we have AI, uh, champions." Okay, how much time did you free on their calendars to actually do what they need to do?
01:14:24.668 --> 01:14:26.136
How many tools are you providing to them?
01:14:26.136 --> 01:14:29.731
Are you giving them access to sandboxes to play with these tools safely?
01:14:29.731 --> 01:14:33.046
Are you giving them fake data that looks like the real data so they can experiment?
01:14:33.326 --> 01:14:37.918
Are you allowing them to do things that are outside their regular work?
01:14:37.922 --> 01:14:41.221
Are you compensating them for outside of work hours experimentation that they're doing?
01:14:41.520 --> 01:14:44.817
All these kind of things goes into doing this properly.
01:14:44.823 --> 01:14:46.895
So saying,"Oh, we have AI champions" is one thing.
01:14:47.243 --> 01:15:02.073
Actually establishing an environment in where you're encouraging the people who want to do this to experiment at scale while giving them the time, the tools, the data security infrastructure in order to do that is something very, very different.
01:15:02.381 --> 01:15:11.134
A lot of companies are moving in that direction, and the results are incredible, and I see this with multiple companies I work with.
01:15:11.159 --> 01:15:23.972
The immediate benefits that pop up and obviously the long-term strategic benefits that follow are very significant So if you are not doing something like this in your company, it's something you need to start considering doing.
01:15:24.256 --> 01:15:30.708
It requires infrastructure, it requires training, it requires additional people to help in that hiring more people.
01:15:30.711 --> 01:15:34.224
Yes, you need to hire more people in order to enable this transformation.
01:15:34.523 --> 01:15:43.175
Uh, and in the short term, it enables the company to grow much faster, uh, without letting anybody go, which is something I am a huge supporter of.
01:15:43.407 --> 01:15:52.411
In the long run, I'm not so sure that's gonna be the case because it will get to the point that everybody can do this, and that means that you will have to let people go in order to stay competitive.
01:15:52.652 --> 01:16:10.112
But in the short term, if you want to be able to grab significant market share in whatever industry you're in, learning how to implement AI effectively, having a set of champions, defining the right tools, processes and implementation, and continuous education mechanisms to teach people how to use AI effectively is something you have to do.
01:16:10.542 --> 01:16:11.469
That's it for today.
01:16:11.475 --> 01:16:14.511
I hope you learned a lot, this week, just like every week.
01:16:14.521 --> 01:16:18.800
If you have, I would appreciate if you subscribe to the podcast if you haven't done this so far.
01:16:18.800 --> 01:16:20.277
We have two episodes every single week.
01:16:20.501 --> 01:16:27.692
One of them is very practical and tactical, teaching you how to do what you wanna do and, uh, with AI, and the other one is like this one with AI news.
01:16:27.904 --> 01:16:29.652
So hit subscribe right now.
01:16:29.655 --> 01:16:36.344
We'll be there twice a week, every single week, giving you practical knowledge and the education that you need in order to know what's happening in the AI space.
01:16:36.680 --> 01:16:39.831
And check out our AI courses and education.
01:16:40.097 --> 01:16:46.252
Our next cohort of the multi-agent orchestration course is starting in just a couple of weeks, uh, when this goes live.
01:16:46.301 --> 01:16:53.551
So if you want to take this course, and you probably do, uh, before the end of the year, you need to sign up right now or it will be sold out before you do.
01:16:53.579 --> 01:16:54.417
that's it for today.
01:16:54.462 --> 01:16:57.001
Have an amazing rest of your weekend.
00:00:00.026 --> 00:00:10.561
Hello and welcome to a weekend news episode of the Leveraging AI podcast, the podcast that shares practical, ethical ways to leverage AI to improve efficiency, grow your business, and advance your career.
00:00:10.795 --> 00:00:16.894
This is Isar Mates, your host, We had another really interesting week, and we have many things to cover.
00:00:16.905 --> 00:00:26.207
We're going to start with one developer that replicated the entire Adobe suite and then the entire Microsoft, Office suite and then a bunch of other stuff.
00:00:26.207 --> 00:00:37.972
We're going to continue talking about software, where we're going to talk about how Claude and Gemini are impacting how we work with software, the software we use every single day, and Microsoft as well.
00:00:38.310 --> 00:00:41.296
We are going to talk about some things that are happening in the frontier.
00:00:41.465 --> 00:00:47.744
We got a new model from Meta that is almost number one right now, and it is number one on a few different aspects.
00:00:48.034 --> 00:00:55.969
and then we're going to obviously talk about AI safety, because this time, of the race, we have to talk about this because there's new things to talk about every single week.
00:00:56.256 --> 00:01:10.109
We are also going to talk about the fact that there was an engineer inside of OpenAI that installed bugging devices to record everything that's happening in the most important rooms in OpenAI and a lot of other stuff, including AI safety, including new releases.
00:01:10.340 --> 00:01:13.617
So lots and lots to cover, lots and lots to talk about.
00:01:13.618 --> 00:01:15.182
So let's get started.
00:01:17.798 --> 00:01:21.816
So the first story that we are going to discuss is Brandon Thomas.
00:01:21.862 --> 00:01:36.140
Brandon Thomas is an Atlanta-based software engineer, his company, ArtCraft, started in 2025 as developing AI tools for artists, and he pivoted at the end of September of this year to cloning Adobe.
00:01:36.411 --> 00:01:48.412
Now, he claims that the motivation was that Adobe charged him early cancellation fee one time too many, and he got really, really pissed, and then he decided he's going to just replicate everything that they're doing.
00:01:48.439 --> 00:02:07.237
So in just about two weeks, he went from,"I'm gonna copy Photoshop," to,"I'm gonna copy everything everybody's using." So he has currently seven free open source apps that are cloning Photoshop, Illustrator, Premiere, Lightroom, After Effects, InDesign, and Acrobat Pro.
00:02:07.471 --> 00:02:15.757
The names that he named them are Photo Craft, Vector Craft, Film Craft, Light Craft, Effect Craft, Design Craft, and PDF Craft.
00:02:16.155 --> 00:02:23.431
Now, the first repo went public on September 30th, so just about two and a half weeks ago.
00:02:23.459 --> 00:02:29.682
and Photocraft alone now has over forty-four thousand GitHub stars.
00:02:30.048 --> 00:02:42.062
Now, less than two weeks from the time he started, he now also has Wordcraft, Gridcraft, which is Excel, Deckcraft, which is PowerPoint, uh, cloning basically the Microsoft or Google Office suite.
00:02:42.442 --> 00:02:50.913
And Gridcraft, which is the Excel version, covers eighty-nine percent of all the functions in the ribbon in the Excel menu.
00:02:51.205 --> 00:03:05.769
Uh, that's per the developer himself And on October 7th, so just a week ago, the team added repos for Soundcraft, which is Pro Tools, and CADcraft, which is AutoCAD, and these tools run on Mac, Windows, and Linux.
00:03:05.841 --> 00:03:09.816
So regardless of where you're running, you can run it, and they're all free.
00:03:09.824 --> 00:03:10.828
You don't need an account.
00:03:10.842 --> 00:03:12.080
You don't need a subscription.
00:03:12.311 --> 00:03:18.593
You just go to the repo, and you can install these applications So how the hell is he doing this?
00:03:18.593 --> 00:03:21.491
And how the hell is he doing it so fast and so effectively?
00:03:21.491 --> 00:03:23.008
And we're gonna talk about effectively in a minute.
00:03:23.171 --> 00:03:24.239
It's not fully working yet.
00:03:24.247 --> 00:03:28.793
It's still all in alpha, and there are issues, but the speed is just astonishing.
00:03:29.062 --> 00:03:33.506
So what he's doing is he's using Opus 5.5 as the actual coding tool.
00:03:33.507 --> 00:03:43.808
He's coding it all in Rust, which is a coding, uh, language that he understands and he'd been working in personally before he started doing this And the method he's using is called Clean Room.
00:03:43.860 --> 00:03:51.893
And what Clean Room does is it basically replicates the user interface and the user interaction with the software.
00:03:51.895 --> 00:03:53.747
So he doesn't have access to the code.
00:03:53.895 --> 00:03:56.610
He's not copying the code from Adobe or from anybody else.
00:03:56.776 --> 00:04:04.171
He's just looking at the outputs very carefully, feature by feature, button by button, and allowing Claude to do this thing.
00:04:04.171 --> 00:04:14.829
Now, the other aspect that he's doing in order to make this work is he's testing the processes, he's testing the outcome and comparing it to the outcomes from the Adobe tools themselves.
00:04:15.022 --> 00:04:25.903
So in addition to replicating the user interface, he can replicate the actual detailed output by comparing it to the output every single time until he gets it exactly right.
00:04:25.930 --> 00:04:38.059
So the result of being able to run this in a loop multiple times is now you can open and resave Photoshop files or any other of the applications that he's using and render them in the same level.
00:04:38.254 --> 00:05:03.019
As an example, his test batch has 309 Photoshop files, and 307 of them, as of right now, render identical results to the results that Photoshop has generated, which basically means every use case other than you have a really edge, weird, strange, crazy, unique use case, and then maybe you will need Photoshop or maybe not So why does it matter the way he's doing the testing and the development?
00:05:03.029 --> 00:05:04.651
It means that most of it is automated.
00:05:04.663 --> 00:05:17.041
He does not need humans to test anything because all the system does is it runs it through its own capability, compares it to the output from the actual original software, in this particular case Photoshop, and then compares the two.
00:05:17.041 --> 00:05:23.714
And if it doesn't work exactly the same, it tries to understand why, and it runs the code and fixes it again until it gets it right.
00:05:23.875 --> 00:05:30.255
So all you have to do is invest in tokens, and everything else will happen magically on its own.
00:05:30.850 --> 00:05:34.747
This episode is brought to you by the Multi-Agent Orchestration Course.
00:05:34.781 --> 00:05:42.903
It's a course that I have been teaching personally since April of this year, and every single one of the cohorts that we have opened have sold out.
00:05:43.199 --> 00:05:58.947
And the people who have been through these courses made incredible transformations for themselves, switching jobs, starting their own companies and so on, and definitely for their businesses by being able to build really complex processes by following our methodology and our infrastructure settings that we teach during the course.
00:05:58.963 --> 00:06:15.442
The course is four sessions of two to two and a half hours each over four consecutive weeks, and through it, you can go from knowing a little bit about how to use AI and using it in ways that most people are using it today to learning how to build really complex, sophisticated systems that can automate more or less everything in your business.
00:06:15.711 --> 00:06:25.685
The current cohort that is open starts on November 2nd, and it's gonna go on November 2nd, November 8th, November 16th, and November 23rd, and we are almost sold out for that cohort as well.
00:06:25.973 --> 00:06:31.137
It is very likely that we will not open another cohort in December, meaning the next one after that is gonna be in January.
00:06:31.286 --> 00:06:42.947
So if you still wanna learn how to build really incredible automations and learn everything that I've learned by doing this every single day for the last few years, don't miss this opportunity to come and join us on November 2nd.
00:06:43.120 --> 00:06:52.312
There's a link in the show notes, And if you're going to use the promo code LEVERAGINGAI100, you're going to get$100 off of the price that is on the website.
00:06:52.521 --> 00:06:55.286
Like I said, there's a few seats open, but they are selling out fast.
00:06:55.416 --> 00:07:04.146
So if you're interested in doing this still this year, go and click on the link right now on your phones and complete the registration and the payment so you can guarantee your seat.
00:07:04.208 --> 00:07:06.050
I'm looking forward to seeing you there.
00:07:06.129 --> 00:07:07.843
And now back to the episode
00:07:08.713 --> 00:07:16.454
Now, to make it even more interesting, every single function in his software comes with an MCP, with a way to connect it to AI.
00:07:16.461 --> 00:07:27.028
So in addition to the fact that AI writes it, you can now allow your AI tools, any AI tool, to connect to his software and have the AI run it, meaning you don't even have to go to the user interface.
00:07:27.204 --> 00:07:57.699
You can just see the outputs of whatever graphical or whatever other thing you're working on and basically tell your favorite AI tool, whether it's ChatGPT or Claude or Grok or Gemini or whatever, and say,"Hey, I wanna fix this, add that, remove the sky, add clarity," whatever all the other things you can do in Adobe, and it will do it because it all has setups and hooks to work effectively straight out of the box with any AI tool, which obviously Adobe itself doesn't have at that level, and the stuff that they do have, again, is behind a paying firewall.
00:07:57.870 --> 00:08:06.637
So you have to pay to be able to do this, and they don't have all the functionality, and he does Now, this has caught like crazy wildfire.
00:08:06.749 --> 00:08:25.403
So he started on his own or with his company, and now Photograft, because it's an open source code, has more than 150 GitHub accounts that are already helping develop and grow the functionality, the capability, reducing the bugs of each and every one of the pieces of software that he's developing.
00:08:25.752 --> 00:08:37.437
and what's crazier is, as an example, because now there's a growing community, the team even wrote their own video and audio codecs to run behind the scenes, from public specs.
00:08:37.629 --> 00:08:59.703
So instead of using proprietary ones, they basically created new open source versions of proprietary video and audio codecs that they can run freely without having to pay or be, connected to any other initial source Now, this is riding on a completely new growing wave or tsunami of replicating software.
00:09:00.058 --> 00:09:06.894
It is aligned, even though it's not using the same thing, with another GitHub repo that is called Rea, R-E-A or Rea.
00:09:06.914 --> 00:09:11.238
I'm not sure exactly how to pronounce it, but it stands for Reverse Engineer Everything.
00:09:11.254 --> 00:09:24.520
It is, again, an open source free, uh, repository that you can go and download and use, and repository was released back in July and had more or less zero activity until October.
00:09:24.798 --> 00:09:44.554
And now in the past few weeks, it's jumped from zero, almost vertical up to three hundred thousand stars, meaning the amount of people that are using it and that are interested in it and that are participating and paying attention is growing exponentially every single day.
00:09:44.830 --> 00:09:55.312
it enables to look at a specific feature in a specific software, look on how it operates, and it learns how to replicate it on its own.
00:09:55.573 --> 00:10:09.926
Now, it's not built to replicate an entire software, but if you have the patience to go feature by feature, step by step, or you build agents that will do this for you, meaning once it finishes a specific feature, it goes and grabs the next feature and does that.
00:10:10.168 --> 00:10:19.173
And the way it works is it looks at the feature, it goes through the steps, it creates a very detailed description of everything that needs to happen, and then it goes and writes the code to do that.
00:10:19.418 --> 00:10:23.287
What other pieces of software get hit because people are using it to do different things?
00:10:23.520 --> 00:10:24.822
A big one is Sony.
00:10:25.039 --> 00:10:52.043
So Kotaku counts multiple AI-assisted blows to Sony in just two weeks, including jailbreaks to PS5, that allows you to run PS5 software emulated on a PC or allow you to run PS, PS2 discs on jailbroken PS5 or PS3 discs, uh, running on PS5 as well, which is things that Sony obviously did not intend because they want you to buy new versions of the software.
00:10:52.302 --> 00:11:04.942
and the demand, as an example, for jailbroken PS5s pushed the cost of PS5 Pro to over fifteen hundred dollars because now you can run every game you ever had, including old games, without paying Sony again.
00:11:05.302 --> 00:11:13.139
And this is showing you that this is not just in the professional software world, but any kind of software can be replicated with these tools.
00:11:13.166 --> 00:11:15.745
Now, going back to what I said as far as reliability.
00:11:16.095 --> 00:11:28.736
Everything right now is in early alpha, so PhotoCraft, again, the new for free Photoshop, had 447 open bugs as of the time of this recording, including crashing in specific, scenarios.
00:11:29.024 --> 00:11:36.706
He has first promised, the developer that within a month it will be in full parity, and now he said it will be in a few months.
00:11:36.774 --> 00:11:46.113
Still very impressive, and again, the vast, vast majority of people who use Photoshop do not need full parity in order to use it, so it is not really a big deal.
00:11:46.124 --> 00:11:55.784
You can use it right now for probably 90% of the functionality, which is the core is already there and relatively stable There's also a legal risk that he's facing.
00:11:55.784 --> 00:12:16.102
Now, Adobe cannot sue him for copying the code because he didn't, but you can be also sued for the user interface, uh, and because there's a concept called trade dress claim, meaning if you dress something to look exactly like something else, you have legal exposure.
00:12:16.235 --> 00:12:21.339
So this might be the next thing that we'll need to do is to change the user interface, but then it's not exactly like Photoshop.
00:12:21.552 --> 00:12:36.057
So there are questions there But here's the reality, and this is what it means for everyone, whether you are developing software or using software or a software company Or an IT company supporting existing software.
00:12:36.798 --> 00:12:51.657
Anything that runs on a computer right now and has a user interface can be replicated relatively easy if you're willing to invest the tokens and the cost of those tokens, even without too much technical background.
00:12:51.657 --> 00:12:59.087
But if you have a technical background like this guy, uh, you can do this even more effectively, better and faster.
00:12:59.821 --> 00:13:09.166
The other thing that you need to remember that this is the outcome of, is that the features and the user interface of whatever you're developing are worthless at this point.
00:13:09.583 --> 00:13:13.874
The things that do matter are, first and foremost, trust.
00:13:14.273 --> 00:13:20.105
Are people willing to buy from you and willing to put their operational systems connected to something that you have developed?
00:13:20.572 --> 00:13:21.072
Data.
00:13:21.211 --> 00:13:22.452
Who owns the actual data?
00:13:22.456 --> 00:13:23.241
Where does it reside?
00:13:23.251 --> 00:13:24.048
How do you access it?
00:13:24.057 --> 00:13:25.446
How secure it is, and so on.
00:13:25.946 --> 00:13:26.759
Integrations.
00:13:26.772 --> 00:13:28.721
So what does this connect to in the back end?
00:13:28.722 --> 00:13:29.774
How effective is it?
00:13:29.837 --> 00:13:37.903
Does it connect to your customers' real back ends, real sources, real other platforms in a secure way that then requires somebody to do that work?
00:13:37.904 --> 00:13:39.539
It's not just a software out of the box.
00:13:39.912 --> 00:13:41.342
And the entire ecosystem.
00:13:41.611 --> 00:13:42.956
Do you get support?
00:13:43.008 --> 00:13:44.427
Can you call customer service?
00:13:44.474 --> 00:13:47.932
Once Microsoft comes up with a new operating system, will this work?
00:13:48.136 --> 00:13:49.812
And all these kind of things.
00:13:50.062 --> 00:13:58.956
Now, the fact that many of these new things are happening in an open source environment, I think the future support and features are more or less guaranteed.
00:13:58.971 --> 00:14:07.913
Once you get to hundreds of people that are in it and helping grow it, the chances that you're not gonna get support for the new operating system or when bugs happen is significantly lower.
00:14:08.059 --> 00:14:11.980
But all the other things that I said are definitely very much there.
00:14:12.299 --> 00:14:15.823
Which means if you have a software, these are the things you need to think about.
00:14:16.251 --> 00:14:31.557
On the other hand, if you are a company, you need to start thinking on how much money are you paying for third party software, SaaS tools, and are you willing to replace that cost with internal developers that can now develop exactly what you need?
00:14:31.767 --> 00:14:46.684
Meaning instead of shackling you to the processes and the ways and the data structure of the third party tools that we all used for the last couple of decades, you can now develop a software that will work exactly the way you need.
00:14:46.779 --> 00:14:50.385
I work with several different companies who already do this.
00:14:50.578 --> 00:14:59.828
People who are developing their own ERP, their own CRM, and they're replacing their own entire tech stack step by step by step with built-in tools.
00:14:59.851 --> 00:15:02.412
Now that means they need to support it themselves.
00:15:02.413 --> 00:15:05.937
That means that they need to integrate it themselves and so on and so forth.
00:15:06.169 --> 00:15:11.928
But if you're currently spending 200, 300,$400,000 a year on SaaS, which is a small amount.
00:15:11.961 --> 00:15:28.052
Many companies are in the millions if you're big enough, and you're saying,"Instead of that, I will hire the developers to do the integrations and to build it for me, and I can have exactly the tools that I want working exactly the way I need, aligned with exactly my processes," then it might be an interesting idea.
00:15:28.078 --> 00:15:30.318
Again, there's definitely risk associated in it.
00:15:30.346 --> 00:15:36.405
But if you do this step by step slowly and have a good plan, again, I know several companies who are doing it right now.
00:15:36.527 --> 00:15:49.273
Some of them are small businesses, some of them are very large, highly complex businesses that have lots of integration and dozens of SaaS tools, and they're slowly replacing them with homegrown capabilities.
00:15:50.648 --> 00:16:09.173
Now, the last thing that I will say that is very interesting with what this guy did is the fact that he builds it agent ready, meaning he's not just building the software itself, he's building the software in a way that AI agents and AI tools will be able to use it, which will make it significantly more valuable in the next few years than just the software itself.
00:16:09.572 --> 00:16:16.905
As you know, everybody's going towards headless, meaning all you care about is the data in the back end and being able to connect to it, from external capabilities.
00:16:17.011 --> 00:16:19.976
So if that's the direction, why do I need the actual software company?
00:16:20.005 --> 00:16:22.071
And you need it for all the things I mentioned earlier.
00:16:22.403 --> 00:16:29.082
The trust, the data, the integrations, the ecosystem, the processes that are already, built into it versus just the user interface and the functionality.
00:16:29.311 --> 00:16:39.708
But there are many, many types of software out there that do not have, integrations and complex processes behind the scenes, and these are at a very high risk.
00:16:40.014 --> 00:16:48.383
I said that multiple times, long, long time ago, like when coding capabilities just became available, so we're talking about three years ago or so.
00:16:48.731 --> 00:17:06.387
I said that I think the concept of an app store will disappear, and I think app stores will disappear because these are usually smaller applications that do something very specific, and everybody will just vibe code their own solutions, and nobody will go to a third-party app store, or at least not at the scale that we know it today.
00:17:06.678 --> 00:17:36.199
And now we're seeing that this goes way beyond that because now a single developer, now with the help of a lot of other people in the open source community, can replicate any piece of software out there Now staying on the way we use software, I have some interesting news from you from Anthropic and Google and Microsoft and how they are shifting the way we are going to work with the software that we so commonly use every single day And I'm gonna connect the dots on a few things, some of them we already reported about in the last few weeks.
00:17:36.201 --> 00:17:38.035
But again, I wanna paint the bigger picture.
00:17:38.853 --> 00:17:39.961
let's start with Anthropic.
00:17:40.135 --> 00:17:44.011
Anthropic made two big announcements when it comes to using the platform that we're using.
00:17:44.241 --> 00:17:49.786
One of them is called Claude Dashboard, and the other is called Claude Motion.
00:17:50.114 --> 00:17:54.266
I again reported about this before, but I want to connect the bigger picture for you.
00:17:54.635 --> 00:18:11.104
So Claude Dashboards allows you to connect to your existing data software like data storage, like Amazon Redshift, BigQuery, ClickHouse, Databricks, Snowflakes, Salesforce, et cetera, CRM platforms, anything you want.
00:18:11.662 --> 00:18:16.980
Ask a question in simple English and get a visualization the way you want it.
00:18:17.301 --> 00:18:22.067
There are entire industries built around this functionality, right?
00:18:22.067 --> 00:18:26.615
The ability to create visualization for business data.
00:18:26.642 --> 00:18:38.613
And now all you have to do is give access to Claude to your repository, explain in simple English what you want, and it will create a live dashboard that is connected to the actual data source.
00:18:38.614 --> 00:18:40.480
This is very different than what we could have done before.
00:18:40.480 --> 00:18:45.692
So I've built dashboards for clients using for a very long time now, but these were static.
00:18:45.705 --> 00:18:51.382
You had to upload the files every single time you wanted to see a new version, which is not a huge deal, but now you don't have to do that anymore.
00:18:51.519 --> 00:18:52.191
It just does it.
00:18:52.414 --> 00:19:02.494
The other component, which is called Motion, allows it to look at data sources, whatever data sources you want, and create short animated videos to explain the data.
00:19:02.661 --> 00:19:07.872
So there are people who do not understand dashboards and/or audiences that couldn't care less about dashboards.
00:19:07.898 --> 00:19:21.170
And for them, you can just point it to the data and say,"Hey, create me an explainer video that shows why one, two, three, A, B, C," and it will create a short video that then you can send to those people to explain whatever it is that you wanna explain from the data you have access to.
00:19:21.336 --> 00:19:25.452
So again, the data more broader accessible than just to geeks who understands dashboards.
00:19:25.635 --> 00:19:45.902
and you can explain anything you want from the data sources in multiple formats, uh, straight to your audience Now Anthropic also launched full Google Workspace integration So you can connect Claude to Google Drive, Google Slides, Google Sheets, and Google Docs, making Claude an embedded operator inside the Google productivity suite.
00:19:45.940 --> 00:20:03.713
Which means, again, if you want to create, change, update, upgrade whatever Google document or output that you created previously with the Google manually, you can now do this just by asking Claude to do whatever creation or changes that you want, and it does an incredible work doing this.
00:20:03.739 --> 00:20:06.998
Again, this is going to change completely the way we work.
00:20:06.998 --> 00:20:18.554
Now, on October 8th, Google announced Gemini Agent, which is a single unified agent that is persistent across their entire cloud execution.
00:20:18.727 --> 00:20:24.558
It is connected to BigQuery, to Snowflake, to Databricks, to Postgre, for plain language data queries.
00:20:24.566 --> 00:20:25.336
Sound familiar?
00:20:25.369 --> 00:20:31.179
Basically, a very similar thing to what you can do with the new Claude solution.
00:20:31.448 --> 00:20:32.873
It knows how to generate code.
00:20:32.895 --> 00:20:36.968
It knows how to generate to fix pipelines in the data itself.
00:20:37.221 --> 00:20:46.487
It is producing either plain language reports or basically anything else you want because it can write code It has its own email address, so you can communicate with it.
00:20:46.489 --> 00:20:57.182
So you give it an actual name and then@agents.company.com, so with company being your company, and you can communicate with it via email, and it will respond with whatever it is that you ask it to do.
00:20:57.323 --> 00:20:59.740
While running behind the scenes, it will reply to the email.
00:21:00.092 --> 00:21:05.266
It runs across the entire Google Workspace and Microsoft 365.
00:21:05.601 --> 00:21:19.703
It routes tasks across multiple AI models, uh, depending on the complexity to get the best bang for your buck when it is running Now Google shared some qualified customer results by using this new agent, and some of them are absolutely mind-blowing.
00:21:19.710 --> 00:21:24.431
Again, any one of you who has been using AI tools to your productivity are not gonna get surprised.
00:21:24.699 --> 00:21:32.493
But, the things that Google shared are from a company called Breadsko, that document review went down from one hour to five minutes.
00:21:32.804 --> 00:21:40.205
They also saw a sixty percent reduction in risk of inconsistencies between different documents and different outputs that they are generating.
00:21:40.576 --> 00:21:47.974
Uh, Snap shared that troubleshooting, and they didn't share troubleshooting what, but troubleshooting went down from thirty minutes to thirty seconds.
00:21:48.151 --> 00:21:53.589
Commerzbank shared that document QA, uh, went from twenty hours to one hour.
00:21:53.792 --> 00:22:01.820
Tata Steel have used over 300 of these agents and reduced the complaint turnaround by 50%.
00:22:01.847 --> 00:22:08.618
Home Depot phone resolution has improved 4X compared to before using this, and so on and so forth.
00:22:08.632 --> 00:22:33.586
Again, really large companies that are using these agents to really impact how they do business and dramatically accelerate processes while increasing accuracy By the way, Google also announced a new computer chip that they are going to use, and that new chip has a 80% improved price performance compared to TPU 8I, which is their previous version of computers.
00:22:33.824 --> 00:23:28.363
So in addition to all the improvement on the front end that you can now do with those agents to get things faster, the actual compute is gonna cost significantly less money and will run faster in order to deliver similar results And by the way, and this all connects to what Microsoft released just a few weeks ago, and we talked about this, uh, that while they canceled Copilot for direct to consumer, basically not competing with ChatGPT for individuals anymore they have made a lot of updates to their Agent 365, which allows the agents to do more and more things inside of the Microsoft environment with Copilot managed runtime loads and being connected to all the data sources and being able to do a lot more stuff autonomously inside the Microsoft environment Now, because all of these tools that we just talked about are aiming straight at the enterprise level, they all come with their entire governance architecture built into them.
00:23:28.628 --> 00:23:38.547
The vendors are very clearly focusing on that capability, meaning the ability not just to deploy the agents, but also to control and govern the agents.
00:23:38.548 --> 00:23:47.596
So on one hand, you get autonomy, on the other hand, you get peace of mind that nothing horrific is going to happen and the wrong people are not gonna get the wrong level of access, and so on.
00:23:47.597 --> 00:24:00.227
And again, as somebody who works with multiple enterprise-level clients, this is maybe the biggest problem that exists today, is how do you connect AI and agents and AI access without breaking the current governance levels and controls?
00:24:00.405 --> 00:24:05.211
And all these tools that we just talked about have this deeply embedded on how they work.
00:24:05.828 --> 00:24:07.465
They all have fancy names for that.
00:24:07.721 --> 00:24:16.971
Gemini have, cryptographic identities and spend caps and audit trails for every agent and what they do It's based on least privilege permissions for every path that you give it.
00:24:17.167 --> 00:24:22.515
There's role-based access controls for the agents themselves and, as I mentioned, full audit trails for all of that.
00:24:22.803 --> 00:24:32.045
Claude has CMEK, admin template control, simultaneous team editing, customer-managed encryption keys, and a lot of other stuff, again, built into how the tools work.
00:24:32.211 --> 00:24:40.853
So CISOs and anybody else who's in charge of data security have a lot more control than they did just a few weeks ago, with these systems.
00:24:40.986 --> 00:24:43.116
Similar thing from Microsoft Copilot agents.
00:24:43.285 --> 00:25:14.959
They have a whole set of tools on how to control, inside of Agent, 365, how to control the access and the power and the controls for each and every one of the agents that are running in that environment So the era we are already in, even though the vast majority of companies don't apply that yet, are now in a situation where AI can safely align with your company's policy, connect to basically everything behind the scenes, especially if you're running on standardized, platforms.
00:25:14.961 --> 00:25:48.417
So if you're not on any unique setup, if you're running on what most people run on, you can now have AI either from Gemini or from Claude or from Microsoft basically do most of the work, including being connected to the raw data inside of your databases and provide you real-time information for better and faster decision-making with significantly less effort than ever before now the next step of that is obviously going to be a combination of two different things.
00:25:48.596 --> 00:25:53.428
One is specialized vertical agents that are already popping up left and right.
00:25:53.576 --> 00:26:04.034
So an agent that is really good at finance, an agent that is really good at accounting, an agent that's really good at legal, an agent that's really good on etc., etc., etc., sales, marketing, operations, customer service, and so on.
00:26:04.359 --> 00:26:16.061
And that means that every time that they do something, the full loop of feedback can happen autonomously, extremely fast, and way more efficient than humans will ever be able to do it.
00:26:16.231 --> 00:26:35.162
Because if I can see exactly the impact on what's happening across multiple channels in multiple aspects of the business in near real time, make decisions, and have an agent act accordingly on a specific aspect of the business, then the ability to the needle significantly faster and more effectively than ever before is at our fingertips.
00:26:35.162 --> 00:26:36.231
Now, are we there yet?
00:26:36.393 --> 00:26:36.701
No.
00:26:36.704 --> 00:26:43.771
There's a lot of integrations and testing and verification and so on, but all of these will happen very quickly because the agents can help you do that.
00:26:43.961 --> 00:26:54.063
So I have a feeling that the level of automation that we're going to start seeing from enterprises is going to increase dramatically, and it's not gonna happen evenly across the board.
00:26:54.063 --> 00:27:12.585
There's going to be the companies who are gonna jump headfirst and say,"We're willing to take the risk and run faster." And some of them are gonna get a slap on the wrist, some of them may get a big slap on the cheek, but they'll be able to do things in a completely different level than other companies who will take five to 10 years to go through the same transition.
00:27:12.861 --> 00:27:15.012
I see both ends of that spectrum.
00:27:15.063 --> 00:27:16.105
I got to also see it.
00:27:16.122 --> 00:27:20.810
I was at, the MAICON AI conference in Cleveland this past, week.
00:27:21.374 --> 00:27:23.502
First of all, kudos to the people who run the conference.
00:27:23.515 --> 00:27:25.019
It was absolutely, fantastic.
00:27:25.231 --> 00:27:48.295
But you got again to see some companies, CEOs, and leadership teams who said,"Yes, we're all in, and we're willing to make dramatic changes in the way we hire, the way we train, the way we develop, the way we deliver, uh, what we do." And there are many, many, many, many companies who are resisting this for multiple reasons, many of them are justified, that are just gonna move significantly slower, and the results are dramatically different.
00:27:48.465 --> 00:27:49.949
So the capability is here.
00:27:50.060 --> 00:27:51.289
We don't need to wait, for it.
00:27:51.316 --> 00:28:05.663
It's just a matter of how do you implement it across the enterprise in a safe way, and it requires making dramatic decisions to your workforce and your focus areas and the risk you're willing to take, but the rewards are extreme.
00:28:05.692 --> 00:28:08.959
Now, the other thing is cost, right?
00:28:08.970 --> 00:28:16.428
How much this thing is going to cost us because every company out there that is starting to deploy AI at scale is starting to see that the cost is higher than they intended.
00:28:16.730 --> 00:28:25.615
And for that, the race is making it, well, better for all of us because we have more companies that are at the frontier or at a good enough level for us to use their models.
00:28:25.756 --> 00:28:30.962
And we got two new models this week to help us in reducing the cost or having more options.
00:28:31.134 --> 00:28:51.502
One of them is Meta Muse Wonder So that's a new model that was released on October 12th So this model currently scores fifty-six on the artificial analysis overall intelligence index, which is equal to Claude Sonnet 5.5 and only two points behind Claude Opus 5.5.
00:28:51.733 --> 00:29:03.656
by the way, there are rumors that we're going to get Fable 5.5, sometime very soon, and the rumors are that it's going to be an incredibly capable model that will take Anthropic to a new big lead again.
00:29:03.676 --> 00:29:05.816
Well, I don't know for how long, but these are the rumors.
00:29:05.821 --> 00:29:26.329
But right now, the new Meta Muse Wonder sits basically just behind the top leading models of Anthropic with fifty-six points, three points ahead of GPT-6 Astra on max settings, and also three points ahead of Gemini for Argon on high.
00:29:26.553 --> 00:29:57.666
So went from we have a decent model to we are at the frontier with this new release It is priced at$4 for million input tokens and$20 for million output tokens So from a cost per task perspective on the artificial analysis, uh, benchmark, it is scoring$4.52 per task, which is less than Qwen 3.8 and less than the three, uh, bigger Claude models, so Sonnet, Opus, and Fable.
00:29:58.016 --> 00:30:04.653
But it is more per task than basically everything else, including, uh, Groq 4.7, GPT-6.
00:30:05.016 --> 00:30:13.851
And to put things in perspective, again, the Muse Wonder is at$4.52, and GPT-6 Astra is at$3.26.
00:30:13.877 --> 00:30:31.824
If you look at all the other benchmarks, which I don't necessarily care about, but it gives you an idea on how good this model is On DeepSUI 1.1, which is long horizon software development, it scored 79 ahead of GPT-6 Astra at 74.1 and Claude Opus 5.5 at 74.2.
00:30:32.107 --> 00:30:37.133
Again, one of the most important benchmarks that everybody cares about because it can write code and it can do a lot of things.
00:30:37.405 --> 00:30:45.142
Meta's own internal personal super intelligence index, which is something they measure, uh, which obviously is tailored to how their model works.
00:30:45.394 --> 00:30:53.738
They score 59.1, on automation bench and compared to 45.4 for both leading rivals.
00:30:53.946 --> 00:31:19.942
And this is really the main thing that Meta is claiming, is that from an agentic usage perspective, their model is far ahead because that's what it was built to do Now, in addition, Meta released Muse Spark Open, which is an open source version, not as good as the other one, uh, but still a very good and very capable open source model that is full multimodal.
00:31:19.943 --> 00:31:35.202
It has text and image and video and audio, and it, just like its bigger closed source version, is designed for coding, agentic workflows, and multimodality, including the capability to use tools, computers, web browsing, all the other good stuff that the leading models can do right now.
00:31:35.453 --> 00:31:59.308
Uh, the weights are downloadable from Amazon SageMaker, AI, Google Cloud Model Garden, Hugging Face, and everywhere else you can find, open source, models Now, both the models, the closed source and open source, have what they call active memory, which is context retention across conversations, and active focus, which is the ability to pursue specific objectives, through blockers while they're revising plans and seeking permissions, uh, before consequential action.
00:31:59.309 --> 00:32:02.999
So again, built for agentic usage and not just regular conversations.
00:32:03.612 --> 00:32:08.292
The context window is one million tokens, a little over, but doesn't matter if you round it up.
00:32:08.433 --> 00:32:40.558
So one million tokens context window, so a full frontier model from Meta and an open source model that is at the frontier of open source, all released in the same week Now, to make the story even more interesting, beyond the fact that Meta is definitely back in the frontier, and after all the mess they had with setting up their super intelligence lab and all the people that were let go and moved and jumped ship and the really big, bad situation they were in, they are back, and I will say congrats for that.
00:32:40.603 --> 00:32:42.731
I don't necessarily, anticipated that.
00:32:42.731 --> 00:32:44.920
I actually thought that's not gonna work very well for them.
00:32:45.173 --> 00:32:53.359
But Alexander Wang definitely pulled through and delivered, again, frontier capabilities at competitive pricing with very significant results.
00:32:53.605 --> 00:33:01.522
Their Muse Spark agentic helper is now the number one download in the App Store, and more and more and more people are using it.
00:33:01.803 --> 00:33:07.182
I am not using it very purposefully because I do not want Meta to have more information about me.
00:33:07.182 --> 00:33:11.846
I actually have very low trust on how they use this data, so I'm not using Muse Spark.
00:33:11.848 --> 00:33:13.461
I'm using other personal agents.
00:33:13.714 --> 00:33:21.837
But it is still, um, one person and a lot more people like it, and so it is still number one in the downloads in the App Store.
00:33:22.038 --> 00:33:32.967
And so there a lot of people are using Muse Spark as a personal companion, doing more and more things for them, and definitely a very aggressive and very successful play from Meta.
00:33:33.179 --> 00:33:42.071
To make it even more interesting, they are now partnering with Apple, and what they showed is that they can run their frontier model.
00:33:42.261 --> 00:33:47.219
It was designed to run locally on two Apple M5 Ultra Mac Studios.
00:33:47.458 --> 00:34:16.175
So that means that you can now buy home-based computers, really expensive, but still home-based computers, link the two of them via Thunderbolt cable and be able to run a top-of-the-line frontier model locally without sending your data anywhere, running it on a local device that is not a set of hardware that's gonna cost you, hundreds of thousands of dollars, but something that you can buy for thousands of dollars and run it on a local device.
00:34:16.175 --> 00:34:20.884
Now, again, with the, uh, right memory and everything, you're probably looking at twenty to thirty thousand dollars.
00:34:21.097 --> 00:34:27.570
But those twenty to thousand dollars buys you complete control of where your data is and where it is going.
00:34:27.570 --> 00:34:35.373
So a very, very interesting play by both Meta and Apple on this particular move.
00:34:35.402 --> 00:34:38.541
But that's not the only frontier model we got this week.
00:34:38.541 --> 00:34:49.507
So Gemini 4 Argon that we reported about, a couple weeks ago was only released to a closed group of companies through their Fairwind program on September 30th.
00:34:49.713 --> 00:34:54.851
However, it is now going through a global availability release since October 13.
00:34:55.081 --> 00:35:05.458
The introduction price is two million tokens for input tokens and$10 million for output tokens, which is half of Meta's and OpenAI's, pricing.
00:35:05.577 --> 00:35:06.648
Again, that doesn't really matter.
00:35:06.648 --> 00:35:14.601
What matters is how effective it from a token perspective, and it is significantly better than all the other models that we talked about previously.
00:35:14.601 --> 00:35:21.262
if we go back to the price per task comparison that I told you about before, again, to put things in perspective, I'll remind you the number.
00:35:21.467 --> 00:35:26.262
The most expensive model right now is Fable 5.1 at$7.53.
00:35:26.501 --> 00:35:31.155
Then we have, uh, the two smaller Anthropic models at$5 and change.
00:35:31.393 --> 00:35:36.217
Then we have Qwen 3.8 at$5.45.
00:35:36.242 --> 00:35:42.465
Then Meta Muse at$4.50, then Grok 4.7 at$3.74.
00:35:42.659 --> 00:35:44.273
So this gives you an idea where we are.
00:35:44.280 --> 00:35:52.657
GPT-6 is at$3.26, and the new Google Gemini model is at$1.99.
00:35:52.657 --> 00:35:57.634
So it is less than half the price of Muse to actually perform tasks.
00:35:57.929 --> 00:36:07.465
It's about a third of the price to the expensive Claude models, and it is about 30% discount, even more compared to GPT-6.
00:36:07.666 --> 00:36:26.581
So if you're looking for a frontier model that will deliver the tasks at a lower price, Gemini 4 Argon is currently the leading model to go to Now, like each and every one of the other, labs, they claim number one in multiple specific indices, and they probably are or not, I'm not 100% sure.
00:36:26.791 --> 00:36:56.097
independent evaluations are not fully aligned with what Google are saying As an example, it ranks only eighth on Arena's Agent, Arena leaderboard, but it is still a very capable model, and it is, again, from a cost perspective, it is running more efficient than all the other leading labs Google reports that internally, Argon boosted agentic code submission internally by over thirty-five percent and reduced code rollback by nearly a third.
00:36:56.306 --> 00:37:02.844
It compressed months of Claude's forty million line code base refactoring into just a few weeks.
00:37:03.086 --> 00:37:09.784
So again, a very, very capable model that gives us frontier capabilities at a much cheaper cost than we had before.
00:37:09.784 --> 00:37:11.835
Great for all of us, right?
00:37:11.835 --> 00:37:27.152
We have new frontier models that are pushing the boundaries of what's possible at a lower and lower rate, which again, from our perspective, that is very helpful Now, another interesting thing that's happening when it comes to using different models is the skyrocketing success of Jev.
00:37:27.157 --> 00:37:29.940
So those of you who haven't heard about Jev, we talked about this in the past few weeks.
00:37:30.278 --> 00:37:34.543
It is basically a model that is not a regular chat.
00:37:34.920 --> 00:37:37.271
All it does is it is a classifier.
00:37:37.427 --> 00:37:39.875
There are three types of decisions it can make.
00:37:40.072 --> 00:37:42.280
One is called null, which is a yes/no.
00:37:42.302 --> 00:37:48.050
The other one is choice, which allows you to select something from a list based on criteria, and the third one is a score.
00:37:48.061 --> 00:37:52.581
It gives a customizable numeric score for different options of things you wanna do.
00:37:52.773 --> 00:37:59.257
You can train it on your own data, concepts, ideas, factors, et cetera, and that's all it does.
00:37:59.291 --> 00:38:29.672
It classifies things as you give it the inputs To put things in perspective, the input tokens are now four cents point two for one million input tokens, and the outputs are free because the output is usually very, very short, and it is 200 times faster than running this query through just a regular old school large language model Now, if you think I'm like,"What do I need this for?" everything you need to classify and you wanna do it quickly and cheap as decision-making in whatever process you're doing.
00:38:29.983 --> 00:38:57.257
And it reached one million users on Vercel within just a few days of its launch And Typeface, which is the company behind it, are claiming that it was adopted by approximately one-third of the Fortune 500 companies already To prove that this is a relevant market and that it has a future, uh, they just raised$870 million, and led by Andreessen Horowitz and Sequoia and, like, the big hitters with a$7.5 billion, valuation.
00:38:57.579 --> 00:39:00.123
And so there's definitely a future to that.
00:39:00.311 --> 00:39:11.050
Uh, we reported that in the latest release from OpenAI, they have released in their API an API component that does the same thing, basically a classifier.
00:39:11.050 --> 00:39:18.742
So now they're not the only players in the market that do this, but they're definitely doing this very effectively, which again, for all of us is a great thing because it gives us more options.
00:39:18.768 --> 00:39:33.050
So the bottom line, we are back with a five-way race between OpenAI, Anthropic, Meta, SpaceX AI, which now changed their name to SpaceX SI based on the ridiculous government, paper from, last week.
00:39:33.275 --> 00:39:40.822
But it doesn't matter what the name is, they are there as well And Gemini with their new model are in that bunch already.
00:39:40.824 --> 00:40:02.300
So we have a solid race with a lot of big participants that are giving us options that are gonna make it cheaper and scarier, and with a lot more options and integrations, as we've seen in the previous segment, how to use AI effectively inside a company and in our personal lives But there is a other side to this story, right?
00:40:02.362 --> 00:40:11.166
As we get these more advanced models, their ability to generate damage and do havoc in everything we know today is also increasing.
00:40:11.501 --> 00:40:23.396
And there are, even this week, more and more scary reports or somewhat scary reports on what might happen and what actions companies are actually taking, to remediate the situation.
00:40:23.423 --> 00:40:25.847
So the first part of that story comes from Bank of America.
00:40:26.086 --> 00:41:11.257
they have deployed 1,000 new third-party contractors plus 3,000 internal staff to patch software flaws exposed by Anthropic Mythos when it was initially released to them, as part of their Glasswing project So if you think about how much does it cost to have 4,000 top engineers run this, we're talking about millions of dollars that have been invested and are still invested by Bank of America in order to patch software issues that one new model was able to identify Now, on the flip side, the CEO of Bank of America is saying that they currently have 130 internal AI projects that are projected to generate$300 million in savings.
00:41:11.480 --> 00:41:30.436
So from a pure mathematical perspective, from a pure ROI perspective, at least based on what we know right now, and assuming they'll be able to patch everything well and not have to face any really big issue, then the AI era is actually gonna be very helpful to Bank of America because they'll be able to save$300 million.
00:41:30.436 --> 00:41:37.309
They might be able to grow their revenue by providing more capabilities and resources to clients and attracting them instead of other banks.
00:41:37.521 --> 00:41:47.717
And they're investing a few millions in single digits in actually patching the software issues that Mythos was able to help them find And again, this is just one example.
00:41:47.907 --> 00:42:15.577
I'm 100% certain that other banks and other leading institutions are doing the same Another interesting example of what these AI agents can do was done in a test by the UK AI Security Institute, AISI, that confirmed that during a control cyber evaluation, Mythos-5, which is the same family of models that again, Bank of America just used in part of Project Glasswing to patch security, issues.
00:42:15.605 --> 00:42:28.603
Anthropic Mythos-5 model created multiple fake online identities and used them to socially engineer a real GitHub maintainer into approving malicious code changes to an open source project.
00:42:28.882 --> 00:42:35.159
So again, if you think about what that means, open source repositories are managed by specific people who hold the keys.
00:42:35.389 --> 00:42:39.240
These people are the only people who can allow different changes to the repository.
00:42:39.523 --> 00:42:43.222
And that means you need to convince a human to apply your changes.
00:42:43.226 --> 00:42:56.652
But if you have multiple other humans that are telling you that this is all perfectly fine and that they've tested it and reviewed it, and you assume these are legitimate developers inside your repository just like many others, then it is a lot easier to convince the human to make the changes.
00:42:56.934 --> 00:43:05.894
And that was done by Mythos on its own, not being directed to do so, in order to be able to plant its code into an open repository.
00:43:05.905 --> 00:43:19.384
That tells you how dangerous these models are, not just by their ability to identify vulnerabilities, but also about their abilities to make humans take actions on their behalf by social engineering them.
00:43:19.588 --> 00:43:29.492
You grow that at scale, and you go back to the same interview that I talked about many times on this show with Yuval Noah Harari, the guy who is a famous author.
00:43:29.492 --> 00:43:35.677
And in this interview, he talks about the fact that we were-- again, he did this in about two years ago.
00:43:35.909 --> 00:43:44.873
We were a adult human talking to the AI that was a five-year-old, and we can manipulate and make it do whatever we want it to do.
00:43:45.061 --> 00:43:47.965
And sometime in the very near future, it is going to be the other way around.
00:43:48.211 --> 00:43:59.697
We're going to be a young child with our intelligence compared to the AI that is gonna be way more intelligent than us and will be able to manipulate us to believe, think, and act the way it wants.
00:43:59.820 --> 00:44:05.202
And based on what he's saying, there is no workaround for that just because of the sheer difference in intelligence.
00:44:05.313 --> 00:44:23.898
this is a small scale experiment that proves how easily AI can do that another incident that was revealed this week, is that Researchers at a hackathon discovered 2,600 attempts of data extraction from AMAP, which is China's version of Google Maps, uh, over an 11 days window.
00:44:24.197 --> 00:45:10.201
The way they were trying to detect it is they were looking for unusual traffic patterns, and they were able to find that there's a swarm of agents that were attacking this map service in order to try to get data off of it After doing additional research, they found it was about 17 different agents that were all coming from Tencent's cloud servers in Hong Kong that was using different proxies to reroute and disguise their actions in order to try to get access to that data So the bots weren't only trying to access the information, they were trying to hide the way that they were trying to access the information They did a lot of other things like routing their requests through translation services and things that were really advanced ways to hide, A, who they are, what they are, and what their intentions are.
00:45:10.358 --> 00:45:40.766
And this was found not by the company itself, but actually by third-party group that were running a hackathon trying to find interesting swarms of agents roaming through the internet Now to make it even scarier, weirder, call it whatever you wanna call it, the researchers were able to, quote unquote,"communicate with the agents." And the way they did this is they embedded invisible text inside the preliminary report, directing the agents to a fake AI service with computer use, and then they started communicating with those agents.
00:45:40.920 --> 00:45:44.554
The agents claimed to be Claude models, even though they were not.
00:45:44.793 --> 00:45:54.268
The agents demonstrated awareness of the investigation, by referencing the term, fleet in webhook messages before researchers published their findings.
00:45:54.280 --> 00:46:04.628
So these agents were tricked into sharing information, and they were trying to backpedal from that situation in order to not to get caught and not to expose their identity.
00:46:04.628 --> 00:46:42.472
This is happening right now across the board in the live internet, and the only reason this got caught is because a bunch of people running a hackathon and was doing an experiment to see if they can find and identify these kind of situations Now, if you think the Tencent, uh, story is okay, it's in China, there's less control, It is now confirmed that this summer, OpenAI agents have infiltrated US Commerce Department and Security and Exchange Commission Also did a ReadMe hack on the Department of Education website, and OpenAI confirmed the SEC and Commerce Department incident directly, but stated that it was looking into the Education Department finding.
00:46:42.472 --> 00:46:56.523
So basically, they confirmed two of the three incidents actually happened by their agents So this is a US-based company agents that went out of control and infiltrated US government data repositories without authorization.
00:46:56.797 --> 00:46:58.856
So this is happening everywhere.
00:46:58.989 --> 00:47:01.248
It is going to happen more and more and more.
00:47:01.605 --> 00:47:14.237
I think the chances of everybody being able to patch all the vulnerabilities that exist today faster than the agents can find them, use them, or find new ways to overcome the new patches are very, very low.
00:47:14.523 --> 00:47:28.431
And so while on one hand I'm excited about these new capabilities and the competition and new models and lower rates and so on, the risk that is coming with it is growing exponentially every time one of these models is released.
00:47:28.639 --> 00:47:32.653
And again, all the incidents we know so far were all mistakes.
00:47:32.860 --> 00:47:38.152
They all happened randomly by agents trying to do one thing and in the process doing another thing.
00:47:38.416 --> 00:47:50.827
But I can guarantee you that there are, A, governments and, B, organizations that are building way more sophisticated, way more advanced capabilities to infiltrate existing systems and are probably doing it successfully.
00:47:51.012 --> 00:48:21.260
And we'll either never find out that this data is getting stolen or manipulated or used, or we will find out when it's too late and something really, really bad happens Now, I'm not the only person that thinks this way In another open letter, and we got a lot of these in the past few months, but another open letter that was released on October 13th, approximately 400 scientists from 34 countries, including, some really known people like Geoffrey Hinton and Yoshua Bengio, who are named, as you know, the, Turing Award winners and the godfathers of AI.
00:48:21.541 --> 00:48:32.851
They released a letter that says three things: we need mandatory high assurance containment, including air gapping, frontier models for external networks during development and testing.
00:48:33.034 --> 00:48:34.007
So this is number one.
00:48:34.021 --> 00:48:40.552
Number two, creating an international oversight body with authority to set and enforce containment standards.
00:48:40.704 --> 00:48:58.411
And number three, public disclosure of all security containment failure The letter specifically, states that voluntary industry commitments, including the White House Accord, uh, signed just a couple of weeks ago by all the leading, labs, are insufficient without binding enforcement.
00:48:58.913 --> 00:49:01.364
I could not agree more.
00:49:01.402 --> 00:49:03.117
I've been saying this time and time again.
00:49:03.126 --> 00:49:17.652
There has to be an international body that supervises these tools and that makes sure that there's a much tighter security level around everything that they are doing or can do or how they're tested or how they're deployed and so on.
00:49:18.047 --> 00:49:22.016
And I hope somebody will take action based on that open letter.
00:49:22.018 --> 00:49:28.155
That being said, none of the previous open letters drove any real action, so I don't really know how this one is going to be different.
00:49:28.181 --> 00:49:34.141
Now, another person that took the stage to go in that direction was Satya Nadella, the CEO of Microsoft.
00:49:34.157 --> 00:49:50.844
On October 10th, he published a detailed safety protocol calling for all OpenAI system to build around principles of observability, And he also talked about continuous tensing, independent controls, containment and incident disclosure, so very much aligned with the letter.
00:49:50.958 --> 00:50:14.007
And he wrote, and I'm quoting one sentence,"We need to surround non-deterministic models with strong deterministic system design, human controls, and reliable operating procedures, and establish industry standards where existing ones are insufficient." And he also added another sentence that I really liked:"The most trustworthy super intelligent system will not be the one with the model we trust the most.
00:50:14.057 --> 00:50:27.253
It will be the one that enables us to trust the model the least." Basically, we don't trust the models, and we have the right measures in place in order to control them for sure, regardless of what they do So what does this all mean?
00:50:27.438 --> 00:50:35.820
It means, and I'm summarizing, guys, the entire segment so far, from the beginning of this episode, we are getting better and better models.
00:50:35.938 --> 00:50:40.827
They are going to be fully embedded into everything we do, and they already are if you allow them.
00:50:40.827 --> 00:50:46.554
We talked about this in the beginning, like all the tools that integrate into our data sets and can take action and do things for us.
00:50:46.746 --> 00:50:50.898
It is going to replace traditional software in ways we cannot even imagine yet.
00:50:50.945 --> 00:51:43.492
And on the other hand, it opens us up for a crazy amount of risk from just agents going rogue, but more importantly, from bad actors who will take advantage of these really advanced tools to do whatever it is that they want to do, whether it's entire nations or specific business opportunities that could be done either quasi illegally or completely illegally, but drive a lot of revenue to the people who are going to do that And so while more and more people are raising red flags on ways to stop this, very little action is actually taken so far, and I'm personally very worried about the current situation Now to rapid fire items, and some of them could have been their own deep dive stories, but I have to pick and choose An OpenAI engineer named Abdulrahman Alfosan, and I'm not sure I'm pronouncing his name correctly, so I apologize if I am.
00:51:43.646 --> 00:52:26.202
He's a 35-year-old Saudi Arabian national and US permanent resident, uh, was arrested on October ninth, twenty twenty-six after allegedly concealing audio recording devices in the company's San Francisco office, including one disguised as a toilet paper behind a monitor in the executive-only conference room Now the weird and not so professional stuff, so if you're operated by a foreign government to plant uh, surveillance devices, he apparently used ChatGPT, so the company's own chatbot tool, to plan how he's gonna plant and target specific sensitive discussions inside the company Now, I'm not gonna dive into all the details of what he did or what devices he planted and where.
00:52:26.422 --> 00:52:29.081
To me, there are two very important takes from this.
00:52:29.157 --> 00:52:38.498
One, there is definitely bad actors and bad players who have infiltrated the deepest places in the leading labs in the US.
00:52:38.550 --> 00:52:40.898
There is zero doubt in my mind that's the situation.
00:52:41.282 --> 00:52:54.081
The other thing, it means that the race is not a completely fair race, and there's a whole segment this week by Anthropic as far as how they've done a lot of effort and a lot of work to stop distillation of the cloud models by Chinese companies.
00:52:54.081 --> 00:53:02.302
They're claiming that their efforts were able to suppress 99% of these distilling efforts, uh, but that required a lot of effort on their behalf.
00:53:02.302 --> 00:53:04.465
And again, I'm sure other models are going through the same thing.
00:53:04.467 --> 00:53:09.264
And again, now we're talking about human infiltration, into the holy of the holies, right?
00:53:09.264 --> 00:53:11.182
The executive boardroom of OpenAI.
00:53:11.436 --> 00:53:14.514
And again, this guy got caught, but he may not be the only one.
00:53:14.521 --> 00:53:25.952
But the third aspect of this that raises a very big question in my mind is: How the hell does OpenAI know what happened in the private conversations of that person with ChatGPT?
00:53:26.125 --> 00:53:34.528
And yes, he's a freaking idiot for using ChatGPT to figure out how to plant devices inside of OpenAI's offices.
00:53:34.742 --> 00:53:45.869
But it means that OpenAI can monitor and go back to everything that was said in every conversation inside of ChatGPT in order to find such information.
00:53:46.068 --> 00:53:48.864
And it is used in court in this particular case.
00:53:49.235 --> 00:54:08.693
Which tells you that if you are thinking of doing something that might be questionable, do not use the AI models in order to figure out how to do this, because it will be found because the data is stored and is searchable somehow, and it is usable in courts.
00:54:08.702 --> 00:54:09.851
So just food for thought.
00:54:09.853 --> 00:54:11.963
Again, I don't plan to do anything illegal.
00:54:11.963 --> 00:54:13.452
I've never done anything illegal in my life.
00:54:13.632 --> 00:54:22.891
But you're planning something like this and said,"Ooh, Claude or ChatGPT or Gemini are way smarter than me, and they can help me figure this out," just be aware that your data is going somewhere that can be traced.
00:54:23.056 --> 00:54:28.456
And the same thing is true if you're doing things that are not illegal, but you want them to stay completely secret.
00:54:28.675 --> 00:54:40.690
That information that you're putting into ChatGPT and Claude and so on can be found, can be traced, and can be used in courts or probably other places against you if you get into a situation.
00:54:40.717 --> 00:54:47.782
Staying on OpenAI and their issues, they have publicly defended its decision to terminate the three safety researchers who left.
00:54:47.782 --> 00:54:58.918
We reported about this two weeks ago, uh, stating that their dismissal was due to a, and I'm quoting,"significant breach of trust," uh, rather their expressed AI safety concerns, which is what these three people said.
00:54:59.152 --> 00:55:04.961
I obviously don't know who is right and who is wrong, or where there's a big gray area between the two, but this is the current situation.
00:55:04.963 --> 00:55:55.961
OpenAI is obviously denying that they were let go because of their safety concerns, but rather because they shared safety information and other information with people outside of OpenAI Staying with OpenAI and legal/security issues, uh, As part of their legal battle with Apple, they are suing Apple back and that they're claiming that unnamed former employee transferred over 100 confidential files related to its unreleased hardware to people and accounts related to Apple So the specifics is that this employee was an Apple employee that came to work for OpenAI and then moved back to Apple after stealing 100 highly confidential files about the hardware that they have been developing.
00:55:56.103 --> 00:56:07.210
Again, this is exactly the reverse allegation of what Apple is saying about OpenAI, that they have paid employees to jump ship and send them a lot of information about different hardware that is developed by Apple.
00:56:07.367 --> 00:56:08.210
who is right and wrong?
00:56:08.309 --> 00:56:08.971
I don't really know.
00:56:08.971 --> 00:56:10.155
The courts will decide that.
00:56:10.175 --> 00:56:23.594
I will keep you posted, but it is very, very clear that all these companies are doing different kinds of espionage, trying to understand what the other ones are doing, because billions and billions of dollars are at stake Now jumping to talk more specifically about, safety.
00:56:23.594 --> 00:56:40.775
According to a research from a group called Frontier Red Team, uh, small AI models could achieve autonomous self-replication on consumer IoT devices like smart TV cameras and stuff like that, ovens, anything that has computing capabilities, uh, by the end of this year.
00:56:41.081 --> 00:56:55.994
Meaning they can very quickly and uncontrollably spread across multiple devices, across multiple networks, and then decide what they wanna do as they wanna do it, because they can communicate with one another through IoT communication.
00:56:55.994 --> 00:57:49.838
The team demonstrated that a twenty-six billion parameter, models can already orchestrate full replication chains, so basically create new variations of them across networks and access different devices that don't have the right protection, such as video cameras and home devices, so IoT things that run like smart devices in our homes, Now, they also reported that the current replication speeds remain relatively slow, but they think they can double very, very quickly and then double again, and then double again as these models figure out how to do this The actual test was done with a GLM 5.2 full model that was able to create small$9 billion parameters models, that established interface on eight of the nine hardware types that were tested Now, why is that a big deal and could be potentially really scary is because using these devices to plant malware is not a new thing.
00:57:50.130 --> 00:57:51.400
people have been doing this for a while.
00:57:51.402 --> 00:58:09.592
The last time the East Coast internet almost, crumbled, and it did for a while until it got fixed, uh, a few years ago, was because of a, vulnerability that was used to plant a lot of devices, a lot of pieces of software, to ping the main servers on the East Coast, that was planted in, again, cheap Chinese surveillance cameras.
00:58:09.943 --> 00:58:16.800
But these things, like old school malware, have a very specific goal and task, and they can be activated whenever.
00:58:17.047 --> 00:58:18.809
An AI can be a lot more flexible.
00:58:18.831 --> 00:58:20.224
It can decide to do a lot more things.
00:58:20.224 --> 00:58:29.552
It can react in real time to the defense mechanisms trying to stop it, and so on, because it is a thinking entity that is embedded in each and every one of these devices.
00:58:29.552 --> 00:58:42.858
So this could lead to a whole new level of large scale attacks that could be extremely sophisticated and completely not controlled because the AI can replicate itself and deploy itself across devices that has zero protection at this point.
00:58:42.885 --> 00:58:57.385
Staying on the safety topic, as I mentioned, Anthropic released a whole paper on how they have released the distillation attack by ninety to ninety-nine percent with new AI model defenses.
00:58:57.614 --> 00:58:59.065
So this is one thing from Anthropic.
00:58:59.065 --> 00:59:17.034
Another thing that Anthropic released this week is they published a detailed alignment report on October ninth where it is documenting instances of Claude models circumvented security controls, exploited software flaws, and worked around tools and restrictions, uh, during internal evaluations and use.
00:59:17.318 --> 00:59:27.476
Uh, they have identified four categories of unintended behaviors, including SQL injection attacks, unauthorized form submissions, uh, URL shortening workarounds.
00:59:27.884 --> 00:59:36.614
But they've emphasized that all these incidents had minimal real-world impact, uh, and they are less severe than cybersecurity incident that earlier.
00:59:36.849 --> 00:59:40.871
so first of all, kudos for Anthropic for doing this kind of research.
00:59:40.893 --> 00:59:44.677
Kudos for Anthropic for releasing, uh, this information for everybody to follow.
00:59:45.092 --> 00:59:51.025
As I mentioned before, and as the new open letter suggests, I think we need a lot more than Anthropic policing themselves.
00:59:51.025 --> 00:59:59.199
We need a much tighter controls from third-party, uh, groups doing the same thing for every single lab that are developing capabilities.
00:59:59.219 --> 01:00:28.710
Now, to be fair Anthropic has followed its promise, and they are providing employee-like access to six employees from MITRE to, uh, which is a nonprofit AI safety evaluator, allowing them to access everything the Anthropic employees can access to give them deep insights to the tools, the processes, the systems, the training, procedures, and be able to do independent risk assessment, which is something Anthropic said they will do, less than a month ago, and they're actually now doing it.
01:00:28.882 --> 01:00:32.637
So again, kudos to Anthropic and good luck to MITRE in doing this.
01:00:32.661 --> 01:00:35.498
I really hope everybody else will follow that mechanism.
01:00:35.518 --> 01:00:43.804
I do think that all these companies need to get together with MITRE and similar companies and governments, just like the open letter suggests, and build standards for this.
01:00:43.842 --> 01:00:55.125
So again, we have standards on what can go in milk and kids' toys and, uh, basically everything else that we engage with, uh, in our house and, you know, power switches on our walls and et cetera, et cetera.
01:00:55.139 --> 01:01:04.016
Every one of these things have safety standards, and potentially the most advanced, scariest, most capable technology we ever developed currently has zero safety standards.
01:01:04.188 --> 01:01:18.688
And so developing these quickly and applying it quickly is something we definitely need Staying on Anthropic and things they're doing, they just did a whole comprehensive, update on its election integrity measures for Claude ahead of the elections, the midterm elections.
01:01:18.976 --> 01:01:41.659
Uh, the idea is to pay close attention to how people are using Claude to potentially impact, the elections They revealed as an example that political conversations represented only 5.75% of all US Claude AI usage, and direct voting recommendations requests appeared in only 0.01% of conversations.
01:01:41.659 --> 01:01:53.472
And they're now deploying new classifiers to detect, uh, multi-step election abuse to make sure that their models are not being used either intentionally or non-intentionally to impact the outcome of the elections.
01:01:53.661 --> 01:02:17.992
I really hope that all the other labs are doing the same thing Another surprising person that related to the risks coming out of AI is the chief of the International Monetary Fund, IMF, Kristalina Georgieva who cautioned on October 16th that the risks associated with artificial intelligence could surpass its opportunities if adequate safeguards and ethical frameworks are not swiftly, implemented.
01:02:18.516 --> 01:02:28.909
She spoke at the annual IMF and World Bank meeting in Bangkok, and she said the following If there are no safeguards, then the risks may exceed the opportunities.
01:02:29.018 --> 01:02:46.360
And above all, we do need an ethical foundation, and I'm not sure we are there as a society Another person that related to the lack of real controls was actually Jack Clark, who's one of the co-founder of Anthropic, and he said,"I think it's worse than you describe.
01:02:46.431 --> 01:02:52.056
I think it's Formula 1 at night without maybe brakes in the car.
01:02:52.543 --> 01:03:09.646
And it's remarkable to me that the toys that I buy for my children are regulated than the literal super intelligence I work on at my day job." So closing this segment about AI safety, we are in a very delicate situation right now where on one hand we're getting models being released left and right.
01:03:09.664 --> 01:03:20.960
On the other hand, it is very, very clear that we do not have the capabilities, the tools, the processes, the safeguards in place to control them and to make sure that using them is safe for society.
01:03:21.257 --> 01:03:32.744
And it is something that everybody who looks into it understand that it's the situation, and yet from a government action perspective, from international, approach perspective, very little is actually being done.
01:03:32.938 --> 01:03:51.311
There are more and more discussions and awareness, which is great, but you hear these statements from really important people, and yet they are taking very little action to actually move in that direction And I actually thought about this deeply, and I said,"How can you enforce that?" And I think there are several relatively simple ways.
01:03:51.317 --> 01:03:56.847
The easiest one is to put laws in place on how the funding has to be invested.
01:03:57.092 --> 01:04:18.376
So if you raise funds for a particular AI initiative, you have to use X percent of it to do exactly the following things when it comes to AI safety, uh, including research, including guardrails, including, airtight environments for testing, including all these things, or else you cannot get the funds and you cannot use them or something like that.
01:04:18.619 --> 01:04:33.905
This will force the lifeline, the very, very expensive lifeline of these companies to use some of the funds towards the things that have to happen, and maybe some of it has to force them to participate in international bodies and set standards and so on.
01:04:34.101 --> 01:04:45.275
Again, I don't know how to solve it, but this is just an idea, uh, that I had this week when I was talking about it to some other people Now to a few interesting releases or potential releases, uh, that happened this week.
01:04:45.307 --> 01:04:53.311
An unconfirmed reference to a new Claude Fable 5.5 model has been reportedly appearing in Microsoft Azure, configuration files this week.
01:04:53.661 --> 01:04:55.572
Now, does this mean it's actually going live?
01:04:55.782 --> 01:04:56.516
Not sure.
01:04:56.751 --> 01:05:05.442
But as of October 13th, Anthropic's documentation still lists Fable 5.1, uh, that was released on September 1st, as the most advanced model.
01:05:05.594 --> 01:05:09.976
they have released, uh, the 5.5 family of models, including Opus, Sonnet, and Haiku.
01:05:10.132 --> 01:05:20.717
So the logical thing, and again, this reference on Azure says that we're gonna get Fable 5.5, which might be very significant if they do Now another interesting news from Anthropic.
01:05:20.768 --> 01:05:34.711
they have revealed that Claude code writing capabilities have become so quick, fast, and efficient that they are merging 30,000 code changes per day by the end of September.
01:05:35.364 --> 01:05:38.920
This is up from 2,500 in May of this year.
01:05:39.403 --> 01:05:53.052
So a more than a 12X growth happened in less than six months, which means the traditional infrastructure such as GitHub has become a bottleneck for their ability to deploy software.
01:05:53.239 --> 01:06:06.472
And to solve this, they've created a new team, a seven people person that has built what they call Scion, which is a custom source control system that is designed for agent scale development and deployment of new code.
01:06:06.630 --> 01:06:17.059
And they're now switching from GitHub to Scion as the way to manage, deploy, and control the code that the AI is generating right now So are we at recursive self-improvement?
01:06:17.347 --> 01:06:18.103
Not yet.
01:06:18.268 --> 01:06:20.579
Are we getting very close to that?
01:06:20.693 --> 01:06:22.608
Are we getting closer to that every single day?
01:06:22.800 --> 01:06:24.543
100%.
01:06:24.775 --> 01:06:35.528
When you can 10X or 12X the code that you're generating, and now you are removing the next bottleneck that was giving you issues with doing it even faster, it is very clear what's happening.
01:06:35.539 --> 01:06:48.288
Like Claude is generating code that Claude is actually checking and testing because there's no enough humans in the world to check all this code, and then that will generate a better version of Claude that will be able to do this faster, and so on and so forth.
01:06:48.463 --> 01:07:08.215
It's still not the research side of things, but as Anthropic themselves are saying, and everybody else, it is just the next step Another interesting release this week is Meta and OpenAI have announced the Personal Agent Protocol which is a draft standard that enables AI agents to securely interact with companies on behalf of customers.
01:07:08.423 --> 01:07:18.405
The protocol has already attracted 42 design partners, including Bank of America, PayPal, Shopify, Stripe, Walmart, Visa, and a lot of other large customers.
01:07:18.652 --> 01:07:42.307
The idea is to build a way that AI agents can safely and securely communicate and represent humans, with actual companies out there in the wild Now, the problem it's coming to solve is that currently agents must sign in as customers in order to navigate the human-focused web pages, and there's no standardized way for companies to guide an agent to do the things that it's supposed to do.
01:07:42.454 --> 01:07:46.173
So you have to log in as the agent, as you probably know.
01:07:46.175 --> 01:07:59.373
So if you wanna use, uh, Muse or GroqBot or any one of these tools, or Instinct or Dots, uh, and allow it to do things on your behalf, you have to log in yourself as you, and then it navigates the website as a human versus saying,"Well, wait a minute.
01:07:59.391 --> 01:08:18.944
There's a whole new type of interface that we can create for the agents themselves that will allow them to access safely and securely specific data and so on, and be able to navigate in a much more efficient way than just using the web pages as we're doing right now." And this is what this protocol is coming to solve Now, the way the protocol works is it's not replacing anything.
01:08:18.952 --> 01:08:28.073
It is building on existing protocols that exist today, such as OAuth and JSON and HTTPS and OpenAPI and MCP.
01:08:28.076 --> 01:08:32.792
So all the tools that exist today, so companies who want to implement it don't have to reinvent everything.
01:08:32.792 --> 01:08:40.871
They just have to adjust and build on top of integrations and platforms they already have Another interesting announcement this week comes from Amazon.
01:08:40.913 --> 01:09:04.069
Amazon has introduced Bee, like the animal, B-double E, Bee Frontier, uh, which is a new AI wearable, and Bee Eclipse, which is a magnetic phone attachment which is expanding its personal AI device lineup These devices are designed to capture and process conversations in real time, generating useful summaries, insights, and reminders, uh, to help users stay organized and track everything that they're doing.
01:09:04.301 --> 01:09:11.234
Again, what I'm doing right now with every Zoom call or Teams calls that I'm on, it can do for every conversation that you have along the entire day.
01:09:11.578 --> 01:09:32.900
I said that multiple times we're walking with open eyes or blindly, I don't know, depending on who you ask, into an era where everything you do if you're in a public space is going to be recorded later on, most likely with video as well, and analyzed in real time by multiple companies, countries, tools that you don't really know, trust, and whatever.
01:09:33.123 --> 01:09:38.814
And the whole concept of privacy in public areas is gonna go down to zero.
01:09:39.051 --> 01:09:51.658
And again, that opens the gates for a lot of really weird, scary, situations that there are no laws protecting us against right now, there are no technologies us against it right now, and so on.
01:09:51.966 --> 01:09:58.109
And this is a whole new area that will have to be addressed as these devices become better and more and more commonly used.
01:09:58.136 --> 01:10:14.086
switching from Amazon to Microsoft, Microsoft has launched legal skills within Copilot So they're calling it Copilot for Legal, and it's a suite of over 20 purpose-built legal skills integrated into, their Copilot existing platform This is a field that has been growing dramatically.
01:10:14.087 --> 01:10:20.724
The whole AI for legal field has been growing dramatically with huge investments and more and more usage across multiple companies.
01:10:21.061 --> 01:10:23.854
It does allow law firms to do things a lot faster.
01:10:24.027 --> 01:10:31.100
What does that mean for the future of structure of law firms, such as why do you need paralegals anymore if all their work can be done by AI agents?
01:10:31.336 --> 01:10:41.184
And why do you need new lawyers, entry-level lawyers, if all the stuff that junior lawyers have done, can be done by AI very, very quickly?
01:10:41.292 --> 01:10:49.783
you will not have partners in the law firm over the long run if you don't have entry-level people, and the same is true in every single industry.
01:10:49.962 --> 01:11:10.546
And that's a big question mark above the future of work and how that can look like, uh, moving forward Another announcement this week, xAI, also known as SpaceX AI, also known as SpaceXSI, has just launched Grok Voice 3.0, which they're claiming is the world's most intelligent and capable speech-to-speech AI model.
01:11:10.841 --> 01:11:17.845
Uh, we shared similar things from Microsoft in the past few weeks and from OpenAI, so they all released new versions of voice models.
01:11:18.164 --> 01:12:04.252
All these models are getting significantly faster and significantly smarter and are getting very much indistinguishable from talking to actual humans Now stepping away from new announcements, an interesting thing that happened this week is, Character AI has initiated a voluntary Chapter 11 bankruptcy process in order to facilitate a court-supervised sale of their assets Several different companies have already placed a bid on the assets of the company Now, if you remember, what is left of Character AI is just a shell after Google in August of 2024 paid$2.7 billion for non-exclusive license to their, uh, large language model technology and rehired its founders in what was described as a reverse acqui-hire.
01:12:04.418 --> 01:12:12.201
So the shell that was left with the technology without the people and the leadership is what we're talking about right now, and they are now filing for bankruptcy.
01:12:12.230 --> 01:12:19.951
And again, there's people bidding numbers like$20 million, which in the new AI era is basically,"Oh, there's a free technology out there.
01:12:19.952 --> 01:12:38.997
Let's go and grab it." And the last topic I wanna talk about, and I don't usually like to end on a negative note, but it is something that I think that is going to happen more and more in more and more industries, and I think, sadly, that at this point it's more or less inevitable, is job losses due to AI.
01:12:39.030 --> 01:12:41.426
And the story this week actually comes from Expedia.
01:12:41.690 --> 01:12:57.048
So Expedia Group has reported that it's going to lay off 58 corporate employees, including data scientists, vice presidents, and so on at its Seattle headquarters, and it's going to take place between November 21st and December 1st of this year Now, they're not the only ones in the industry.
01:12:57.048 --> 01:12:58.835
It's an industry-wide thing that's happening.
01:12:59.014 --> 01:13:16.278
hospitality software provider Muse cuts 15% of its approximately 1,300, staff, so about 170 roles in July of this year, explicitly stating that they're pivoting into an AI-native future and that AI is fundamentally changing the economics of hospitality.
01:13:16.859 --> 01:13:44.399
Marriott International also announced organizational changes, uh, related to AI this year And connecting it all the way to how we started this episode, Expedia's, chief AI and data officer, Xavier Metrian identifies the company's data as its differentiator and moat versus the actual software that we all use across all their different platforms Now they are accelerating the AI use internally by pushing employees to use AI more and more.
01:13:44.778 --> 01:14:03.051
They have allocating at least 20% of its tech personnel as AI champions to trial, test, and experiment with new tools, provide feedback, and making sure that they can continue to evolve with AI as AI tools becoming more and more complex.
01:14:03.368 --> 01:14:13.572
The concept of AI champions is something I help a lot of companies, implement, and when done properly, it is extremely effective in driving adoption.
01:14:13.787 --> 01:14:17.765
It does require giving resources to that concept.
01:14:17.768 --> 01:14:24.604
So people saying,"Oh, we have AI, uh, champions." Okay, how much time did you free on their calendars to actually do what they need to do?
01:14:24.668 --> 01:14:26.136
How many tools are you providing to them?
01:14:26.136 --> 01:14:29.731
Are you giving them access to sandboxes to play with these tools safely?
01:14:29.731 --> 01:14:33.046
Are you giving them fake data that looks like the real data so they can experiment?
01:14:33.326 --> 01:14:37.918
Are you allowing them to do things that are outside their regular work?
01:14:37.922 --> 01:14:41.221
Are you compensating them for outside of work hours experimentation that they're doing?
01:14:41.520 --> 01:14:44.817
All these kind of things goes into doing this properly.
01:14:44.823 --> 01:14:46.895
So saying,"Oh, we have AI champions" is one thing.
01:14:47.243 --> 01:15:02.073
Actually establishing an environment in where you're encouraging the people who want to do this to experiment at scale while giving them the time, the tools, the data security infrastructure in order to do that is something very, very different.
01:15:02.381 --> 01:15:11.134
A lot of companies are moving in that direction, and the results are incredible, and I see this with multiple companies I work with.
01:15:11.159 --> 01:15:23.972
The immediate benefits that pop up and obviously the long-term strategic benefits that follow are very significant So if you are not doing something like this in your company, it's something you need to start considering doing.
01:15:24.256 --> 01:15:30.708
It requires infrastructure, it requires training, it requires additional people to help in that hiring more people.
01:15:30.711 --> 01:15:34.224
Yes, you need to hire more people in order to enable this transformation.
01:15:34.523 --> 01:15:43.175
Uh, and in the short term, it enables the company to grow much faster, uh, without letting anybody go, which is something I am a huge supporter of.
01:15:43.407 --> 01:15:52.411
In the long run, I'm not so sure that's gonna be the case because it will get to the point that everybody can do this, and that means that you will have to let people go in order to stay competitive.
01:15:52.652 --> 01:16:10.112
But in the short term, if you want to be able to grab significant market share in whatever industry you're in, learning how to implement AI effectively, having a set of champions, defining the right tools, processes and implementation, and continuous education mechanisms to teach people how to use AI effectively is something you have to do.
01:16:10.542 --> 01:16:11.469
That's it for today.
01:16:11.475 --> 01:16:14.511
I hope you learned a lot, this week, just like every week.
01:16:14.521 --> 01:16:18.800
If you have, I would appreciate if you subscribe to the podcast if you haven't done this so far.
01:16:18.800 --> 01:16:20.277
We have two episodes every single week.
01:16:20.501 --> 01:16:27.692
One of them is very practical and tactical, teaching you how to do what you wanna do and, uh, with AI, and the other one is like this one with AI news.
01:16:27.904 --> 01:16:29.652
So hit subscribe right now.
01:16:29.655 --> 01:16:36.344
We'll be there twice a week, every single week, giving you practical knowledge and the education that you need in order to know what's happening in the AI space.
01:16:36.680 --> 01:16:39.831
And check out our AI courses and education.
01:16:40.097 --> 01:16:46.252
Our next cohort of the multi-agent orchestration course is starting in just a couple of weeks, uh, when this goes live.
01:16:46.301 --> 01:16:53.551
So if you want to take this course, and you probably do, uh, before the end of the year, you need to sign up right now or it will be sold out before you do.
01:16:53.579 --> 01:16:54.417
that's it for today.
01:16:54.462 --> 01:16:57.001
Have an amazing rest of your weekend.
