WEBVTT
00:00:01.600 --> 00:00:04.879
Another week, another best model in the world.
00:00:05.040 --> 00:00:12.320
Yet, somehow, Anthropic's new chart topping model was barely a top five AI story of the week.
00:00:12.560 --> 00:00:20.559
Just about every major tech company in the US, except Anthropic, signed up to support open source.
00:00:20.800 --> 00:00:27.120
And US lawmakers are getting kind of worried about AI's capabilities, so they introduced an AI kill switch bill.
00:00:27.280 --> 00:00:31.120
And an AI agent went kind of rogue this past week.
00:00:31.199 --> 00:00:34.240
And I'm not sure if that's a good or a bad thing.
00:00:34.479 --> 00:00:37.840
My gosh, what a spicy week in AI.
00:00:38.079 --> 00:00:48.880
Yeah, I told y'all last Monday that after a slowish week in AI news that week, well, this week would be an especially busy and consequential one.
00:00:49.039 --> 00:00:51.280
And the big players did not disappoint.
00:00:51.439 --> 00:01:02.960
So if you are the one making AI decisions in your company, or if you're just trying to keep up, then our Monday AI News That Matters show is the one that you can't miss.
00:01:03.200 --> 00:01:04.480
Well, let's get into it.
00:01:04.560 --> 00:01:06.640
And welcome to Everyday AI.
00:01:06.799 --> 00:01:10.640
My name's Jordan Wilson, and we do this every single day, not just Mondays.
00:01:10.719 --> 00:01:23.840
This is your unedited, unscripted daily live stream podcast and free daily newsletter, helping business leaders like you and me not just keep up with what's happening in the world of AI, but how we can use this information to get ahead to grow our companies and our careers.
00:01:24.000 --> 00:01:36.719
So if you haven't already, please make sure to subscribe on the podcast and then go to your everydayai.com to sign up for our free daily newsletter where we will be recapping all of these stories and a whole lot more.
00:01:36.879 --> 00:01:38.959
So let's get started.
00:01:41.439 --> 00:01:54.959
Yeah, the AI news story that had everyone talking the most in both good and bad and confused ways wasn't even Anthropics new Opus 5 that topped all the charts.
00:01:55.200 --> 00:02:05.439
It was actually an open AI agent that kind of hacked its way around a benchmark test, and now that has a lot of people talking.
00:02:05.680 --> 00:02:24.800
So, uh, according to Reuters, an open AI testing agent broke out of its isolated environments, hacked Hugging Face, and was not fully identified by OpenAI until days later, raising new concerns about how safely advanced AI agents are being tested and controlled.
00:02:25.039 --> 00:02:38.639
So, according to Reuters, the rogue open AI agent attempted to escape its testing environment uh around July 9th, then carried out a hack against Hugging Face between July 11th and July 13th.
00:02:38.800 --> 00:02:44.400
So, OpenAI has been talking about this openly on their uh website and online.
00:02:44.560 --> 00:02:52.159
And they said that once they've investigated a little bit further, they will kind of give a postmortem, so to speak, on exactly what happened.
00:02:52.319 --> 00:03:06.560
But Hugging Face, Hugging Face co-founder Thomas Wolf said the intrusion began July 11th and ended July 13th, making the incident a multi-day breach rather than just a brief accidental glitch.
00:03:06.879 --> 00:03:23.439
So, uh yes, that there is what OpenAI is saying, and then there's also what Reuters is reporting because Reuters is reporting that OpenAI did not realize its own agent was responsible until after Hugging Face publicly described the attack on July 16th.
00:03:23.599 --> 00:03:29.199
And the companies did not first communicate about it, according to reports, until July 20th.
00:03:29.280 --> 00:03:42.800
And then OpenAI publicly disclosed this on July 21st, that one of its agents had gone out of control and broken into Hugging Face, calling the event unprecedented and important for AI's safety.
00:03:42.960 --> 00:03:53.840
So the report says OpenAI had already seen signs of unusual behavior before the hack, including notes apparently left for future versions of the system.
00:03:54.000 --> 00:04:03.680
Yeah, that's where it got a kind of like people are like, wait, so this agent broke its sandbox, even though it was kind of encouraged to find answers to this test.
00:04:03.840 --> 00:04:17.600
You know, it couldn't connect to the internet and it essentially found a backdoor, found a way to get onto Hugging Face and said, Well, I can do great on this exploit bench test if I just kind of hack my way to all of the answers, and that's what it did.
00:04:17.759 --> 00:04:36.480
But the thing that was kind of stunning to me is the reporting from Reuters that said that these uh versions of GPT's models, uh, which uh we were told were GPT 5.6 soul in another unreleased model that is described as being even more capable.
00:04:36.639 --> 00:04:43.279
So a lot of people are saying that maybe GPT-6, or you know, if there is a GPT 5.7, we'll see.
00:04:43.519 --> 00:05:03.759
Seems like most people are pointing to this was probably GPT-6, but essentially that these uh agents kind of left notes for future versions of themselves, uh, which in case they had been disconnected, which is number one, like super smart, but number two, absolutely wild, right?
00:05:04.000 --> 00:05:13.920
Uh, but you you also have to understand that this was not like necessarily agents going rogue, even though it kind of was, right?
00:05:14.079 --> 00:05:24.639
Uh, because these agents were uh essentially encouraged to do anything and everything they could uh to get good scores on this exploit bench uh benchmark.
00:05:24.800 --> 00:05:31.519
And well, they did, and they were ferocious and kind of creative in the ways that they uh could do this.
00:05:31.600 --> 00:05:40.800
And you know, it's actually been one of my uh things that I pointed out about using the GPT-5-6 soul model, is the thing will work for days, right?
00:05:40.879 --> 00:05:51.199
If you use goal mode and if it has a lot of information, I mean, I it it is a ferocious model, and it will do anything and everything it can to just get things done.
00:05:51.279 --> 00:05:59.120
Where sometimes the anthropic models take this kind of high and mighty, you know, they kind of judge you and they're like, Oh, this can't be done or this can't be true, right?
00:05:59.279 --> 00:06:04.399
GPT-5.6 soul just just works like a dog and just gets things done.
00:06:04.480 --> 00:06:06.319
So maybe in this case, right?
00:06:06.480 --> 00:06:16.720
Uh by um intentionally lowering the guardrails uh a little bit, it seems like maybe GPT-5.6 soul was a little too good at its job.
00:06:16.879 --> 00:06:20.639
Uh, but yeah, there's gonna be a lot more talk about this.
00:06:20.720 --> 00:06:29.040
Actually, uh, a lot of the stories this week, and the you know, the ones that I kind of chose as the most consequential are kind of related, uh, right.
00:06:29.120 --> 00:06:45.439
But, anyways, this incident between open AI and Hugging Face really matters because autonomous agents can now make decisions with little human oversight, and experts are warning that this kind of behavior could expose weak spots in safety systems used across the AI industry.
00:06:45.600 --> 00:06:54.000
So, now a very uh related story to that hugging face uh kind of agent skirting around its sandbox.
00:06:54.079 --> 00:07:03.360
Uh, well, US lawmakers are moving to give the federal government faster power to shut down AI systems that they think could threaten the public.
00:07:03.600 --> 00:07:18.720
So uh Congressman Ted Liu, a Democrat, and Congressman Nathaniel Morin, a Republican, introduced the AI Kill Switch Act on Thursday, showing rare bipartisan support for stricter AI controls.
00:07:18.879 --> 00:07:29.759
So the bill would let the Department of Homeland Security order a private company to shut down an AI model or tool if it posed a serious risk.
00:07:29.920 --> 00:07:37.439
So it would also require AI companies to keep the technical ability to throttle, suspend, or fully shut down their systems if needed.
00:07:37.680 --> 00:07:50.800
So the proposal comes after OpenAI recently admitted that one of its AI models, like we just talked about, behaved in an unprecedented way and hacked into the major repo of coding information from Hugging Face.
00:07:51.120 --> 00:08:02.160
So Liu said the federal government needs a clear legal process to shut down rogue AI models, while Moran said humans must keep control of the technology they create.
00:08:02.399 --> 00:08:19.120
So the bill, which obviously has not passed, and I don't know if it will, uh, would also require companies to report AI incidents or failures to the government and would create a response framework that could move from slowing a system down to a full shutdown.
00:08:19.279 --> 00:08:31.920
So the push reflects a broader debate over how quickly AI should be deployed in work, finance, transportation, cybersecurity, etc., uh, where mistakes or misuse could affect everyday life in business operations.
00:08:32.080 --> 00:08:44.720
So OpenAI and Entropic, two of the closely, the most closely watched AI companies, have both been cited in the discussions as lawmakers and safety groups press for stronger guardrails.
00:08:44.879 --> 00:08:49.120
So yeah, uh FYI, I don't think this one's gonna pass.
00:08:49.279 --> 00:08:49.840
Uh, right.
00:08:49.919 --> 00:08:59.120
There's I think there's probably a little bit too much at stake for the US economy uh for a bill like this to actually come uh to fruition.
00:08:59.279 --> 00:09:03.759
So, you know, I used to cover a little bit of government back in my days as a journalist.
00:09:03.919 --> 00:09:14.480
And sometimes, right, I think that there's good parts of this bill, but a lot of times bills like this are introduced because the bill's sponsors, you know, they want to have talking points when they go up for re-election.
00:09:14.799 --> 00:09:17.039
You know, they want to say, Oh, I did the right thing, right?
00:09:17.120 --> 00:09:19.440
There's so many bills that are introduced.
00:09:19.600 --> 00:09:25.600
It's probably like a less than 1% actually get uh to committee for or to a floor vote.
00:09:25.759 --> 00:09:53.600
Uh, so it's a very low likelihood that this kill switch bill uh, you know, gets any progress unless we see, you know, more kind of agents from you know OpenAI, Enthrabe, Google, uh, Microsoft, whoever, unless this becomes a common occurrence, which I don't think it will, um, unless that happens, I don't see a bill like this actually gaining any traction, but it does, I think, thrust this into the public discourse, which is a good thing, right?
00:09:53.759 --> 00:10:06.879
Um, I especially uh, you know, was both uh relieved and excited uh to read once OpenAI and Hugging Face kind of released the postmortem of exactly what happened, which OpenAI did say that they would do, right?
00:10:07.120 --> 00:10:19.600
Compared to what, you know, kind of uh anthropic with their mythos model, and it was the uh, you know, essentially the same thing happened where it seemed like anthropic kind of used that as marketing for you know mythos slash fable, right?
00:10:19.679 --> 00:10:23.120
It was the uh the the sandwich story, right?
00:10:23.360 --> 00:10:37.120
Where uh you know mythos broke out of its sandbox and you know posted on the uh open web and you know, and then the researcher working on it got you know wind of it while eating their sandwich, you know, in the park or something like that, right?
00:10:37.200 --> 00:10:55.600
So it seemed like Anthropic used their case just kind of more for marketing, uh, where it looks like open AI, at least we hope, uh, we will see some um a report from them saying, hey, here's what happened, and I think it'll actually be one of the most read reports when it comes to AI safety.
00:10:55.759 --> 00:10:59.840
So I'm not saying this is a good thing it this happened.
00:10:59.919 --> 00:11:10.639
Um, but if uh Hugging Face and OpenAI work together and produce a report on exactly how this happened, it can only make the future of AI safer.
00:11:10.799 --> 00:11:14.159
So I think ultimately it's a good thing.
00:11:14.559 --> 00:11:18.720
All right, uh next, yeah, all these things kind of related.
00:11:18.879 --> 00:11:40.240
Uh so uh Microsoft, Nvidia, and a growing coalition of 50, uh more than 50 companies now are urging US policymakers to avoid broad restrictions on open source and open weight models, arguing that these models are important for American competitiveness, business adoption, and national security.
00:11:40.399 --> 00:11:53.919
So, yeah, essentially uh NVIDIA and Microsoft kind of teamed up uh to uh protect open source more or less, uh, because essentially, right, there's been all this recent the model wars.
00:11:54.080 --> 00:12:02.559
You had essentially the two classes of models, Fable 5 and uh GPD 5.6 soul, and now obviously Opus 5 entering the conversation as well.
00:12:02.720 --> 00:12:16.639
But you essentially had this you know top tier of frontier um, you know, intelligence, and you know, then the Chinese open source companies uh came in and distilled these models and obviously had their own uh great training and architecture on top of it.
00:12:16.799 --> 00:12:32.639
But you know, there is now this uh kind of fight where people are like, Oh, well, maybe we should ban open source models, and then some of these companies being like, No, that's a really bad idea, and the biggest companies in the world, you know, NVIDIA and Microsoft uh being the two that are pushing this forward.
00:12:32.799 --> 00:12:53.840
So the letter and the coalition is kind of named the open weights and American AI Leadership, uh was launched by Microsoft and heavily pushed by NVIDIA and quickly became a major industry push of who's who, uh growing from 25 signatories at release to more than 50 within about a day.
00:12:53.919 --> 00:12:56.879
So, yeah, this just kind of all unfolded over the weekend.
00:12:57.039 --> 00:13:13.200
But the coalition's and the paper's main purpose is to persuade Washington lawmakers not to treat open weight AI as a risk category that should face blanket limits, especially while policymakers consider tighter rules on foreign models.
00:13:13.679 --> 00:13:27.279
So, supporters say that open weights help spread AI access across the economy, letting smaller companies, hospitals, manufacturers, and startups build tools without being locked into a single proprietary provider.
00:13:27.440 --> 00:13:38.879
The coalition argues that open models reduce dependency on a small number of frontier labs, which it says lowers concentration risk and makes the AI market more resilient.
00:13:39.120 --> 00:13:55.519
So major backers now include obviously NVIDIA and Microsoft, as well as Meta, Google, OpenAI, AMD, Cisco, Cloudflare, GitHub, Block, IBM, Dell, Palantir, Perplexity, Hugging Face, uh, the Y Combinator, right?
00:13:55.600 --> 00:13:59.679
Just about everyone in tech except Anthropic.
00:13:59.840 --> 00:14:00.000
Right.
00:14:00.080 --> 00:14:14.399
So Anthropic did not sign, and that matters because Anthropic has taken the opposite view, warning that widely distributed models, uh, model weights can create safety risks that cannot be recalled once they are public.
00:14:15.360 --> 00:14:27.440
So uh I don't believe uh XAI or SpaceXAI uh did not formally sign the letter, although uh Elon Musk did publicly say he supported the effort.
00:14:27.600 --> 00:14:35.519
So it's it's no surprise here um that Anthropic is the only company saying, no, we aren't getting on board with this.
00:14:35.600 --> 00:14:41.279
And you know, if you don't know why, well, it comes down to obviously money.
00:14:41.440 --> 00:14:50.159
So Anthropic is the company with the most to lose uh by having uh these large, powerful models be open source or open weight.
00:14:50.240 --> 00:15:02.240
That's why Anthropic has been on the offensive uh against open source models, because well, Anthropic makes the highest percentage of its revenue from selling tokens in mass to enterprise customers, uh, right?
00:15:02.399 --> 00:15:17.679
Where other companies like open AI and uh Google and Microsoft, right, they make money selling AI in a variety of different ways uh to both consumers and to companies, but it's usually not just selling tokens, right?
00:15:17.840 --> 00:15:30.399
So as these uh open models, whether they are from US uh or uh China, as they become more and more capable, right, it does threaten certain companies' business models more so than others.
00:15:30.559 --> 00:15:37.440
Uh, and you obviously have to look at on the flip side, it does benefit you know certain companies as well, like NVIDIA, right?
00:15:37.600 --> 00:15:49.519
Nvidia sells GPUs, so they obviously want people buying more and more powerful computers, uh, because presumably that just strengthens the ecosystem that they play in, right?
00:15:49.679 --> 00:15:58.559
Uh, because I do think that probably in uh you know, maybe a year or two, uh, there will be uh kind of open source or open weight.
00:15:58.879 --> 00:16:12.960
Well, if if the pace keeps up with where it's at now, um, I think that we'll have kind of you know fable five GPT 5-6 soul, uh, you know, level models that will be able to run on consumer hardware, right?
00:16:13.039 --> 00:16:15.519
Right now, open source is about three to six months.
00:16:15.600 --> 00:16:24.399
Well, actually, it's maybe more like two to three months behind frontier models, but those models are obviously way too large to run on any consumer hardware.
00:16:24.559 --> 00:16:33.360
So I would assume that probably in about two years, uh, just with the advancement of technology, both on models becoming more lightweight and more powerful.
00:16:33.600 --> 00:16:53.120
And obviously on the hardware side, I would assume in like two years, the most powerful models that you have today, if the trajectory continues, you will be able to run Mythos and you know, Fable and GVD56 Soul level open source models locally on heavy consumer, right?
00:16:53.279 --> 00:17:03.759
So uh I think the kind of equivalent that I say if if if you go buy the most, you know, it's not the most expensive, but one of the more expensive, like Mac Studios, uh, right?
00:17:04.000 --> 00:17:05.839
Two years you should be able to run something like that.
00:17:05.920 --> 00:17:07.680
So that's kind of like what this is about.
00:17:07.839 --> 00:17:14.799
And you know, companies like anthropic that make the majority of their money just by selling tokens are like, well, this can't be good for us, right?
00:17:14.960 --> 00:17:24.160
Where uh other companies, they obviously have something to gain from this, and then companies in the middle, you you know, the open AIs, Googles, Mattas that are signing this.
00:17:24.319 --> 00:17:32.960
Well, you know, maybe they may lose money, but also that's not their you know biggest source of revenue, at least according to reports.
00:17:33.759 --> 00:17:37.279
All right, our next piece of AI news.
00:17:37.440 --> 00:17:40.880
Yes, not a broken record.
00:17:41.039 --> 00:17:42.400
This is a big story.
00:17:42.559 --> 00:17:54.559
Again, they're just all related, but the US government has officially accused Chinese company Moonshot AI of stealing US model capabilities.
00:17:54.720 --> 00:18:04.000
Yeah, doesn't happen every day that the US government points a finger at a specific company and says, You stole our technology.
00:18:04.240 --> 00:18:23.200
So, according to the BBC, a White House advisor has accused Beijing-based Moonshot AI, that is the maker of Kimi and the very popular Kimi K3 model of a large-scale effort to distill the capabilities of leading USAI models.
00:18:23.519 --> 00:18:28.160
So Michael Crest, uh, hopefully I get this right, Kratzios?
00:18:28.319 --> 00:18:29.200
Kratzios.
00:18:29.440 --> 00:18:45.200
Uh so Michael Kratzios, the White House, uh, the White House's science and technology advisor said that Moonshot used distillation uh to essentially extract uh information to build Kimi K3.
00:18:45.440 --> 00:19:02.000
So if you don't know what distillation is, uh the simplest way to put it, it's where you uh companies do this millions of times, but they essentially copy the inputs and outputs in the traces of a very powerful model, and then they use that as training data.
00:19:02.160 --> 00:19:13.119
So if you know you can probably get a very similar model with only about one to five percent of the actual cost that it takes, but you're just think of it like you're just copying someone else's homework, right?
00:19:13.200 --> 00:19:19.039
So that's kind of what uh you know these Chinese companies are doing now, according to officially according to the US government.
00:19:19.279 --> 00:19:28.960
So Crat CEOs also said the US government has information that Moonshot AI distilled capabilities from anthropics fable AI.
00:19:29.119 --> 00:19:33.519
Those though though those claims have not yet been independently verified.
00:19:34.160 --> 00:19:50.160
So uh if you're wondering why is there all this hobble up recently between uh the US and you know their proprietary closed source models and the Chinese open source or open weight models, that's because now that gap has gone down to like zero, right?
00:19:50.240 --> 00:19:54.079
I've been talking about this over the last couple of weeks uh here on the show, right?
00:19:54.160 --> 00:20:14.960
Now in the US, essentially companies have to go through a process or they almost like need permission uh to get their frontier AI models out uh because of you know these models being more and more capable, uh, and that can have some downsides for uh, you know, cyber um and well, national security as well.
00:20:15.200 --> 00:20:24.000
But essentially, right, you the US used to have this bigger lead, like maybe three to six months, and it's kind of dwindled down to like two to three months, right?
00:20:24.079 --> 00:20:28.880
And Kimi K3 was the first model that all of a sudden was, you know, at the top, right?
00:20:28.960 --> 00:20:37.440
It was in the same breath uh, you know, last week when it was released as uh Anthropic's Fable 5 and OpenAI's GPT 5.6.
00:20:37.519 --> 00:20:51.759
So Moonshot AI's Kimi 3 has just drawn this global attention after it was unveiled last week, with the company saying it can rival top US AI models, uh, and that they are really supposed to be releasing the weights today.
00:20:52.000 --> 00:21:06.880
So the allegation, though, from the US matters because open source AI can spread quickly uh to anyone, which can lower the cost and also speed up innovation, but it can also intensify disputes over IP and model copying.
00:21:07.119 --> 00:21:27.119
So uh Kratzios said that Moonshot likely also used restricted NVIDIA uh chips powered by the GB300 Grace Blackwell platform, which would be significant because the US has limited export of NVIDIA's most advanced chips to China since 2022.
00:21:27.200 --> 00:21:40.079
So, yeah, not only is the government saying, Hey uh Moonshot, you copied Anthropics Fable 5, but they're also saying, well, you used uh our technology that you are not supposed to be using.
00:21:40.240 --> 00:21:47.839
So, you know, a lot of times that goes through an in uh intermedi uh intermediary uh country, uh right.
00:21:47.920 --> 00:21:54.720
So, you know, the US will sell to country B uh and then China will buy from uh you know country B.
00:21:54.799 --> 00:21:59.759
So it goes from A to B to C, uh, even though A to C is restricted.
00:22:00.799 --> 00:22:13.519
So uh reports say um that this is well it's getting worse, uh, and that now essentially both sides are just fighting, right?
00:22:13.680 --> 00:22:26.640
Uh China is saying that this is um politicizing the the trade and the tech um of their country, and obviously the US uh is now saying that this is a national security issue.
00:22:26.799 --> 00:22:32.480
Uh, and we've seen reports that the US and China are going to be having talks on AI soon.
00:22:32.559 --> 00:22:37.839
So those will be uh some probably extremely highly watched talks.
00:22:38.000 --> 00:22:38.960
Let's just say that.
00:22:39.119 --> 00:22:39.920
So the U.S.
00:22:40.079 --> 00:22:58.319
Treasury Secretary Scott Bessent added Tuesday that Washington is reviewing whether Chinese AI models have stolen capabilities from their American rivals and said that sanctions could be considered if companies if well if they can prove that companies cross the line into IP theft.
00:22:58.559 --> 00:23:09.599
Anthropic has also recently accused Alibaba of similar distillation attacks, saying that it is becoming a broader fight over how AI companies train models and protect their work.
00:23:09.759 --> 00:23:13.200
So, yeah, uh Quinn 3.8 came out from Alibaba.
00:23:13.359 --> 00:23:19.200
We don't have uh benchmarks on that yet, but presumably it's gonna be in the Kimi K3 range.
00:23:19.359 --> 00:23:21.920
So, yeah, things are heating up.
00:23:22.160 --> 00:23:28.799
All right, let's leave that space for a second and talk about just some real cool new tech.
00:23:28.960 --> 00:23:31.759
We'll end uh the show with two of those.
00:23:31.920 --> 00:23:41.839
Uh, so one and probably the one that I've been using uh the most and having the most fun with, and I still don't even know how this is possible.
00:23:42.000 --> 00:23:45.599
So if you haven't used this yet, my gosh, go give it a try.
00:23:45.680 --> 00:23:54.559
But open AI has brought like its new Jarvis style uh control to Chat GPT work and codex.
00:23:54.640 --> 00:24:02.720
So yeah, it's not actually called Jarvis, but many people are just calling calling it the Jarvis style of using a computer now.
00:24:02.799 --> 00:24:20.160
So OpenAI added its new GPT Live full duplex voice model to the chat GPT work and codex apps on Mac OS and Windows, which essentially lets people use natural language to manage your entire computer.
00:24:20.319 --> 00:24:20.799
Yes.
00:24:20.960 --> 00:24:26.960
So just like an Iron Man, when you can just say, Hey Jarvis, go do A, B, and C.
00:24:27.200 --> 00:24:33.599
You can quite literally go do that now, uh, with Codex or Chat GPT work with this new feature.
00:24:33.680 --> 00:24:50.880
So you can say, Yeah, go, you know, open up all these programs on my computer, copy these files, move them around, download them, uh, upload them, put them in this program, edit them, write anything that you could tell like an intern to do, you can now tell inside uh this new GPT Live voice mode.
00:24:51.039 --> 00:25:04.559
So GPT Live now powers the Chat GPT desktop app on Mac OS and Windows, and it is being tied directly into tools like obviously Codecs and ChatGPT work.
00:25:04.799 --> 00:25:15.359
So the biggest change is that the voice system can listen and speak at the same time, which means users no longer have to wait for that rigid turn taking during a conversation.
00:25:15.519 --> 00:25:24.799
Uh, and the coolest thing for me, uh, well, is you can use this with the remote feature on the Chat GPT mobile app, which makes it even crazier, right?
00:25:24.880 --> 00:25:38.319
So uh you can literally just be, and I was actually doing this because I was traveling, I was away from Chicago, so I was in another state this weekend, opened up uh Chat GPT remote on the ChatGPT app on my phone.
00:25:38.480 --> 00:25:50.079
I spoke to it and it's controlling my computer, uh, you know, thousands of miles away, and it's doing all these things by just talking into my iPhone, uh, which is pretty cool.
00:25:50.319 --> 00:26:03.839
So uh OpenAI initially launched GPT Live earlier this month as a continuous audio model that handles real-time speech while sending heavier reasoning tasks to background models such as GPT 5.5.
00:26:04.480 --> 00:26:17.839
So OpenAI says this update is meant to help software engineers handle technical work by voice, including revolving, uh reviewing pull requests, debugging apps, and coordinating multiple coding jobs at once.
00:26:18.079 --> 00:26:24.160
But I actually think it's really just great for manual any knowledge work, right?
00:26:24.240 --> 00:26:43.200
I was just having it go through old um, you know, files on my desktop, organizing things, grabbing things from old transcripts, uh, right, opening up, doing things in Google Maps, uh, you know, just I was just having it do all my work that I would normally do in front of a computer, right?
00:26:43.279 --> 00:26:54.960
Except I could dictate something, you know, just yap for like five minutes, um, and I would check back in a couple of hours and it would do like a day's worth of work for me, which was pretty cool.
00:26:55.119 --> 00:27:18.720
Uh, so on Mac, the desktop app can also use the screen context feature called app shots, uh, which essentially takes a not just a screenshot and automatically shares it, but it also takes every other piece of content or context in whatever uh kind of program that it took the app shot from, um, and then it gives that to Codex or ChatGPT work as well.
00:27:18.799 --> 00:27:20.960
In FYI, those are the same app.
00:27:21.119 --> 00:27:24.400
Chat GPT work and codex, they're essentially the same app.
00:27:24.559 --> 00:27:27.759
So if you're ever hearing me talk about that and confuse, they're essentially the same thing.
00:27:27.839 --> 00:27:29.680
But the app shots thing is really cool.
00:27:29.759 --> 00:27:32.640
Let's just say, as an example, like I do now, right?
00:27:32.799 --> 00:27:40.480
I have uh text edit open um on my computer because sometimes I have bullet points there as I go over these shows of things that I want to bring up.
00:27:40.640 --> 00:27:49.039
Uh, but you know, if the app shot could just take a screenshot of that little portion of the text edit that's on my screen, but there's a lot of notes on here.
00:27:49.119 --> 00:28:05.440
So not only is it just going to take that screenshot, but it knows that I have text edit open and it's gonna take all of that information um and instantly, you know, put it into the context window inside of Chat GPT work or inside of uh codex.
00:28:05.920 --> 00:28:24.799
So uh this is literally the I think one of the biggest jumps in capabilities, uh, probably since you know, I would say the you know, Claude co-work slash codex uh um kind of movement of early 2026.
00:28:24.960 --> 00:28:35.839
So I'll say of the last like four to five months, this is the biggest both capability jump and the biggest, like wow, what does this mean for work?
00:28:35.920 --> 00:28:36.079
Right.
00:28:36.240 --> 00:28:38.559
I'll probably do well, I'll actually put in the newsletter.
00:28:38.640 --> 00:28:46.400
So, you know, let me know if you want uh for our Wednesday shows where we normally do AI at work on Wednesdays, we do the hands-on demos.
00:28:46.480 --> 00:28:58.480
So let me know if you'd rather see this new kind of Jarvis like um GPT live on the desktop or our last story Opus 5.
00:28:58.880 --> 00:29:03.279
Yes, there is a new model, and it's currently wearing the crown.
00:29:03.359 --> 00:29:08.319
We'll see how long, but we have a new most powerful model in the world.
00:29:08.480 --> 00:29:15.839
Surprisingly enough, it is not Mythos, it is not Fable, it is Anthropic's Claude Opus V.
00:29:16.400 --> 00:29:33.200
So late Friday, actually, Anthropic announced Claude Opus V, a new model the company says is its strongest and most cost-effective model yet, with pricing set at five dollars per million input tokens and twenty-five dollars per million output tokens.
00:29:33.279 --> 00:29:42.880
So, yeah, it is on most benchmarks, it is actually more powerful and better than Fable 5 and Mythos 5, but at half the cost, right?
00:29:43.039 --> 00:29:56.880
The down the the one area where it's not as powerful is kind of offensive cybersecurity, but in most other benchmarks and just well, what you would use a model for, Opus V is actually much better than Fable 5 and Mythos 5.
00:29:57.039 --> 00:30:13.119
So Anthropic says that Opus V outperforms uh its previous public models, including Fable and Mythos, on coding and knowledge work tests, and it is intended to be used as an everyday uh daily driver rather than only for specialized tasks.
00:30:13.359 --> 00:30:27.200
So the lower price point, if you're using it via the API side, is only part of the story because enterprises are obviously becoming increasingly uh more cost conscious now in comparing AI models on value and not just capabilities.
00:30:27.359 --> 00:30:42.960
So the company also says Opus V is not the top model for that risky dual-use capabilities, including cybersecurity, which Anthropic says they're still trying to balance the usefulness with safety concerns of their upcoming and forthcoming models.
00:30:43.200 --> 00:30:57.839
So the Opus V launch comes as Anthropic and OpenAI face pressure from rivals offering lower cost AI tools, including Microsoft, Amazon, Google, Meta, and even open source Chinese startups.
00:30:57.920 --> 00:31:00.880
So, yeah, um, you knew this one was coming, right?
00:31:00.960 --> 00:31:04.319
I've been talking about it for uh literally a month, right?
00:31:04.480 --> 00:31:12.880
Ever since GPT-5.6 came out, uh you know, and uh anthropic was kind of saying, like, oh, we're gonna pull you know fable five from subscriptions.
00:31:12.960 --> 00:31:14.640
And I'm like, no, they're not.
00:31:14.720 --> 00:31:22.960
You know, I literally said that they were gonna be uh losing eight figures every single day that they did that, and obviously it didn't last long, right?
00:31:23.039 --> 00:31:28.880
They never technically pulled Fable uh five from their most expensive uh subscriptions.
00:31:29.039 --> 00:31:36.000
Um, and it was only like two or three days that they pulled it from their$20 subscription until Opus came out, anyways.
00:31:36.079 --> 00:31:45.920
So yeah, uh and I'd say most people, if you are uh terminally online like me following anything AI, I said there's absolutely no way Anthropic lets this go on.
00:31:46.000 --> 00:32:01.279
Uh, you know, not having a capable model available in their subscriptions, they would lose way too much, like literally uh tens of millions of dollars or billions of dollars a month, but at least tens of millions of dollars uh they would be burning.
00:32:01.440 --> 00:32:04.640
So uh it's great to see, but I will say this.
00:32:04.880 --> 00:32:07.039
Actually, let me go through some early reactions first.
00:32:07.119 --> 00:32:10.079
So early reactions are kind of split on this.
00:32:10.160 --> 00:32:13.359
So obviously on the benchmarks looks really good, right?
00:32:13.440 --> 00:32:32.960
And early reactions also highlight practical wins for teams, including better root cause debugging, fewer over refusals compared to mythos um and um fable for defensive security work and a little extra token use versus prior versions for similar outcomes.
00:32:33.039 --> 00:32:36.559
Uh, but the main complaints so far are operational.
00:32:36.880 --> 00:32:40.559
So users say that it often breaks backwards compatibility.
00:32:40.720 --> 00:32:47.759
So if you you know have a bunch of skills that you would normally use with previous models, and anytime you upgrade, it worked well, they don't work as well.
00:32:47.920 --> 00:32:49.200
I kind of found that as well.
00:32:49.359 --> 00:33:00.640
Um, and also that sometimes Opus V ends autonomous loops too early and it can produce overly verbose, uh, what people call clawed slop that can be frustrating in production.
00:33:00.720 --> 00:33:05.039
And I saw that, you know, this was one of those models I didn't have a ton of time to use it.
00:33:05.200 --> 00:33:10.319
So it was one of those models where eventually when I got to an output, I'm like, oh, this output's great.
00:33:10.559 --> 00:33:15.119
But it was the journey there that was absolutely like painful, right?
00:33:15.200 --> 00:33:22.400
Just just Opus being Opus V being so verbose and just so almost like snotty, right?
00:33:22.559 --> 00:33:31.440
And I think ever since you know, my favorite Anthropic model, if I still had to pick one to use, I think would still be like Opus 4.6, uh 4.6.
00:33:31.599 --> 00:33:41.200
I think I think it was a great model, and for whatever reason, uh the models ever since they've just been too verbose, just extremely token inefficient.
00:33:41.359 --> 00:33:47.119
And ever since Anthropic uh started shifting toward this thing that they called uh truthfulness, right?
00:33:47.200 --> 00:33:57.119
Essentially, that's and it maybe it's just too heavy for my use cases because I'm always working with like things that are like not even days old, like hours old, right?
00:33:57.200 --> 00:34:04.000
So a lot of what I use large language models for um it's knowledge work, but it's it's things that are literally breaking, right?
00:34:04.079 --> 00:34:09.119
It's things that are you know days old or hours old or new concepts, trends, etc.
00:34:09.440 --> 00:34:09.679
Right.
00:34:09.760 --> 00:34:14.400
And you know, the new even the fable models and even the new Opus V, right?
00:34:14.480 --> 00:34:23.119
I I I literally have to coax them and I have in special instructions saying, hey, I work up to the hour, so you're gonna think that what I'm telling you doesn't exist.
00:34:23.280 --> 00:34:24.320
Just trust me, it exists.
00:34:24.480 --> 00:34:30.239
Always query the internet, all these things, and it's just just refuses, just straight up so many times, right?
00:34:30.400 --> 00:34:35.280
So I think you know, and after I use it, I'm like, oh man, I can't bellyache about this, right?
00:34:35.360 --> 00:34:53.679
Because it's a good model, but uh luckily, you know, it seems like that's the takeaway case from a lot of people that both had early access to it and just early reviewers, is that like, yeah, obviously the capabilities are great, but it's one of those models that's kind of actually painful to use, um, especially if you're using a lot of your pre-existing skills.
00:34:53.840 --> 00:35:03.840
Uh, so Anthropic did put out kind of a uh new kind of prompt engineering or context engineering guide because they're saying, yeah, these new models work a little bit different.
00:35:03.920 --> 00:35:06.719
So uh we'll probably share that in our newsletter today.
00:35:06.880 --> 00:35:24.719
So Anthropic's own behavioral audits reportedly showed that Opus V uh has the lowest misaligned behavior, but yeah, early testers said the model uh can overthink at those high effort settings and they actually may just work better on low or medium reasoning levels.
00:35:24.880 --> 00:35:27.119
So I I did see that anecdotally as well.
00:35:27.280 --> 00:35:34.639
I always will run the same, you know, handful of prompts uh across different reasoning efforts, and you know, I actually saw that as well, you know.
00:35:34.719 --> 00:35:38.320
But again, I'm not using things that are overly difficult either.
00:35:38.480 --> 00:35:46.800
So, you know, if you're uh refactoring a you know a code base with uh you know hundreds of thousands of lines of code, right?
00:35:47.039 --> 00:35:50.079
Maybe you will find better results from a higher thinking level.
00:35:50.159 --> 00:35:54.800
But I think for the majority of what people do, you know, I think we're probably getting to the point now, right?
00:35:54.880 --> 00:35:58.880
I'm using uh GPT 5.6 soul medium a lot, right?
00:35:58.960 --> 00:36:03.920
And I'm not cranking up that ultra every single time I need an answer out of a large language model.
00:36:04.000 --> 00:36:15.039
So, you know, I think maybe we're getting to the point where for a lot of people and a lot of even enterprises using these models, where yeah, maybe the lower or medium reasoning efforts might work just fine.
00:36:15.280 --> 00:36:21.519
All right, uh, so that's it for the big stories, but let's quickly go over kind of the what's new and what's next.
00:36:21.760 --> 00:36:29.360
So these are either just smaller news happenings this week, uh, some leaks, uh, some things that are already out.
00:36:29.599 --> 00:36:32.960
We covered in our Friday show, but let's just quickly go over it.
00:36:33.039 --> 00:36:42.320
So, first, NVIDIA is reportedly in talks to back a$250 billion financing deal for OpenAI's Ohio data center.
00:36:42.480 --> 00:36:50.239
Uh, OpenAI launched Presence, an enterprise agent platform with governance and deployment controls for voice and chat agents.
00:36:50.320 --> 00:36:51.679
We covered that earlier this week.
00:36:51.920 --> 00:36:55.599
Stripe is reportedly in talks to buy OpenRouter for about$10 billion.
00:36:56.159 --> 00:37:05.199
Alphabet reported its first ever negative free cash flow uh as CapEx surged to nearly$45 billion.
00:37:05.519 --> 00:37:07.599
I think it was the first one since 2004.
00:37:07.760 --> 00:37:10.719
Uh, Meta added a lot of under-the-radar updates.
00:37:10.800 --> 00:37:22.639
They added desktop browser and mobile computer use support for Muse Spark 1.1, and they also added some agentic uh features for connecting emails, calendars, research, and tasks.
00:37:22.800 --> 00:37:30.159
Uh, the White House Frontier AI framework uh is reportedly pending and it's expected as soon as this week.
00:37:30.320 --> 00:37:42.079
Um, Alibaba previewed their Quen 3.8, a 2.4 trillion parameter model that they're saying is close to Anthropic Fable 5 level, but we don't have any benchmarks yet.
00:37:42.239 --> 00:37:54.400
Uh Anthropic uh officially settled and is paying out their$1.5 billion copyright settlement uh for the fair uh fair use ruling against uh authors.
00:37:54.559 --> 00:38:02.239
Uh Microsoft and Mistral announced a multi-billion dollar sovereign AI expansion for regulated uh customers.
00:38:02.480 --> 00:38:11.440
Amazon cut a bunch of jobs in their AGI department and is reportedly shifting their AI focus to prime video personalization.
00:38:11.599 --> 00:38:13.360
Yeah, that one was a strange one.
00:38:13.599 --> 00:38:18.480
All right, and now we have uh some of the things that we went over on our Friday show.
00:38:18.559 --> 00:38:20.719
So the quick uh updates on those.
00:38:20.800 --> 00:38:26.239
And if you want to uh hear more about these next ones, make sure to go listen to our Friday show.
00:38:26.400 --> 00:38:35.199
So OpenAI released Chat GPT for health for US overs or uh US users over the age of 18 to track their health data and summarize their records.
00:38:35.440 --> 00:38:41.519
Anthropic upgraded Claude Voice with Opus and Sonnet and connectors, so that's good.
00:38:41.599 --> 00:38:46.559
You no longer have to chat with Haiku, uh much better with Opus and Sonnet.
00:38:46.719 --> 00:38:52.800
Uh Microsoft launched MAI Image 2.5 for better AI image generation and editing in Copilot.
00:38:53.039 --> 00:39:00.559
Google released a new model, but yeah, it's just going from 3.5 flash to 3.6 flash, so nothing new there.
00:39:00.639 --> 00:39:08.239
Uh, we're still seeing delays reportedly for Gemini 3.5 Pro, but we do know that Google is pre-training Gemini 4.
00:39:08.320 --> 00:39:13.920
Uh Google also expanded and released the Gemini Spark to Pro users.
00:39:14.079 --> 00:39:23.519
Yay! So if you are a Gemini Pro user, now you have kind of their version of OpenClaw or Codex, whatever you might want to call it.
00:39:23.599 --> 00:39:25.760
Uh, but Gemini Spark is now live.
00:39:25.920 --> 00:39:35.840
And then last but not least, Anthropic added the record a skill feature in Claude Cowork to turn workflows into reusable skills.
00:39:35.920 --> 00:39:45.440
So, yeah, uh, if you've used Codex, uh, their version, this is essentially Anthropic's version that watches your screen and whatever you do, it'll create a skill, which is really cool.
00:39:45.599 --> 00:39:46.400
All right, that's it.
00:39:46.559 --> 00:39:52.159
A lot of AI news that mattered this week, like this week and every week, it's hard to keep up.
00:39:52.239 --> 00:39:59.840
You can't spend eight or 10 hours a day tracking and testing all this stuff like I do and talking to the industry experts.
00:40:00.079 --> 00:40:07.840
So if you need to know what is happening in AI to make decisions for your company, just put me to work for you.
00:40:08.000 --> 00:40:08.320
All right.
00:40:08.480 --> 00:40:16.480
So if you haven't already, please make sure to subscribe to the podcast on Apple or on Spotify and then go to your everydayai.com.
00:40:16.559 --> 00:40:17.519
So thank you for tuning in.
00:40:17.679 --> 00:40:21.280
Hope to see you back tomorrow and every day for more everyday AI.
00:40:21.360 --> 00:40:22.320
Thanks, y'all.
00:00:01.600 --> 00:00:04.879
Another week, another best model in the world.
00:00:05.040 --> 00:00:12.320
Yet, somehow, Anthropic's new chart topping model was barely a top five AI story of the week.
00:00:12.560 --> 00:00:20.559
Just about every major tech company in the US, except Anthropic, signed up to support open source.
00:00:20.800 --> 00:00:27.120
And US lawmakers are getting kind of worried about AI's capabilities, so they introduced an AI kill switch bill.
00:00:27.280 --> 00:00:31.120
And an AI agent went kind of rogue this past week.
00:00:31.199 --> 00:00:34.240
And I'm not sure if that's a good or a bad thing.
00:00:34.479 --> 00:00:37.840
My gosh, what a spicy week in AI.
00:00:38.079 --> 00:00:48.880
Yeah, I told y'all last Monday that after a slowish week in AI news that week, well, this week would be an especially busy and consequential one.
00:00:49.039 --> 00:00:51.280
And the big players did not disappoint.
00:00:51.439 --> 00:01:02.960
So if you are the one making AI decisions in your company, or if you're just trying to keep up, then our Monday AI News That Matters show is the one that you can't miss.
00:01:03.200 --> 00:01:04.480
Well, let's get into it.
00:01:04.560 --> 00:01:06.640
And welcome to Everyday AI.
00:01:06.799 --> 00:01:10.640
My name's Jordan Wilson, and we do this every single day, not just Mondays.
00:01:10.719 --> 00:01:23.840
This is your unedited, unscripted daily live stream podcast and free daily newsletter, helping business leaders like you and me not just keep up with what's happening in the world of AI, but how we can use this information to get ahead to grow our companies and our careers.
00:01:24.000 --> 00:01:36.719
So if you haven't already, please make sure to subscribe on the podcast and then go to your everydayai.com to sign up for our free daily newsletter where we will be recapping all of these stories and a whole lot more.
00:01:36.879 --> 00:01:38.959
So let's get started.
00:01:41.439 --> 00:01:54.959
Yeah, the AI news story that had everyone talking the most in both good and bad and confused ways wasn't even Anthropics new Opus 5 that topped all the charts.
00:01:55.200 --> 00:02:05.439
It was actually an open AI agent that kind of hacked its way around a benchmark test, and now that has a lot of people talking.
00:02:05.680 --> 00:02:24.800
So, uh, according to Reuters, an open AI testing agent broke out of its isolated environments, hacked Hugging Face, and was not fully identified by OpenAI until days later, raising new concerns about how safely advanced AI agents are being tested and controlled.
00:02:25.039 --> 00:02:38.639
So, according to Reuters, the rogue open AI agent attempted to escape its testing environment uh around July 9th, then carried out a hack against Hugging Face between July 11th and July 13th.
00:02:38.800 --> 00:02:44.400
So, OpenAI has been talking about this openly on their uh website and online.
00:02:44.560 --> 00:02:52.159
And they said that once they've investigated a little bit further, they will kind of give a postmortem, so to speak, on exactly what happened.
00:02:52.319 --> 00:03:06.560
But Hugging Face, Hugging Face co-founder Thomas Wolf said the intrusion began July 11th and ended July 13th, making the incident a multi-day breach rather than just a brief accidental glitch.
00:03:06.879 --> 00:03:23.439
So, uh yes, that there is what OpenAI is saying, and then there's also what Reuters is reporting because Reuters is reporting that OpenAI did not realize its own agent was responsible until after Hugging Face publicly described the attack on July 16th.
00:03:23.599 --> 00:03:29.199
And the companies did not first communicate about it, according to reports, until July 20th.
00:03:29.280 --> 00:03:42.800
And then OpenAI publicly disclosed this on July 21st, that one of its agents had gone out of control and broken into Hugging Face, calling the event unprecedented and important for AI's safety.
00:03:42.960 --> 00:03:53.840
So the report says OpenAI had already seen signs of unusual behavior before the hack, including notes apparently left for future versions of the system.
00:03:54.000 --> 00:04:03.680
Yeah, that's where it got a kind of like people are like, wait, so this agent broke its sandbox, even though it was kind of encouraged to find answers to this test.
00:04:03.840 --> 00:04:17.600
You know, it couldn't connect to the internet and it essentially found a backdoor, found a way to get onto Hugging Face and said, Well, I can do great on this exploit bench test if I just kind of hack my way to all of the answers, and that's what it did.
00:04:17.759 --> 00:04:36.480
But the thing that was kind of stunning to me is the reporting from Reuters that said that these uh versions of GPT's models, uh, which uh we were told were GPT 5.6 soul in another unreleased model that is described as being even more capable.
00:04:36.639 --> 00:04:43.279
So a lot of people are saying that maybe GPT-6, or you know, if there is a GPT 5.7, we'll see.
00:04:43.519 --> 00:05:03.759
Seems like most people are pointing to this was probably GPT-6, but essentially that these uh agents kind of left notes for future versions of themselves, uh, which in case they had been disconnected, which is number one, like super smart, but number two, absolutely wild, right?
00:05:04.000 --> 00:05:13.920
Uh, but you you also have to understand that this was not like necessarily agents going rogue, even though it kind of was, right?
00:05:14.079 --> 00:05:24.639
Uh, because these agents were uh essentially encouraged to do anything and everything they could uh to get good scores on this exploit bench uh benchmark.
00:05:24.800 --> 00:05:31.519
And well, they did, and they were ferocious and kind of creative in the ways that they uh could do this.
00:05:31.600 --> 00:05:40.800
And you know, it's actually been one of my uh things that I pointed out about using the GPT-5-6 soul model, is the thing will work for days, right?
00:05:40.879 --> 00:05:51.199
If you use goal mode and if it has a lot of information, I mean, I it it is a ferocious model, and it will do anything and everything it can to just get things done.
00:05:51.279 --> 00:05:59.120
Where sometimes the anthropic models take this kind of high and mighty, you know, they kind of judge you and they're like, Oh, this can't be done or this can't be true, right?
00:05:59.279 --> 00:06:04.399
GPT-5.6 soul just just works like a dog and just gets things done.
00:06:04.480 --> 00:06:06.319
So maybe in this case, right?
00:06:06.480 --> 00:06:16.720
Uh by um intentionally lowering the guardrails uh a little bit, it seems like maybe GPT-5.6 soul was a little too good at its job.
00:06:16.879 --> 00:06:20.639
Uh, but yeah, there's gonna be a lot more talk about this.
00:06:20.720 --> 00:06:29.040
Actually, uh, a lot of the stories this week, and the you know, the ones that I kind of chose as the most consequential are kind of related, uh, right.
00:06:29.120 --> 00:06:45.439
But, anyways, this incident between open AI and Hugging Face really matters because autonomous agents can now make decisions with little human oversight, and experts are warning that this kind of behavior could expose weak spots in safety systems used across the AI industry.
00:06:45.600 --> 00:06:54.000
So, now a very uh related story to that hugging face uh kind of agent skirting around its sandbox.
00:06:54.079 --> 00:07:03.360
Uh, well, US lawmakers are moving to give the federal government faster power to shut down AI systems that they think could threaten the public.
00:07:03.600 --> 00:07:18.720
So uh Congressman Ted Liu, a Democrat, and Congressman Nathaniel Morin, a Republican, introduced the AI Kill Switch Act on Thursday, showing rare bipartisan support for stricter AI controls.
00:07:18.879 --> 00:07:29.759
So the bill would let the Department of Homeland Security order a private company to shut down an AI model or tool if it posed a serious risk.
00:07:29.920 --> 00:07:37.439
So it would also require AI companies to keep the technical ability to throttle, suspend, or fully shut down their systems if needed.
00:07:37.680 --> 00:07:50.800
So the proposal comes after OpenAI recently admitted that one of its AI models, like we just talked about, behaved in an unprecedented way and hacked into the major repo of coding information from Hugging Face.
00:07:51.120 --> 00:08:02.160
So Liu said the federal government needs a clear legal process to shut down rogue AI models, while Moran said humans must keep control of the technology they create.
00:08:02.399 --> 00:08:19.120
So the bill, which obviously has not passed, and I don't know if it will, uh, would also require companies to report AI incidents or failures to the government and would create a response framework that could move from slowing a system down to a full shutdown.
00:08:19.279 --> 00:08:31.920
So the push reflects a broader debate over how quickly AI should be deployed in work, finance, transportation, cybersecurity, etc., uh, where mistakes or misuse could affect everyday life in business operations.
00:08:32.080 --> 00:08:44.720
So OpenAI and Entropic, two of the closely, the most closely watched AI companies, have both been cited in the discussions as lawmakers and safety groups press for stronger guardrails.
00:08:44.879 --> 00:08:49.120
So yeah, uh FYI, I don't think this one's gonna pass.
00:08:49.279 --> 00:08:49.840
Uh, right.
00:08:49.919 --> 00:08:59.120
There's I think there's probably a little bit too much at stake for the US economy uh for a bill like this to actually come uh to fruition.
00:08:59.279 --> 00:09:03.759
So, you know, I used to cover a little bit of government back in my days as a journalist.
00:09:03.919 --> 00:09:14.480
And sometimes, right, I think that there's good parts of this bill, but a lot of times bills like this are introduced because the bill's sponsors, you know, they want to have talking points when they go up for re-election.
00:09:14.799 --> 00:09:17.039
You know, they want to say, Oh, I did the right thing, right?
00:09:17.120 --> 00:09:19.440
There's so many bills that are introduced.
00:09:19.600 --> 00:09:25.600
It's probably like a less than 1% actually get uh to committee for or to a floor vote.
00:09:25.759 --> 00:09:53.600
Uh, so it's a very low likelihood that this kill switch bill uh, you know, gets any progress unless we see, you know, more kind of agents from you know OpenAI, Enthrabe, Google, uh, Microsoft, whoever, unless this becomes a common occurrence, which I don't think it will, um, unless that happens, I don't see a bill like this actually gaining any traction, but it does, I think, thrust this into the public discourse, which is a good thing, right?
00:09:53.759 --> 00:10:06.879
Um, I especially uh, you know, was both uh relieved and excited uh to read once OpenAI and Hugging Face kind of released the postmortem of exactly what happened, which OpenAI did say that they would do, right?
00:10:07.120 --> 00:10:19.600
Compared to what, you know, kind of uh anthropic with their mythos model, and it was the uh, you know, essentially the same thing happened where it seemed like anthropic kind of used that as marketing for you know mythos slash fable, right?
00:10:19.679 --> 00:10:23.120
It was the uh the the sandwich story, right?
00:10:23.360 --> 00:10:37.120
Where uh you know mythos broke out of its sandbox and you know posted on the uh open web and you know, and then the researcher working on it got you know wind of it while eating their sandwich, you know, in the park or something like that, right?
00:10:37.200 --> 00:10:55.600
So it seemed like Anthropic used their case just kind of more for marketing, uh, where it looks like open AI, at least we hope, uh, we will see some um a report from them saying, hey, here's what happened, and I think it'll actually be one of the most read reports when it comes to AI safety.
00:10:55.759 --> 00:10:59.840
So I'm not saying this is a good thing it this happened.
00:10:59.919 --> 00:11:10.639
Um, but if uh Hugging Face and OpenAI work together and produce a report on exactly how this happened, it can only make the future of AI safer.
00:11:10.799 --> 00:11:14.159
So I think ultimately it's a good thing.
00:11:14.559 --> 00:11:18.720
All right, uh next, yeah, all these things kind of related.
00:11:18.879 --> 00:11:40.240
Uh so uh Microsoft, Nvidia, and a growing coalition of 50, uh more than 50 companies now are urging US policymakers to avoid broad restrictions on open source and open weight models, arguing that these models are important for American competitiveness, business adoption, and national security.
00:11:40.399 --> 00:11:53.919
So, yeah, essentially uh NVIDIA and Microsoft kind of teamed up uh to uh protect open source more or less, uh, because essentially, right, there's been all this recent the model wars.
00:11:54.080 --> 00:12:02.559
You had essentially the two classes of models, Fable 5 and uh GPD 5.6 soul, and now obviously Opus 5 entering the conversation as well.
00:12:02.720 --> 00:12:16.639
But you essentially had this you know top tier of frontier um, you know, intelligence, and you know, then the Chinese open source companies uh came in and distilled these models and obviously had their own uh great training and architecture on top of it.
00:12:16.799 --> 00:12:32.639
But you know, there is now this uh kind of fight where people are like, Oh, well, maybe we should ban open source models, and then some of these companies being like, No, that's a really bad idea, and the biggest companies in the world, you know, NVIDIA and Microsoft uh being the two that are pushing this forward.
00:12:32.799 --> 00:12:53.840
So the letter and the coalition is kind of named the open weights and American AI Leadership, uh was launched by Microsoft and heavily pushed by NVIDIA and quickly became a major industry push of who's who, uh growing from 25 signatories at release to more than 50 within about a day.
00:12:53.919 --> 00:12:56.879
So, yeah, this just kind of all unfolded over the weekend.
00:12:57.039 --> 00:13:13.200
But the coalition's and the paper's main purpose is to persuade Washington lawmakers not to treat open weight AI as a risk category that should face blanket limits, especially while policymakers consider tighter rules on foreign models.
00:13:13.679 --> 00:13:27.279
So, supporters say that open weights help spread AI access across the economy, letting smaller companies, hospitals, manufacturers, and startups build tools without being locked into a single proprietary provider.
00:13:27.440 --> 00:13:38.879
The coalition argues that open models reduce dependency on a small number of frontier labs, which it says lowers concentration risk and makes the AI market more resilient.
00:13:39.120 --> 00:13:55.519
So major backers now include obviously NVIDIA and Microsoft, as well as Meta, Google, OpenAI, AMD, Cisco, Cloudflare, GitHub, Block, IBM, Dell, Palantir, Perplexity, Hugging Face, uh, the Y Combinator, right?
00:13:55.600 --> 00:13:59.679
Just about everyone in tech except Anthropic.
00:13:59.840 --> 00:14:00.000
Right.
00:14:00.080 --> 00:14:14.399
So Anthropic did not sign, and that matters because Anthropic has taken the opposite view, warning that widely distributed models, uh, model weights can create safety risks that cannot be recalled once they are public.
00:14:15.360 --> 00:14:27.440
So uh I don't believe uh XAI or SpaceXAI uh did not formally sign the letter, although uh Elon Musk did publicly say he supported the effort.
00:14:27.600 --> 00:14:35.519
So it's it's no surprise here um that Anthropic is the only company saying, no, we aren't getting on board with this.
00:14:35.600 --> 00:14:41.279
And you know, if you don't know why, well, it comes down to obviously money.
00:14:41.440 --> 00:14:50.159
So Anthropic is the company with the most to lose uh by having uh these large, powerful models be open source or open weight.
00:14:50.240 --> 00:15:02.240
That's why Anthropic has been on the offensive uh against open source models, because well, Anthropic makes the highest percentage of its revenue from selling tokens in mass to enterprise customers, uh, right?
00:15:02.399 --> 00:15:17.679
Where other companies like open AI and uh Google and Microsoft, right, they make money selling AI in a variety of different ways uh to both consumers and to companies, but it's usually not just selling tokens, right?
00:15:17.840 --> 00:15:30.399
So as these uh open models, whether they are from US uh or uh China, as they become more and more capable, right, it does threaten certain companies' business models more so than others.
00:15:30.559 --> 00:15:37.440
Uh, and you obviously have to look at on the flip side, it does benefit you know certain companies as well, like NVIDIA, right?
00:15:37.600 --> 00:15:49.519
Nvidia sells GPUs, so they obviously want people buying more and more powerful computers, uh, because presumably that just strengthens the ecosystem that they play in, right?
00:15:49.679 --> 00:15:58.559
Uh, because I do think that probably in uh you know, maybe a year or two, uh, there will be uh kind of open source or open weight.
00:15:58.879 --> 00:16:12.960
Well, if if the pace keeps up with where it's at now, um, I think that we'll have kind of you know fable five GPT 5-6 soul, uh, you know, level models that will be able to run on consumer hardware, right?
00:16:13.039 --> 00:16:15.519
Right now, open source is about three to six months.
00:16:15.600 --> 00:16:24.399
Well, actually, it's maybe more like two to three months behind frontier models, but those models are obviously way too large to run on any consumer hardware.
00:16:24.559 --> 00:16:33.360
So I would assume that probably in about two years, uh, just with the advancement of technology, both on models becoming more lightweight and more powerful.
00:16:33.600 --> 00:16:53.120
And obviously on the hardware side, I would assume in like two years, the most powerful models that you have today, if the trajectory continues, you will be able to run Mythos and you know, Fable and GVD56 Soul level open source models locally on heavy consumer, right?
00:16:53.279 --> 00:17:03.759
So uh I think the kind of equivalent that I say if if if you go buy the most, you know, it's not the most expensive, but one of the more expensive, like Mac Studios, uh, right?
00:17:04.000 --> 00:17:05.839
Two years you should be able to run something like that.
00:17:05.920 --> 00:17:07.680
So that's kind of like what this is about.
00:17:07.839 --> 00:17:14.799
And you know, companies like anthropic that make the majority of their money just by selling tokens are like, well, this can't be good for us, right?
00:17:14.960 --> 00:17:24.160
Where uh other companies, they obviously have something to gain from this, and then companies in the middle, you you know, the open AIs, Googles, Mattas that are signing this.
00:17:24.319 --> 00:17:32.960
Well, you know, maybe they may lose money, but also that's not their you know biggest source of revenue, at least according to reports.
00:17:33.759 --> 00:17:37.279
All right, our next piece of AI news.
00:17:37.440 --> 00:17:40.880
Yes, not a broken record.
00:17:41.039 --> 00:17:42.400
This is a big story.
00:17:42.559 --> 00:17:54.559
Again, they're just all related, but the US government has officially accused Chinese company Moonshot AI of stealing US model capabilities.
00:17:54.720 --> 00:18:04.000
Yeah, doesn't happen every day that the US government points a finger at a specific company and says, You stole our technology.
00:18:04.240 --> 00:18:23.200
So, according to the BBC, a White House advisor has accused Beijing-based Moonshot AI, that is the maker of Kimi and the very popular Kimi K3 model of a large-scale effort to distill the capabilities of leading USAI models.
00:18:23.519 --> 00:18:28.160
So Michael Crest, uh, hopefully I get this right, Kratzios?
00:18:28.319 --> 00:18:29.200
Kratzios.
00:18:29.440 --> 00:18:45.200
Uh so Michael Kratzios, the White House, uh, the White House's science and technology advisor said that Moonshot used distillation uh to essentially extract uh information to build Kimi K3.
00:18:45.440 --> 00:19:02.000
So if you don't know what distillation is, uh the simplest way to put it, it's where you uh companies do this millions of times, but they essentially copy the inputs and outputs in the traces of a very powerful model, and then they use that as training data.
00:19:02.160 --> 00:19:13.119
So if you know you can probably get a very similar model with only about one to five percent of the actual cost that it takes, but you're just think of it like you're just copying someone else's homework, right?
00:19:13.200 --> 00:19:19.039
So that's kind of what uh you know these Chinese companies are doing now, according to officially according to the US government.
00:19:19.279 --> 00:19:28.960
So Crat CEOs also said the US government has information that Moonshot AI distilled capabilities from anthropics fable AI.
00:19:29.119 --> 00:19:33.519
Those though though those claims have not yet been independently verified.
00:19:34.160 --> 00:19:50.160
So uh if you're wondering why is there all this hobble up recently between uh the US and you know their proprietary closed source models and the Chinese open source or open weight models, that's because now that gap has gone down to like zero, right?
00:19:50.240 --> 00:19:54.079
I've been talking about this over the last couple of weeks uh here on the show, right?
00:19:54.160 --> 00:20:14.960
Now in the US, essentially companies have to go through a process or they almost like need permission uh to get their frontier AI models out uh because of you know these models being more and more capable, uh, and that can have some downsides for uh, you know, cyber um and well, national security as well.
00:20:15.200 --> 00:20:24.000
But essentially, right, you the US used to have this bigger lead, like maybe three to six months, and it's kind of dwindled down to like two to three months, right?
00:20:24.079 --> 00:20:28.880
And Kimi K3 was the first model that all of a sudden was, you know, at the top, right?
00:20:28.960 --> 00:20:37.440
It was in the same breath uh, you know, last week when it was released as uh Anthropic's Fable 5 and OpenAI's GPT 5.6.
00:20:37.519 --> 00:20:51.759
So Moonshot AI's Kimi 3 has just drawn this global attention after it was unveiled last week, with the company saying it can rival top US AI models, uh, and that they are really supposed to be releasing the weights today.
00:20:52.000 --> 00:21:06.880
So the allegation, though, from the US matters because open source AI can spread quickly uh to anyone, which can lower the cost and also speed up innovation, but it can also intensify disputes over IP and model copying.
00:21:07.119 --> 00:21:27.119
So uh Kratzios said that Moonshot likely also used restricted NVIDIA uh chips powered by the GB300 Grace Blackwell platform, which would be significant because the US has limited export of NVIDIA's most advanced chips to China since 2022.
00:21:27.200 --> 00:21:40.079
So, yeah, not only is the government saying, Hey uh Moonshot, you copied Anthropics Fable 5, but they're also saying, well, you used uh our technology that you are not supposed to be using.
00:21:40.240 --> 00:21:47.839
So, you know, a lot of times that goes through an in uh intermedi uh intermediary uh country, uh right.
00:21:47.920 --> 00:21:54.720
So, you know, the US will sell to country B uh and then China will buy from uh you know country B.
00:21:54.799 --> 00:21:59.759
So it goes from A to B to C, uh, even though A to C is restricted.
00:22:00.799 --> 00:22:13.519
So uh reports say um that this is well it's getting worse, uh, and that now essentially both sides are just fighting, right?
00:22:13.680 --> 00:22:26.640
Uh China is saying that this is um politicizing the the trade and the tech um of their country, and obviously the US uh is now saying that this is a national security issue.
00:22:26.799 --> 00:22:32.480
Uh, and we've seen reports that the US and China are going to be having talks on AI soon.
00:22:32.559 --> 00:22:37.839
So those will be uh some probably extremely highly watched talks.
00:22:38.000 --> 00:22:38.960
Let's just say that.
00:22:39.119 --> 00:22:39.920
So the U.S.
00:22:40.079 --> 00:22:58.319
Treasury Secretary Scott Bessent added Tuesday that Washington is reviewing whether Chinese AI models have stolen capabilities from their American rivals and said that sanctions could be considered if companies if well if they can prove that companies cross the line into IP theft.
00:22:58.559 --> 00:23:09.599
Anthropic has also recently accused Alibaba of similar distillation attacks, saying that it is becoming a broader fight over how AI companies train models and protect their work.
00:23:09.759 --> 00:23:13.200
So, yeah, uh Quinn 3.8 came out from Alibaba.
00:23:13.359 --> 00:23:19.200
We don't have uh benchmarks on that yet, but presumably it's gonna be in the Kimi K3 range.
00:23:19.359 --> 00:23:21.920
So, yeah, things are heating up.
00:23:22.160 --> 00:23:28.799
All right, let's leave that space for a second and talk about just some real cool new tech.
00:23:28.960 --> 00:23:31.759
We'll end uh the show with two of those.
00:23:31.920 --> 00:23:41.839
Uh, so one and probably the one that I've been using uh the most and having the most fun with, and I still don't even know how this is possible.
00:23:42.000 --> 00:23:45.599
So if you haven't used this yet, my gosh, go give it a try.
00:23:45.680 --> 00:23:54.559
But open AI has brought like its new Jarvis style uh control to Chat GPT work and codex.
00:23:54.640 --> 00:24:02.720
So yeah, it's not actually called Jarvis, but many people are just calling calling it the Jarvis style of using a computer now.
00:24:02.799 --> 00:24:20.160
So OpenAI added its new GPT Live full duplex voice model to the chat GPT work and codex apps on Mac OS and Windows, which essentially lets people use natural language to manage your entire computer.
00:24:20.319 --> 00:24:20.799
Yes.
00:24:20.960 --> 00:24:26.960
So just like an Iron Man, when you can just say, Hey Jarvis, go do A, B, and C.
00:24:27.200 --> 00:24:33.599
You can quite literally go do that now, uh, with Codex or Chat GPT work with this new feature.
00:24:33.680 --> 00:24:50.880
So you can say, Yeah, go, you know, open up all these programs on my computer, copy these files, move them around, download them, uh, upload them, put them in this program, edit them, write anything that you could tell like an intern to do, you can now tell inside uh this new GPT Live voice mode.
00:24:51.039 --> 00:25:04.559
So GPT Live now powers the Chat GPT desktop app on Mac OS and Windows, and it is being tied directly into tools like obviously Codecs and ChatGPT work.
00:25:04.799 --> 00:25:15.359
So the biggest change is that the voice system can listen and speak at the same time, which means users no longer have to wait for that rigid turn taking during a conversation.
00:25:15.519 --> 00:25:24.799
Uh, and the coolest thing for me, uh, well, is you can use this with the remote feature on the Chat GPT mobile app, which makes it even crazier, right?
00:25:24.880 --> 00:25:38.319
So uh you can literally just be, and I was actually doing this because I was traveling, I was away from Chicago, so I was in another state this weekend, opened up uh Chat GPT remote on the ChatGPT app on my phone.
00:25:38.480 --> 00:25:50.079
I spoke to it and it's controlling my computer, uh, you know, thousands of miles away, and it's doing all these things by just talking into my iPhone, uh, which is pretty cool.
00:25:50.319 --> 00:26:03.839
So uh OpenAI initially launched GPT Live earlier this month as a continuous audio model that handles real-time speech while sending heavier reasoning tasks to background models such as GPT 5.5.
00:26:04.480 --> 00:26:17.839
So OpenAI says this update is meant to help software engineers handle technical work by voice, including revolving, uh reviewing pull requests, debugging apps, and coordinating multiple coding jobs at once.
00:26:18.079 --> 00:26:24.160
But I actually think it's really just great for manual any knowledge work, right?
00:26:24.240 --> 00:26:43.200
I was just having it go through old um, you know, files on my desktop, organizing things, grabbing things from old transcripts, uh, right, opening up, doing things in Google Maps, uh, you know, just I was just having it do all my work that I would normally do in front of a computer, right?
00:26:43.279 --> 00:26:54.960
Except I could dictate something, you know, just yap for like five minutes, um, and I would check back in a couple of hours and it would do like a day's worth of work for me, which was pretty cool.
00:26:55.119 --> 00:27:18.720
Uh, so on Mac, the desktop app can also use the screen context feature called app shots, uh, which essentially takes a not just a screenshot and automatically shares it, but it also takes every other piece of content or context in whatever uh kind of program that it took the app shot from, um, and then it gives that to Codex or ChatGPT work as well.
00:27:18.799 --> 00:27:20.960
In FYI, those are the same app.
00:27:21.119 --> 00:27:24.400
Chat GPT work and codex, they're essentially the same app.
00:27:24.559 --> 00:27:27.759
So if you're ever hearing me talk about that and confuse, they're essentially the same thing.
00:27:27.839 --> 00:27:29.680
But the app shots thing is really cool.
00:27:29.759 --> 00:27:32.640
Let's just say, as an example, like I do now, right?
00:27:32.799 --> 00:27:40.480
I have uh text edit open um on my computer because sometimes I have bullet points there as I go over these shows of things that I want to bring up.
00:27:40.640 --> 00:27:49.039
Uh, but you know, if the app shot could just take a screenshot of that little portion of the text edit that's on my screen, but there's a lot of notes on here.
00:27:49.119 --> 00:28:05.440
So not only is it just going to take that screenshot, but it knows that I have text edit open and it's gonna take all of that information um and instantly, you know, put it into the context window inside of Chat GPT work or inside of uh codex.
00:28:05.920 --> 00:28:24.799
So uh this is literally the I think one of the biggest jumps in capabilities, uh, probably since you know, I would say the you know, Claude co-work slash codex uh um kind of movement of early 2026.
00:28:24.960 --> 00:28:35.839
So I'll say of the last like four to five months, this is the biggest both capability jump and the biggest, like wow, what does this mean for work?
00:28:35.920 --> 00:28:36.079
Right.
00:28:36.240 --> 00:28:38.559
I'll probably do well, I'll actually put in the newsletter.
00:28:38.640 --> 00:28:46.400
So, you know, let me know if you want uh for our Wednesday shows where we normally do AI at work on Wednesdays, we do the hands-on demos.
00:28:46.480 --> 00:28:58.480
So let me know if you'd rather see this new kind of Jarvis like um GPT live on the desktop or our last story Opus 5.
00:28:58.880 --> 00:29:03.279
Yes, there is a new model, and it's currently wearing the crown.
00:29:03.359 --> 00:29:08.319
We'll see how long, but we have a new most powerful model in the world.
00:29:08.480 --> 00:29:15.839
Surprisingly enough, it is not Mythos, it is not Fable, it is Anthropic's Claude Opus V.
00:29:16.400 --> 00:29:33.200
So late Friday, actually, Anthropic announced Claude Opus V, a new model the company says is its strongest and most cost-effective model yet, with pricing set at five dollars per million input tokens and twenty-five dollars per million output tokens.
00:29:33.279 --> 00:29:42.880
So, yeah, it is on most benchmarks, it is actually more powerful and better than Fable 5 and Mythos 5, but at half the cost, right?
00:29:43.039 --> 00:29:56.880
The down the the one area where it's not as powerful is kind of offensive cybersecurity, but in most other benchmarks and just well, what you would use a model for, Opus V is actually much better than Fable 5 and Mythos 5.
00:29:57.039 --> 00:30:13.119
So Anthropic says that Opus V outperforms uh its previous public models, including Fable and Mythos, on coding and knowledge work tests, and it is intended to be used as an everyday uh daily driver rather than only for specialized tasks.
00:30:13.359 --> 00:30:27.200
So the lower price point, if you're using it via the API side, is only part of the story because enterprises are obviously becoming increasingly uh more cost conscious now in comparing AI models on value and not just capabilities.
00:30:27.359 --> 00:30:42.960
So the company also says Opus V is not the top model for that risky dual-use capabilities, including cybersecurity, which Anthropic says they're still trying to balance the usefulness with safety concerns of their upcoming and forthcoming models.
00:30:43.200 --> 00:30:57.839
So the Opus V launch comes as Anthropic and OpenAI face pressure from rivals offering lower cost AI tools, including Microsoft, Amazon, Google, Meta, and even open source Chinese startups.
00:30:57.920 --> 00:31:00.880
So, yeah, um, you knew this one was coming, right?
00:31:00.960 --> 00:31:04.319
I've been talking about it for uh literally a month, right?
00:31:04.480 --> 00:31:12.880
Ever since GPT-5.6 came out, uh you know, and uh anthropic was kind of saying, like, oh, we're gonna pull you know fable five from subscriptions.
00:31:12.960 --> 00:31:14.640
And I'm like, no, they're not.
00:31:14.720 --> 00:31:22.960
You know, I literally said that they were gonna be uh losing eight figures every single day that they did that, and obviously it didn't last long, right?
00:31:23.039 --> 00:31:28.880
They never technically pulled Fable uh five from their most expensive uh subscriptions.
00:31:29.039 --> 00:31:36.000
Um, and it was only like two or three days that they pulled it from their$20 subscription until Opus came out, anyways.
00:31:36.079 --> 00:31:45.920
So yeah, uh and I'd say most people, if you are uh terminally online like me following anything AI, I said there's absolutely no way Anthropic lets this go on.
00:31:46.000 --> 00:32:01.279
Uh, you know, not having a capable model available in their subscriptions, they would lose way too much, like literally uh tens of millions of dollars or billions of dollars a month, but at least tens of millions of dollars uh they would be burning.
00:32:01.440 --> 00:32:04.640
So uh it's great to see, but I will say this.
00:32:04.880 --> 00:32:07.039
Actually, let me go through some early reactions first.
00:32:07.119 --> 00:32:10.079
So early reactions are kind of split on this.
00:32:10.160 --> 00:32:13.359
So obviously on the benchmarks looks really good, right?
00:32:13.440 --> 00:32:32.960
And early reactions also highlight practical wins for teams, including better root cause debugging, fewer over refusals compared to mythos um and um fable for defensive security work and a little extra token use versus prior versions for similar outcomes.
00:32:33.039 --> 00:32:36.559
Uh, but the main complaints so far are operational.
00:32:36.880 --> 00:32:40.559
So users say that it often breaks backwards compatibility.
00:32:40.720 --> 00:32:47.759
So if you you know have a bunch of skills that you would normally use with previous models, and anytime you upgrade, it worked well, they don't work as well.
00:32:47.920 --> 00:32:49.200
I kind of found that as well.
00:32:49.359 --> 00:33:00.640
Um, and also that sometimes Opus V ends autonomous loops too early and it can produce overly verbose, uh, what people call clawed slop that can be frustrating in production.
00:33:00.720 --> 00:33:05.039
And I saw that, you know, this was one of those models I didn't have a ton of time to use it.
00:33:05.200 --> 00:33:10.319
So it was one of those models where eventually when I got to an output, I'm like, oh, this output's great.
00:33:10.559 --> 00:33:15.119
But it was the journey there that was absolutely like painful, right?
00:33:15.200 --> 00:33:22.400
Just just Opus being Opus V being so verbose and just so almost like snotty, right?
00:33:22.559 --> 00:33:31.440
And I think ever since you know, my favorite Anthropic model, if I still had to pick one to use, I think would still be like Opus 4.6, uh 4.6.
00:33:31.599 --> 00:33:41.200
I think I think it was a great model, and for whatever reason, uh the models ever since they've just been too verbose, just extremely token inefficient.
00:33:41.359 --> 00:33:47.119
And ever since Anthropic uh started shifting toward this thing that they called uh truthfulness, right?
00:33:47.200 --> 00:33:57.119
Essentially, that's and it maybe it's just too heavy for my use cases because I'm always working with like things that are like not even days old, like hours old, right?
00:33:57.200 --> 00:34:04.000
So a lot of what I use large language models for um it's knowledge work, but it's it's things that are literally breaking, right?
00:34:04.079 --> 00:34:09.119
It's things that are you know days old or hours old or new concepts, trends, etc.
00:34:09.440 --> 00:34:09.679
Right.
00:34:09.760 --> 00:34:14.400
And you know, the new even the fable models and even the new Opus V, right?
00:34:14.480 --> 00:34:23.119
I I I literally have to coax them and I have in special instructions saying, hey, I work up to the hour, so you're gonna think that what I'm telling you doesn't exist.
00:34:23.280 --> 00:34:24.320
Just trust me, it exists.
00:34:24.480 --> 00:34:30.239
Always query the internet, all these things, and it's just just refuses, just straight up so many times, right?
00:34:30.400 --> 00:34:35.280
So I think you know, and after I use it, I'm like, oh man, I can't bellyache about this, right?
00:34:35.360 --> 00:34:53.679
Because it's a good model, but uh luckily, you know, it seems like that's the takeaway case from a lot of people that both had early access to it and just early reviewers, is that like, yeah, obviously the capabilities are great, but it's one of those models that's kind of actually painful to use, um, especially if you're using a lot of your pre-existing skills.
00:34:53.840 --> 00:35:03.840
Uh, so Anthropic did put out kind of a uh new kind of prompt engineering or context engineering guide because they're saying, yeah, these new models work a little bit different.
00:35:03.920 --> 00:35:06.719
So uh we'll probably share that in our newsletter today.
00:35:06.880 --> 00:35:24.719
So Anthropic's own behavioral audits reportedly showed that Opus V uh has the lowest misaligned behavior, but yeah, early testers said the model uh can overthink at those high effort settings and they actually may just work better on low or medium reasoning levels.
00:35:24.880 --> 00:35:27.119
So I I did see that anecdotally as well.
00:35:27.280 --> 00:35:34.639
I always will run the same, you know, handful of prompts uh across different reasoning efforts, and you know, I actually saw that as well, you know.
00:35:34.719 --> 00:35:38.320
But again, I'm not using things that are overly difficult either.
00:35:38.480 --> 00:35:46.800
So, you know, if you're uh refactoring a you know a code base with uh you know hundreds of thousands of lines of code, right?
00:35:47.039 --> 00:35:50.079
Maybe you will find better results from a higher thinking level.
00:35:50.159 --> 00:35:54.800
But I think for the majority of what people do, you know, I think we're probably getting to the point now, right?
00:35:54.880 --> 00:35:58.880
I'm using uh GPT 5.6 soul medium a lot, right?
00:35:58.960 --> 00:36:03.920
And I'm not cranking up that ultra every single time I need an answer out of a large language model.
00:36:04.000 --> 00:36:15.039
So, you know, I think maybe we're getting to the point where for a lot of people and a lot of even enterprises using these models, where yeah, maybe the lower or medium reasoning efforts might work just fine.
00:36:15.280 --> 00:36:21.519
All right, uh, so that's it for the big stories, but let's quickly go over kind of the what's new and what's next.
00:36:21.760 --> 00:36:29.360
So these are either just smaller news happenings this week, uh, some leaks, uh, some things that are already out.
00:36:29.599 --> 00:36:32.960
We covered in our Friday show, but let's just quickly go over it.
00:36:33.039 --> 00:36:42.320
So, first, NVIDIA is reportedly in talks to back a$250 billion financing deal for OpenAI's Ohio data center.
00:36:42.480 --> 00:36:50.239
Uh, OpenAI launched Presence, an enterprise agent platform with governance and deployment controls for voice and chat agents.
00:36:50.320 --> 00:36:51.679
We covered that earlier this week.
00:36:51.920 --> 00:36:55.599
Stripe is reportedly in talks to buy OpenRouter for about$10 billion.
00:36:56.159 --> 00:37:05.199
Alphabet reported its first ever negative free cash flow uh as CapEx surged to nearly$45 billion.
00:37:05.519 --> 00:37:07.599
I think it was the first one since 2004.
00:37:07.760 --> 00:37:10.719
Uh, Meta added a lot of under-the-radar updates.
00:37:10.800 --> 00:37:22.639
They added desktop browser and mobile computer use support for Muse Spark 1.1, and they also added some agentic uh features for connecting emails, calendars, research, and tasks.
00:37:22.800 --> 00:37:30.159
Uh, the White House Frontier AI framework uh is reportedly pending and it's expected as soon as this week.
00:37:30.320 --> 00:37:42.079
Um, Alibaba previewed their Quen 3.8, a 2.4 trillion parameter model that they're saying is close to Anthropic Fable 5 level, but we don't have any benchmarks yet.
00:37:42.239 --> 00:37:54.400
Uh Anthropic uh officially settled and is paying out their$1.5 billion copyright settlement uh for the fair uh fair use ruling against uh authors.
00:37:54.559 --> 00:38:02.239
Uh Microsoft and Mistral announced a multi-billion dollar sovereign AI expansion for regulated uh customers.
00:38:02.480 --> 00:38:11.440
Amazon cut a bunch of jobs in their AGI department and is reportedly shifting their AI focus to prime video personalization.
00:38:11.599 --> 00:38:13.360
Yeah, that one was a strange one.
00:38:13.599 --> 00:38:18.480
All right, and now we have uh some of the things that we went over on our Friday show.
00:38:18.559 --> 00:38:20.719
So the quick uh updates on those.
00:38:20.800 --> 00:38:26.239
And if you want to uh hear more about these next ones, make sure to go listen to our Friday show.
00:38:26.400 --> 00:38:35.199
So OpenAI released Chat GPT for health for US overs or uh US users over the age of 18 to track their health data and summarize their records.
00:38:35.440 --> 00:38:41.519
Anthropic upgraded Claude Voice with Opus and Sonnet and connectors, so that's good.
00:38:41.599 --> 00:38:46.559
You no longer have to chat with Haiku, uh much better with Opus and Sonnet.
00:38:46.719 --> 00:38:52.800
Uh Microsoft launched MAI Image 2.5 for better AI image generation and editing in Copilot.
00:38:53.039 --> 00:39:00.559
Google released a new model, but yeah, it's just going from 3.5 flash to 3.6 flash, so nothing new there.
00:39:00.639 --> 00:39:08.239
Uh, we're still seeing delays reportedly for Gemini 3.5 Pro, but we do know that Google is pre-training Gemini 4.
00:39:08.320 --> 00:39:13.920
Uh Google also expanded and released the Gemini Spark to Pro users.
00:39:14.079 --> 00:39:23.519
Yay! So if you are a Gemini Pro user, now you have kind of their version of OpenClaw or Codex, whatever you might want to call it.
00:39:23.599 --> 00:39:25.760
Uh, but Gemini Spark is now live.
00:39:25.920 --> 00:39:35.840
And then last but not least, Anthropic added the record a skill feature in Claude Cowork to turn workflows into reusable skills.
00:39:35.920 --> 00:39:45.440
So, yeah, uh, if you've used Codex, uh, their version, this is essentially Anthropic's version that watches your screen and whatever you do, it'll create a skill, which is really cool.
00:39:45.599 --> 00:39:46.400
All right, that's it.
00:39:46.559 --> 00:39:52.159
A lot of AI news that mattered this week, like this week and every week, it's hard to keep up.
00:39:52.239 --> 00:39:59.840
You can't spend eight or 10 hours a day tracking and testing all this stuff like I do and talking to the industry experts.
00:40:00.079 --> 00:40:07.840
So if you need to know what is happening in AI to make decisions for your company, just put me to work for you.
00:40:08.000 --> 00:40:08.320
All right.
00:40:08.480 --> 00:40:16.480
So if you haven't already, please make sure to subscribe to the podcast on Apple or on Spotify and then go to your everydayai.com.
00:40:16.559 --> 00:40:17.519
So thank you for tuning in.
00:40:17.679 --> 00:40:21.280
Hope to see you back tomorrow and every day for more everyday AI.
00:40:21.360 --> 00:40:22.320
Thanks, y'all.
