WEBVTT
00:00:00.959 --> 00:00:04.959
Anthropic had a really, really bad week.
00:00:05.120 --> 00:00:06.639
So, some context first.
00:00:06.799 --> 00:00:23.440
When Anthropic released its very popular fable mythos 5 models, they warned the world that they were too dangerous, hyped them for months, released it, got it pulled by the government, fought with the White House, and dragged its reputation with developers through the mud in the process.
00:00:23.679 --> 00:00:30.640
And its biggest competitor in OpenAI this past week, well, worked with the government to release its new GPT-5.
00:00:32.079 --> 00:00:43.039
Without months of trumpeting its own horn, it stole all the benchmarks back from Anthropic, and now Anthropic is having to enter the week with its tail between its legs.
00:00:43.280 --> 00:00:46.399
And what made matters even worse for Anthropic?
00:00:46.560 --> 00:01:03.600
Yes, OpenAI released even more with its new live voice mode and chat GPT for work, but Anthropic actually got two new serious competitors for its coding API business in Grok 4.5 in Meta's Muse Spark 1.1.
00:01:03.840 --> 00:01:14.000
Yep, it was one of those busy weeks in AI where positions at the top seemed to shift every day, and your head was probably left spinning trying to keep up.
00:01:14.159 --> 00:01:20.799
Don't worry, we'll get your head on straight as we unwrap it all on our weekly AI News That Matters segments.
00:01:20.959 --> 00:01:23.120
All right, welcome to Everyday AI.
00:01:23.200 --> 00:01:26.239
If you're new here, my name's Jordan, and well, we do this every day.
00:01:26.400 --> 00:01:39.680
It's your daily, unedited, unscripted live stream podcast and free daily newsletter helping business leaders like you and me not just keep up, but how we can make sense of all these happenings and use that information to get ahead to grow our company and our career.
00:01:39.840 --> 00:01:50.719
So it starts here, but please, if you haven't already, subscribe to the podcast uh on Apple or Spotify or wherever you're listening and make sure you go to our website at your everydayai.com.
00:01:50.799 --> 00:01:56.159
We're gonna highlight everything we talk about here as well as everything else that you may have missed today.
00:01:56.400 --> 00:02:01.359
All right, so y'all just don't don't pin, don't do this every day.
00:02:01.519 --> 00:02:05.840
I'm just saying that now because it's gonna start really heating up.
00:02:06.000 --> 00:02:07.840
Acceleration is accelerating.
00:02:08.000 --> 00:02:12.960
There's only gonna be new models quicker, new features quicker, new drama, new partnerships.
00:02:13.120 --> 00:02:18.400
Y'all, you have work to do, so just tune in on Mondays and well every day, and we'll keep you up to date.
00:02:18.479 --> 00:02:18.800
All right.
00:02:18.960 --> 00:02:22.000
Uh, first story, one that a lot of people didn't see coming.
00:02:22.159 --> 00:02:32.080
Well, Apple sued OpenAI on Friday, alleging that its consumer hardware team stole trade secrets through two former Apple employees.
00:02:32.240 --> 00:02:41.360
So Apple says OpenAI encouraged recruits to share confidential information and avoid scrutiny during hiring, according to reports.
00:02:41.520 --> 00:02:46.319
So its complaints named Tang Tan and uh Chang Luai.
00:02:46.560 --> 00:02:54.960
So OpenAI uh denies any wrongdoing, and it says the company focused on building its own technology rather than using Apple's secrets.
00:02:55.120 --> 00:02:57.919
So that leaves the two companies in a strange position.
00:02:58.159 --> 00:03:10.639
Apple still integrates ChatGPT with OpenAI, uh uh, and while OpenAI is also developing hardware that is likely going to compete at one point with Apple's core interface.
00:03:10.800 --> 00:03:22.240
So the filing also gave Elon Musk some fresh ammunition against Sam Altman, and Musk quickly turned Apple's allegations into another public attack on X.
00:03:22.719 --> 00:03:34.560
And for those not keeping score at home, two months earlier, Musk lost his open AI case uh against the company after a jury found that he filed too late and he has promised an appeal.
00:03:34.639 --> 00:03:40.080
So Musk took the opportunity to call Altman a scammer across several posts.
00:03:40.159 --> 00:03:51.280
And Altman, while I was grabbing my popcorn reading the uh replies, uh answered by suggesting Musk's renewed fixation showed that GBT-5 must be really good.
00:03:51.520 --> 00:03:54.960
Uh, neither man, though, is a party to Apple's case.
00:03:55.039 --> 00:04:02.479
Uh, their exchange, though, is in reaction to the foul filing, not evidence that uh Apple's allegations are true.
00:04:02.639 --> 00:04:05.120
Uh, the timing obviously adds another layer.
00:04:05.199 --> 00:04:15.840
Uh, Elon Musk's SpaceX debuted at a whopping and impressive 2.5 trillion valuation as they went public a couple of weeks ago.
00:04:15.919 --> 00:04:21.199
Well, open AI submitted its confidential IPO paperwork in June.
00:04:21.600 --> 00:04:25.199
So uh we'll see what happens with this lawsuit.
00:04:25.360 --> 00:04:33.279
It seems like it's not really anything to do with OpenAI as a company, but just the employees at OpenAI.
00:04:33.439 --> 00:04:37.439
And it could maybe impact OpenAI's hardware plans.
00:04:37.600 --> 00:04:40.560
Uh depends on what happens with this suit.
00:04:40.720 --> 00:04:50.240
And well, it puts also one of Apple's most uh powerful, at least so far, consumer partnerships under the spotlight.
00:04:50.399 --> 00:05:02.480
So obviously, um Apple's uh Siri moving forward is going to be powered in part by Google's Gemini models, but uh up until now it was powered by Chat GPT.
00:05:02.560 --> 00:05:12.079
So essentially, any uh query that you talked to the not so smart Siri uh over the past you know year and a half or so got routed to ChatGPT.
00:05:12.160 --> 00:05:14.560
Essentially, Apple said, Yeah, our Siri can't handle this.
00:05:14.720 --> 00:05:17.360
We're gonna partner with ChatGPT to give people the answers they need.
00:05:17.439 --> 00:05:24.480
So, yeah, it's one of those like awkward moments where it's like, wait, you guys are still kind of friends, but definitely enemies.
00:05:24.639 --> 00:05:33.279
And now this is uh at least the first time that we've seen a public lawsuit going from Apple against Open AI.
00:05:33.360 --> 00:05:43.040
And we have heard that they've uh open AI has hired quite a few talented uh engineers, not just on the hardware side, uh, from across Apple.
00:05:43.680 --> 00:05:53.839
All right, our next piece of AI news a big release for Apple, or sorry, that for OpenAI that makes Apple's hardware so much better.
00:05:53.920 --> 00:05:58.319
Uh my gosh, I probably played just and by played used.
00:05:58.560 --> 00:06:09.279
I probably used the new GPT live uh feature from OpenAI almost as much as I used the new GPT 5.6 soul model.
00:06:09.519 --> 00:06:20.639
So let's talk about OpenAI's new launch as they released GPT Live this past week as ChatGPT's new default voice model for paid plans.
00:06:20.800 --> 00:06:26.240
Uh, and I do think that they have some, although very limited, use for free plans as well.
00:06:26.399 --> 00:06:32.240
So previous voice conversations before this, well, they still felt kind of turn-based, right?
00:06:32.319 --> 00:06:48.560
It was walkie-talkie, and this was whether it was OpenAI's advanced voice model, Google's Gemini Live, and there's some others, but I would say those were the two leaders, but none were really good because essentially, you know, someone would speak, the model kind of waited, and then it answers.
00:06:48.720 --> 00:06:53.519
So GPT Live has this uh, you know, kind of dual mode, right?
00:06:53.680 --> 00:06:58.079
Where it can listen to you and speak at the same time, right?
00:06:58.160 --> 00:07:05.199
So it feels much more natural, and not only that, but it can actually query things in real time, which is something, right?
00:07:05.279 --> 00:07:11.519
When Gemini Live first came out, I was super excited that it had that capability, but it was always very delayed.
00:07:11.680 --> 00:07:13.199
Um, right, it couldn't do it live.
00:07:13.279 --> 00:07:22.560
So the cool thing about this new uh GPT live voice mode is you can say, like, hey, you know, what what's the biggest, you know, what's happening in my city this weekend?
00:07:22.639 --> 00:07:29.360
And instantly, you know, within like a second or two, it's gonna start talking, uh, right, about something that's factual.
00:07:29.519 --> 00:07:38.079
Whereas before, you know, with like Gemini Live as an example, it could do that, but it would take a good like four, five, 10 seconds, uh, right.
00:07:38.160 --> 00:07:45.199
And you didn't have this ability to have uh the conversation be too dual in tone, uh, right.
00:07:45.279 --> 00:07:54.319
It this this uh you know, full duplex mode, essentially, where you know you can be talking and it can interrupt you and it doesn't break the flow, and vice versa.
00:07:54.480 --> 00:07:57.920
So it's been a pretty impressive model so far.
00:07:58.560 --> 00:08:08.079
So essentially it allows the model to listen while also speaking, it can better handle interruptions, stay quiet while someone thinks, and decide when a tool is needed.
00:08:08.240 --> 00:08:15.040
But harder questions then move to the GPT-5.5 model in the background without ending the live conversation.
00:08:15.120 --> 00:08:19.360
So then the interface stays responsive while another model works.
00:08:19.519 --> 00:08:26.480
So OpenAI says more than 150 people use voice or dictation this uh each week.
00:08:26.639 --> 00:08:30.800
So this is not just a small audience, this is pretty big.
00:08:31.040 --> 00:08:43.519
Um, so right now uh it is the paid plans uh that uh have um access, although I believe like uh enterprise and edu counts uh don't have this just yet.
00:08:43.679 --> 00:08:50.559
And there should be video and screen sharing coming soon to go along with this live conversation mode.
00:08:50.720 --> 00:08:55.440
So, yeah, like I said, if you've tried the other kind of voice modes, right?
00:08:55.519 --> 00:09:03.200
So open AI had its standard voice mode, uh, open AI had its advanced voice mode, neither of those were very good.
00:09:03.360 --> 00:09:16.320
Although when advanced voice mode came out, it seemed better, but you know, for the most part, the you know, AI voice modes that came out in you know 24 and 25 were essentially just sounded more human and the latency was a little less, right?
00:09:16.399 --> 00:09:19.039
But it they couldn't call tools essentially, right?
00:09:19.200 --> 00:09:23.120
The thing that I did with this, I had a 60-page document, right?
00:09:23.279 --> 00:09:26.879
And so I just opened up my phone, I had the document saved in there.
00:09:27.039 --> 00:09:38.000
I did that, I had to do some driving, and I just talked with Chat GPT live voice for you know the 20 or so minutes that I was in the car and had a really good conversation.
00:09:38.080 --> 00:09:49.200
And that's something that none of the other voice modes have been able to do, is to consistently, you know, chat over a document, which is one of the big reasons why I always want to use voice mode, right?
00:09:49.279 --> 00:09:55.200
It's maybe one of the reasons I use the interactive join feature in Notebook LM's audio overview.
00:09:55.279 --> 00:09:58.559
So if you haven't used that, it's really good, but this is much better, right?
00:09:58.639 --> 00:10:03.840
Uh that one, it always kind of goes back to the same track, where obviously with GPT Live, right?
00:10:04.000 --> 00:10:14.879
Think uh the way I use it, I had this 60-page doc, and I'm like, hey, ask me questions, and you know, depending on my answer, you know, go and keep asking me more questions, and then I can go back and use the transcript.
00:10:14.960 --> 00:10:16.399
And it was actually super helpful.
00:10:16.639 --> 00:10:18.320
All right, our next piece of AI news.
00:10:18.480 --> 00:10:20.720
Yeah, we had a lot of releases this week, y'all.
00:10:20.879 --> 00:10:25.679
So Grok 4.5 from XAI is out.
00:10:26.000 --> 00:10:28.720
So Cursor and Space XAI.
00:10:28.799 --> 00:10:31.440
So, yes, it's called SpaceXAI.
00:10:31.840 --> 00:10:42.399
Uh, and cursor obviously is part of SpaceXAI, but they jointly trained and released Grok 4.5, giving the coding platform a model.
00:10:42.559 --> 00:10:45.200
It helped shape around its own workflow.
00:10:45.360 --> 00:10:54.720
So, cursor matters here because it owns the application and controls the model routing, and it doesn't need users to choose grok as a general chatbot.
00:10:54.879 --> 00:11:01.519
So, Anthropic has built substantial commercial momentum around coding and agentic work.
00:11:01.840 --> 00:11:07.519
So, this is part of that premium that a cheaper rival can attack.
00:11:07.679 --> 00:11:32.399
So, yeah, this is another part of the bad news for anthropic that grok 4.5 on a lot of benchmarks because you know, uh, on things like Sui Marathon, uh, there were some other, I think uh T3, um, uh there was uh terminal coding, a couple other benchmarks that Grok either tied Fable in or was actually better than Fable and Opus, obviously, as well.
00:11:32.639 --> 00:11:35.039
So it's pretty pretty interesting, right?
00:11:35.120 --> 00:11:41.360
It's not great on all benchmarks, Grok 4.5, but I think for cheap coding tasks, right?
00:11:41.440 --> 00:11:45.759
Or if your company is just coding software developments, right?
00:11:45.919 --> 00:11:50.159
I think you know, two years ago, it's like hands down, everyone's using Anthropic.
00:11:50.240 --> 00:11:53.279
Uh 18 to 12 months ago, right?
00:11:53.440 --> 00:11:54.879
Google came into the mix.
00:11:54.960 --> 00:12:04.559
Uh, and then over the past, you know, three to four months, we've seen open models now with you know open AI's uh five six um uh Luna.
00:12:04.799 --> 00:12:07.600
It's like that's a fantastic model, cost efficient.
00:12:07.759 --> 00:12:12.000
And now we have Grok 4.5 joining uh the foray.
00:12:12.320 --> 00:12:14.480
So the price comparison, yeah.
00:12:14.639 --> 00:12:31.519
So grok is two dollars per million input and six dollars per million output versus fable that is uh literally five times cheaper on the inputs and nearly 10 or sorry, five times more expensive.
00:12:31.600 --> 00:12:34.799
Uh so fable is five times more expensive than grok.
00:12:35.120 --> 00:12:40.320
Uh, and on the output side, it is nearly 10 times more expensive.
00:12:40.399 --> 00:12:48.320
So yeah, grok 4.5,$2 per million in,$6 per million out versus Fables,$10 per million in,$50 per million out.
00:12:48.480 --> 00:12:56.879
So SpaceX also reports four plus times fewer output tokens than Opus 4.8 on one test.
00:12:57.120 --> 00:13:03.759
So saying it's not only a lot cheaper, well, it's also just way more token efficient.
00:13:04.080 --> 00:13:14.799
So uh we'll see what happens here because one thing I've I've I've always said, I don't think the general public is really ever gonna use Grok.
00:13:15.279 --> 00:13:16.480
I don't think that.
00:13:16.639 --> 00:13:21.360
Um, will they use uh cursors, composer models?
00:13:21.519 --> 00:13:42.480
Maybe, but I do think that certain um types of organizations, certain people within certain departments will use groc's models, and I think maybe we'll see the first kind of mass uh enterprise adoption with Grok 4.5 just for this, for people who are just refactoring entired code bases, right?
00:13:42.559 --> 00:13:49.679
When you're talking about in those cases burning billions of tokens an hour, probably way more, right?
00:13:49.840 --> 00:13:59.759
That's when you have to start looking at like, hey, maybe our team should be off this, you know, this expensive model and be looking at something like maybe a Grok 4.5.
00:14:00.000 --> 00:14:04.879
All right, speaking of models and routing them to other places, this one's interesting.
00:14:05.039 --> 00:14:17.840
According to reports, Microsoft is now starting to use its own internal models, the Microsoft AI models, for some Excel and Outlook features, according to a Bloomberg report.
00:14:18.480 --> 00:14:26.399
So, uh, according to the report, for many employees, Copilot is not just one among several.
00:14:27.519 --> 00:14:31.440
It may be the only AI that their company can use.
00:14:31.519 --> 00:14:34.159
That's not according to reports, that's just in general, right?
00:14:34.799 --> 00:14:41.679
Because so many enterprises out there, if they just use Copilot, that's all that's approved, right?
00:14:41.840 --> 00:14:48.879
And luckily, uh, right, because Microsoft's own models aren't very good yet, they're getting better.
00:14:49.039 --> 00:15:07.519
Um, right, but the the reason I think Copilot has still been fairly uh decent to above average versus you know the dozens of other competitors out there as well, because they've been able to use the anthropic and open AI models, right?
00:15:07.679 --> 00:15:08.879
Powering Copilot.
00:15:09.039 --> 00:15:16.879
But now, yeah, according to reports, they're starting to try to save money and swap it out and use their own MAI models.
00:15:17.039 --> 00:15:20.960
So Microsoft can still switch the intelligence underneath the product.
00:15:21.120 --> 00:15:28.480
So most employees will never know whether it's an MAI, open AI, or whether it's anthropics that answered.
00:15:28.720 --> 00:15:39.440
So tens of thousands of Excel and Outlook prompts already are reportedly running through MAI each week, making this a production change rather than a future promise.
00:15:39.679 --> 00:15:42.159
So the likely architecture, though, is tiered.
00:15:42.320 --> 00:15:49.279
Microsoft can still use kind of this cheaper internal models for routine prompts and then reserve, you know, GPT-5.
00:15:55.200 --> 00:15:58.240
Quality is also the unresolved issue here.
00:15:58.320 --> 00:15:58.879
Uh, right.
00:15:58.960 --> 00:16:10.559
So one recent large government trial found that 44% of participants using other generated generative eye gen AI tools preferred them versus co-pilot.
00:16:10.639 --> 00:16:12.799
Yeah, Kat's got my tongue today, apparently.
00:16:12.960 --> 00:16:15.279
So essentially, here's what it means, right?
00:16:15.440 --> 00:16:26.879
Uh, I think Microsoft, we've seen reports that they're looking at potentially uh using Deep Seek models for some of their um co-pilot co-work.
00:16:27.039 --> 00:16:32.559
So it's not surprising necessarily that they're gonna start to use some of their own internal models.
00:16:32.639 --> 00:16:41.200
But I think ultimately what this is, which I've been talking about this now for a couple of months, it's kind of this this overhang, right?
00:16:41.600 --> 00:16:46.879
Like 90-ish percent of what these models, right?
00:16:47.120 --> 00:16:51.440
GPT 5.6, Claude Fable 5, 90 of their capabilities.
00:16:51.759 --> 00:16:54.000
Most people don't need it, right?
00:16:54.240 --> 00:16:58.480
Until our day-to-day work catches up with where the models are.
00:16:58.720 --> 00:17:04.160
A lot of what these more frontier models are, well, it's overkill, right?
00:17:04.240 --> 00:17:06.480
That's why even within the own systems, right?
00:17:06.559 --> 00:17:10.559
One thing I set up over the week is well, I set up a lot of fun things, by the way.
00:17:10.640 --> 00:17:17.519
Now I'm you know, using um, you know, using uh soul, uh 5.6 soul inside of codex.
00:17:17.759 --> 00:17:31.680
I'm using fable inside of or wait, I'm using, let me get this right, uh, right, using Claude Code and then pulling in uh GPT 5.6 soul and then vice versa inside codex, um, you know, using Fable 5.
00:17:31.759 --> 00:17:40.240
But I think this is also why now internally people are setting up skills using the same models within one company, right?
00:17:40.480 --> 00:17:50.079
Where it's hey, we're just gonna use Fable in Claude Code or in Claude Cowork as the planner, and then we're gonna use something like Sonnet 5 or something like that.
00:17:50.160 --> 00:18:08.559
So, even not just for people that are paying for the tokens via the API, but also people, you know, if you're on a$20 a month plan, you're gonna have to start using these things because as the models become more and more powerful, well, all that means is they're just using more tokens, they're calling more tools, they're becoming more agentic.
00:18:08.720 --> 00:18:16.000
So, you know, six, nine months ago, a$20 a month AI plan subscription felt great.
00:18:16.160 --> 00:18:18.079
So it's these two things happening at once.
00:18:18.240 --> 00:18:30.400
All these companies, you know, cutting costs, and they're saying, well, the token subsidy is over, and the fact that, well, agentic systems are using more tokens, and obviously those things go hand in foot.
00:18:30.640 --> 00:18:42.640
But uh it goes to tell you here that Microsoft, both looking at Deep Seek and their own internal models, but I think this is actually indicative of a bigger shift moving ahead.
00:18:42.799 --> 00:18:52.319
So, you know, if you're a Microsoft copilot organization, I wouldn't look at this in a bad way necessarily, because I'm sure they're doing it extremely meticulously and strategically.
00:18:52.480 --> 00:18:55.519
So ultimately, users don't notice, right?
00:18:55.599 --> 00:19:02.000
Because if all of a sudden people that don't already, right, the general consensus around copilot is it's not the best, right?
00:19:02.160 --> 00:19:21.119
A lot of people, you know, that at least I talk to at larger organizations, once they get approval for something like you know, Claude or OpenAI, they're usually happy because you know, uh Copilot, Microsoft Copilot, as good as it is, it's just the learning curve is outrageous just because of the permissions and the security, right?
00:19:21.200 --> 00:19:30.960
It's so hard where things like you know, Claude and um oh you know, open AI, ChatGPT codecs are kind of like plug and play and very easy, less of a learning curve.
00:19:31.119 --> 00:19:38.400
Uh, but yeah, even within their own ecosystem, I think you're gonna start seeing you know people setting up these model routers, right?
00:19:38.480 --> 00:19:48.720
Like, yeah, using Fable Five as the you know front-end strategist and then kicking some of the actual work to like a sonnet or an Opus or maybe even a haiku.
00:19:49.359 --> 00:19:49.759
All right.
00:19:50.000 --> 00:19:59.680
Our next piece of AI news, a new study, is saying that well, most Americans don't think AI companies should own their own companies.
00:19:59.920 --> 00:20:14.400
So a new VeraSite survey found that 69% of those surveyed supported requiring major AI companies to transfer half of their stock into a publicly owned fund.
00:20:14.720 --> 00:20:22.799
Uh, so decent size survey, it covered about 1700 US adults, and it ran uh this past June.
00:20:22.880 --> 00:20:28.640
So a fairly recent poll, and it did support different party lines.
00:20:28.799 --> 00:20:35.599
So uh this was interesting because I thought originally seeing these kind of studies, I'm like, oh, this is gonna be across party lines, and it wasn't.
00:20:35.759 --> 00:20:49.039
So supports crossed party lines as 82 percent of those respondents were Democrats, 77 independents, and 59 of Republicans initially backed the proposal.
00:20:49.200 --> 00:20:56.240
Uh, then you know, when they uh survey people told people a little bit more information when they attached Bernie Sanders' name.
00:20:56.400 --> 00:21:03.119
Uh overall, the support fell to 64, uh, although nearly half of the Republicans still supported it.
00:21:03.200 --> 00:21:09.200
So, yeah, former uh presidential candidate, Democrat, uh Democrat independent uh Bernie Sanders.
00:21:09.279 --> 00:21:24.000
So, Sanders did introduce that related legislation about a month and a half ago, but forcing companies to surrender half their equity obviously faces enormous political, legal, and implementation barriers.
00:21:24.240 --> 00:21:42.240
And this new survey arrives as Frontier Labs prepare for public markets and technology companies cut jobs and the public debates who will capture AI's financial upside as they're looking to move into people's backyards for the needed data centers.
00:21:42.480 --> 00:21:46.960
So, also Sanders bill that he put out, it's not gonna pass, right?
00:21:47.119 --> 00:21:48.400
Uh, this is right.
00:21:48.480 --> 00:21:57.680
I used to cover politics like 95% of the stuff that you read about uh you know in the news, it never actually comes to a vote, right?
00:21:57.839 --> 00:22:06.880
Um, it's a way for Politicians to you know kick some dust and to you know push certain agendas.
00:22:07.039 --> 00:22:14.079
But you know what's interesting here is although I think that the 50% thing will never happen, right?
00:22:14.240 --> 00:22:33.359
Uh what's interesting is we talked about this on the show, I think last week is well, reportedly there might be uh some sort of a deal between open AI uh and the White House, giving the White House slash maybe the American people a cut of open AI when the company goes public.
00:22:33.680 --> 00:22:44.240
So, I mean, if employment disruption grows, a wealth distribution fund could become a recurring part of AI policy rather than just a fringe idea.
00:22:44.480 --> 00:22:52.720
But if you're thinking that, you know, uh Americans are gonna own 50% of AI companies, I'm gonna tell you that's not gonna happen.
00:22:52.880 --> 00:23:07.119
Might there be one of these smaller sovereign wealth funds uh where certain companies either uh voluntarily or are forced into giving a small percentage, you know, maybe uh, you know, one to five percent.
00:23:07.279 --> 00:23:18.880
It's a possibility, and we've seen that so far uh with open AI, according to reports, going to the table uh with the White House and seemingly being in favor of something like this, right?
00:23:18.960 --> 00:23:28.160
Not necessarily, there was no report saying uh that they were agreeable to the terms that were uh reported, but at least that they're in talks.
00:23:28.880 --> 00:23:33.279
All right, let's go to our next AI news story.
00:23:33.599 --> 00:23:42.079
A big week for Meta, and well, another uh maybe punch in the gut for anthropic with this one as well.
00:23:42.400 --> 00:24:00.000
So Meta compressed its year-long AI catch up into three days as the company released Muse Image, a Muse video preview, then its Muse Spark 1.1 model, and also its first public API.
00:24:00.400 --> 00:24:06.799
So Muse Image can search, write code, refine its own work, and coordinate with Muse Spark.
00:24:06.960 --> 00:24:14.079
So it launched across Meta AI, Instagram stories, and select what apps markets, uh, what app markets.
00:24:14.319 --> 00:24:35.759
So one rollout choice though, immediately backfired uh because meta previously include the ability to use AI to uh essentially kind of like the cameo feature feature that openai had in Sora, where you could use AI to just make uh any photos of you know public Instagram accounts, and meta pulled that out after a few a few days.
00:24:35.920 --> 00:24:38.880
All right, so that's meta or sorry, muse image.
00:24:38.960 --> 00:24:43.359
It actually looked pretty impressive with the uh kind of results that I saw.
00:24:43.599 --> 00:24:52.880
Then they have Muse Video, so this remains an early preview, so this is not a broad release, unlike the uh Muse image.
00:24:52.960 --> 00:25:01.519
So meta says it supports uh Muse Video supports native audio while acknowledging gaps in synchronization and fast motion.
00:25:01.599 --> 00:25:05.440
So, but the samples they did show looked pretty good, right?
00:25:05.599 --> 00:25:25.440
But uh it looks like it might be a while until there's a public-facing product on the video side, and then meta followed it up with Spark 1.1, uh, their newest large language model, which adds stronger coding, computer use, multi-agent orchestration, and a 1 million token context window.
00:25:25.599 --> 00:25:30.720
And that version now powers Meta AI's thinking mode.
00:25:31.440 --> 00:25:32.880
So here we go again.
00:25:33.119 --> 00:25:40.480
Zuck came out and essentially said, All the other companies are milking you when you're paying for tokens.
00:25:40.640 --> 00:25:44.559
We're gonna offer you a good model for very cheap.
00:25:44.720 --> 00:25:50.640
So yeah, uh, this is very much in the Grok 4.5 price range.
00:25:50.799 --> 00:26:03.359
So uh right now you can access it in public preview for a dollar twenty-five per million input tokens and four twenty-five per million output tokens.
00:26:03.519 --> 00:26:12.559
So yeah, uh very much uh they're in line uh with the grok prices at two and six, uh, right.
00:26:12.640 --> 00:26:18.880
So even a little less than that at a dollar twenty-five and four dollars and twenty-five cents.
00:26:19.039 --> 00:26:31.519
So uh I think overall, right, if you're just comparing the two, because if you're a company that just as an example, right, software development company, uh something in um you know, financial analysis, right?
00:26:31.599 --> 00:26:42.160
But essentially, if your company has become a uh token burning institution, but on the dev side, writing code, you know, you might be looking at some of these models licking your chops.
00:26:42.240 --> 00:27:08.319
I do think, at least according to most benchmarks, that it seems like Rock 4.5 is a little bit better on the coding side, uh, but at least Meta's new Muse Spark 1.1 is a viable alternative, which I most people were not expecting uh Meta to be on the bingo card of you know top four or five uh coding agents in mid-2026.
00:27:08.640 --> 00:27:24.319
But here we are, at least when it comes to price per performance, because like I said, it seems like the rest of 2026 and heading into 2027, price per performance is going to be the big factor, not just what is the best benchmark that they can throw out there.
00:27:24.559 --> 00:27:37.279
And speaking of uh Meta's good week, uh meta investors also learned that Meta plans to begin producing Iris, its first proprietary AI chip this September.
00:27:37.440 --> 00:27:44.000
So it will supplement rather than replace NVIDIA's A and AMD's GPUs.
00:27:44.079 --> 00:27:50.480
Uh so yeah, uh Meta is obviously one of NVIDIA's biggest customers, so they said that they just need more compute.
00:27:50.640 --> 00:28:22.240
So Wall Street liked the combined picture of the new photo model, the new video model, uh, the new text model, and the new iris chip, as their shares gained more than five percent on Friday, and they've already erased their 2026 losses, and Meta finished the week about 15 higher, so an overall huge week for Meta, but their week was not the best week, and that was open AI.
00:28:22.400 --> 00:28:43.200
So uh, we already alluded to the new GPT 5.6 soul, and we already kind of talked a little bit about the GPT live one model, and we did FYI do a dedicated episode on the new GPT 5.6 soul and the chat GPT work on Friday.
00:28:43.279 --> 00:28:44.720
So make sure you just go listen to that.
00:28:44.799 --> 00:28:46.400
That's episode 816.
00:28:46.720 --> 00:28:53.279
But let's do the quick recap here of what open AI introduced, and it is good.
00:28:53.440 --> 00:29:07.599
So open AI launched GPT 5.6 across three different varieties, as well as Chat GPT work, pairing a frontier model release with a new desktop agent for everyday knowledge work.
00:29:07.759 --> 00:29:10.160
So GPT 5.6 comes in three tiers.
00:29:10.319 --> 00:29:18.079
Soul is the flagship, while Terra is the medium model, and then low Luna is the cheaper and faster version.
00:29:18.559 --> 00:29:24.640
So Soul leads, uh, I mean, in just pretty much every single benchmark.
00:29:24.720 --> 00:29:28.480
So uh LM Arena, it's tied with Fable 5 at the top.
00:29:28.640 --> 00:29:32.799
If you look at artificial analysis, they have three main uh indexes.
00:29:32.960 --> 00:29:39.279
So yeah, GPT-5 Seoul leads in two of the three, and the other three, it's like one point behind.
00:29:39.519 --> 00:29:46.960
Uh so at least if you look at the uh four kind of major indexes that bring all these different votes and benchmarks together.
00:29:47.039 --> 00:29:51.359
Uh, open AI is tied or leading in three of the four.
00:29:51.440 --> 00:29:53.519
In the other one, they're only behind by one point.
00:29:53.680 --> 00:29:57.279
But the surprising thing here is right, GPT-5.
00:30:01.599 --> 00:30:08.559
Which is crazy because right, everyone said, Oh, it's gonna take years for you know open AI to ever catch anthropic when it comes to design.
00:30:08.720 --> 00:30:16.880
So, in a third party, uh, the most respected third party uh model where the people uh and experts vote on it.
00:30:17.039 --> 00:30:22.319
Yes, uh GPT 5.6 is uh great at front-end design.
00:30:22.480 --> 00:30:39.839
Uh so open AI's pricing, though, is the bigger pressure points because they list it at five dollars per million input and thirty dollars per million token output, uh, versus that 10 in 50 for Fable.
00:30:40.079 --> 00:30:46.559
So, yeah, it's about half of the price for more of the performance.
00:30:46.799 --> 00:30:51.279
So, yeah, that's why I said Anthropics had a really bad week.
00:30:51.359 --> 00:30:55.680
So, not only did they get knocked off at the frontier, uh, right, and we'll see.
00:30:55.759 --> 00:31:01.119
I'm sure Fable 5-1 uh or Opus 5 is going to be here any hour.
00:31:01.279 --> 00:31:04.000
And I have a guess on when that's actually going to be coming.
00:31:04.160 --> 00:31:29.200
So, not only did they get knocked off at the top end of their business, well, they also got maybe not knocked off, but they got chipped away at the bottom end because uh the last couple of months, uh, open weight models have actually become good enough for a lot of production use, and then you came in with the proprietary options that are very cheap and very good in Grock 4.5 and Muse Spark 1.1.
00:31:29.359 --> 00:31:34.480
So then all also from OpenAI is Chat GPT Work.
00:31:35.119 --> 00:31:38.079
So this turns the same model family, right?
00:31:38.160 --> 00:31:43.440
So the GPT 5.6 models into finished documents, right?
00:31:43.599 --> 00:31:47.359
Spreadsheets, presentations, analysis, websites, and other professional work.
00:31:47.599 --> 00:31:52.559
So uh the simplest way to put it, chat GPT work, because there is a little bit of confusion.
00:31:52.720 --> 00:31:54.240
I was confused as well.
00:31:54.400 --> 00:32:05.519
So essentially, the if you go to download the chat GPT desktop app, now you get well, it's called Chat GPT Work, which is essentially like codex light.
00:32:05.759 --> 00:32:12.000
Okay, so uh ChatGPT work is just codex, but a little simpler, right?
00:32:12.079 --> 00:32:18.480
It doesn't have you know as much of the uh coding tools accessible, it's just a cleaner, sleeker look.
00:32:18.640 --> 00:32:26.160
So it's if you're just used to the web version of Chat GPT, uh, it might be like 10% overwhelming at first.
00:32:26.319 --> 00:32:32.480
But what's strange, y'all know I'm a huge codecs lover, and I'm like Codecs is still available, FYI.
00:32:32.559 --> 00:32:43.680
There's been a lot of confusion on that, but essentially when you download the new Chat GPT desktop app now, ChatGPT Work, or if you go to download codecs, they're the same thing, and then there's a toggle at the top, right?
00:32:43.759 --> 00:32:51.839
So you either choose Chat GPT Work, and that gives you all the same features, just in a more uh minimal, non-technical interface.
00:32:51.920 --> 00:32:58.160
So it does all the same things, or you can choose codecs in the same app and then get some of those more technical views, right?
00:32:58.319 --> 00:33:04.799
So, as an example, in codecs now, there's a dedicated tab for pull requests, right?
00:33:05.440 --> 00:33:11.920
You don't have that in uh in in chat GPT work, and or you don't have the the diff view, right?
00:33:12.000 --> 00:33:15.680
So you can look at the difference between two generated pieces of codes.
00:33:15.759 --> 00:33:26.640
Uh, so there's a lot of actually since this was released on Thursday, there's actually been a handful of updates, and the team did announce that another big update is coming to the desktop app on Monday.
00:33:26.799 --> 00:33:35.599
This is great timing, it just so happens I'm doing a uh training on this, you know, a couple hours after this next big release is coming out.
00:33:35.759 --> 00:33:46.880
But uh, regardless, if you haven't checked out the new Chat GPT Work desktop app, it's really good because it is essentially just an easier, simpler version of Codex.
00:33:47.200 --> 00:33:54.319
All right, so that is a wrap for the big stories, but let's quickly go over the what's new and what's next.
00:33:54.480 --> 00:34:03.599
So uh Meta withdrew their Muse Image Instagram account reference feature, like I referenced, uh, after privacy backlash and criticism.
00:34:03.680 --> 00:34:08.079
Uh Infropic updated Claude Code with an in-app browser.
00:34:08.239 --> 00:34:08.639
All right.
00:34:08.719 --> 00:34:12.800
So took them a while to bring that after Codex has had it for a couple of months.
00:34:12.880 --> 00:34:18.719
Uh OpenAI's GPT-5.6 soul model set an ARC AGI3 benchmark record.
00:34:18.880 --> 00:34:25.039
Uh Meta Muse, uh, we talked about that already, but to recap, new image and video models.
00:34:25.119 --> 00:34:39.280
Uh rumors are suggesting that OpenAI will fast track GPT 6 with a new pre-train for potentially a summer launch to uh counter Anthropic's upcoming Fable 5 Mythos 5.1.
00:34:40.159 --> 00:34:42.000
Uh Meta discontinued.
00:34:42.159 --> 00:34:43.039
Oh, talked about that.
00:34:43.119 --> 00:34:44.400
I got some duplicates here.
00:34:44.480 --> 00:34:49.280
This is how you know this is put together by a human, and AI would never put together duplicates, right?
00:34:49.440 --> 00:34:55.440
Uh, CNBC reported a China-linked backdoor security threat tied to Claude Code.
00:34:55.599 --> 00:35:05.840
Uh OpenAI and new OpenAI entry solved all five at coder algorithm contest problems, outperforming top human competitors.
00:35:06.000 --> 00:35:07.920
Yes, the first time that's ever happened.
00:35:08.000 --> 00:35:15.280
It's a little above my head, but uh pretty impressive that OpenAI was able to solve all five at coder algorithm contest problems.
00:35:15.519 --> 00:35:20.800
All right, next, open AI also released their new GPT real-time 2.1.
00:35:20.880 --> 00:35:26.000
So kind of a version of the GPT live one that they released for ChatGPT.
00:35:26.320 --> 00:35:30.239
Um Notion Agents uh came out.
00:35:30.320 --> 00:35:33.840
That's a standalone iPhone app where you can use Notion Agents.
00:35:34.000 --> 00:35:40.320
Uh Google Voice AI, that's Gemini powered automatic call transcription and summary for businesses.
00:35:40.480 --> 00:35:43.599
Uh Bite Dances, C Dream 5.0 Pro.
00:35:43.840 --> 00:35:48.719
That's their new advanced image model with design, infographics, and layer editing.
00:35:48.800 --> 00:35:52.239
It is out, and their new video model should be coming out any day.
00:35:52.400 --> 00:36:01.519
Uh Anthropics 5, yeah, Enthropics Fable 5 got the access extended through July 19th.
00:36:01.679 --> 00:36:02.639
Oh, weird.
00:36:02.719 --> 00:36:03.519
What happened there?
00:36:03.599 --> 00:36:06.320
Well, that's a good place to end the show.
00:36:06.480 --> 00:36:35.679
And you know, I kind of alluded to this, I think, on the podcast um on Thursday and Friday, and I posted something on Twitter because originally the uh anthropic's Fable 5 was only supposed to be in uh subscription plans through yesterday, through Sunday, July 12th, and after GBD56 Soul came out, I said there's a zero percent chance that Anthropic gets rid of Fable Five in subscriptions.
00:36:35.760 --> 00:36:37.920
They will be bleeding eight figures a day.
00:36:38.000 --> 00:36:42.559
There's literally a zero percent chance, and sure enough, they extended it through next week.
00:36:42.639 --> 00:36:43.760
So, what does that mean?
00:36:43.920 --> 00:36:44.880
Well, that means this.
00:36:45.039 --> 00:36:46.239
This is my bold prediction.
00:36:46.400 --> 00:36:48.639
Join me next week, and we'll see if I was right or not.
00:36:48.800 --> 00:36:52.000
They're gonna do one of a couple of things, which is gonna be exciting to watch.
00:36:52.159 --> 00:37:06.880
You're either gonna come out with a mythos uh fable 5.1, which I would put a low probability, a higher probability is we see maybe an Opus 5 that would be at that GPT 5.6 soul level.
00:37:07.039 --> 00:37:20.559
Um, or we're gonna see a price cut for uh Fable 5 because I've heard like respected people in the industry essentially say that Fable 5 is a dead model, right?
00:37:20.639 --> 00:37:21.679
Not me saying it.
00:37:21.840 --> 00:37:29.920
Literally, I've heard it and seen it from a lot of respected people that are said we're never gonna touch Fable Five again because of the pricing, right?
00:37:30.000 --> 00:37:42.480
The capabilities are there, but when open AI comes and they undercut it, and you say, Oh, Fable Five, this 5.6 soul is as good, but actually better, and it costs like half.
00:37:43.039 --> 00:37:45.119
And depending on what benchmark you look at, right?
00:37:45.199 --> 00:37:48.320
Like the artificial analysis cost index, it's like a third.
00:37:48.480 --> 00:37:50.559
Yeah, it's not gonna that's what I said.
00:37:50.719 --> 00:37:52.239
Very bad week for anthropic.
00:37:52.320 --> 00:38:00.159
But what makes that fun is we do know that we're gonna get either new models or cheaper prices from anthropic probably in the next week.
00:38:00.320 --> 00:38:01.360
That would be my guess.
00:38:01.519 --> 00:38:03.119
All right, so that's a wrap.
00:38:03.199 --> 00:38:04.079
I hope this was helpful.
00:38:04.159 --> 00:38:09.199
If so, if you haven't already, please subscribe to the podcast and then go to your everydayai.com.
00:38:09.280 --> 00:38:10.639
Sign up for the free daily newsletter.
00:38:10.719 --> 00:38:13.280
We'll see you back tomorrow in every day for more everyday AI.
00:38:13.360 --> 00:38:14.320
Thanks, y'all.
00:00:00.959 --> 00:00:04.959
Anthropic had a really, really bad week.
00:00:05.120 --> 00:00:06.639
So, some context first.
00:00:06.799 --> 00:00:23.440
When Anthropic released its very popular fable mythos 5 models, they warned the world that they were too dangerous, hyped them for months, released it, got it pulled by the government, fought with the White House, and dragged its reputation with developers through the mud in the process.
00:00:23.679 --> 00:00:30.640
And its biggest competitor in OpenAI this past week, well, worked with the government to release its new GPT-5.
00:00:32.079 --> 00:00:43.039
Without months of trumpeting its own horn, it stole all the benchmarks back from Anthropic, and now Anthropic is having to enter the week with its tail between its legs.
00:00:43.280 --> 00:00:46.399
And what made matters even worse for Anthropic?
00:00:46.560 --> 00:01:03.600
Yes, OpenAI released even more with its new live voice mode and chat GPT for work, but Anthropic actually got two new serious competitors for its coding API business in Grok 4.5 in Meta's Muse Spark 1.1.
00:01:03.840 --> 00:01:14.000
Yep, it was one of those busy weeks in AI where positions at the top seemed to shift every day, and your head was probably left spinning trying to keep up.
00:01:14.159 --> 00:01:20.799
Don't worry, we'll get your head on straight as we unwrap it all on our weekly AI News That Matters segments.
00:01:20.959 --> 00:01:23.120
All right, welcome to Everyday AI.
00:01:23.200 --> 00:01:26.239
If you're new here, my name's Jordan, and well, we do this every day.
00:01:26.400 --> 00:01:39.680
It's your daily, unedited, unscripted live stream podcast and free daily newsletter helping business leaders like you and me not just keep up, but how we can make sense of all these happenings and use that information to get ahead to grow our company and our career.
00:01:39.840 --> 00:01:50.719
So it starts here, but please, if you haven't already, subscribe to the podcast uh on Apple or Spotify or wherever you're listening and make sure you go to our website at your everydayai.com.
00:01:50.799 --> 00:01:56.159
We're gonna highlight everything we talk about here as well as everything else that you may have missed today.
00:01:56.400 --> 00:02:01.359
All right, so y'all just don't don't pin, don't do this every day.
00:02:01.519 --> 00:02:05.840
I'm just saying that now because it's gonna start really heating up.
00:02:06.000 --> 00:02:07.840
Acceleration is accelerating.
00:02:08.000 --> 00:02:12.960
There's only gonna be new models quicker, new features quicker, new drama, new partnerships.
00:02:13.120 --> 00:02:18.400
Y'all, you have work to do, so just tune in on Mondays and well every day, and we'll keep you up to date.
00:02:18.479 --> 00:02:18.800
All right.
00:02:18.960 --> 00:02:22.000
Uh, first story, one that a lot of people didn't see coming.
00:02:22.159 --> 00:02:32.080
Well, Apple sued OpenAI on Friday, alleging that its consumer hardware team stole trade secrets through two former Apple employees.
00:02:32.240 --> 00:02:41.360
So Apple says OpenAI encouraged recruits to share confidential information and avoid scrutiny during hiring, according to reports.
00:02:41.520 --> 00:02:46.319
So its complaints named Tang Tan and uh Chang Luai.
00:02:46.560 --> 00:02:54.960
So OpenAI uh denies any wrongdoing, and it says the company focused on building its own technology rather than using Apple's secrets.
00:02:55.120 --> 00:02:57.919
So that leaves the two companies in a strange position.
00:02:58.159 --> 00:03:10.639
Apple still integrates ChatGPT with OpenAI, uh uh, and while OpenAI is also developing hardware that is likely going to compete at one point with Apple's core interface.
00:03:10.800 --> 00:03:22.240
So the filing also gave Elon Musk some fresh ammunition against Sam Altman, and Musk quickly turned Apple's allegations into another public attack on X.
00:03:22.719 --> 00:03:34.560
And for those not keeping score at home, two months earlier, Musk lost his open AI case uh against the company after a jury found that he filed too late and he has promised an appeal.
00:03:34.639 --> 00:03:40.080
So Musk took the opportunity to call Altman a scammer across several posts.
00:03:40.159 --> 00:03:51.280
And Altman, while I was grabbing my popcorn reading the uh replies, uh answered by suggesting Musk's renewed fixation showed that GBT-5 must be really good.
00:03:51.520 --> 00:03:54.960
Uh, neither man, though, is a party to Apple's case.
00:03:55.039 --> 00:04:02.479
Uh, their exchange, though, is in reaction to the foul filing, not evidence that uh Apple's allegations are true.
00:04:02.639 --> 00:04:05.120
Uh, the timing obviously adds another layer.
00:04:05.199 --> 00:04:15.840
Uh, Elon Musk's SpaceX debuted at a whopping and impressive 2.5 trillion valuation as they went public a couple of weeks ago.
00:04:15.919 --> 00:04:21.199
Well, open AI submitted its confidential IPO paperwork in June.
00:04:21.600 --> 00:04:25.199
So uh we'll see what happens with this lawsuit.
00:04:25.360 --> 00:04:33.279
It seems like it's not really anything to do with OpenAI as a company, but just the employees at OpenAI.
00:04:33.439 --> 00:04:37.439
And it could maybe impact OpenAI's hardware plans.
00:04:37.600 --> 00:04:40.560
Uh depends on what happens with this suit.
00:04:40.720 --> 00:04:50.240
And well, it puts also one of Apple's most uh powerful, at least so far, consumer partnerships under the spotlight.
00:04:50.399 --> 00:05:02.480
So obviously, um Apple's uh Siri moving forward is going to be powered in part by Google's Gemini models, but uh up until now it was powered by Chat GPT.
00:05:02.560 --> 00:05:12.079
So essentially, any uh query that you talked to the not so smart Siri uh over the past you know year and a half or so got routed to ChatGPT.
00:05:12.160 --> 00:05:14.560
Essentially, Apple said, Yeah, our Siri can't handle this.
00:05:14.720 --> 00:05:17.360
We're gonna partner with ChatGPT to give people the answers they need.
00:05:17.439 --> 00:05:24.480
So, yeah, it's one of those like awkward moments where it's like, wait, you guys are still kind of friends, but definitely enemies.
00:05:24.639 --> 00:05:33.279
And now this is uh at least the first time that we've seen a public lawsuit going from Apple against Open AI.
00:05:33.360 --> 00:05:43.040
And we have heard that they've uh open AI has hired quite a few talented uh engineers, not just on the hardware side, uh, from across Apple.
00:05:43.680 --> 00:05:53.839
All right, our next piece of AI news a big release for Apple, or sorry, that for OpenAI that makes Apple's hardware so much better.
00:05:53.920 --> 00:05:58.319
Uh my gosh, I probably played just and by played used.
00:05:58.560 --> 00:06:09.279
I probably used the new GPT live uh feature from OpenAI almost as much as I used the new GPT 5.6 soul model.
00:06:09.519 --> 00:06:20.639
So let's talk about OpenAI's new launch as they released GPT Live this past week as ChatGPT's new default voice model for paid plans.
00:06:20.800 --> 00:06:26.240
Uh, and I do think that they have some, although very limited, use for free plans as well.
00:06:26.399 --> 00:06:32.240
So previous voice conversations before this, well, they still felt kind of turn-based, right?
00:06:32.319 --> 00:06:48.560
It was walkie-talkie, and this was whether it was OpenAI's advanced voice model, Google's Gemini Live, and there's some others, but I would say those were the two leaders, but none were really good because essentially, you know, someone would speak, the model kind of waited, and then it answers.
00:06:48.720 --> 00:06:53.519
So GPT Live has this uh, you know, kind of dual mode, right?
00:06:53.680 --> 00:06:58.079
Where it can listen to you and speak at the same time, right?
00:06:58.160 --> 00:07:05.199
So it feels much more natural, and not only that, but it can actually query things in real time, which is something, right?
00:07:05.279 --> 00:07:11.519
When Gemini Live first came out, I was super excited that it had that capability, but it was always very delayed.
00:07:11.680 --> 00:07:13.199
Um, right, it couldn't do it live.
00:07:13.279 --> 00:07:22.560
So the cool thing about this new uh GPT live voice mode is you can say, like, hey, you know, what what's the biggest, you know, what's happening in my city this weekend?
00:07:22.639 --> 00:07:29.360
And instantly, you know, within like a second or two, it's gonna start talking, uh, right, about something that's factual.
00:07:29.519 --> 00:07:38.079
Whereas before, you know, with like Gemini Live as an example, it could do that, but it would take a good like four, five, 10 seconds, uh, right.
00:07:38.160 --> 00:07:45.199
And you didn't have this ability to have uh the conversation be too dual in tone, uh, right.
00:07:45.279 --> 00:07:54.319
It this this uh you know, full duplex mode, essentially, where you know you can be talking and it can interrupt you and it doesn't break the flow, and vice versa.
00:07:54.480 --> 00:07:57.920
So it's been a pretty impressive model so far.
00:07:58.560 --> 00:08:08.079
So essentially it allows the model to listen while also speaking, it can better handle interruptions, stay quiet while someone thinks, and decide when a tool is needed.
00:08:08.240 --> 00:08:15.040
But harder questions then move to the GPT-5.5 model in the background without ending the live conversation.
00:08:15.120 --> 00:08:19.360
So then the interface stays responsive while another model works.
00:08:19.519 --> 00:08:26.480
So OpenAI says more than 150 people use voice or dictation this uh each week.
00:08:26.639 --> 00:08:30.800
So this is not just a small audience, this is pretty big.
00:08:31.040 --> 00:08:43.519
Um, so right now uh it is the paid plans uh that uh have um access, although I believe like uh enterprise and edu counts uh don't have this just yet.
00:08:43.679 --> 00:08:50.559
And there should be video and screen sharing coming soon to go along with this live conversation mode.
00:08:50.720 --> 00:08:55.440
So, yeah, like I said, if you've tried the other kind of voice modes, right?
00:08:55.519 --> 00:09:03.200
So open AI had its standard voice mode, uh, open AI had its advanced voice mode, neither of those were very good.
00:09:03.360 --> 00:09:16.320
Although when advanced voice mode came out, it seemed better, but you know, for the most part, the you know, AI voice modes that came out in you know 24 and 25 were essentially just sounded more human and the latency was a little less, right?
00:09:16.399 --> 00:09:19.039
But it they couldn't call tools essentially, right?
00:09:19.200 --> 00:09:23.120
The thing that I did with this, I had a 60-page document, right?
00:09:23.279 --> 00:09:26.879
And so I just opened up my phone, I had the document saved in there.
00:09:27.039 --> 00:09:38.000
I did that, I had to do some driving, and I just talked with Chat GPT live voice for you know the 20 or so minutes that I was in the car and had a really good conversation.
00:09:38.080 --> 00:09:49.200
And that's something that none of the other voice modes have been able to do, is to consistently, you know, chat over a document, which is one of the big reasons why I always want to use voice mode, right?
00:09:49.279 --> 00:09:55.200
It's maybe one of the reasons I use the interactive join feature in Notebook LM's audio overview.
00:09:55.279 --> 00:09:58.559
So if you haven't used that, it's really good, but this is much better, right?
00:09:58.639 --> 00:10:03.840
Uh that one, it always kind of goes back to the same track, where obviously with GPT Live, right?
00:10:04.000 --> 00:10:14.879
Think uh the way I use it, I had this 60-page doc, and I'm like, hey, ask me questions, and you know, depending on my answer, you know, go and keep asking me more questions, and then I can go back and use the transcript.
00:10:14.960 --> 00:10:16.399
And it was actually super helpful.
00:10:16.639 --> 00:10:18.320
All right, our next piece of AI news.
00:10:18.480 --> 00:10:20.720
Yeah, we had a lot of releases this week, y'all.
00:10:20.879 --> 00:10:25.679
So Grok 4.5 from XAI is out.
00:10:26.000 --> 00:10:28.720
So Cursor and Space XAI.
00:10:28.799 --> 00:10:31.440
So, yes, it's called SpaceXAI.
00:10:31.840 --> 00:10:42.399
Uh, and cursor obviously is part of SpaceXAI, but they jointly trained and released Grok 4.5, giving the coding platform a model.
00:10:42.559 --> 00:10:45.200
It helped shape around its own workflow.
00:10:45.360 --> 00:10:54.720
So, cursor matters here because it owns the application and controls the model routing, and it doesn't need users to choose grok as a general chatbot.
00:10:54.879 --> 00:11:01.519
So, Anthropic has built substantial commercial momentum around coding and agentic work.
00:11:01.840 --> 00:11:07.519
So, this is part of that premium that a cheaper rival can attack.
00:11:07.679 --> 00:11:32.399
So, yeah, this is another part of the bad news for anthropic that grok 4.5 on a lot of benchmarks because you know, uh, on things like Sui Marathon, uh, there were some other, I think uh T3, um, uh there was uh terminal coding, a couple other benchmarks that Grok either tied Fable in or was actually better than Fable and Opus, obviously, as well.
00:11:32.639 --> 00:11:35.039
So it's pretty pretty interesting, right?
00:11:35.120 --> 00:11:41.360
It's not great on all benchmarks, Grok 4.5, but I think for cheap coding tasks, right?
00:11:41.440 --> 00:11:45.759
Or if your company is just coding software developments, right?
00:11:45.919 --> 00:11:50.159
I think you know, two years ago, it's like hands down, everyone's using Anthropic.
00:11:50.240 --> 00:11:53.279
Uh 18 to 12 months ago, right?
00:11:53.440 --> 00:11:54.879
Google came into the mix.
00:11:54.960 --> 00:12:04.559
Uh, and then over the past, you know, three to four months, we've seen open models now with you know open AI's uh five six um uh Luna.
00:12:04.799 --> 00:12:07.600
It's like that's a fantastic model, cost efficient.
00:12:07.759 --> 00:12:12.000
And now we have Grok 4.5 joining uh the foray.
00:12:12.320 --> 00:12:14.480
So the price comparison, yeah.
00:12:14.639 --> 00:12:31.519
So grok is two dollars per million input and six dollars per million output versus fable that is uh literally five times cheaper on the inputs and nearly 10 or sorry, five times more expensive.
00:12:31.600 --> 00:12:34.799
Uh so fable is five times more expensive than grok.
00:12:35.120 --> 00:12:40.320
Uh, and on the output side, it is nearly 10 times more expensive.
00:12:40.399 --> 00:12:48.320
So yeah, grok 4.5,$2 per million in,$6 per million out versus Fables,$10 per million in,$50 per million out.
00:12:48.480 --> 00:12:56.879
So SpaceX also reports four plus times fewer output tokens than Opus 4.8 on one test.
00:12:57.120 --> 00:13:03.759
So saying it's not only a lot cheaper, well, it's also just way more token efficient.
00:13:04.080 --> 00:13:14.799
So uh we'll see what happens here because one thing I've I've I've always said, I don't think the general public is really ever gonna use Grok.
00:13:15.279 --> 00:13:16.480
I don't think that.
00:13:16.639 --> 00:13:21.360
Um, will they use uh cursors, composer models?
00:13:21.519 --> 00:13:42.480
Maybe, but I do think that certain um types of organizations, certain people within certain departments will use groc's models, and I think maybe we'll see the first kind of mass uh enterprise adoption with Grok 4.5 just for this, for people who are just refactoring entired code bases, right?
00:13:42.559 --> 00:13:49.679
When you're talking about in those cases burning billions of tokens an hour, probably way more, right?
00:13:49.840 --> 00:13:59.759
That's when you have to start looking at like, hey, maybe our team should be off this, you know, this expensive model and be looking at something like maybe a Grok 4.5.
00:14:00.000 --> 00:14:04.879
All right, speaking of models and routing them to other places, this one's interesting.
00:14:05.039 --> 00:14:17.840
According to reports, Microsoft is now starting to use its own internal models, the Microsoft AI models, for some Excel and Outlook features, according to a Bloomberg report.
00:14:18.480 --> 00:14:26.399
So, uh, according to the report, for many employees, Copilot is not just one among several.
00:14:27.519 --> 00:14:31.440
It may be the only AI that their company can use.
00:14:31.519 --> 00:14:34.159
That's not according to reports, that's just in general, right?
00:14:34.799 --> 00:14:41.679
Because so many enterprises out there, if they just use Copilot, that's all that's approved, right?
00:14:41.840 --> 00:14:48.879
And luckily, uh, right, because Microsoft's own models aren't very good yet, they're getting better.
00:14:49.039 --> 00:15:07.519
Um, right, but the the reason I think Copilot has still been fairly uh decent to above average versus you know the dozens of other competitors out there as well, because they've been able to use the anthropic and open AI models, right?
00:15:07.679 --> 00:15:08.879
Powering Copilot.
00:15:09.039 --> 00:15:16.879
But now, yeah, according to reports, they're starting to try to save money and swap it out and use their own MAI models.
00:15:17.039 --> 00:15:20.960
So Microsoft can still switch the intelligence underneath the product.
00:15:21.120 --> 00:15:28.480
So most employees will never know whether it's an MAI, open AI, or whether it's anthropics that answered.
00:15:28.720 --> 00:15:39.440
So tens of thousands of Excel and Outlook prompts already are reportedly running through MAI each week, making this a production change rather than a future promise.
00:15:39.679 --> 00:15:42.159
So the likely architecture, though, is tiered.
00:15:42.320 --> 00:15:49.279
Microsoft can still use kind of this cheaper internal models for routine prompts and then reserve, you know, GPT-5.
00:15:55.200 --> 00:15:58.240
Quality is also the unresolved issue here.
00:15:58.320 --> 00:15:58.879
Uh, right.
00:15:58.960 --> 00:16:10.559
So one recent large government trial found that 44% of participants using other generated generative eye gen AI tools preferred them versus co-pilot.
00:16:10.639 --> 00:16:12.799
Yeah, Kat's got my tongue today, apparently.
00:16:12.960 --> 00:16:15.279
So essentially, here's what it means, right?
00:16:15.440 --> 00:16:26.879
Uh, I think Microsoft, we've seen reports that they're looking at potentially uh using Deep Seek models for some of their um co-pilot co-work.
00:16:27.039 --> 00:16:32.559
So it's not surprising necessarily that they're gonna start to use some of their own internal models.
00:16:32.639 --> 00:16:41.200
But I think ultimately what this is, which I've been talking about this now for a couple of months, it's kind of this this overhang, right?
00:16:41.600 --> 00:16:46.879
Like 90-ish percent of what these models, right?
00:16:47.120 --> 00:16:51.440
GPT 5.6, Claude Fable 5, 90 of their capabilities.
00:16:51.759 --> 00:16:54.000
Most people don't need it, right?
00:16:54.240 --> 00:16:58.480
Until our day-to-day work catches up with where the models are.
00:16:58.720 --> 00:17:04.160
A lot of what these more frontier models are, well, it's overkill, right?
00:17:04.240 --> 00:17:06.480
That's why even within the own systems, right?
00:17:06.559 --> 00:17:10.559
One thing I set up over the week is well, I set up a lot of fun things, by the way.
00:17:10.640 --> 00:17:17.519
Now I'm you know, using um, you know, using uh soul, uh 5.6 soul inside of codex.
00:17:17.759 --> 00:17:31.680
I'm using fable inside of or wait, I'm using, let me get this right, uh, right, using Claude Code and then pulling in uh GPT 5.6 soul and then vice versa inside codex, um, you know, using Fable 5.
00:17:31.759 --> 00:17:40.240
But I think this is also why now internally people are setting up skills using the same models within one company, right?
00:17:40.480 --> 00:17:50.079
Where it's hey, we're just gonna use Fable in Claude Code or in Claude Cowork as the planner, and then we're gonna use something like Sonnet 5 or something like that.
00:17:50.160 --> 00:18:08.559
So, even not just for people that are paying for the tokens via the API, but also people, you know, if you're on a$20 a month plan, you're gonna have to start using these things because as the models become more and more powerful, well, all that means is they're just using more tokens, they're calling more tools, they're becoming more agentic.
00:18:08.720 --> 00:18:16.000
So, you know, six, nine months ago, a$20 a month AI plan subscription felt great.
00:18:16.160 --> 00:18:18.079
So it's these two things happening at once.
00:18:18.240 --> 00:18:30.400
All these companies, you know, cutting costs, and they're saying, well, the token subsidy is over, and the fact that, well, agentic systems are using more tokens, and obviously those things go hand in foot.
00:18:30.640 --> 00:18:42.640
But uh it goes to tell you here that Microsoft, both looking at Deep Seek and their own internal models, but I think this is actually indicative of a bigger shift moving ahead.
00:18:42.799 --> 00:18:52.319
So, you know, if you're a Microsoft copilot organization, I wouldn't look at this in a bad way necessarily, because I'm sure they're doing it extremely meticulously and strategically.
00:18:52.480 --> 00:18:55.519
So ultimately, users don't notice, right?
00:18:55.599 --> 00:19:02.000
Because if all of a sudden people that don't already, right, the general consensus around copilot is it's not the best, right?
00:19:02.160 --> 00:19:21.119
A lot of people, you know, that at least I talk to at larger organizations, once they get approval for something like you know, Claude or OpenAI, they're usually happy because you know, uh Copilot, Microsoft Copilot, as good as it is, it's just the learning curve is outrageous just because of the permissions and the security, right?
00:19:21.200 --> 00:19:30.960
It's so hard where things like you know, Claude and um oh you know, open AI, ChatGPT codecs are kind of like plug and play and very easy, less of a learning curve.
00:19:31.119 --> 00:19:38.400
Uh, but yeah, even within their own ecosystem, I think you're gonna start seeing you know people setting up these model routers, right?
00:19:38.480 --> 00:19:48.720
Like, yeah, using Fable Five as the you know front-end strategist and then kicking some of the actual work to like a sonnet or an Opus or maybe even a haiku.
00:19:49.359 --> 00:19:49.759
All right.
00:19:50.000 --> 00:19:59.680
Our next piece of AI news, a new study, is saying that well, most Americans don't think AI companies should own their own companies.
00:19:59.920 --> 00:20:14.400
So a new VeraSite survey found that 69% of those surveyed supported requiring major AI companies to transfer half of their stock into a publicly owned fund.
00:20:14.720 --> 00:20:22.799
Uh, so decent size survey, it covered about 1700 US adults, and it ran uh this past June.
00:20:22.880 --> 00:20:28.640
So a fairly recent poll, and it did support different party lines.
00:20:28.799 --> 00:20:35.599
So uh this was interesting because I thought originally seeing these kind of studies, I'm like, oh, this is gonna be across party lines, and it wasn't.
00:20:35.759 --> 00:20:49.039
So supports crossed party lines as 82 percent of those respondents were Democrats, 77 independents, and 59 of Republicans initially backed the proposal.
00:20:49.200 --> 00:20:56.240
Uh, then you know, when they uh survey people told people a little bit more information when they attached Bernie Sanders' name.
00:20:56.400 --> 00:21:03.119
Uh overall, the support fell to 64, uh, although nearly half of the Republicans still supported it.
00:21:03.200 --> 00:21:09.200
So, yeah, former uh presidential candidate, Democrat, uh Democrat independent uh Bernie Sanders.
00:21:09.279 --> 00:21:24.000
So, Sanders did introduce that related legislation about a month and a half ago, but forcing companies to surrender half their equity obviously faces enormous political, legal, and implementation barriers.
00:21:24.240 --> 00:21:42.240
And this new survey arrives as Frontier Labs prepare for public markets and technology companies cut jobs and the public debates who will capture AI's financial upside as they're looking to move into people's backyards for the needed data centers.
00:21:42.480 --> 00:21:46.960
So, also Sanders bill that he put out, it's not gonna pass, right?
00:21:47.119 --> 00:21:48.400
Uh, this is right.
00:21:48.480 --> 00:21:57.680
I used to cover politics like 95% of the stuff that you read about uh you know in the news, it never actually comes to a vote, right?
00:21:57.839 --> 00:22:06.880
Um, it's a way for Politicians to you know kick some dust and to you know push certain agendas.
00:22:07.039 --> 00:22:14.079
But you know what's interesting here is although I think that the 50% thing will never happen, right?
00:22:14.240 --> 00:22:33.359
Uh what's interesting is we talked about this on the show, I think last week is well, reportedly there might be uh some sort of a deal between open AI uh and the White House, giving the White House slash maybe the American people a cut of open AI when the company goes public.
00:22:33.680 --> 00:22:44.240
So, I mean, if employment disruption grows, a wealth distribution fund could become a recurring part of AI policy rather than just a fringe idea.
00:22:44.480 --> 00:22:52.720
But if you're thinking that, you know, uh Americans are gonna own 50% of AI companies, I'm gonna tell you that's not gonna happen.
00:22:52.880 --> 00:23:07.119
Might there be one of these smaller sovereign wealth funds uh where certain companies either uh voluntarily or are forced into giving a small percentage, you know, maybe uh, you know, one to five percent.
00:23:07.279 --> 00:23:18.880
It's a possibility, and we've seen that so far uh with open AI, according to reports, going to the table uh with the White House and seemingly being in favor of something like this, right?
00:23:18.960 --> 00:23:28.160
Not necessarily, there was no report saying uh that they were agreeable to the terms that were uh reported, but at least that they're in talks.
00:23:28.880 --> 00:23:33.279
All right, let's go to our next AI news story.
00:23:33.599 --> 00:23:42.079
A big week for Meta, and well, another uh maybe punch in the gut for anthropic with this one as well.
00:23:42.400 --> 00:24:00.000
So Meta compressed its year-long AI catch up into three days as the company released Muse Image, a Muse video preview, then its Muse Spark 1.1 model, and also its first public API.
00:24:00.400 --> 00:24:06.799
So Muse Image can search, write code, refine its own work, and coordinate with Muse Spark.
00:24:06.960 --> 00:24:14.079
So it launched across Meta AI, Instagram stories, and select what apps markets, uh, what app markets.
00:24:14.319 --> 00:24:35.759
So one rollout choice though, immediately backfired uh because meta previously include the ability to use AI to uh essentially kind of like the cameo feature feature that openai had in Sora, where you could use AI to just make uh any photos of you know public Instagram accounts, and meta pulled that out after a few a few days.
00:24:35.920 --> 00:24:38.880
All right, so that's meta or sorry, muse image.
00:24:38.960 --> 00:24:43.359
It actually looked pretty impressive with the uh kind of results that I saw.
00:24:43.599 --> 00:24:52.880
Then they have Muse Video, so this remains an early preview, so this is not a broad release, unlike the uh Muse image.
00:24:52.960 --> 00:25:01.519
So meta says it supports uh Muse Video supports native audio while acknowledging gaps in synchronization and fast motion.
00:25:01.599 --> 00:25:05.440
So, but the samples they did show looked pretty good, right?
00:25:05.599 --> 00:25:25.440
But uh it looks like it might be a while until there's a public-facing product on the video side, and then meta followed it up with Spark 1.1, uh, their newest large language model, which adds stronger coding, computer use, multi-agent orchestration, and a 1 million token context window.
00:25:25.599 --> 00:25:30.720
And that version now powers Meta AI's thinking mode.
00:25:31.440 --> 00:25:32.880
So here we go again.
00:25:33.119 --> 00:25:40.480
Zuck came out and essentially said, All the other companies are milking you when you're paying for tokens.
00:25:40.640 --> 00:25:44.559
We're gonna offer you a good model for very cheap.
00:25:44.720 --> 00:25:50.640
So yeah, uh, this is very much in the Grok 4.5 price range.
00:25:50.799 --> 00:26:03.359
So uh right now you can access it in public preview for a dollar twenty-five per million input tokens and four twenty-five per million output tokens.
00:26:03.519 --> 00:26:12.559
So yeah, uh very much uh they're in line uh with the grok prices at two and six, uh, right.
00:26:12.640 --> 00:26:18.880
So even a little less than that at a dollar twenty-five and four dollars and twenty-five cents.
00:26:19.039 --> 00:26:31.519
So uh I think overall, right, if you're just comparing the two, because if you're a company that just as an example, right, software development company, uh something in um you know, financial analysis, right?
00:26:31.599 --> 00:26:42.160
But essentially, if your company has become a uh token burning institution, but on the dev side, writing code, you know, you might be looking at some of these models licking your chops.
00:26:42.240 --> 00:27:08.319
I do think, at least according to most benchmarks, that it seems like Rock 4.5 is a little bit better on the coding side, uh, but at least Meta's new Muse Spark 1.1 is a viable alternative, which I most people were not expecting uh Meta to be on the bingo card of you know top four or five uh coding agents in mid-2026.
00:27:08.640 --> 00:27:24.319
But here we are, at least when it comes to price per performance, because like I said, it seems like the rest of 2026 and heading into 2027, price per performance is going to be the big factor, not just what is the best benchmark that they can throw out there.
00:27:24.559 --> 00:27:37.279
And speaking of uh Meta's good week, uh meta investors also learned that Meta plans to begin producing Iris, its first proprietary AI chip this September.
00:27:37.440 --> 00:27:44.000
So it will supplement rather than replace NVIDIA's A and AMD's GPUs.
00:27:44.079 --> 00:27:50.480
Uh so yeah, uh Meta is obviously one of NVIDIA's biggest customers, so they said that they just need more compute.
00:27:50.640 --> 00:28:22.240
So Wall Street liked the combined picture of the new photo model, the new video model, uh, the new text model, and the new iris chip, as their shares gained more than five percent on Friday, and they've already erased their 2026 losses, and Meta finished the week about 15 higher, so an overall huge week for Meta, but their week was not the best week, and that was open AI.
00:28:22.400 --> 00:28:43.200
So uh, we already alluded to the new GPT 5.6 soul, and we already kind of talked a little bit about the GPT live one model, and we did FYI do a dedicated episode on the new GPT 5.6 soul and the chat GPT work on Friday.
00:28:43.279 --> 00:28:44.720
So make sure you just go listen to that.
00:28:44.799 --> 00:28:46.400
That's episode 816.
00:28:46.720 --> 00:28:53.279
But let's do the quick recap here of what open AI introduced, and it is good.
00:28:53.440 --> 00:29:07.599
So open AI launched GPT 5.6 across three different varieties, as well as Chat GPT work, pairing a frontier model release with a new desktop agent for everyday knowledge work.
00:29:07.759 --> 00:29:10.160
So GPT 5.6 comes in three tiers.
00:29:10.319 --> 00:29:18.079
Soul is the flagship, while Terra is the medium model, and then low Luna is the cheaper and faster version.
00:29:18.559 --> 00:29:24.640
So Soul leads, uh, I mean, in just pretty much every single benchmark.
00:29:24.720 --> 00:29:28.480
So uh LM Arena, it's tied with Fable 5 at the top.
00:29:28.640 --> 00:29:32.799
If you look at artificial analysis, they have three main uh indexes.
00:29:32.960 --> 00:29:39.279
So yeah, GPT-5 Seoul leads in two of the three, and the other three, it's like one point behind.
00:29:39.519 --> 00:29:46.960
Uh so at least if you look at the uh four kind of major indexes that bring all these different votes and benchmarks together.
00:29:47.039 --> 00:29:51.359
Uh, open AI is tied or leading in three of the four.
00:29:51.440 --> 00:29:53.519
In the other one, they're only behind by one point.
00:29:53.680 --> 00:29:57.279
But the surprising thing here is right, GPT-5.
00:30:01.599 --> 00:30:08.559
Which is crazy because right, everyone said, Oh, it's gonna take years for you know open AI to ever catch anthropic when it comes to design.
00:30:08.720 --> 00:30:16.880
So, in a third party, uh, the most respected third party uh model where the people uh and experts vote on it.
00:30:17.039 --> 00:30:22.319
Yes, uh GPT 5.6 is uh great at front-end design.
00:30:22.480 --> 00:30:39.839
Uh so open AI's pricing, though, is the bigger pressure points because they list it at five dollars per million input and thirty dollars per million token output, uh, versus that 10 in 50 for Fable.
00:30:40.079 --> 00:30:46.559
So, yeah, it's about half of the price for more of the performance.
00:30:46.799 --> 00:30:51.279
So, yeah, that's why I said Anthropics had a really bad week.
00:30:51.359 --> 00:30:55.680
So, not only did they get knocked off at the frontier, uh, right, and we'll see.
00:30:55.759 --> 00:31:01.119
I'm sure Fable 5-1 uh or Opus 5 is going to be here any hour.
00:31:01.279 --> 00:31:04.000
And I have a guess on when that's actually going to be coming.
00:31:04.160 --> 00:31:29.200
So, not only did they get knocked off at the top end of their business, well, they also got maybe not knocked off, but they got chipped away at the bottom end because uh the last couple of months, uh, open weight models have actually become good enough for a lot of production use, and then you came in with the proprietary options that are very cheap and very good in Grock 4.5 and Muse Spark 1.1.
00:31:29.359 --> 00:31:34.480
So then all also from OpenAI is Chat GPT Work.
00:31:35.119 --> 00:31:38.079
So this turns the same model family, right?
00:31:38.160 --> 00:31:43.440
So the GPT 5.6 models into finished documents, right?
00:31:43.599 --> 00:31:47.359
Spreadsheets, presentations, analysis, websites, and other professional work.
00:31:47.599 --> 00:31:52.559
So uh the simplest way to put it, chat GPT work, because there is a little bit of confusion.
00:31:52.720 --> 00:31:54.240
I was confused as well.
00:31:54.400 --> 00:32:05.519
So essentially, the if you go to download the chat GPT desktop app, now you get well, it's called Chat GPT Work, which is essentially like codex light.
00:32:05.759 --> 00:32:12.000
Okay, so uh ChatGPT work is just codex, but a little simpler, right?
00:32:12.079 --> 00:32:18.480
It doesn't have you know as much of the uh coding tools accessible, it's just a cleaner, sleeker look.
00:32:18.640 --> 00:32:26.160
So it's if you're just used to the web version of Chat GPT, uh, it might be like 10% overwhelming at first.
00:32:26.319 --> 00:32:32.480
But what's strange, y'all know I'm a huge codecs lover, and I'm like Codecs is still available, FYI.
00:32:32.559 --> 00:32:43.680
There's been a lot of confusion on that, but essentially when you download the new Chat GPT desktop app now, ChatGPT Work, or if you go to download codecs, they're the same thing, and then there's a toggle at the top, right?
00:32:43.759 --> 00:32:51.839
So you either choose Chat GPT Work, and that gives you all the same features, just in a more uh minimal, non-technical interface.
00:32:51.920 --> 00:32:58.160
So it does all the same things, or you can choose codecs in the same app and then get some of those more technical views, right?
00:32:58.319 --> 00:33:04.799
So, as an example, in codecs now, there's a dedicated tab for pull requests, right?
00:33:05.440 --> 00:33:11.920
You don't have that in uh in in chat GPT work, and or you don't have the the diff view, right?
00:33:12.000 --> 00:33:15.680
So you can look at the difference between two generated pieces of codes.
00:33:15.759 --> 00:33:26.640
Uh, so there's a lot of actually since this was released on Thursday, there's actually been a handful of updates, and the team did announce that another big update is coming to the desktop app on Monday.
00:33:26.799 --> 00:33:35.599
This is great timing, it just so happens I'm doing a uh training on this, you know, a couple hours after this next big release is coming out.
00:33:35.759 --> 00:33:46.880
But uh, regardless, if you haven't checked out the new Chat GPT Work desktop app, it's really good because it is essentially just an easier, simpler version of Codex.
00:33:47.200 --> 00:33:54.319
All right, so that is a wrap for the big stories, but let's quickly go over the what's new and what's next.
00:33:54.480 --> 00:34:03.599
So uh Meta withdrew their Muse Image Instagram account reference feature, like I referenced, uh, after privacy backlash and criticism.
00:34:03.680 --> 00:34:08.079
Uh Infropic updated Claude Code with an in-app browser.
00:34:08.239 --> 00:34:08.639
All right.
00:34:08.719 --> 00:34:12.800
So took them a while to bring that after Codex has had it for a couple of months.
00:34:12.880 --> 00:34:18.719
Uh OpenAI's GPT-5.6 soul model set an ARC AGI3 benchmark record.
00:34:18.880 --> 00:34:25.039
Uh Meta Muse, uh, we talked about that already, but to recap, new image and video models.
00:34:25.119 --> 00:34:39.280
Uh rumors are suggesting that OpenAI will fast track GPT 6 with a new pre-train for potentially a summer launch to uh counter Anthropic's upcoming Fable 5 Mythos 5.1.
00:34:40.159 --> 00:34:42.000
Uh Meta discontinued.
00:34:42.159 --> 00:34:43.039
Oh, talked about that.
00:34:43.119 --> 00:34:44.400
I got some duplicates here.
00:34:44.480 --> 00:34:49.280
This is how you know this is put together by a human, and AI would never put together duplicates, right?
00:34:49.440 --> 00:34:55.440
Uh, CNBC reported a China-linked backdoor security threat tied to Claude Code.
00:34:55.599 --> 00:35:05.840
Uh OpenAI and new OpenAI entry solved all five at coder algorithm contest problems, outperforming top human competitors.
00:35:06.000 --> 00:35:07.920
Yes, the first time that's ever happened.
00:35:08.000 --> 00:35:15.280
It's a little above my head, but uh pretty impressive that OpenAI was able to solve all five at coder algorithm contest problems.
00:35:15.519 --> 00:35:20.800
All right, next, open AI also released their new GPT real-time 2.1.
00:35:20.880 --> 00:35:26.000
So kind of a version of the GPT live one that they released for ChatGPT.
00:35:26.320 --> 00:35:30.239
Um Notion Agents uh came out.
00:35:30.320 --> 00:35:33.840
That's a standalone iPhone app where you can use Notion Agents.
00:35:34.000 --> 00:35:40.320
Uh Google Voice AI, that's Gemini powered automatic call transcription and summary for businesses.
00:35:40.480 --> 00:35:43.599
Uh Bite Dances, C Dream 5.0 Pro.
00:35:43.840 --> 00:35:48.719
That's their new advanced image model with design, infographics, and layer editing.
00:35:48.800 --> 00:35:52.239
It is out, and their new video model should be coming out any day.
00:35:52.400 --> 00:36:01.519
Uh Anthropics 5, yeah, Enthropics Fable 5 got the access extended through July 19th.
00:36:01.679 --> 00:36:02.639
Oh, weird.
00:36:02.719 --> 00:36:03.519
What happened there?
00:36:03.599 --> 00:36:06.320
Well, that's a good place to end the show.
00:36:06.480 --> 00:36:35.679
And you know, I kind of alluded to this, I think, on the podcast um on Thursday and Friday, and I posted something on Twitter because originally the uh anthropic's Fable 5 was only supposed to be in uh subscription plans through yesterday, through Sunday, July 12th, and after GBD56 Soul came out, I said there's a zero percent chance that Anthropic gets rid of Fable Five in subscriptions.
00:36:35.760 --> 00:36:37.920
They will be bleeding eight figures a day.
00:36:38.000 --> 00:36:42.559
There's literally a zero percent chance, and sure enough, they extended it through next week.
00:36:42.639 --> 00:36:43.760
So, what does that mean?
00:36:43.920 --> 00:36:44.880
Well, that means this.
00:36:45.039 --> 00:36:46.239
This is my bold prediction.
00:36:46.400 --> 00:36:48.639
Join me next week, and we'll see if I was right or not.
00:36:48.800 --> 00:36:52.000
They're gonna do one of a couple of things, which is gonna be exciting to watch.
00:36:52.159 --> 00:37:06.880
You're either gonna come out with a mythos uh fable 5.1, which I would put a low probability, a higher probability is we see maybe an Opus 5 that would be at that GPT 5.6 soul level.
00:37:07.039 --> 00:37:20.559
Um, or we're gonna see a price cut for uh Fable 5 because I've heard like respected people in the industry essentially say that Fable 5 is a dead model, right?
00:37:20.639 --> 00:37:21.679
Not me saying it.
00:37:21.840 --> 00:37:29.920
Literally, I've heard it and seen it from a lot of respected people that are said we're never gonna touch Fable Five again because of the pricing, right?
00:37:30.000 --> 00:37:42.480
The capabilities are there, but when open AI comes and they undercut it, and you say, Oh, Fable Five, this 5.6 soul is as good, but actually better, and it costs like half.
00:37:43.039 --> 00:37:45.119
And depending on what benchmark you look at, right?
00:37:45.199 --> 00:37:48.320
Like the artificial analysis cost index, it's like a third.
00:37:48.480 --> 00:37:50.559
Yeah, it's not gonna that's what I said.
00:37:50.719 --> 00:37:52.239
Very bad week for anthropic.
00:37:52.320 --> 00:38:00.159
But what makes that fun is we do know that we're gonna get either new models or cheaper prices from anthropic probably in the next week.
00:38:00.320 --> 00:38:01.360
That would be my guess.
00:38:01.519 --> 00:38:03.119
All right, so that's a wrap.
00:38:03.199 --> 00:38:04.079
I hope this was helpful.
00:38:04.159 --> 00:38:09.199
If so, if you haven't already, please subscribe to the podcast and then go to your everydayai.com.
00:38:09.280 --> 00:38:10.639
Sign up for the free daily newsletter.
00:38:10.719 --> 00:38:13.280
We'll see you back tomorrow in every day for more everyday AI.
00:38:13.360 --> 00:38:14.320
Thanks, y'all.
