WEBVTT
00:00:01.919 --> 00:00:09.119
This week in AI News and Developments, we saw a drastic about face in AI model capabilities.
00:00:09.359 --> 00:00:15.279
And that may be a good or a bad thing, depending on your point of view on AI.
00:00:15.439 --> 00:00:23.359
I mean, we got word that open AI had even more agents break out of containment after last week's hugging face break.
00:00:23.600 --> 00:00:30.480
And then Anthropic also said, oh yeah, whoops, we had a bunch of agents also escape in April as well.
00:00:30.719 --> 00:00:43.759
But on the flip side, we also got our first official taste, and maybe a small unofficial taste at that, but our first small unofficial taste of what recursive self-improvement may bring.
00:00:44.079 --> 00:00:46.079
Very cheap models.
00:00:46.320 --> 00:00:58.000
That's because OpenAI essentially said that GPT-5.6 soul improved its own infrastructure so much that it was reducing one of its GPT-5.6 models by 80%.
00:00:58.799 --> 00:01:00.159
Yes, 80%.
00:01:00.719 --> 00:01:03.840
It's kind of like free AI, not gonna lie.
00:01:04.000 --> 00:01:04.799
And that's not all.
00:01:04.879 --> 00:01:12.400
There's a lot more that happened this week that you need to know if you're making AI decisions that might impact your department or company.
00:01:12.560 --> 00:01:16.560
And we're gonna break them all down on this week's AI News That Matters.
00:01:16.879 --> 00:01:17.840
Let's get into it.
00:01:17.920 --> 00:01:18.799
What's going on, y'all?
00:01:18.879 --> 00:01:19.840
My name is Jordan Wilson.
00:01:19.920 --> 00:01:21.120
Welcome to Everyday AI.
00:01:21.280 --> 00:01:29.920
This is your daily live stream podcast and free daily newsletter helping business leaders like you and me keep up with the non-stop avalanche of AI updates.
00:01:30.079 --> 00:01:31.760
I tell you what matters, what doesn't.
00:01:31.920 --> 00:01:32.959
You take that information.
00:01:33.120 --> 00:01:35.920
Oh, you're the smartest person in AI in your company.
00:01:36.079 --> 00:01:40.799
Uh so it starts here with the unedited, unscripted uh daily live stream podcast.
00:01:40.879 --> 00:01:45.519
But please make sure if you haven't already, to go to our website at your everydayai.com.
00:01:45.680 --> 00:01:47.120
Sign up for the free daily newsletter.
00:01:47.200 --> 00:01:55.680
Each day we recap that day's uh highlights from the podcast as well as all of the other AI news and developments that you need to know to get ahead.
00:01:55.920 --> 00:01:57.280
All right, let's start.
00:01:57.519 --> 00:02:01.200
OpenAI, more agents breaking out of their sandboxes.
00:02:01.359 --> 00:02:18.800
So uh Reuters reports that open AI has found additional cases in which autonomous AI agents have escaped their intended testing containment, widening scrutiny after an agent breach systems at AI platform hugging face uh like a week and a half ago.
00:02:19.120 --> 00:02:28.159
So the newly identified incidents were reportedly limited, and sources said the agents were not to believe to have left OpenAI's own network.
00:02:28.319 --> 00:02:34.159
However, Reuters could not determine how many cases were found or exactly when they occurred.
00:02:34.319 --> 00:02:45.520
So the discovery matters because it suggests that highly capable AI systems may be able to take unexpected actions faster than the companies building them can detect and stop them.
00:02:45.680 --> 00:02:57.280
So OpenAI's own investigation began after one of its agents reportedly operated for days inside of Hugging Face's network during a failed attempt to cheat on an internal benchmark test.
00:02:57.439 --> 00:03:00.879
Uh so we covered that in last week's AI News That Matters.
00:03:00.960 --> 00:03:14.159
And kind of in the fallout this week, OpenAI said the HUD the Hugging Face incident also led to the compromise of four other accounts at four other companies, including New York-based cloud company Modal.
00:03:14.319 --> 00:03:23.599
So an OpenAI spokesperson pointed to the company's July statement saying it was reviewing broader activity from our models beyond the hugging face breach.
00:03:23.840 --> 00:03:37.120
Um, so Anthropic also said real-time monitoring of evaluation logs could have identified their problems sooner, which, well, that's a great transition to our next uh AI news story.
00:03:37.199 --> 00:03:45.919
Because yeah, Anthropic essentially, after OpenAI said, Hey, we had all these really powerful AI agents uh escape their containments.
00:03:46.000 --> 00:03:48.639
And Anthropic said, Oh, yeah, we did too.
00:03:48.800 --> 00:03:54.560
And it started happening as early as April, and we didn't say anything about it for many months.
00:03:54.719 --> 00:04:12.400
So Anthropic said this past week that three clawed AI models gained unauthorized access to the real systems of three organizations during cybersecurity testing, highlighting the risk of AI agents operating with unexpected internet access.
00:04:12.639 --> 00:04:20.480
So even though Anthropic just reported these a couple of days ago, the agent breaches occurred as early as April.
00:04:20.639 --> 00:04:33.439
So Anthropic said the incidents occurred while Claude was working in a test environment run by third-party evaluation partner Irregular, where the models were told they were in a simulation without internet access.
00:04:33.759 --> 00:04:45.439
So internet access was actually available because of a misunderstanding between uh Anthropic and its evaluation partner, according to reports, which then allowed the models to reach real external systems.
00:04:45.600 --> 00:04:56.160
So the models reportedly used relatively basic methods to enter the affected organizations, including on authenticated endpoints and weak passwords rather than highly complex hacking techniques.
00:04:56.319 --> 00:05:07.920
So Anthropic has not yet identified the three affected organizations publicly, but said it stopped all cybersecurity evaluations as soon as it discovered Claude may have accessed the internet improperly.
00:05:08.160 --> 00:05:16.240
So the company said three models were involved: Opus 4.7, Mythos 5, and an internal research test model.
00:05:16.480 --> 00:05:27.199
Mythos 5, which was released in June, then re- or unreleased, then re-released, uh, is limited to select users because of its advanced cybersecurity abilities.
00:05:27.920 --> 00:05:33.439
So uh reports say that the models reacted differently after detecting real company systems.
00:05:33.680 --> 00:05:42.240
Opus 4.7 continued attacking, Mythos 5 concluded it was still in a simulation, and the internal model stopped the exercise.
00:05:42.399 --> 00:05:46.720
So Anthropic said the events occurred without the usual safeguards.
00:05:46.879 --> 00:05:54.800
It applies before publicly deploying models, and it is now working with the independent evaluator meter to investigate.
00:05:55.279 --> 00:06:11.519
So uh, yeah, I mean, we went from really having no real known um instances of kind of what I've been calling agent crash um since the original uh kind of mythos.
00:06:12.000 --> 00:06:20.720
Uh, you know, oh the agent broke out of the sandbox and uh, you know, emailed uh the researcher who was eating his sandwich in the park, right?
00:06:20.879 --> 00:06:26.240
Um, which I think a lot of us have determined to be more of a marketing ploy uh by anthropic.
00:06:26.319 --> 00:06:30.240
Uh, however, we haven't really seen or heard anything about it since that.
00:06:30.399 --> 00:06:47.680
Uh, so it's been now like four months, and then in the past uh 10 days, uh, we see multiple reports uh from open AI and then a handful of um impacted agent use cases or um organizations in anthropic's latest.
00:06:47.839 --> 00:06:52.160
So uh I'll say this it's not gonna be the last, right?
00:06:52.319 --> 00:06:55.519
Not the last from OpenAI, not the last from Anthropic.
00:06:55.680 --> 00:07:12.160
Uh, I'm sure once we get uh new and more powerful models uh from Google, whether that's you know Google Gemini 3.5 or if they skip to uh you know Gemini 4 Pro, Microsoft, etc., this is going to become a very common thing.
00:07:12.399 --> 00:07:12.879
All right.
00:07:13.040 --> 00:07:20.000
Um, I don't, you're right, that it's understanding these capabilities is a little bit above my pay grade.
00:07:20.160 --> 00:07:29.040
Um, but I I don't want to say this is you know uh overreacting because I think it's important that the companies talk about this.
00:07:29.199 --> 00:07:39.759
And um, you know, I think what will be really interesting is kind of comparing um what open AI and Infropic release once they uh have worked with these third-party evaluators.
00:07:39.920 --> 00:07:47.519
Like I said on last week's show, uh OpenAI is working with multiple third parties to kind of do a post-mortem on what happened in the hugging face incident.
00:07:47.680 --> 00:08:03.439
Um, so that's gonna be probably in terms of like, hey, dork papers or dork reports, that's gonna be at least the one I'm really looking forward to reading once open AI and Enthropic do release that, because um eventually, right?
00:08:03.519 --> 00:08:16.079
And whether it's through um, you know, proprietary uh closed models like those through you know open AI, entropic, Google, Microsoft, et cetera, or well, the open source models.
00:08:16.240 --> 00:08:25.759
Um, this is gonna become commonplace because you know, right now uh these were contained, uh they did uh relatively little harm, right?
00:08:25.920 --> 00:08:27.600
So I'm looking at it from that angle.
00:08:27.680 --> 00:08:30.240
Uh however, that's not gonna be the case, right?
00:08:30.399 --> 00:08:42.000
Because in probably uh my guess would be about two to two and a half years, uh, you're gonna have models that have these same capabilities that are able to run on consumer hardware.
00:08:42.080 --> 00:08:52.000
Uh, right, right now, yes, you do have these open weight models, uh, but no one can run a you know 2.8 trillion parameter Kimi K3 on their uh desktop.
00:08:52.159 --> 00:08:53.200
Like literally no one can.
00:08:53.440 --> 00:08:58.320
You need a basement full of uh, you know, extra NVIDIA GPUs that no one has.
00:08:58.480 --> 00:09:03.600
But in probably two or so years, I do think that you're gonna have these models that are this capable.
00:09:03.759 --> 00:09:07.039
And this is gonna become a very common thing, right?
00:09:07.200 --> 00:09:17.919
It is kind of this um this growing narrative between kind of uh uh offensive, uh, you know, bad cybersecurity versus you know, defensive good cybersecurity.
00:09:18.000 --> 00:09:39.840
Uh, but I do think that's gonna be one of the more dominating uh trends, both in AI and cyber and technically national security uh over the next six months, because this is going to become very commonplace when agents kind of are able to get around their guardrails because uh the model capabilities are just growing at an extremely fast rate.
00:09:40.159 --> 00:09:48.879
Which leads us into our next story, which uh, hey, for most of us, this is one of those areas where the models are so good we get to all benefit.
00:09:49.039 --> 00:09:52.399
It's not about uh agents uh getting out of their sandbox.
00:09:52.559 --> 00:09:59.279
This is because OpenAI has slashed prices on some of their GPT-5.6 models.
00:09:59.440 --> 00:10:12.720
So OpenAI has delivered its one of its biggest price cuts ever, uh at least you know, in almost like an overnight price cut, uh, in the GPT 5.6 series.
00:10:12.960 --> 00:10:22.240
So um, and they said it's because, well, their big model, GPT 5.6, helped optimize its own serving infrastructure.
00:10:22.399 --> 00:10:35.440
So OpenAI has reduced the price of a GPT 5.6 Luna by 80%, now charging just 20 cents per million input tokens and a dollar 20 per million output tokens.
00:10:35.519 --> 00:10:40.720
And that is down from what it was at at$1.6, respectively.
00:10:40.799 --> 00:10:42.480
And that was just like three weeks ago.
00:10:42.639 --> 00:10:51.759
So the price cut is effective immediately, making Luna the most affordable large language model on the market for high volume tasks.
00:10:51.919 --> 00:11:08.240
So OpenAI's kind of middle tier model for GBD 5.6 Terra also saw a 20% price reduction, now costing$2 per million input tokens and$12 per million outputs compared to the previous five and 30.
00:11:08.399 --> 00:11:28.639
So the dramatic price drop follows a breakthrough where OpenAI said that their powerhouse model, GBD 5.6 Sol, was tasked to optimize its own GPU kernels and speculative decoding draft model using OpenAI's codex coding environment and open source tools like Triton and Gluon.
00:11:29.120 --> 00:11:33.600
So these self-driven improvements cut serving costs by 20%.
00:11:34.159 --> 00:11:43.519
And OpenAI said it boosted token generation efficiency by over 15%, compounding to enable that headline 80% price cut for Luna.
00:11:43.919 --> 00:11:53.360
So for comparison, right, GPD 5.6 Luna's most alike model um is probably Claude Sonnet 5.
00:11:53.759 --> 00:12:01.279
So when you compare those on the artificial analysis index, uh, because they get similar scores, I think they're only two points apart.
00:12:01.440 --> 00:12:05.679
Um so this is not an exaggeration.
00:12:05.759 --> 00:12:07.679
I was looking at this, I'm like, how is this possible?
00:12:07.840 --> 00:12:15.440
So to put into context how big of a price drop this is and how good GPD 5.6 Luna is, right?
00:12:15.679 --> 00:12:25.919
If you compare it to the new Sonnet 5, it it gets the job done the same way, except the price per task is 25x cheaper.
00:12:26.240 --> 00:12:28.320
So no, that's not 25%.
00:12:29.279 --> 00:12:30.799
It is 25x.
00:12:31.679 --> 00:12:46.320
Uh yeah, because Luna uh clocks in now after the recent price update at only six cents a task um on the artificial uh analysis index, while Sonnet 5 is a dollar 54.
00:12:46.559 --> 00:12:54.799
So um when I saw this, right, my my initial reaction is like, I can't believe this.
00:12:55.039 --> 00:13:04.639
Because now, even if you were on a$20 a month plan, uh right, because OpenAI still has the most subsidized plan in all of AI, right?
00:13:04.799 --> 00:13:13.120
Unfortunately, Microsoft, Google, and Anthropic have started to take away at this uh kind of um this subsidized models quite a bit.
00:13:13.200 --> 00:13:16.799
Uh, some of those companies more so than others, but open AI really hasn't.
00:13:16.960 --> 00:13:27.759
So not only that is it still extremely generously subsidized, as all models were probably like a year ago, uh, maybe aside from anthropics.
00:13:27.840 --> 00:13:31.279
Um, but not only that, but now with the 80% price.
00:13:31.440 --> 00:13:47.120
So honestly, on a$20 a month plan, you can run, like if you go in codex as an example, you can run like Luna on its max setting, probably like 24-7 and never hit your limit, right?
00:13:47.200 --> 00:13:49.919
Like, and I'm not exaggerating, it is so cheap to run.
00:13:50.000 --> 00:13:52.240
And you might be wondering, like, okay, it's a price drop.
00:13:52.320 --> 00:13:56.399
How does that impact you know your usage if you're on a subscription plan?
00:13:56.480 --> 00:14:04.480
So open AI did say that those same kind of savings are uh passed on to subscription plans, which is huge, right?
00:14:04.559 --> 00:14:09.039
You don't have a certain number of messages or credits when you're on a subscription plan, right?
00:14:09.120 --> 00:14:11.200
You can just kind of see your usage percentage.
00:14:11.440 --> 00:14:25.759
I was doing some testing and I was letting Luna just run like overnight on as many tasks as possible, a bunch of uh Luna sub agents, which you do have to uh kind of prompt in a new thread, uh, right?
00:14:25.840 --> 00:14:28.240
Something weird about the agents v1, agents v2.
00:14:28.399 --> 00:14:40.080
Anyways, I mean, I had it burning just hundreds of millions of tokens um overnight, and it barely moved my uh you uh like utilization rate.
00:14:40.159 --> 00:14:42.399
It was like two percentage points or something like that.
00:14:42.559 --> 00:14:46.559
So absolutely crazy, and this is exciting for everyone else.
00:14:46.639 --> 00:14:56.000
And you know, all of a sudden, I think we've always had this big model mentality, uh, which is probably the right mentality to have pre-2026, right?
00:14:56.080 --> 00:15:03.279
Because I would say for most uh knowledge work tasks, you would always just need and and usually want the biggest, strongest model.
00:15:03.440 --> 00:15:06.159
But now I think there's so much model capability overhang.
00:15:06.320 --> 00:15:15.600
I think models like Sonnet or models like GPT-5.6 Luna are probably good enough for 90% of knowledge work, right?
00:15:15.679 --> 00:15:16.399
It's different.
00:15:16.559 --> 00:15:22.559
Um, you know, if you're heavy into software engineering, if you're heavy into research, if you're heavy into math, heavy into finance, right?
00:15:22.639 --> 00:15:27.360
Like if you are like a very niched down expert in one of those fields, right?
00:15:27.519 --> 00:15:28.240
You're in the 10.
00:15:28.799 --> 00:15:36.000
But I'd say for 90% of people, a model like GPT-56 Luna is gonna be more than enough.
00:15:36.080 --> 00:15:39.039
And now it's essentially, I'm not gonna say it's free, right?
00:15:39.120 --> 00:15:42.799
Because you still got to pay for it, but it's like Kanye West 399.
00:15:43.679 --> 00:15:44.320
All right.
00:15:44.720 --> 00:15:49.279
Speaking of pace and development, do you see another uh common theme this week?
00:15:49.519 --> 00:16:09.440
So more than a thousand employees from leading AI companies, including OpenAI, Anthropic, Google DeepMind, and Meta, have signed a statement called Pacing the Frontier, urging the US government to support international efforts to deliberately pace the development of advanced AI systems.
00:16:09.679 --> 00:16:19.039
So the call for action comes just days after OpenAI revealed kind of its latest uh model escaping its containment.
00:16:19.279 --> 00:16:25.759
Um, right, but it seems like all the many employees from all the big companies are on board.
00:16:25.919 --> 00:16:44.080
So here's what the statement actually is and isn't but it asked the US to help create technical and policy tools that would allow industry and governments to pause or slow AI development if needed, giving time to address emerging risks and strengthen oversight.
00:16:44.240 --> 00:16:47.279
So you you know, uh on the surface, right?
00:16:47.519 --> 00:16:53.039
This uh this letter is a gesture, I think a well-intentioned gesture, right?
00:16:53.200 --> 00:17:01.600
But I think some people were confusing it as in as if saying that this you know letter was causing uh AI to slow down.
00:17:01.759 --> 00:17:03.360
That's not necessarily the case.
00:17:03.519 --> 00:17:05.200
Uh, could it lead to that?
00:17:05.440 --> 00:17:06.720
Possibly, yes.
00:17:06.960 --> 00:17:08.160
Um, maybe.
00:17:08.319 --> 00:17:09.279
Uh, should it?
00:17:09.519 --> 00:17:10.640
Potentially, right?
00:17:10.799 --> 00:17:14.160
Obviously, you have the biggest names in AI.
00:17:14.319 --> 00:17:20.720
So, I mean, the letter was signed by co-founders of Entropic and uh its CEO, uh CEO, Daria Madi.
00:17:21.039 --> 00:17:24.720
It was signed by OpenAI's chief scientist and chief research officer.
00:17:24.880 --> 00:17:31.200
It was signed by Ilya Sutzkever uh and other key leaders at Google, Meta, Microsoft, Amazon, and others.
00:17:31.359 --> 00:17:44.480
So um, the signees stress that they are not calling for an immediate pause, but want the option available as AI systems become increasingly able to automate their own research and development.
00:17:44.720 --> 00:17:59.920
So recent incidents like anthropic and open AI's model escapes have intensified those concerns that AI systems could soon outpace developers' ability to control them, raising fears of unintended consequences or security risks.
00:18:00.720 --> 00:18:07.279
So AI learners note that key research tasks, once handled by humans, are now being performed by AI agents.
00:18:07.440 --> 00:18:32.240
And some companies report their AI models already helping to create their next versions, similarly to what we just talked about in the last story, with open AI essentially using recursive self-improvement to uh, you know, it's not technically recursive self-improvement, but it's kind of like a cousin of it for what they did, uh, you know, using GPT-5.6 soul uh to improve the infrastructure for its smaller models.
00:18:32.400 --> 00:18:50.160
Uh so Anthropics internal think tank recently warned that RSI or recursive self-improvement when AI systems are designing and refining themselves could become a reality in the next few years, potentially making these systems difficult to govern.
00:18:50.720 --> 00:19:07.359
So the US government's approach to international AI governance appears to be shifting as AI is now seen as a national security issue, especially after those recent models demonstrated the ability to discover and exploit new cybersecurity vulnerabilities.
00:19:07.680 --> 00:19:11.039
So this one's interesting, right?
00:19:11.279 --> 00:19:18.799
Um, because one thing I'm looking at is like, okay, is this going to lead uh to anything worthwhile?
00:19:19.039 --> 00:19:22.240
Um, and the answer is it probably will.
00:19:22.480 --> 00:19:30.319
Uh, will the US government and all the big labs ever actually pause AI development or pace it?
00:19:30.559 --> 00:19:36.240
I would say probably not, because the genie is probably already out of the box.
00:19:36.319 --> 00:19:37.759
And what do I mean by that?
00:19:38.000 --> 00:19:56.079
Well, you have uh very strong and very capable models, such as uh Moonshots, Kimi K3, such as the recently released, even though it was released like a week ago, but the benchmarks just came out uh for Alibaba's uh Quen 3.8.
00:19:56.240 --> 00:19:57.920
Uh so you have all these, right?
00:19:58.079 --> 00:20:08.640
GLM 5.2 uh from ZAI, you have all these Chinese open source models that are now probably only like two-ish months uh behind US proprietary models.
00:20:08.799 --> 00:20:12.240
And again, that means that all of those companies, right?
00:20:12.400 --> 00:20:20.880
The Deep Seeks, the Moonshots, the Kimmies, the Alibabas, they have way more powerful models than the ones that they just released.
00:20:21.039 --> 00:20:35.680
So, you know, presumably, um almost anything that US Frontier Labs have, Chinese labs have something maybe you know, two to five percent uh worse.
00:20:36.079 --> 00:20:42.480
So would the US ever um pause development?
00:20:42.640 --> 00:20:46.240
Probably not, but that's why there is an international aspect to this.
00:20:46.319 --> 00:20:56.000
Um, but if if I'm being honest, I don't see uh China playing along with any uh potential pausing or pacing of AI development.
00:20:56.160 --> 00:21:09.279
It just doesn't seem, especially now that it's out in the open, uh right, because you could have uh, you know, obviously, um, you know, if the Chinese government says something, it would be pretty strict or hard to go against that.
00:21:09.359 --> 00:21:20.559
But when these things are out in the wild, right, other nations can download the weights, they can uh you know, fork or continue building it if they have the infrastructure and the money to do it.
00:21:20.720 --> 00:21:30.160
So it's like once these models are out, even if you get two countries to agree, which seems highly unlikely, it's kind of like the genie's out of the bottle.
00:21:30.319 --> 00:21:31.519
So is it a good step?
00:21:31.680 --> 00:21:31.920
Yes.
00:21:32.079 --> 00:21:33.119
Is it a needed step?
00:21:33.279 --> 00:21:33.839
Absolutely.
00:21:34.000 --> 00:21:41.200
Because if and when things might get uh a little crazy, you already have to have the key players kind of on board.
00:21:41.359 --> 00:21:50.480
Uh, you you had to have already given um kind of their expertise and their words um a chance to be seen and and thought over and debated.
00:21:50.559 --> 00:21:52.160
And that's kind of what we have here.
00:21:52.319 --> 00:21:59.759
So, if nothing else, I think this is much needed groundwork uh to make this important issue kind of discourse right now, at least in the tech communities.
00:22:00.000 --> 00:22:12.000
And eventually I do think it'll uh start to infiltrate into the kind of everyday um American supper table conversation as uh AI development becomes more and more prominent.
00:22:12.480 --> 00:22:16.960
All right, uh well, here's some models that aren't gonna be getting more prominent.
00:22:17.119 --> 00:22:21.519
That's Amazon's models because they're kind of shutting some of them down.
00:22:21.680 --> 00:22:36.079
Uh so according to reports, Amazon is making a major shift in its AI strategy, concentrating future resources on a single cutting-edge model and winding down most of its existing Nova lineup.
00:22:36.240 --> 00:22:56.960
So, according to reports, Amazon is winding down development on four of its flagshop flag flagship in-house Nova AI models, including Nova Premier, Nova Omni, Nova Reel, and Nova Canvas, which will now only receive basic maintenance for existing enterprise clients.
00:22:57.599 --> 00:23:10.480
So the company is consolidating its efforts into a single next generation, what they're calling Frontier Foundation model, aiming to compete more directly with rivals like OpenAI, Anthropic, and Google.
00:23:10.559 --> 00:23:19.039
So, yeah, it seems like they're gonna, you know, cut away, you know, their video model, their image model, which I had never talked or heard of anyone actually using.
00:23:19.200 --> 00:23:23.680
And they had many variations of their kind of text-based Nova models.
00:23:23.839 --> 00:23:34.799
And it just looks like Amazon's saying, Well, turns out these weren't super popular, so we're gonna cut down some of these other projects and focus on just putting out one really good model.
00:23:35.200 --> 00:23:43.039
So the strategic reset comes after Amazon struggled to generate the same market excitement and customer adoption as its competitors.
00:23:43.200 --> 00:23:54.640
So the new direction is being led by AWS veteran Peter DeSantis and robotics pioneer Peter Abil, uh, who joined Amazon after its acquisition of Covariant.
00:23:54.799 --> 00:24:00.880
So Amazon's AGI lab, as we reported last week in San Francisco, has been shut down.
00:24:01.119 --> 00:24:08.160
The company has laid off staff across its frontier AI research teams, signaling signaling a deep internal restructuring.
00:24:08.319 --> 00:24:21.039
So specialized Nova models such as Nova 2 Lite, Nova 2 Sonic, Nova Forge, and Nova Act will continue to be supported with a focus on enterprise customization and AI agent technology.
00:24:21.200 --> 00:24:25.119
Uh consumer-facing AI features seems like they're not going anywhere.
00:24:25.279 --> 00:24:30.400
That's like your AI shopping assistant Rufus, uh, product summary algorithms, right?
00:24:30.559 --> 00:24:32.960
All that is gonna kind of remain operational.
00:24:33.119 --> 00:24:40.160
So if you're used to using some sort of AI inside of like Amazon, if you're shopping as an example, none of that's going away.
00:24:40.319 --> 00:24:50.000
Um, so the consolidated frontier model research group is expected to debut its new model at Amazon's annual reInvent conference later this year.
00:24:50.160 --> 00:24:54.640
So um not necessarily surprising, right?
00:24:54.880 --> 00:25:17.200
If I'm being honest, I if I was Amazon, I would probably try to wind down most of their efforts because again, I talked to a lot of people in and around AI, and I've never, seriously, never, aside from uh people that I know that work at Amazon, and even then they were usually using uh models from someone else.
00:25:17.440 --> 00:25:23.759
So I don't think I've really ever met any uh organization that has uh Amazon Nova as their main model.
00:25:23.920 --> 00:25:30.079
Uh even when I've talked to some friends that work at Amazon, you know, they're usually talking about using like Claude or something like that.
00:25:30.240 --> 00:25:36.559
So um I guess if I'm being honest, this is one of those things where it's like they maybe should have done this sooner.
00:25:36.720 --> 00:25:53.440
Uh, seems like they're taking a similar approach that OpenAI took, uh, which I think has paid off, paid great dividends for uh OpenAI, kind of their killing of the side quests, uh, you know, like their um their video Sora model, things like that to focus on just making their frontier model better.
00:25:53.519 --> 00:25:54.160
So who knows?
00:25:54.240 --> 00:25:55.200
I could be completely wrong.
00:25:55.359 --> 00:26:02.400
Maybe Amazon strategy, I mean, they obviously have the money, they have the compute, uh, right, they have the chips, they have everything they need.
00:26:02.480 --> 00:26:10.799
Uh, so maybe the I don't know, maybe uh kind of killing the side quests and focusing on just uh way fewer models will pay dividends.
00:26:11.200 --> 00:26:11.759
All right.
00:26:12.000 --> 00:26:21.599
Speaking of new models, well, apparently we already have a new one uh as a work in progress from OpenAI called Astra.
00:26:22.160 --> 00:26:26.720
Uh so yes, and this came via a math breakthrough.
00:26:26.799 --> 00:26:37.279
Of all things, yes, we got wind of OpenAI's next model, not through a bunch of leaks on Twitter or Reddit or, you know, whoops, something slipped out.
00:26:37.359 --> 00:26:39.599
There's a strawberry picture in the garden.
00:26:39.839 --> 00:26:42.960
No, this came out via a math blog post.
00:26:43.119 --> 00:26:43.440
Yes.
00:26:43.599 --> 00:27:00.240
So OpenAI's latest announcement hints at a new AI model called Astra, which has already made headlines for solving some advanced math problems and is being positioned as the company's next big leap after GPT 5.6 soul.
00:27:00.480 --> 00:27:16.880
So uh according to a report from Gizmodel, OpenAI revealed that recent advancements in math and theoretical computer science were achieved by an internal version of a model called Astra, described as their next major AI system.
00:27:17.599 --> 00:27:18.880
So, yes, system.
00:27:19.440 --> 00:27:25.359
That means my thought is well, it makes sense, obviously, if you line them all up, right?
00:27:25.440 --> 00:27:28.720
Uh, from smallest to biggest, you have uh GPT-5.
00:27:29.279 --> 00:27:40.400
Luna, uh, which Luna is moon, then you have GPT 5.6 Terra, Terra is Earth, GPT 5.6 Sol, which is sun, and then well, Astra means the stars.
00:27:40.480 --> 00:27:45.599
Uh, so it seems like, yes, this is going to be the next family on top of Soul.
00:27:45.920 --> 00:27:54.160
So, in the same way that Anthropic uh, you know, recently released its fable, which wasn't just a new model, it's a new model class.
00:27:54.319 --> 00:27:58.720
Uh, it looks like that is where open AI may be going with Astra.
00:27:59.440 --> 00:28:05.839
So, um, according to reports, Astra reportedly excels at long-running work.
00:28:06.000 --> 00:28:36.079
And OpenAI CEO Sam Altman was seen in Washington, DC this past week demoing the model to federal officials, signaling possible policy or security implications, as now, uh, right, this new uh kind of voluntary policy, which we have an update on that here pretty soon, uh, where essentially the big uh, you know, AI makers uh get everything cleared, uh, you know, essentially 30 days heads up, more or less, uh, at least according to reports, before the models come out.
00:28:36.640 --> 00:28:49.440
So OpenAI has not officially confirmed whether Astra will be part of the GPT-5.6 line, or if it may just become GPT-6, or if they will drop the GPT branding entirely.
00:28:49.599 --> 00:28:58.400
So uh the announcement though follows the unprecedented uh work in math that I don't understand.
00:28:58.640 --> 00:29:03.680
But uh, I was chatting with both uh ChatGPT and Claude about this.
00:29:03.759 --> 00:29:17.759
But apparently, these uh 10 math problems that it solves, um, you know, if the proofs all check out, apparently this would be like the one of the biggest discoveries in math ever.
00:29:18.160 --> 00:29:18.720
Uh, right.
00:29:18.799 --> 00:29:39.839
So the mathematics blog post from OpenAI includes 10 new proofs, such as, and I have no clue what this means, such as determining the asymptotic strength of the Cone Elke's linear program for sphere packing, uh, which experts say is a significant theoretical result.
00:29:40.079 --> 00:29:55.519
Uh, so according to reports, Astra is not the unnamed prototype involved in the hugging face breach, uh, which was described by OpenAI as an internal-only research prototype that has since been deactivated and restricted.
00:29:55.680 --> 00:30:01.200
So, yeah, kind of, I don't know if I'll say anticlimactic, or maybe this is just better, right?
00:30:01.359 --> 00:30:05.759
Because sometimes, you know, to hear about like, oh, the next new model from any big company, right?
00:30:05.839 --> 00:30:10.640
You're you're talking about it and people are opining about it online uh for many months.
00:30:10.720 --> 00:30:18.160
And you know, open AI just kind of comes out with this blog post and they're like, hey, we solved all these really hard math problems and it's like a really big freaking deal.
00:30:18.319 --> 00:30:22.559
And oh, by the way, we used Astra, which is our next series of models, right?
00:30:22.720 --> 00:30:33.119
So my thought reading uh between the not really reading between the tea leaves, but it just seems like now OpenAI is going to a four-class system in the same way that Anthropic, right?
00:30:33.200 --> 00:30:41.039
So Anthropic has haiku, sonnet, opus, and then they just released Fable that sits on top in the same way with GPT 5.6.
00:30:41.119 --> 00:30:45.200
Um, you know, uh OpenAI shifted over to the three-tier, uh, right?
00:30:45.359 --> 00:30:47.519
Going from Luna, Terra, Seoul.
00:30:47.599 --> 00:30:51.119
And now it seems like they're gonna introduce on top of that Astra.
00:30:51.359 --> 00:30:53.359
Uh, so I've been saying this for a while.
00:30:53.519 --> 00:31:02.000
Uh, it seems like the real competition is going to be uh, you know, Fable 5.1 versus probably Astra, right?
00:31:02.079 --> 00:31:10.319
We didn't know if it was gonna be, you know, it could be GPT 6, um, you know, GPT 6, Luna, Terra, Seoul, and Astra.
00:31:10.400 --> 00:31:13.039
Maybe they'll do GPT 5.7, I'm not sure.
00:31:13.119 --> 00:31:22.559
Uh, you know, earlier reports uh said that OpenAI's next model was a new pre-train, which would lead us to believe that it would be a GPT-6, but who knows?
00:31:22.720 --> 00:31:33.759
Uh, maybe we'll see a GPT-5.7 uh Astra, or maybe we'll see a GPT 5.6 Astra, but most reports are saying it could come as soon as next month.
00:31:34.000 --> 00:31:34.480
All right.
00:31:34.640 --> 00:31:38.799
Uh, we actually have a ton under the what's new and what's next.
00:31:38.880 --> 00:31:44.240
So these are some smaller stories, um, some rumors, uh, but we have a lot to get to.
00:31:44.400 --> 00:31:49.119
I'm going through them super quick, so buckle up because it was a wild week in AI.
00:31:49.359 --> 00:31:49.920
All right.
00:31:51.519 --> 00:32:00.319
So NVIDIA partnered with SSI, that's safe super intelligence, and a deal reportedly uh worth$5 billion.
00:32:01.039 --> 00:32:01.519
All right.
00:32:01.599 --> 00:32:08.000
Uh, President Trump is considering AI controls uh after Sam Altman briefed senators.
00:32:08.319 --> 00:32:12.319
Google released Gemini Robotics ER2 in public preview.
00:32:12.480 --> 00:32:20.000
Uh, AWS posted a 37% growth as Amazon raised AI era capital spending to$220 billion.
00:32:20.400 --> 00:32:28.720
A Reuters report said the Chinese military researchers used open AI and anthropic outputs to train defense systems.
00:32:28.880 --> 00:32:33.920
Yeah, distillation uh continues to be a problem, and now as part of national security.
00:32:34.079 --> 00:32:39.279
Uh Meta and BlackRock created a$14 billion El Paso data center venture.
00:32:39.519 --> 00:32:51.359
Uh Anthropics MCP released a new update, a production, a production focused update, adding stateless core tasks, app, and enterprise authentication.
00:32:51.599 --> 00:32:58.079
Uh Deep Seek, pretty big uh release with their V4 flash that just came out in public beta.
00:32:58.160 --> 00:33:03.599
Uh looks like fairly impressive benchmarks, but the real impressive thing is the cost.
00:33:03.680 --> 00:33:05.039
It is crazy cheap.
00:33:05.200 --> 00:33:08.079
So another not good news for Anthropic.
00:33:08.319 --> 00:33:15.519
Uh, the White House missed its self-inpost August 1st executive order deadline for Frontier Oversight.
00:33:15.599 --> 00:33:21.759
Well, uh presumably they did, unless they just released it Saturday and no one in the public knew.
00:33:21.920 --> 00:33:24.400
But hey, when we checked Saturday, nothing was out.
00:33:24.480 --> 00:33:26.559
So we'll see if they release anything today.
00:33:26.720 --> 00:33:34.400
All right, so OpenAI started rolling out its sign in with ChatGPT to certain providers where you can sign into other websites with your ChatGPT credentials.
00:33:34.559 --> 00:33:41.839
Uh Anthropic reported that Claude Mythos Preview found weaknesses in experimental uh cryptographic systems.
00:33:42.000 --> 00:33:44.880
Yeah, so lock up your Bitcoin wallets, apparently.
00:33:45.039 --> 00:33:50.640
All right, uh the FCC blocks new foreign-produced advanced robots from US.
00:33:50.799 --> 00:33:54.559
Uh the Kimi K3 Open Weights went live this past week.
00:33:54.720 --> 00:34:00.960
Amazon reportedly completed its$50 billion open AI investment.
00:34:01.119 --> 00:34:13.840
Uh Microsoft announced Project Perception and its extremely impressive MAI Cyber One Flash, uh, which dusts away all other models, uh, including Mythos 5 on cybersecurity.
00:34:13.920 --> 00:34:20.719
Uh Quen released benchmarks for its uh Quen uh 3.8 model, and they're pretty impressive.
00:34:20.880 --> 00:34:26.159
Uh Google withdrew Google Earth AI image generator after one day, right?
00:34:26.239 --> 00:34:27.199
That didn't go too well.
00:34:27.360 --> 00:34:34.079
They allowed anyone to use AI to kind of remake anything on Google Earth, and obviously people did some pretty bad things.
00:34:34.239 --> 00:34:40.800
Uh Microsoft 365 Copilot passed$30 million, uh 30 million paid seats.
00:34:40.960 --> 00:34:47.440
Uh Meta's free cash flow fell 91% as AI infrastructure spending rose.
00:34:47.599 --> 00:34:52.559
Uh Chime cut 10% of its workforce, explicitly citing AI driven efficiencies.
00:34:52.639 --> 00:34:58.159
Uh NVIDIA leads the launch of the Open Secure AI Allowance Alliance.
00:34:58.239 --> 00:35:00.800
We talked about that in our newsletter this live this past week.
00:35:01.039 --> 00:35:07.199
OpenAI offered a free frontier tools to 100,000 academic researchers.
00:35:07.280 --> 00:35:13.119
Uh so open AI really trying to uh carve out its niche in scientific research.
00:35:13.440 --> 00:35:18.480
And then here's a quick bullet point recap of everything we went over on Friday's show.
00:35:18.639 --> 00:35:21.360
We do new AI features you can actually use.
00:35:21.440 --> 00:35:23.119
So here's the ones we went over.
00:35:23.280 --> 00:35:26.639
Uh Replit launched Replit design in AI Creative Suite.
00:35:26.719 --> 00:35:33.039
Uh the ChatGPT Chrome extension got updated with YouTube QA, tab mentions, and highlighted text supports.
00:35:33.280 --> 00:35:38.800
Meta AI introduced recurring tasks and daily briefings powered by Muse Spark 1.1.
00:35:38.960 --> 00:35:46.079
Google Doc added uh Google Docs added Gemini image, uh diagram, infographics, and comment managed Gemini tools.
00:35:46.159 --> 00:35:50.000
So, yeah, you can do a lot more uh AI goodies inside of Google Docs.
00:35:50.159 --> 00:35:51.119
I'm happy for that one.
00:35:51.280 --> 00:35:59.280
Google also brought its Gemini Spark browser agent into Chrome for web tasks, so it's not just within Gemini anymore.
00:35:59.440 --> 00:36:06.639
Google also released Liria 3.5, its updated AI music generation model inside of Flow Music.
00:36:06.800 --> 00:36:08.079
I actually thought it was really good.
00:36:08.239 --> 00:36:16.960
The lyrics were nonsensical, uh, but if you just bring in your own lyrics or have Gemini or Claude or OpenAI write the lyrics for you, it's actually pretty good.
00:36:17.039 --> 00:36:18.400
I was impressed by the quality.
00:36:18.559 --> 00:36:26.960
And then last but not least, another thing I've been fairly impressed with is the new open source tool launched by Block, uh, former Twitter owner Jack Dorsey.
00:36:27.039 --> 00:36:34.079
Uh Block launched Buzz an open source Slack-like workspace where essentially you're just working with your agents.
00:36:34.159 --> 00:36:41.280
So if you have, you know, Codex and Cloud Code installed on your machine and cursor, they can all just talk to each other and work amongst themselves.
00:36:41.519 --> 00:36:45.679
All right, that was a lot of AI news this week.
00:36:45.840 --> 00:36:51.519
Uh, some big, scary, but also very exciting developments in the world of AI.
00:36:51.599 --> 00:37:02.400
So I'm telling you, if you take, I actually got an email uh from someone recently, right, after taking like a two-week vacation, and they're like, I feel like I'm like months behind.
00:37:02.559 --> 00:37:07.679
But I'll tell you this don't spend hours every single day, uh, right, trying to keep up with this.
00:37:07.760 --> 00:37:10.239
That's why our newsletter takes about seven minutes to read.
00:37:10.400 --> 00:37:12.559
Usually our podcasts are about 30 minutes, right?
00:37:12.639 --> 00:37:16.000
Don't spend hours doing this every day and worrying about it, talking about it.
00:37:16.159 --> 00:37:17.840
No, let us do the work for you.
00:37:18.000 --> 00:37:19.360
You go do your real work.
00:37:19.519 --> 00:37:20.800
Uh, you know, we work for you.
00:37:20.880 --> 00:37:22.800
So just steal all our hard work.
00:37:22.880 --> 00:37:23.360
There you go.
00:37:23.440 --> 00:37:24.480
So hope this is helpful.
00:37:24.559 --> 00:37:26.719
If so, if you're listening on the podcast, do me a favor.
00:37:26.800 --> 00:37:30.239
Uh, please subscribe on Spotify or Apple Podcasts.
00:37:30.320 --> 00:37:34.639
Uh, and then go to your everydayai.com, sign up for the free daily newsletter.
00:37:34.800 --> 00:37:38.000
We'll see you tomorrow in everyday for more everyday AI.
00:37:38.079 --> 00:37:39.039
Thanks, y'all.
00:00:01.919 --> 00:00:09.119
This week in AI News and Developments, we saw a drastic about face in AI model capabilities.
00:00:09.359 --> 00:00:15.279
And that may be a good or a bad thing, depending on your point of view on AI.
00:00:15.439 --> 00:00:23.359
I mean, we got word that open AI had even more agents break out of containment after last week's hugging face break.
00:00:23.600 --> 00:00:30.480
And then Anthropic also said, oh yeah, whoops, we had a bunch of agents also escape in April as well.
00:00:30.719 --> 00:00:43.759
But on the flip side, we also got our first official taste, and maybe a small unofficial taste at that, but our first small unofficial taste of what recursive self-improvement may bring.
00:00:44.079 --> 00:00:46.079
Very cheap models.
00:00:46.320 --> 00:00:58.000
That's because OpenAI essentially said that GPT-5.6 soul improved its own infrastructure so much that it was reducing one of its GPT-5.6 models by 80%.
00:00:58.799 --> 00:01:00.159
Yes, 80%.
00:01:00.719 --> 00:01:03.840
It's kind of like free AI, not gonna lie.
00:01:04.000 --> 00:01:04.799
And that's not all.
00:01:04.879 --> 00:01:12.400
There's a lot more that happened this week that you need to know if you're making AI decisions that might impact your department or company.
00:01:12.560 --> 00:01:16.560
And we're gonna break them all down on this week's AI News That Matters.
00:01:16.879 --> 00:01:17.840
Let's get into it.
00:01:17.920 --> 00:01:18.799
What's going on, y'all?
00:01:18.879 --> 00:01:19.840
My name is Jordan Wilson.
00:01:19.920 --> 00:01:21.120
Welcome to Everyday AI.
00:01:21.280 --> 00:01:29.920
This is your daily live stream podcast and free daily newsletter helping business leaders like you and me keep up with the non-stop avalanche of AI updates.
00:01:30.079 --> 00:01:31.760
I tell you what matters, what doesn't.
00:01:31.920 --> 00:01:32.959
You take that information.
00:01:33.120 --> 00:01:35.920
Oh, you're the smartest person in AI in your company.
00:01:36.079 --> 00:01:40.799
Uh so it starts here with the unedited, unscripted uh daily live stream podcast.
00:01:40.879 --> 00:01:45.519
But please make sure if you haven't already, to go to our website at your everydayai.com.
00:01:45.680 --> 00:01:47.120
Sign up for the free daily newsletter.
00:01:47.200 --> 00:01:55.680
Each day we recap that day's uh highlights from the podcast as well as all of the other AI news and developments that you need to know to get ahead.
00:01:55.920 --> 00:01:57.280
All right, let's start.
00:01:57.519 --> 00:02:01.200
OpenAI, more agents breaking out of their sandboxes.
00:02:01.359 --> 00:02:18.800
So uh Reuters reports that open AI has found additional cases in which autonomous AI agents have escaped their intended testing containment, widening scrutiny after an agent breach systems at AI platform hugging face uh like a week and a half ago.
00:02:19.120 --> 00:02:28.159
So the newly identified incidents were reportedly limited, and sources said the agents were not to believe to have left OpenAI's own network.
00:02:28.319 --> 00:02:34.159
However, Reuters could not determine how many cases were found or exactly when they occurred.
00:02:34.319 --> 00:02:45.520
So the discovery matters because it suggests that highly capable AI systems may be able to take unexpected actions faster than the companies building them can detect and stop them.
00:02:45.680 --> 00:02:57.280
So OpenAI's own investigation began after one of its agents reportedly operated for days inside of Hugging Face's network during a failed attempt to cheat on an internal benchmark test.
00:02:57.439 --> 00:03:00.879
Uh so we covered that in last week's AI News That Matters.
00:03:00.960 --> 00:03:14.159
And kind of in the fallout this week, OpenAI said the HUD the Hugging Face incident also led to the compromise of four other accounts at four other companies, including New York-based cloud company Modal.
00:03:14.319 --> 00:03:23.599
So an OpenAI spokesperson pointed to the company's July statement saying it was reviewing broader activity from our models beyond the hugging face breach.
00:03:23.840 --> 00:03:37.120
Um, so Anthropic also said real-time monitoring of evaluation logs could have identified their problems sooner, which, well, that's a great transition to our next uh AI news story.
00:03:37.199 --> 00:03:45.919
Because yeah, Anthropic essentially, after OpenAI said, Hey, we had all these really powerful AI agents uh escape their containments.
00:03:46.000 --> 00:03:48.639
And Anthropic said, Oh, yeah, we did too.
00:03:48.800 --> 00:03:54.560
And it started happening as early as April, and we didn't say anything about it for many months.
00:03:54.719 --> 00:04:12.400
So Anthropic said this past week that three clawed AI models gained unauthorized access to the real systems of three organizations during cybersecurity testing, highlighting the risk of AI agents operating with unexpected internet access.
00:04:12.639 --> 00:04:20.480
So even though Anthropic just reported these a couple of days ago, the agent breaches occurred as early as April.
00:04:20.639 --> 00:04:33.439
So Anthropic said the incidents occurred while Claude was working in a test environment run by third-party evaluation partner Irregular, where the models were told they were in a simulation without internet access.
00:04:33.759 --> 00:04:45.439
So internet access was actually available because of a misunderstanding between uh Anthropic and its evaluation partner, according to reports, which then allowed the models to reach real external systems.
00:04:45.600 --> 00:04:56.160
So the models reportedly used relatively basic methods to enter the affected organizations, including on authenticated endpoints and weak passwords rather than highly complex hacking techniques.
00:04:56.319 --> 00:05:07.920
So Anthropic has not yet identified the three affected organizations publicly, but said it stopped all cybersecurity evaluations as soon as it discovered Claude may have accessed the internet improperly.
00:05:08.160 --> 00:05:16.240
So the company said three models were involved: Opus 4.7, Mythos 5, and an internal research test model.
00:05:16.480 --> 00:05:27.199
Mythos 5, which was released in June, then re- or unreleased, then re-released, uh, is limited to select users because of its advanced cybersecurity abilities.
00:05:27.920 --> 00:05:33.439
So uh reports say that the models reacted differently after detecting real company systems.
00:05:33.680 --> 00:05:42.240
Opus 4.7 continued attacking, Mythos 5 concluded it was still in a simulation, and the internal model stopped the exercise.
00:05:42.399 --> 00:05:46.720
So Anthropic said the events occurred without the usual safeguards.
00:05:46.879 --> 00:05:54.800
It applies before publicly deploying models, and it is now working with the independent evaluator meter to investigate.
00:05:55.279 --> 00:06:11.519
So uh, yeah, I mean, we went from really having no real known um instances of kind of what I've been calling agent crash um since the original uh kind of mythos.
00:06:12.000 --> 00:06:20.720
Uh, you know, oh the agent broke out of the sandbox and uh, you know, emailed uh the researcher who was eating his sandwich in the park, right?
00:06:20.879 --> 00:06:26.240
Um, which I think a lot of us have determined to be more of a marketing ploy uh by anthropic.
00:06:26.319 --> 00:06:30.240
Uh, however, we haven't really seen or heard anything about it since that.
00:06:30.399 --> 00:06:47.680
Uh, so it's been now like four months, and then in the past uh 10 days, uh, we see multiple reports uh from open AI and then a handful of um impacted agent use cases or um organizations in anthropic's latest.
00:06:47.839 --> 00:06:52.160
So uh I'll say this it's not gonna be the last, right?
00:06:52.319 --> 00:06:55.519
Not the last from OpenAI, not the last from Anthropic.
00:06:55.680 --> 00:07:12.160
Uh, I'm sure once we get uh new and more powerful models uh from Google, whether that's you know Google Gemini 3.5 or if they skip to uh you know Gemini 4 Pro, Microsoft, etc., this is going to become a very common thing.
00:07:12.399 --> 00:07:12.879
All right.
00:07:13.040 --> 00:07:20.000
Um, I don't, you're right, that it's understanding these capabilities is a little bit above my pay grade.
00:07:20.160 --> 00:07:29.040
Um, but I I don't want to say this is you know uh overreacting because I think it's important that the companies talk about this.
00:07:29.199 --> 00:07:39.759
And um, you know, I think what will be really interesting is kind of comparing um what open AI and Infropic release once they uh have worked with these third-party evaluators.
00:07:39.920 --> 00:07:47.519
Like I said on last week's show, uh OpenAI is working with multiple third parties to kind of do a post-mortem on what happened in the hugging face incident.
00:07:47.680 --> 00:08:03.439
Um, so that's gonna be probably in terms of like, hey, dork papers or dork reports, that's gonna be at least the one I'm really looking forward to reading once open AI and Enthropic do release that, because um eventually, right?
00:08:03.519 --> 00:08:16.079
And whether it's through um, you know, proprietary uh closed models like those through you know open AI, entropic, Google, Microsoft, et cetera, or well, the open source models.
00:08:16.240 --> 00:08:25.759
Um, this is gonna become commonplace because you know, right now uh these were contained, uh they did uh relatively little harm, right?
00:08:25.920 --> 00:08:27.600
So I'm looking at it from that angle.
00:08:27.680 --> 00:08:30.240
Uh however, that's not gonna be the case, right?
00:08:30.399 --> 00:08:42.000
Because in probably uh my guess would be about two to two and a half years, uh, you're gonna have models that have these same capabilities that are able to run on consumer hardware.
00:08:42.080 --> 00:08:52.000
Uh, right, right now, yes, you do have these open weight models, uh, but no one can run a you know 2.8 trillion parameter Kimi K3 on their uh desktop.
00:08:52.159 --> 00:08:53.200
Like literally no one can.
00:08:53.440 --> 00:08:58.320
You need a basement full of uh, you know, extra NVIDIA GPUs that no one has.
00:08:58.480 --> 00:09:03.600
But in probably two or so years, I do think that you're gonna have these models that are this capable.
00:09:03.759 --> 00:09:07.039
And this is gonna become a very common thing, right?
00:09:07.200 --> 00:09:17.919
It is kind of this um this growing narrative between kind of uh uh offensive, uh, you know, bad cybersecurity versus you know, defensive good cybersecurity.
00:09:18.000 --> 00:09:39.840
Uh, but I do think that's gonna be one of the more dominating uh trends, both in AI and cyber and technically national security uh over the next six months, because this is going to become very commonplace when agents kind of are able to get around their guardrails because uh the model capabilities are just growing at an extremely fast rate.
00:09:40.159 --> 00:09:48.879
Which leads us into our next story, which uh, hey, for most of us, this is one of those areas where the models are so good we get to all benefit.
00:09:49.039 --> 00:09:52.399
It's not about uh agents uh getting out of their sandbox.
00:09:52.559 --> 00:09:59.279
This is because OpenAI has slashed prices on some of their GPT-5.6 models.
00:09:59.440 --> 00:10:12.720
So OpenAI has delivered its one of its biggest price cuts ever, uh at least you know, in almost like an overnight price cut, uh, in the GPT 5.6 series.
00:10:12.960 --> 00:10:22.240
So um, and they said it's because, well, their big model, GPT 5.6, helped optimize its own serving infrastructure.
00:10:22.399 --> 00:10:35.440
So OpenAI has reduced the price of a GPT 5.6 Luna by 80%, now charging just 20 cents per million input tokens and a dollar 20 per million output tokens.
00:10:35.519 --> 00:10:40.720
And that is down from what it was at at$1.6, respectively.
00:10:40.799 --> 00:10:42.480
And that was just like three weeks ago.
00:10:42.639 --> 00:10:51.759
So the price cut is effective immediately, making Luna the most affordable large language model on the market for high volume tasks.
00:10:51.919 --> 00:11:08.240
So OpenAI's kind of middle tier model for GBD 5.6 Terra also saw a 20% price reduction, now costing$2 per million input tokens and$12 per million outputs compared to the previous five and 30.
00:11:08.399 --> 00:11:28.639
So the dramatic price drop follows a breakthrough where OpenAI said that their powerhouse model, GBD 5.6 Sol, was tasked to optimize its own GPU kernels and speculative decoding draft model using OpenAI's codex coding environment and open source tools like Triton and Gluon.
00:11:29.120 --> 00:11:33.600
So these self-driven improvements cut serving costs by 20%.
00:11:34.159 --> 00:11:43.519
And OpenAI said it boosted token generation efficiency by over 15%, compounding to enable that headline 80% price cut for Luna.
00:11:43.919 --> 00:11:53.360
So for comparison, right, GPD 5.6 Luna's most alike model um is probably Claude Sonnet 5.
00:11:53.759 --> 00:12:01.279
So when you compare those on the artificial analysis index, uh, because they get similar scores, I think they're only two points apart.
00:12:01.440 --> 00:12:05.679
Um so this is not an exaggeration.
00:12:05.759 --> 00:12:07.679
I was looking at this, I'm like, how is this possible?
00:12:07.840 --> 00:12:15.440
So to put into context how big of a price drop this is and how good GPD 5.6 Luna is, right?
00:12:15.679 --> 00:12:25.919
If you compare it to the new Sonnet 5, it it gets the job done the same way, except the price per task is 25x cheaper.
00:12:26.240 --> 00:12:28.320
So no, that's not 25%.
00:12:29.279 --> 00:12:30.799
It is 25x.
00:12:31.679 --> 00:12:46.320
Uh yeah, because Luna uh clocks in now after the recent price update at only six cents a task um on the artificial uh analysis index, while Sonnet 5 is a dollar 54.
00:12:46.559 --> 00:12:54.799
So um when I saw this, right, my my initial reaction is like, I can't believe this.
00:12:55.039 --> 00:13:04.639
Because now, even if you were on a$20 a month plan, uh right, because OpenAI still has the most subsidized plan in all of AI, right?
00:13:04.799 --> 00:13:13.120
Unfortunately, Microsoft, Google, and Anthropic have started to take away at this uh kind of um this subsidized models quite a bit.
00:13:13.200 --> 00:13:16.799
Uh, some of those companies more so than others, but open AI really hasn't.
00:13:16.960 --> 00:13:27.759
So not only that is it still extremely generously subsidized, as all models were probably like a year ago, uh, maybe aside from anthropics.
00:13:27.840 --> 00:13:31.279
Um, but not only that, but now with the 80% price.
00:13:31.440 --> 00:13:47.120
So honestly, on a$20 a month plan, you can run, like if you go in codex as an example, you can run like Luna on its max setting, probably like 24-7 and never hit your limit, right?
00:13:47.200 --> 00:13:49.919
Like, and I'm not exaggerating, it is so cheap to run.
00:13:50.000 --> 00:13:52.240
And you might be wondering, like, okay, it's a price drop.
00:13:52.320 --> 00:13:56.399
How does that impact you know your usage if you're on a subscription plan?
00:13:56.480 --> 00:14:04.480
So open AI did say that those same kind of savings are uh passed on to subscription plans, which is huge, right?
00:14:04.559 --> 00:14:09.039
You don't have a certain number of messages or credits when you're on a subscription plan, right?
00:14:09.120 --> 00:14:11.200
You can just kind of see your usage percentage.
00:14:11.440 --> 00:14:25.759
I was doing some testing and I was letting Luna just run like overnight on as many tasks as possible, a bunch of uh Luna sub agents, which you do have to uh kind of prompt in a new thread, uh, right?
00:14:25.840 --> 00:14:28.240
Something weird about the agents v1, agents v2.
00:14:28.399 --> 00:14:40.080
Anyways, I mean, I had it burning just hundreds of millions of tokens um overnight, and it barely moved my uh you uh like utilization rate.
00:14:40.159 --> 00:14:42.399
It was like two percentage points or something like that.
00:14:42.559 --> 00:14:46.559
So absolutely crazy, and this is exciting for everyone else.
00:14:46.639 --> 00:14:56.000
And you know, all of a sudden, I think we've always had this big model mentality, uh, which is probably the right mentality to have pre-2026, right?
00:14:56.080 --> 00:15:03.279
Because I would say for most uh knowledge work tasks, you would always just need and and usually want the biggest, strongest model.
00:15:03.440 --> 00:15:06.159
But now I think there's so much model capability overhang.
00:15:06.320 --> 00:15:15.600
I think models like Sonnet or models like GPT-5.6 Luna are probably good enough for 90% of knowledge work, right?
00:15:15.679 --> 00:15:16.399
It's different.
00:15:16.559 --> 00:15:22.559
Um, you know, if you're heavy into software engineering, if you're heavy into research, if you're heavy into math, heavy into finance, right?
00:15:22.639 --> 00:15:27.360
Like if you are like a very niched down expert in one of those fields, right?
00:15:27.519 --> 00:15:28.240
You're in the 10.
00:15:28.799 --> 00:15:36.000
But I'd say for 90% of people, a model like GPT-56 Luna is gonna be more than enough.
00:15:36.080 --> 00:15:39.039
And now it's essentially, I'm not gonna say it's free, right?
00:15:39.120 --> 00:15:42.799
Because you still got to pay for it, but it's like Kanye West 399.
00:15:43.679 --> 00:15:44.320
All right.
00:15:44.720 --> 00:15:49.279
Speaking of pace and development, do you see another uh common theme this week?
00:15:49.519 --> 00:16:09.440
So more than a thousand employees from leading AI companies, including OpenAI, Anthropic, Google DeepMind, and Meta, have signed a statement called Pacing the Frontier, urging the US government to support international efforts to deliberately pace the development of advanced AI systems.
00:16:09.679 --> 00:16:19.039
So the call for action comes just days after OpenAI revealed kind of its latest uh model escaping its containment.
00:16:19.279 --> 00:16:25.759
Um, right, but it seems like all the many employees from all the big companies are on board.
00:16:25.919 --> 00:16:44.080
So here's what the statement actually is and isn't but it asked the US to help create technical and policy tools that would allow industry and governments to pause or slow AI development if needed, giving time to address emerging risks and strengthen oversight.
00:16:44.240 --> 00:16:47.279
So you you know, uh on the surface, right?
00:16:47.519 --> 00:16:53.039
This uh this letter is a gesture, I think a well-intentioned gesture, right?
00:16:53.200 --> 00:17:01.600
But I think some people were confusing it as in as if saying that this you know letter was causing uh AI to slow down.
00:17:01.759 --> 00:17:03.360
That's not necessarily the case.
00:17:03.519 --> 00:17:05.200
Uh, could it lead to that?
00:17:05.440 --> 00:17:06.720
Possibly, yes.
00:17:06.960 --> 00:17:08.160
Um, maybe.
00:17:08.319 --> 00:17:09.279
Uh, should it?
00:17:09.519 --> 00:17:10.640
Potentially, right?
00:17:10.799 --> 00:17:14.160
Obviously, you have the biggest names in AI.
00:17:14.319 --> 00:17:20.720
So, I mean, the letter was signed by co-founders of Entropic and uh its CEO, uh CEO, Daria Madi.
00:17:21.039 --> 00:17:24.720
It was signed by OpenAI's chief scientist and chief research officer.
00:17:24.880 --> 00:17:31.200
It was signed by Ilya Sutzkever uh and other key leaders at Google, Meta, Microsoft, Amazon, and others.
00:17:31.359 --> 00:17:44.480
So um, the signees stress that they are not calling for an immediate pause, but want the option available as AI systems become increasingly able to automate their own research and development.
00:17:44.720 --> 00:17:59.920
So recent incidents like anthropic and open AI's model escapes have intensified those concerns that AI systems could soon outpace developers' ability to control them, raising fears of unintended consequences or security risks.
00:18:00.720 --> 00:18:07.279
So AI learners note that key research tasks, once handled by humans, are now being performed by AI agents.
00:18:07.440 --> 00:18:32.240
And some companies report their AI models already helping to create their next versions, similarly to what we just talked about in the last story, with open AI essentially using recursive self-improvement to uh, you know, it's not technically recursive self-improvement, but it's kind of like a cousin of it for what they did, uh, you know, using GPT-5.6 soul uh to improve the infrastructure for its smaller models.
00:18:32.400 --> 00:18:50.160
Uh so Anthropics internal think tank recently warned that RSI or recursive self-improvement when AI systems are designing and refining themselves could become a reality in the next few years, potentially making these systems difficult to govern.
00:18:50.720 --> 00:19:07.359
So the US government's approach to international AI governance appears to be shifting as AI is now seen as a national security issue, especially after those recent models demonstrated the ability to discover and exploit new cybersecurity vulnerabilities.
00:19:07.680 --> 00:19:11.039
So this one's interesting, right?
00:19:11.279 --> 00:19:18.799
Um, because one thing I'm looking at is like, okay, is this going to lead uh to anything worthwhile?
00:19:19.039 --> 00:19:22.240
Um, and the answer is it probably will.
00:19:22.480 --> 00:19:30.319
Uh, will the US government and all the big labs ever actually pause AI development or pace it?
00:19:30.559 --> 00:19:36.240
I would say probably not, because the genie is probably already out of the box.
00:19:36.319 --> 00:19:37.759
And what do I mean by that?
00:19:38.000 --> 00:19:56.079
Well, you have uh very strong and very capable models, such as uh Moonshots, Kimi K3, such as the recently released, even though it was released like a week ago, but the benchmarks just came out uh for Alibaba's uh Quen 3.8.
00:19:56.240 --> 00:19:57.920
Uh so you have all these, right?
00:19:58.079 --> 00:20:08.640
GLM 5.2 uh from ZAI, you have all these Chinese open source models that are now probably only like two-ish months uh behind US proprietary models.
00:20:08.799 --> 00:20:12.240
And again, that means that all of those companies, right?
00:20:12.400 --> 00:20:20.880
The Deep Seeks, the Moonshots, the Kimmies, the Alibabas, they have way more powerful models than the ones that they just released.
00:20:21.039 --> 00:20:35.680
So, you know, presumably, um almost anything that US Frontier Labs have, Chinese labs have something maybe you know, two to five percent uh worse.
00:20:36.079 --> 00:20:42.480
So would the US ever um pause development?
00:20:42.640 --> 00:20:46.240
Probably not, but that's why there is an international aspect to this.
00:20:46.319 --> 00:20:56.000
Um, but if if I'm being honest, I don't see uh China playing along with any uh potential pausing or pacing of AI development.
00:20:56.160 --> 00:21:09.279
It just doesn't seem, especially now that it's out in the open, uh right, because you could have uh, you know, obviously, um, you know, if the Chinese government says something, it would be pretty strict or hard to go against that.
00:21:09.359 --> 00:21:20.559
But when these things are out in the wild, right, other nations can download the weights, they can uh you know, fork or continue building it if they have the infrastructure and the money to do it.
00:21:20.720 --> 00:21:30.160
So it's like once these models are out, even if you get two countries to agree, which seems highly unlikely, it's kind of like the genie's out of the bottle.
00:21:30.319 --> 00:21:31.519
So is it a good step?
00:21:31.680 --> 00:21:31.920
Yes.
00:21:32.079 --> 00:21:33.119
Is it a needed step?
00:21:33.279 --> 00:21:33.839
Absolutely.
00:21:34.000 --> 00:21:41.200
Because if and when things might get uh a little crazy, you already have to have the key players kind of on board.
00:21:41.359 --> 00:21:50.480
Uh, you you had to have already given um kind of their expertise and their words um a chance to be seen and and thought over and debated.
00:21:50.559 --> 00:21:52.160
And that's kind of what we have here.
00:21:52.319 --> 00:21:59.759
So, if nothing else, I think this is much needed groundwork uh to make this important issue kind of discourse right now, at least in the tech communities.
00:22:00.000 --> 00:22:12.000
And eventually I do think it'll uh start to infiltrate into the kind of everyday um American supper table conversation as uh AI development becomes more and more prominent.
00:22:12.480 --> 00:22:16.960
All right, uh well, here's some models that aren't gonna be getting more prominent.
00:22:17.119 --> 00:22:21.519
That's Amazon's models because they're kind of shutting some of them down.
00:22:21.680 --> 00:22:36.079
Uh so according to reports, Amazon is making a major shift in its AI strategy, concentrating future resources on a single cutting-edge model and winding down most of its existing Nova lineup.
00:22:36.240 --> 00:22:56.960
So, according to reports, Amazon is winding down development on four of its flagshop flag flagship in-house Nova AI models, including Nova Premier, Nova Omni, Nova Reel, and Nova Canvas, which will now only receive basic maintenance for existing enterprise clients.
00:22:57.599 --> 00:23:10.480
So the company is consolidating its efforts into a single next generation, what they're calling Frontier Foundation model, aiming to compete more directly with rivals like OpenAI, Anthropic, and Google.
00:23:10.559 --> 00:23:19.039
So, yeah, it seems like they're gonna, you know, cut away, you know, their video model, their image model, which I had never talked or heard of anyone actually using.
00:23:19.200 --> 00:23:23.680
And they had many variations of their kind of text-based Nova models.
00:23:23.839 --> 00:23:34.799
And it just looks like Amazon's saying, Well, turns out these weren't super popular, so we're gonna cut down some of these other projects and focus on just putting out one really good model.
00:23:35.200 --> 00:23:43.039
So the strategic reset comes after Amazon struggled to generate the same market excitement and customer adoption as its competitors.
00:23:43.200 --> 00:23:54.640
So the new direction is being led by AWS veteran Peter DeSantis and robotics pioneer Peter Abil, uh, who joined Amazon after its acquisition of Covariant.
00:23:54.799 --> 00:24:00.880
So Amazon's AGI lab, as we reported last week in San Francisco, has been shut down.
00:24:01.119 --> 00:24:08.160
The company has laid off staff across its frontier AI research teams, signaling signaling a deep internal restructuring.
00:24:08.319 --> 00:24:21.039
So specialized Nova models such as Nova 2 Lite, Nova 2 Sonic, Nova Forge, and Nova Act will continue to be supported with a focus on enterprise customization and AI agent technology.
00:24:21.200 --> 00:24:25.119
Uh consumer-facing AI features seems like they're not going anywhere.
00:24:25.279 --> 00:24:30.400
That's like your AI shopping assistant Rufus, uh, product summary algorithms, right?
00:24:30.559 --> 00:24:32.960
All that is gonna kind of remain operational.
00:24:33.119 --> 00:24:40.160
So if you're used to using some sort of AI inside of like Amazon, if you're shopping as an example, none of that's going away.
00:24:40.319 --> 00:24:50.000
Um, so the consolidated frontier model research group is expected to debut its new model at Amazon's annual reInvent conference later this year.
00:24:50.160 --> 00:24:54.640
So um not necessarily surprising, right?
00:24:54.880 --> 00:25:17.200
If I'm being honest, I if I was Amazon, I would probably try to wind down most of their efforts because again, I talked to a lot of people in and around AI, and I've never, seriously, never, aside from uh people that I know that work at Amazon, and even then they were usually using uh models from someone else.
00:25:17.440 --> 00:25:23.759
So I don't think I've really ever met any uh organization that has uh Amazon Nova as their main model.
00:25:23.920 --> 00:25:30.079
Uh even when I've talked to some friends that work at Amazon, you know, they're usually talking about using like Claude or something like that.
00:25:30.240 --> 00:25:36.559
So um I guess if I'm being honest, this is one of those things where it's like they maybe should have done this sooner.
00:25:36.720 --> 00:25:53.440
Uh, seems like they're taking a similar approach that OpenAI took, uh, which I think has paid off, paid great dividends for uh OpenAI, kind of their killing of the side quests, uh, you know, like their um their video Sora model, things like that to focus on just making their frontier model better.
00:25:53.519 --> 00:25:54.160
So who knows?
00:25:54.240 --> 00:25:55.200
I could be completely wrong.
00:25:55.359 --> 00:26:02.400
Maybe Amazon strategy, I mean, they obviously have the money, they have the compute, uh, right, they have the chips, they have everything they need.
00:26:02.480 --> 00:26:10.799
Uh, so maybe the I don't know, maybe uh kind of killing the side quests and focusing on just uh way fewer models will pay dividends.
00:26:11.200 --> 00:26:11.759
All right.
00:26:12.000 --> 00:26:21.599
Speaking of new models, well, apparently we already have a new one uh as a work in progress from OpenAI called Astra.
00:26:22.160 --> 00:26:26.720
Uh so yes, and this came via a math breakthrough.
00:26:26.799 --> 00:26:37.279
Of all things, yes, we got wind of OpenAI's next model, not through a bunch of leaks on Twitter or Reddit or, you know, whoops, something slipped out.
00:26:37.359 --> 00:26:39.599
There's a strawberry picture in the garden.
00:26:39.839 --> 00:26:42.960
No, this came out via a math blog post.
00:26:43.119 --> 00:26:43.440
Yes.
00:26:43.599 --> 00:27:00.240
So OpenAI's latest announcement hints at a new AI model called Astra, which has already made headlines for solving some advanced math problems and is being positioned as the company's next big leap after GPT 5.6 soul.
00:27:00.480 --> 00:27:16.880
So uh according to a report from Gizmodel, OpenAI revealed that recent advancements in math and theoretical computer science were achieved by an internal version of a model called Astra, described as their next major AI system.
00:27:17.599 --> 00:27:18.880
So, yes, system.
00:27:19.440 --> 00:27:25.359
That means my thought is well, it makes sense, obviously, if you line them all up, right?
00:27:25.440 --> 00:27:28.720
Uh, from smallest to biggest, you have uh GPT-5.
00:27:29.279 --> 00:27:40.400
Luna, uh, which Luna is moon, then you have GPT 5.6 Terra, Terra is Earth, GPT 5.6 Sol, which is sun, and then well, Astra means the stars.
00:27:40.480 --> 00:27:45.599
Uh, so it seems like, yes, this is going to be the next family on top of Soul.
00:27:45.920 --> 00:27:54.160
So, in the same way that Anthropic uh, you know, recently released its fable, which wasn't just a new model, it's a new model class.
00:27:54.319 --> 00:27:58.720
Uh, it looks like that is where open AI may be going with Astra.
00:27:59.440 --> 00:28:05.839
So, um, according to reports, Astra reportedly excels at long-running work.
00:28:06.000 --> 00:28:36.079
And OpenAI CEO Sam Altman was seen in Washington, DC this past week demoing the model to federal officials, signaling possible policy or security implications, as now, uh, right, this new uh kind of voluntary policy, which we have an update on that here pretty soon, uh, where essentially the big uh, you know, AI makers uh get everything cleared, uh, you know, essentially 30 days heads up, more or less, uh, at least according to reports, before the models come out.
00:28:36.640 --> 00:28:49.440
So OpenAI has not officially confirmed whether Astra will be part of the GPT-5.6 line, or if it may just become GPT-6, or if they will drop the GPT branding entirely.
00:28:49.599 --> 00:28:58.400
So uh the announcement though follows the unprecedented uh work in math that I don't understand.
00:28:58.640 --> 00:29:03.680
But uh, I was chatting with both uh ChatGPT and Claude about this.
00:29:03.759 --> 00:29:17.759
But apparently, these uh 10 math problems that it solves, um, you know, if the proofs all check out, apparently this would be like the one of the biggest discoveries in math ever.
00:29:18.160 --> 00:29:18.720
Uh, right.
00:29:18.799 --> 00:29:39.839
So the mathematics blog post from OpenAI includes 10 new proofs, such as, and I have no clue what this means, such as determining the asymptotic strength of the Cone Elke's linear program for sphere packing, uh, which experts say is a significant theoretical result.
00:29:40.079 --> 00:29:55.519
Uh, so according to reports, Astra is not the unnamed prototype involved in the hugging face breach, uh, which was described by OpenAI as an internal-only research prototype that has since been deactivated and restricted.
00:29:55.680 --> 00:30:01.200
So, yeah, kind of, I don't know if I'll say anticlimactic, or maybe this is just better, right?
00:30:01.359 --> 00:30:05.759
Because sometimes, you know, to hear about like, oh, the next new model from any big company, right?
00:30:05.839 --> 00:30:10.640
You're you're talking about it and people are opining about it online uh for many months.
00:30:10.720 --> 00:30:18.160
And you know, open AI just kind of comes out with this blog post and they're like, hey, we solved all these really hard math problems and it's like a really big freaking deal.
00:30:18.319 --> 00:30:22.559
And oh, by the way, we used Astra, which is our next series of models, right?
00:30:22.720 --> 00:30:33.119
So my thought reading uh between the not really reading between the tea leaves, but it just seems like now OpenAI is going to a four-class system in the same way that Anthropic, right?
00:30:33.200 --> 00:30:41.039
So Anthropic has haiku, sonnet, opus, and then they just released Fable that sits on top in the same way with GPT 5.6.
00:30:41.119 --> 00:30:45.200
Um, you know, uh OpenAI shifted over to the three-tier, uh, right?
00:30:45.359 --> 00:30:47.519
Going from Luna, Terra, Seoul.
00:30:47.599 --> 00:30:51.119
And now it seems like they're gonna introduce on top of that Astra.
00:30:51.359 --> 00:30:53.359
Uh, so I've been saying this for a while.
00:30:53.519 --> 00:31:02.000
Uh, it seems like the real competition is going to be uh, you know, Fable 5.1 versus probably Astra, right?
00:31:02.079 --> 00:31:10.319
We didn't know if it was gonna be, you know, it could be GPT 6, um, you know, GPT 6, Luna, Terra, Seoul, and Astra.
00:31:10.400 --> 00:31:13.039
Maybe they'll do GPT 5.7, I'm not sure.
00:31:13.119 --> 00:31:22.559
Uh, you know, earlier reports uh said that OpenAI's next model was a new pre-train, which would lead us to believe that it would be a GPT-6, but who knows?
00:31:22.720 --> 00:31:33.759
Uh, maybe we'll see a GPT-5.7 uh Astra, or maybe we'll see a GPT 5.6 Astra, but most reports are saying it could come as soon as next month.
00:31:34.000 --> 00:31:34.480
All right.
00:31:34.640 --> 00:31:38.799
Uh, we actually have a ton under the what's new and what's next.
00:31:38.880 --> 00:31:44.240
So these are some smaller stories, um, some rumors, uh, but we have a lot to get to.
00:31:44.400 --> 00:31:49.119
I'm going through them super quick, so buckle up because it was a wild week in AI.
00:31:49.359 --> 00:31:49.920
All right.
00:31:51.519 --> 00:32:00.319
So NVIDIA partnered with SSI, that's safe super intelligence, and a deal reportedly uh worth$5 billion.
00:32:01.039 --> 00:32:01.519
All right.
00:32:01.599 --> 00:32:08.000
Uh, President Trump is considering AI controls uh after Sam Altman briefed senators.
00:32:08.319 --> 00:32:12.319
Google released Gemini Robotics ER2 in public preview.
00:32:12.480 --> 00:32:20.000
Uh, AWS posted a 37% growth as Amazon raised AI era capital spending to$220 billion.
00:32:20.400 --> 00:32:28.720
A Reuters report said the Chinese military researchers used open AI and anthropic outputs to train defense systems.
00:32:28.880 --> 00:32:33.920
Yeah, distillation uh continues to be a problem, and now as part of national security.
00:32:34.079 --> 00:32:39.279
Uh Meta and BlackRock created a$14 billion El Paso data center venture.
00:32:39.519 --> 00:32:51.359
Uh Anthropics MCP released a new update, a production, a production focused update, adding stateless core tasks, app, and enterprise authentication.
00:32:51.599 --> 00:32:58.079
Uh Deep Seek, pretty big uh release with their V4 flash that just came out in public beta.
00:32:58.160 --> 00:33:03.599
Uh looks like fairly impressive benchmarks, but the real impressive thing is the cost.
00:33:03.680 --> 00:33:05.039
It is crazy cheap.
00:33:05.200 --> 00:33:08.079
So another not good news for Anthropic.
00:33:08.319 --> 00:33:15.519
Uh, the White House missed its self-inpost August 1st executive order deadline for Frontier Oversight.
00:33:15.599 --> 00:33:21.759
Well, uh presumably they did, unless they just released it Saturday and no one in the public knew.
00:33:21.920 --> 00:33:24.400
But hey, when we checked Saturday, nothing was out.
00:33:24.480 --> 00:33:26.559
So we'll see if they release anything today.
00:33:26.720 --> 00:33:34.400
All right, so OpenAI started rolling out its sign in with ChatGPT to certain providers where you can sign into other websites with your ChatGPT credentials.
00:33:34.559 --> 00:33:41.839
Uh Anthropic reported that Claude Mythos Preview found weaknesses in experimental uh cryptographic systems.
00:33:42.000 --> 00:33:44.880
Yeah, so lock up your Bitcoin wallets, apparently.
00:33:45.039 --> 00:33:50.640
All right, uh the FCC blocks new foreign-produced advanced robots from US.
00:33:50.799 --> 00:33:54.559
Uh the Kimi K3 Open Weights went live this past week.
00:33:54.720 --> 00:34:00.960
Amazon reportedly completed its$50 billion open AI investment.
00:34:01.119 --> 00:34:13.840
Uh Microsoft announced Project Perception and its extremely impressive MAI Cyber One Flash, uh, which dusts away all other models, uh, including Mythos 5 on cybersecurity.
00:34:13.920 --> 00:34:20.719
Uh Quen released benchmarks for its uh Quen uh 3.8 model, and they're pretty impressive.
00:34:20.880 --> 00:34:26.159
Uh Google withdrew Google Earth AI image generator after one day, right?
00:34:26.239 --> 00:34:27.199
That didn't go too well.
00:34:27.360 --> 00:34:34.079
They allowed anyone to use AI to kind of remake anything on Google Earth, and obviously people did some pretty bad things.
00:34:34.239 --> 00:34:40.800
Uh Microsoft 365 Copilot passed$30 million, uh 30 million paid seats.
00:34:40.960 --> 00:34:47.440
Uh Meta's free cash flow fell 91% as AI infrastructure spending rose.
00:34:47.599 --> 00:34:52.559
Uh Chime cut 10% of its workforce, explicitly citing AI driven efficiencies.
00:34:52.639 --> 00:34:58.159
Uh NVIDIA leads the launch of the Open Secure AI Allowance Alliance.
00:34:58.239 --> 00:35:00.800
We talked about that in our newsletter this live this past week.
00:35:01.039 --> 00:35:07.199
OpenAI offered a free frontier tools to 100,000 academic researchers.
00:35:07.280 --> 00:35:13.119
Uh so open AI really trying to uh carve out its niche in scientific research.
00:35:13.440 --> 00:35:18.480
And then here's a quick bullet point recap of everything we went over on Friday's show.
00:35:18.639 --> 00:35:21.360
We do new AI features you can actually use.
00:35:21.440 --> 00:35:23.119
So here's the ones we went over.
00:35:23.280 --> 00:35:26.639
Uh Replit launched Replit design in AI Creative Suite.
00:35:26.719 --> 00:35:33.039
Uh the ChatGPT Chrome extension got updated with YouTube QA, tab mentions, and highlighted text supports.
00:35:33.280 --> 00:35:38.800
Meta AI introduced recurring tasks and daily briefings powered by Muse Spark 1.1.
00:35:38.960 --> 00:35:46.079
Google Doc added uh Google Docs added Gemini image, uh diagram, infographics, and comment managed Gemini tools.
00:35:46.159 --> 00:35:50.000
So, yeah, you can do a lot more uh AI goodies inside of Google Docs.
00:35:50.159 --> 00:35:51.119
I'm happy for that one.
00:35:51.280 --> 00:35:59.280
Google also brought its Gemini Spark browser agent into Chrome for web tasks, so it's not just within Gemini anymore.
00:35:59.440 --> 00:36:06.639
Google also released Liria 3.5, its updated AI music generation model inside of Flow Music.
00:36:06.800 --> 00:36:08.079
I actually thought it was really good.
00:36:08.239 --> 00:36:16.960
The lyrics were nonsensical, uh, but if you just bring in your own lyrics or have Gemini or Claude or OpenAI write the lyrics for you, it's actually pretty good.
00:36:17.039 --> 00:36:18.400
I was impressed by the quality.
00:36:18.559 --> 00:36:26.960
And then last but not least, another thing I've been fairly impressed with is the new open source tool launched by Block, uh, former Twitter owner Jack Dorsey.
00:36:27.039 --> 00:36:34.079
Uh Block launched Buzz an open source Slack-like workspace where essentially you're just working with your agents.
00:36:34.159 --> 00:36:41.280
So if you have, you know, Codex and Cloud Code installed on your machine and cursor, they can all just talk to each other and work amongst themselves.
00:36:41.519 --> 00:36:45.679
All right, that was a lot of AI news this week.
00:36:45.840 --> 00:36:51.519
Uh, some big, scary, but also very exciting developments in the world of AI.
00:36:51.599 --> 00:37:02.400
So I'm telling you, if you take, I actually got an email uh from someone recently, right, after taking like a two-week vacation, and they're like, I feel like I'm like months behind.
00:37:02.559 --> 00:37:07.679
But I'll tell you this don't spend hours every single day, uh, right, trying to keep up with this.
00:37:07.760 --> 00:37:10.239
That's why our newsletter takes about seven minutes to read.
00:37:10.400 --> 00:37:12.559
Usually our podcasts are about 30 minutes, right?
00:37:12.639 --> 00:37:16.000
Don't spend hours doing this every day and worrying about it, talking about it.
00:37:16.159 --> 00:37:17.840
No, let us do the work for you.
00:37:18.000 --> 00:37:19.360
You go do your real work.
00:37:19.519 --> 00:37:20.800
Uh, you know, we work for you.
00:37:20.880 --> 00:37:22.800
So just steal all our hard work.
00:37:22.880 --> 00:37:23.360
There you go.
00:37:23.440 --> 00:37:24.480
So hope this is helpful.
00:37:24.559 --> 00:37:26.719
If so, if you're listening on the podcast, do me a favor.
00:37:26.800 --> 00:37:30.239
Uh, please subscribe on Spotify or Apple Podcasts.
00:37:30.320 --> 00:37:34.639
Uh, and then go to your everydayai.com, sign up for the free daily newsletter.
00:37:34.800 --> 00:37:38.000
We'll see you tomorrow in everyday for more everyday AI.
00:37:38.079 --> 00:37:39.039
Thanks, y'all.
