WEBVTT
00:00:02.000 --> 00:00:11.039
While the US government and Anthropic continued trying to make up and be friends again, the rest of the AI world didn't stop.
00:00:11.199 --> 00:00:23.359
And it was actually some smaller players that were making big splashes this week, potentially signaling more competition at the top for the big four of Anthropic, OpenAI, Google, and Microsoft.
00:00:23.519 --> 00:00:38.320
That's because open source may finally be having its ChatGPT moment, as a new Chinese model might actually be closing the gap between proprietary US models and open source or open weight options.
00:00:38.560 --> 00:00:41.520
Oh, and Midjourney made some news as well.
00:00:41.679 --> 00:00:45.679
Yeah, that mid journey that you probably haven't heard from in years.
00:00:45.920 --> 00:00:50.399
That's because they may be trying to disrupt the medical imaging space.
00:00:50.640 --> 00:00:55.359
Yeah, there's a lot going on if you don't follow every single day.
00:00:55.600 --> 00:00:58.240
And don't worry, I do that for you.
00:00:58.399 --> 00:01:00.320
So welcome to Everyday AI.
00:01:00.399 --> 00:01:07.280
And this is our Monday segment where we bring you only the most relevant news updates.
00:01:07.439 --> 00:01:11.920
I tell you what matters, what doesn't, and how you can use that information to grow your company and career.
00:01:12.000 --> 00:01:13.280
This is Everyday AI.
00:01:13.359 --> 00:01:20.159
Uh, so if you're new here, yeah, we do this every single day, daily, unedited, unscripted, live stream podcasts, and free daily newsletter.
00:01:20.239 --> 00:01:26.159
So uh if you are new, please make sure to subscribe to the show uh on Spotify or Apple Podcasts.
00:01:26.239 --> 00:01:26.799
I'd appreciate that.
00:01:26.879 --> 00:01:32.480
And then make sure you go to our website at your everydayai.com, sign up for the free daily newsletter each and every day.
00:01:32.560 --> 00:01:39.680
Uh, we recap the highlights from today's show as well as give you all the other AI news you need to know just in a quick digestible form.
00:01:39.920 --> 00:01:47.120
All right, uh, so without further ado, let's go over the biggest stories of the week that you might have missed.
00:01:47.280 --> 00:01:49.840
And I'm gonna break it down for you simply.
00:01:50.159 --> 00:02:02.000
All right, let's start at the top because yes, at least at the time of this recording, we still don't have access to Fable Five or Mythos V at whole.
00:02:02.079 --> 00:02:07.920
That's because the Trump White House and Anthropic are still trying to make things work.
00:02:08.159 --> 00:02:30.000
So, uh, actually a little bit of a controversial statement, maybe, uh, from President Trump as he told the Axios show on Friday that he considered Anthropic previously and or its CEO, Dario Amati, to be a national security threat, though he now believes the situation is improving.
00:02:30.159 --> 00:02:31.439
Yeah, he did say that.
00:02:31.520 --> 00:02:34.159
You might go want to listen to the interview yourself.
00:02:34.319 --> 00:02:46.800
We'll be linking it in today's newsletter, but he definitely said that he considered them a few weeks ago, whether he was talking about Anthropic as a whole or Dario Amati to be a national security threat, though it's now he says it's improving.
00:02:46.960 --> 00:02:47.599
So the U.S.
00:02:47.680 --> 00:03:05.520
government, if you haven't been following along, the US government recently imposed strict export controls on Anthropic, barring any country outside of the US and foreign nationals within the US from accessing uh Anthropic's most recent models called Fable V and Mythos V.
00:03:06.159 --> 00:03:24.319
So Anthropic has completely disabled in response to that, they've disabled public access uh to those models, Fable V, which is the version of Mythos V with more guardrails, uh so yeah, but it's been down and gone now for almost a week since that export control.
00:03:24.719 --> 00:03:34.560
Uh so the directive bars any foreign nationals, including uh foreign national anthropic employees, from accessing uh these frontier systems.
00:03:34.719 --> 00:03:47.680
But because anthropic can't instantly verify the exact citizens uh citizenship of every single user on its platform, it was forced to yank both those models offline globally to remain compliant.
00:03:47.840 --> 00:03:49.520
And that's been about a week.
00:03:49.680 --> 00:04:01.280
So, according to Axios in this new interview and story, the Trump administration's concerns were triggered by the Amazon report detailing the vulnerability in Anthropics technology.
00:04:01.439 --> 00:04:09.680
Uh, right, and a lot of reports said that was their CEO, Andy Jassy, and that led to the urgent discussions with the company leadership.
00:04:10.000 --> 00:04:25.120
Trump said he was initially alarmed by Anthropics' response, but was reassured after Amoti uh acted quickly and responsibly, calling him nice and smart after beating him at the G7 summit this past week.
00:04:25.279 --> 00:04:32.079
So the Pentagon has labeled Anthropic a supply chain risk, a designation typically reserved for foreign adversaries.
00:04:32.160 --> 00:04:33.759
And it was the first time a U.S.
00:04:33.839 --> 00:04:36.560
company had gotten that label uh from the U.S.
00:04:36.639 --> 00:04:37.120
government.
00:04:37.279 --> 00:04:50.160
So Trump did not rule out using emergency powers under the Defense Production Act if Anthropic does not comply with new standards, but stated he doesn't think that step will be necessary this time.
00:04:50.399 --> 00:05:00.800
Uh, Trump revealed that it was a competitor and part owner identified later as Amazon that alerted his administration about Anthropic's uh practice.
00:05:00.879 --> 00:05:03.680
So we went over this uh in detail, FYI.
00:05:03.920 --> 00:05:08.319
If you missed the story, let me just go ahead and pull the uh the episode number.
00:05:08.399 --> 00:05:10.560
That was from Thursday, episode 801.
00:05:10.720 --> 00:05:16.319
So if you want more on uh the background on those developments, you go listen to that episode 801.
00:05:16.399 --> 00:05:20.560
But at least for now, Mythos 5, still offline.
00:05:20.720 --> 00:05:25.839
Uh, but the latest here, uh, President Trump sat down with an interview with Axios.
00:05:25.920 --> 00:05:29.199
So, yeah, if you missed that, uh, definitely worth checking out.
00:05:29.600 --> 00:05:41.920
All right, our next story, and this seems to be an underlying theme this week on open source models, uh, apparently being considered at the highest levels.
00:05:42.160 --> 00:05:49.120
So, Microsoft is moving its co-pilot co-work AI assistant to a usage-based pricing model.
00:05:49.279 --> 00:05:51.439
So, uh kind of a couple pieces of news here.
00:05:51.519 --> 00:06:01.759
Number one, that Microsoft is going usage based on co-pilot cowork instead of just including it or going uh with you know a certain number that's included in a monthly subscription.
00:06:01.920 --> 00:06:02.639
That's number one.
00:06:02.800 --> 00:06:06.319
Number two is it's generally available, whereas before it was in preview.
00:06:06.480 --> 00:06:22.480
And number three, the most recent news, according to reports, is that they're actually looking at models like China's Deep Seek V4, a version that Microsoft may fine-tune uh just for co-pilot co-work to bring down costs.
00:06:23.199 --> 00:06:35.920
So, according to reports, the company is exploring a lower cost self-hosted AI model as an alternative to the more expensive anthropic models that are currently powering co-pilot co-work by default.
00:06:36.079 --> 00:06:56.240
And Axios did report that Microsoft is considering, among others, Deep uh Deep Seek V4, a Chinese developed open source model, which could dramatically reduce costs, potentially charging just about 87 cents per million tokens compared to Anthropic's$50 per million tokens.
00:06:56.399 --> 00:07:03.759
So, yeah, not the best at math, but that's like 18x cheaper, uh, or something like that.
00:07:03.920 --> 00:07:16.720
Uh, so Microsoft expects to announce its final model choice and deployment details within weeks with new uh with any new option to be hosted on Azure to keep customer data within Microsoft's cloud infrastructure.
00:07:16.879 --> 00:07:24.720
So the shift responds to the high compute cost that we've been hearing about over the last month or two uh of agentic AI tools.
00:07:24.800 --> 00:07:31.040
Uh, specifically, we've seen a lot of reports from enterprises uh, you know, getting claud bills that are pretty high, right?
00:07:31.120 --> 00:07:34.000
We saw that report uh a couple couple of weeks ago.
00:07:34.160 --> 00:07:45.519
We don't know if it was legitimate or not, but apparently an enterprise company had a uh nine-figure uh bill from Anthropic, which is a lot, uh, right?
00:07:45.680 --> 00:07:51.680
Hundreds of millions of dollars uh because they weren't keeping uh an eye on usage caps.
00:07:51.839 --> 00:08:06.000
So that's why open source models like Deep Seek V4 can cut costs, but they do require companies like Microsoft to handle additional hosting, optimization, and safety engineering, shifting some burdens away from commercial API providers.
00:08:06.160 --> 00:08:18.639
So there are obviously a ton of political and regulatory concerns as DeepSeek's Chinese origins have attracted scrutiny in Washington, raising questions about data security and compliance for U.S.
00:08:18.800 --> 00:08:19.680
companies.
00:08:20.000 --> 00:08:32.480
Uh, Microsoft has stated that any Deep Seek-based solution would be optional and subject to additional safety and compliance filters, aiming to address enterprise risk and regulatory requirements.
00:08:32.639 --> 00:08:48.799
Uh, so industry observers are watching for Microsoft's final decision, noting that total uh cost of ownership will depend not just on it per token prices, but also on model performance, reliability, and the need for robust monitoring and safety systems.
00:08:48.960 --> 00:09:01.039
So, this one it is kind of uh intriguing, I'll say that to say the least, uh, with Microsoft making this move and specifically, according to reports, looking at Deep Seek.
00:09:01.120 --> 00:09:02.639
So, why is that noteworthy?
00:09:02.799 --> 00:09:10.799
Well, uh, Microsoft obviously has good relationships and huge ownership stakes in both Anthropic and OpenAI.
00:09:10.879 --> 00:09:11.360
Uh, right.
00:09:11.519 --> 00:09:17.919
So uh traditionally, uh it's been the GBT models from OpenAI that have powered uh Copilot.
00:09:18.000 --> 00:09:25.120
They recently started offering some Claude models as well, but the Copilot Cowork was running Claude models by default.
00:09:25.279 --> 00:09:36.159
Uh, so co-work is kind of the original technology from Anthropic that Microsoft essentially had their own version that was powered under the hood by Anthropic's Copilot technology.
00:09:36.399 --> 00:09:58.240
So, what's interesting here with Microsoft looking at model providers like DeepSeq is well, OpenAI and Anthropic have launched some serious concerns in uh regulatory um steps that they've taken, uh, you know, going to the US government, accusing DeepSeq and other Chinese companies of distilling their models.
00:09:58.399 --> 00:10:00.480
So that's what's gonna make this really juicy.
00:10:00.639 --> 00:10:31.759
Uh, pretty, if I'm being honest, a pretty bold uh step here from Microsoft, even though it is just according to reports uh to do this, considering uh two of their biggest and most important partners in Open AI and Anthropic probably don't have the best view of Deep Seek for those exact reasons because they've shown uh right, Anthropic came out with a little bit more of a report how DeepSeq and other Chinese companies are essentially taking all of their hard work to make these open source models.
00:10:31.919 --> 00:10:35.519
All right, so our next AI story staying in the same vein.
00:10:35.679 --> 00:10:39.600
Yeah, we have a new king of the open hill.
00:10:39.759 --> 00:10:50.240
Uh that's Chinese startup ZAI is making a ton of headlines this week with their new open weights model GLM52.
00:10:50.720 --> 00:11:04.720
So ZAI has released that model, GLM5.2, a massive 753 billion parameter language model, now available with unrestricted MIT open source licensing.
00:11:04.879 --> 00:11:19.440
So GLM 5.2 is designed for long horizon autonomous coding and engineering tasks, featuring a highly stable 1 million token context window, allowing it to handle extremely long documents and complex workflows.
00:11:19.600 --> 00:11:23.440
So the model is immediately accessible for download on Hugging Face.
00:11:23.679 --> 00:11:33.360
But yeah, you kind of have to be an enterprise with a little bit of compute to be able to do that because it is a large model, unless you're going to run a highly quantized version of it.
00:11:33.519 --> 00:11:42.639
Uh, but you can also use it through Hugging Face, uh, through the ZAI API, and in over 20 third-party coding environments.
00:11:42.799 --> 00:11:50.879
Uh, so uh the new model, which is pretty interesting here, uh, it does look like it's very benchmaxed FYI, uh right.
00:11:50.960 --> 00:12:15.440
That's when a company maybe puts a little too much priority, uh priority on getting certain benchmarks, uh, but still it outperforms most other open source competitors, and it is even now challenging the top proprietary models on several industry benchmarks, including Suitebench Pro, which I think is thankfully on its way out of being an important model for software engineering.
00:12:15.600 --> 00:12:27.919
Yet still, uh, it is batting on some of these benchmarks, including SuiteBench Pro, Frontier Suite, uh, and others around GPT-5.5 and Anthropics Claude Opus 4.8.
00:12:28.000 --> 00:12:31.120
So with Fable 5 off the shelf, right?
00:12:31.279 --> 00:12:37.440
I mean, oh GPT-5.5 and uh Opus 4.8 are the undisputed leaders.
00:12:37.600 --> 00:12:53.840
So uh pretty significant, even if it is maybe a little bit oversaturated and a little bit bench maxed, uh, that an open source model, uh, I I won't say they've closed the gap, but they're starting to close the gap on uh frontier models.
00:12:53.919 --> 00:13:01.440
So there's obviously some use cases that you would absolutely not want to use uh GLM5.2, mainly because it's not multimodal.
00:13:01.519 --> 00:13:02.720
So it's text only.
00:13:02.879 --> 00:13:04.480
So that's a big downside, right?
00:13:04.559 --> 00:13:15.759
So for someone like me, I'm constantly uh doing multimodal inputs, you know, whether that's in Google's AI Studio, whether it's in uh, you know, Anthropic Claude, Codex, ChatGPT, etc.
00:13:16.000 --> 00:13:16.240
Right.
00:13:16.320 --> 00:13:19.919
So you can't have multimodal input or multimodal output.
00:13:20.000 --> 00:13:39.120
But with text only, uh, you know, GLM5.2 does seem to be a competitor, and it is technically state of the art, even beating uh in a couple benchmarks, I kid you not, Fable 5, Opus 4.8, and GPT-55 on some more niche uh kind of front-end coding design, some certain benchmarks out there.
00:13:39.279 --> 00:13:52.159
So uh pretty surprising uh considering that even the best open source models prior to this were generally, I would say, about four to eight months behind the frontier counterpart.
00:13:52.399 --> 00:13:58.960
So still GLM52 is far behind, mainly in terms of capabilities because it's not multimodal.
00:13:59.120 --> 00:14:06.799
But for text only, you know, at this point, it is only like two-ish, two to three months behind, uh, which is fairly impressive.
00:14:06.879 --> 00:14:18.240
And I think it does change the narrative, especially when uh, you know, kind of all of the rage in the second quarter in the summer so far has been shifting to token efficiency.
00:14:18.399 --> 00:14:45.919
I do think most companies, the biggest decision that they're gonna make in 2026, similarly, similarly to what Microsoft is doing, uh, you know, looking at switching over from some anthropic models uh to deep seek is well, most companies are gonna have to start looking at their uh their whether you're talking about token budgets, your AI spend, but you know, switching over uh from token maxing to token efficiency, which we did a story on that in our start here series.
00:14:46.000 --> 00:14:47.120
So make sure you go check that out.
00:14:47.200 --> 00:15:00.720
But essentially, uh, right, some of these agents run in loops nonstop, and as companies have kind of started to uh pluck away the uh you know, um them kind of subsidizing, right?
00:15:00.879 --> 00:15:14.159
Because you see different reports, you know, a$200 a month plan if you were paying API costs might cost actually, if you're paying for those tokens, anywhere from you know 10 to 15 to 20,000 uh, you know, for 200.
00:15:14.559 --> 00:15:22.960
So, you know, essentially these companies have been eating uh those costs just to get users in and happy and sticky and to you know get it rolled out into their organization.
00:15:23.039 --> 00:15:26.720
So a lot of companies are going to be have to be looking at token efficiency.
00:15:26.879 --> 00:15:30.240
So uh pretty big news, both with the last two stories.
00:15:30.399 --> 00:15:43.840
Microsoft looking at open alternatives to power its co-pilot cowork as they switch to usage base and with the new ZAI 5 uh GLM 5.2 model, definitely one worth keeping an eye on.
00:15:44.000 --> 00:15:57.360
All right, our next AI news story the CEOs of Anthropic and Google DeepMind uh and OpenAI met with G7 leaders and top US officials at the G7 Summit in France.
00:15:58.000 --> 00:16:07.519
So the main news out of here, at least when it comes to an artificial intelligence standpoint, was a push for a US-led coalition on artificial intelligence.
00:16:07.759 --> 00:16:23.440
So uh Dario Amati and Google DeepMind CEO Damas uh Demas Hasabas urged the creation of a US-led international coalition to set rules and standards for AI, citing urgent risks and the need for global cooperation.
00:16:23.679 --> 00:16:25.519
So, President Donald Trump, uh U.S.
00:16:25.600 --> 00:16:33.200
Secretary of State Marco Rubio, Treasury Secretary Scott Bessant, um, and Commerce Secretary Howard Lutnick represented the U.S.
00:16:33.279 --> 00:16:36.639
at the meeting, highlighting the high-level attention on AI governance.
00:16:36.879 --> 00:16:42.639
We also had Canadian Prime Minister uh Mark Carney, who expressed support for U.S.
00:16:42.720 --> 00:16:48.799
leadership in forming an AI standards coalition, signaling growing international alignment on the issue.
00:16:48.960 --> 00:16:51.679
So the call for action comes after the U.S.
00:16:51.759 --> 00:17:05.359
government imposed export controls on Anthropic's latest AI models, Fable V and Mythos V, which led the company to disable access due to the national security concerns that the government labeled.
00:17:05.519 --> 00:17:17.920
Uh, so emoti recommended that international cooperation should cover structured access to advanced AI models and trade of chips and critical components, specifically excluding China from these exchanges.
00:17:18.079 --> 00:17:29.839
So the tech leaders emphasized the urgent need to address AI risks in cybersecurity, bioterrorism, and intelligence, warning that unchecked AI could lead to major disasters if misused.
00:17:30.160 --> 00:17:44.400
OpenAI CEO Sam Altman advocated for a global forum to set standards for testing, risk analysis, and international collaboration, while OpenAI's Chris Lehan noted broad recognition that the US should take the lead.
00:17:44.640 --> 00:17:51.519
So the big development out of G7 Summit, well, number one is that AI was front and center, uh.
00:17:51.920 --> 00:17:59.920
So maybe it was just timing, circumstantial timing, uh, with the mythos uh pull down that this happened in the middle of it.
00:18:00.079 --> 00:18:23.200
So uh, you know, especially when it comes to the rise of Chinese open source models, uh, obviously, from an adversarial perspective, the G7 uh collectively wants to keep an eye on what China is doing in AI, how they're using this technology, because, like I've been saying for years, uh ultimately this will play out one way or another on the battlefield, right?
00:18:23.279 --> 00:18:47.920
Uh, this is, I think, why uh the G7 and world leaders uh are starting to pay more and more attention to AI, not just because of how it can be used um in cyber warfare, but also on the actual battlefield, which is uh actually goes back to the initial uh kind of feud between Anthropic and the US government for how anthropic did not want the military to use its models.
00:18:48.160 --> 00:19:00.799
So uh the other big takeaway here is obviously seemingly some international support uh for the US to take a lead in this G7 AI coalition, uh, which makes sense, right?
00:19:00.960 --> 00:19:07.920
A lot of times other world leaders, you know, might not want to, you know, step aside and let the US take a lead on something like this.
00:19:08.079 --> 00:19:17.599
But obviously, uh, with the uh the biggest and most powerful systems in the world residing in the US, it does make a lot of sense.
00:19:18.319 --> 00:19:18.720
All right.
00:19:18.880 --> 00:19:22.319
Here's one that maybe didn't make as much sense.
00:19:22.559 --> 00:19:24.319
That's mid-journey.
00:19:24.640 --> 00:19:29.119
They've got an interesting pivot as I take a sip here of my coffee.
00:19:30.880 --> 00:19:33.200
Uh love the headline here.
00:19:33.279 --> 00:19:35.279
Uh, I I believe this was from The Verge.
00:19:35.440 --> 00:19:41.039
They said Midjourney goes from generating cat images to full-body ultrasound scans.
00:19:41.200 --> 00:19:42.480
Yeah, that's where we're at.
00:19:43.039 --> 00:19:51.680
So if you missed it, pretty uh, I mean, this could either end up being one of the biggest pivots or one of just the most confusing.
00:19:51.920 --> 00:20:04.960
But Midjourney CEO David Holes has revealed the company's first hardware product, marking obviously a very significant shift from its well AI well-known AI image generator.
00:20:05.039 --> 00:20:09.119
Uh, so here's what they announced at their events late last week.
00:20:09.359 --> 00:20:22.559
So they announced the Mid Journey Scanner, which is a full-body ultrasound device designed to capture detailed internal images of muscle, fat, bone, and organs using a ring of sensors.
00:20:22.720 --> 00:20:33.680
So Holse claims the scanner aims for image quality comparable to an actual MRI, though the scans should only take about 60 seconds, according to the company.
00:20:33.839 --> 00:20:46.160
So this pivot, well, you might be like being like, okay, how did they literally go from, you know, one of the first AI image models, you know, generating silly images, you know, especially early on that didn't look very good, right?
00:20:46.240 --> 00:20:52.400
But Midjourney, I think, was uh, you know, the big player early on, before Dolly and all these other players.
00:20:52.480 --> 00:21:01.279
Uh, but before Midjourney, uh, their CEO, um, David Holes, actually had a relevant background.
00:21:01.359 --> 00:21:12.240
So he spent a decade as the CTO of Leap Motion, which is a company entirely focused on advanced hardware, optical spatial computing, and tracking human anatomy.
00:21:12.319 --> 00:21:21.359
Uh, and the other reason why this pivot maybe makes sense is well, because Mid Journey doesn't have any investors, so they can do what they want.
00:21:21.519 --> 00:21:29.519
So, presumably, aside from there's a lot of lawsuits out there that are still pending, so we'll see where they uh land or settle in those.
00:21:29.599 --> 00:21:39.680
But presumably, Mid Journey has a nice stack of cash, and they don't have you know, let alone companies like you know, SpaceX and Cerebrus that just went public.
00:21:39.839 --> 00:21:43.519
Obviously, anthropic and open AI will be going public uh fairly soon.
00:21:43.599 --> 00:21:47.039
So that changes what you can and can't do, and even in the time leading up.
00:21:47.119 --> 00:21:53.680
But Mid Journey, as a company that's essentially bootstrapped in uh at a time at least, they were printing money.
00:21:53.839 --> 00:21:55.920
Uh, they can make these kind of pivots.
00:21:56.000 --> 00:21:59.759
And with Holz's background, it maybe makes a lot more sense than people.
00:22:00.160 --> 00:22:02.640
People initially thought when they just saw the headlines.
00:22:02.799 --> 00:22:04.880
So let's talk a little bit more about the device.
00:22:05.039 --> 00:22:17.759
So it uses 40 butterfly ultrasound on chip imaging modules and two petaflops of processing power developed in network uh in partnership with the butterfly network.
00:22:17.839 --> 00:22:33.759
So the scanning process involves the subject stepping into a shallow pool of light and descending into water, where thousands of underwater sensors use ultrasonic waves to create 3D images of the person that is in the machine.
00:22:34.000 --> 00:22:36.160
So here's the other interesting part.
00:22:36.400 --> 00:22:43.279
Well, you're not gonna, well, maybe you will, but at least that's not the plan to find these things in a uh a doctor's office.
00:22:43.359 --> 00:22:58.880
That's because Midjourney First plans to open a spa in San Francisco's Union Square by the end of next year, featuring 10 scanners, a gym, saunas, cold plunges, and hot tub equipped scanning rooms.
00:22:59.119 --> 00:23:06.720
So yeah, not something that they're you know going to be rolling out to uh, you know, AI equipped doctor offices.
00:23:06.799 --> 00:23:10.640
It seems like they're gonna be opening spas with these scanners.
00:23:10.799 --> 00:23:16.720
So the company is currently focused on body composition maps, which do not require FDA clearance.
00:23:16.799 --> 00:23:17.759
That's the key there.
00:23:17.920 --> 00:23:22.559
But future medical applications would need regulatory approval.
00:23:22.720 --> 00:23:28.079
So users will be able to share their scan library with doctors and AI health tools.
00:23:28.160 --> 00:23:33.119
And Midjourney promised to prioritize data privacy as the launch approaches.
00:23:33.279 --> 00:23:46.559
So Holse envisions the scanner becoming a faster, safer alternative to an MRI without radiation or magnets in hopes for future FDA device classifications to enable broader data collection.
00:23:47.119 --> 00:23:53.680
So, yeah, a strange one here, but uh the more you think about it, it kind of makes sense.
00:23:53.920 --> 00:23:54.240
All right.
00:23:54.480 --> 00:24:00.160
MRIs are expensive, they're not easy to get into.
00:24:00.319 --> 00:24:04.480
So we'll see uh how this is initially received, and we'll see.
00:24:04.640 --> 00:24:18.480
I think the telling thing will be how this is received outside of San Francisco, uh, because let's be honest, there's a decent chance no one knows how much this is gonna cost, but this just seems like it's made for tech bros, right?
00:24:18.559 --> 00:24:19.200
Let's be honest.
00:24:19.279 --> 00:24:29.839
Let's it it it it sounds like it's made for people with you know who uh put you know use stacks of cash to, you know, as a booster seat, essentially.
00:24:30.000 --> 00:24:38.720
So I think people who have you know millions of dollars of disposable income who live in Silicon Valley might see something like this and say, oh, absolutely.
00:24:38.880 --> 00:24:44.480
Uh outside of San Francisco, I mean, we'll see how this catches on, right?
00:24:44.559 --> 00:24:45.680
I'm from Chicago.
00:24:45.839 --> 00:24:47.920
I don't necessarily see this catching on.
00:24:48.079 --> 00:24:49.680
Again, it depends on price.
00:24:49.839 --> 00:24:51.359
Yes, there's all these other things.
00:24:51.599 --> 00:24:55.279
The uh the spa, the plunge pool sounds great and all.
00:24:55.359 --> 00:25:12.319
Uh, but I don't know how many people would be interested in, you know, in their kind of thought on this is the more that you scan, the more scans you get, uh, the more likely you are to pick something up and detect a potential uh disease or abnormal uh abnormality sooner, which makes total sense.
00:25:12.400 --> 00:25:26.160
But I just don't know the amount of people who can afford something like that, or if there's an appetite outside of tight tech circles for people to be going, I don't know, monthly or even quarterly uh to get these scans.
00:25:26.240 --> 00:25:29.920
But regardless, hey, you got to swing for the fences to hit a home run.
00:25:30.000 --> 00:25:33.680
You're not gonna put runs on the board by bunting every single time at bat.
00:25:34.160 --> 00:25:34.640
All right.
00:25:34.799 --> 00:25:43.119
Speaking of swinging for the fences, this one did not get a lot of coverage, but I decided to make it a main story because I think it's really important.
00:25:43.359 --> 00:25:59.200
So Cursor announced a new, well, they didn't actually announce it, but they kind of previewed it that they're working on their own from scratch, 1.5 trillion parameter model, and they are releasing a GitHub competitor.
00:25:59.359 --> 00:26:03.680
So two big pieces from uh Cursor.
00:26:03.839 --> 00:26:14.480
So days after officially being acquired by the now public SpaceX, Cursor dropped some major announcements at their compile conference.
00:26:14.640 --> 00:26:30.400
So those are that they're launching a new made-from scratch AI model with over 1.5 trillion parameters, trained on scratch on more than 100,000 GPUs, making it potentially one of the largest and most powerful models to date.
00:26:30.640 --> 00:26:48.160
So that's what happens when you uh are now owned by SpaceX and you have access uh to all these uh you know compute factories, uh Colossus one and two, uh, that uh Groc slash XAI slash SpaceX, right?
00:26:48.240 --> 00:26:50.720
They're now just one thing they've invested heavily in.
00:26:50.880 --> 00:27:04.640
So uh cursor says this new model is designed to be a generally intelligent assistant moving beyond code generation to handle complex engineering tasks like planning, testing, and interacting with user interfaces.
00:27:04.880 --> 00:27:08.000
So, a couple of reasons why I think just the model is important.
00:27:08.160 --> 00:27:19.920
Well, number one, their last model that they up uh that they came out with that I believe was based on uh Kimmy, so it was just a fine-tuned or built on top of an open source model.
00:27:20.000 --> 00:27:23.759
So Composer 2.5 was actually a really good model.
00:27:23.920 --> 00:27:28.400
Uh, it was the first time that Cursor, I think, had a model that a lot of people were starting to consider.
00:27:28.559 --> 00:27:32.880
And as we start to shift into the super app phase, all right.
00:27:32.960 --> 00:27:40.480
So make sure you go listen to episode 799 where we talked about AI super apps, and I did mention Cursor is a player uh now.
00:27:40.799 --> 00:27:43.839
Um, you like you have to look at the cost efficiency.
00:27:44.000 --> 00:27:57.200
So, you know, essentially with Composer 2.5, you got about maybe 90% of the power that you would get from something like Opus uh 4.8 at like 10% of the cost.
00:27:57.440 --> 00:28:22.240
Uh so with cursor going from scratch and now having that relationship with XAI, Grok, SpaceX, uh, I do think we have to start paying a little bit of attention, a little bit more attention uh to cursor, especially because I do think on the harness side, uh it is 1A and 1B in terms of I do think still, uh Codex on the Harness side, 1A and Cursor 1B.
00:28:22.400 --> 00:28:27.359
They have a great uh harness in the um in the cursor platform.
00:28:27.920 --> 00:28:38.559
So uh cursor's CEO Michael Truel emphasized the scale and fresh training approach, allows for better control over model behavior, and supports a wider range of engineering workloads.
00:28:38.720 --> 00:28:57.839
The other big news is they did announce their origin feature, a new Git native code hosting platform built for AI agents, which aims to challenge Microsoft's GitHub by supporting thousands of automated code operations daily in featuring automated conflict uh resolution.
00:28:58.000 --> 00:29:17.039
So cursor's acquisition gives it access to SpaceX's vast computational resources, with analysts saying the combination of Cursor's reinforcement learning expertise and SpaceX's compute power could quickly make Cursor a top competitor to Claude Code, Codex, and even GitHub.
00:29:17.200 --> 00:29:30.640
Uh so the GitHub play is interesting because we saw reports about a year and a half ago that OpenAI was developing their own uh GitHub uh kind of um competitor, but it never came to fruition.
00:29:30.720 --> 00:29:34.240
So maybe that was one of those you know side quests that was tabled.
00:29:34.319 --> 00:29:38.880
Uh so cursor maybe could start putting itself into the conversation.
00:29:38.960 --> 00:29:45.599
We'll see if we start shifting from talking about the big four of anthropic Google, open AI, and Microsoft.
00:29:45.680 --> 00:29:50.319
And maybe Cursor might thrust itself up into that first tier.
00:29:50.640 --> 00:29:55.279
All right, our last big piece of AI news.
00:29:55.519 --> 00:30:02.240
Yeah, it seems GPT-5.6 from OpenAI could be days away.
00:30:02.480 --> 00:30:15.759
So, according to reports, leaks, sleuths online, all that stuff, open AI is preparing to release its new GPT-5.6 family of AI models as soon as next week.
00:30:15.920 --> 00:30:31.359
Uh, so according to reports and leaks, uh, you might see a couple different variations of the GPT 5.6 model, including the standard, a new mini, and pro versions that could debut together.
00:30:31.519 --> 00:30:46.720
Uh, early access to GPT-5 Pro has reportedly already reached some pro subscribers who report noticeably noticeably better understanding on technical prompts, uh, though certain web development issues remain unresolved.
00:30:46.799 --> 00:30:54.079
So, yeah, uh, if you've been on you know Twitter at all over the last like five days, you'll see a lot of people.
00:30:54.160 --> 00:30:59.759
And I need to do a better job of you know coming up with these uh use cases that I actually save.
00:31:00.000 --> 00:31:15.599
The thing is, I try to not take on as many of the visual use cases because I think there's plenty of people out there, and I don't think that for the most part, these are things aside from you know creating things like decks, PowerPoints, um, you know, Word documents, you know, business dashboards.
00:31:15.680 --> 00:31:18.799
I think a lot of those um are easy to compare.
00:31:18.960 --> 00:31:28.000
But I think you know, what a lot of people are doing is just like you know, 3GS, uh, you know, 3D worlds, uh, you know, coding games.
00:31:28.160 --> 00:31:40.400
I don't think for the majority of our audience that's uh uh something that people are excited about, but that's why I think we've seen all these reports that apparently OpenAI maybe testing, A B testing uh GVD5.
00:31:44.880 --> 00:31:50.799
You know, create an interactive world model that you can walk through, you know, just using 3GS or something like that.
00:31:50.960 --> 00:31:56.240
And people have those kind of saved, and all of a sudden it's wow, like three, four times better.
00:31:56.400 --> 00:31:59.039
Uh, so a couple more details.
00:31:59.200 --> 00:32:13.839
The new GPT-5.6 is rumored to push the context window to 1.5 million tokens up from 1 million tokens in GPT-5.5, allowing for significantly longer and more complex conversations and coding sessions.
00:32:14.000 --> 00:32:30.319
So, developers who think or believe that they have access to GPT-5.6 Pro say that it outperforms anthropic's mythos models in agentic coding tasks, raising the stakes in the competition for the AI-powered software development.
00:32:30.720 --> 00:32:35.359
Uh, so uh make sure to keep an eye on the newsletter uh this week.
00:32:35.440 --> 00:32:51.519
And obviously, if it does come out, you better believe we will be having a show because there's a good chance that uh maybe as soon as today, maybe by the time that you uh you know hear this podcast, that we might have Fable Five from Anthropic back.
00:32:51.839 --> 00:32:57.039
Uh again, according to reports, we could have GVD5.6 as soon as this week.
00:32:57.200 --> 00:33:01.359
And we did earlier uh from Google themselves at their I.O.
00:33:01.599 --> 00:33:03.279
conference last month.
00:33:03.359 --> 00:33:11.759
Uh, they did say that next month, which would be this month in June, they did say that they're going to be releasing their Gemini 3.5 Pro model.
00:33:12.079 --> 00:33:19.519
So, in theory, this week we could get access or get access again to three state-of-the-art models.
00:33:19.599 --> 00:33:21.119
So, uh pretty big week.
00:33:21.200 --> 00:33:32.400
Uh, buckle in, make sure you got your tokens ready or your API uh budgets in check because there could be a lot of new model access rolling down the pipeline.
00:33:32.559 --> 00:33:36.160
All right, so that is it for our main AI news stories.
00:33:36.319 --> 00:33:38.480
So let's quickly go over everything else.
00:33:38.640 --> 00:33:50.480
This is the what's new and what's next, just stories that didn't make our top cut, some rumors, uh, some leaks, uh, but let's go ahead and go through them quickly, bullet point style.
00:33:51.359 --> 00:33:52.480
All right, here we go.
00:33:52.640 --> 00:34:09.280
So Amazon, Nvidia, and AMD invested$310 million in Odyssey ML to advance real-world AI understanding, anthropic upgraded claw and design, adding custom design systems, code sync, and export features.
00:34:09.519 --> 00:34:10.800
That one really good.
00:34:10.960 --> 00:34:11.840
FYI.
00:34:12.000 --> 00:34:19.039
OpenAI is reportedly preparing to integrate the library feature from ChatGPT into Codex.
00:34:19.199 --> 00:34:20.000
I would like that.
00:34:20.079 --> 00:34:22.400
I spent a lot of time trying to find all those files.
00:34:22.480 --> 00:34:27.039
Uh, XAI released their Grok Imagine Video 1.5 model, looks pretty good.
00:34:27.280 --> 00:34:38.239
Adobe expanded their Firefly AI assistant to other uh tools in their stack, including Premiere, Illustrator, InDesign, and Frame IO with some new features.
00:34:38.400 --> 00:34:44.480
OpenAI released unified usage analytics and spend controls for ChatGPT Enterprise.
00:34:44.639 --> 00:34:53.519
Uh, some new leaks show that Anthropic is working on a schedules feature for its always-on upcoming Claude Conway agent.
00:34:53.760 --> 00:35:01.280
Google rolled out their Ask Ad Manager Beta, an AI agent for custom insights and troubleshooting for Google Ads.
00:35:01.840 --> 00:35:06.480
Uh, like I said, GBT5.6 and Gemini 3.5 are likely around the corner.
00:35:06.719 --> 00:35:12.960
Codex, this is the big one, released their record and replay workflow capture for Mac.
00:35:13.280 --> 00:35:15.119
Uh, so excluding some countries.
00:35:15.280 --> 00:35:18.719
And I do believe I'll have to check again at the uh the newsletter.
00:35:18.880 --> 00:35:25.119
Most of the times, either on Friday or Monday, I'll ask you guys what you want for our Wednesday show, which is a hands-on demo.
00:35:25.280 --> 00:35:31.280
And I do believe that the new Codex uh record and replay uh one, which I was surprised.
00:35:31.360 --> 00:35:39.440
I thought Claude Design was gonna win our poll, uh, but it seems like everyone wanted the new uh record and replay feature, which is really cool.
00:35:39.599 --> 00:35:50.239
Uh Google Gemini co-lead Noam Shazir left for Open AI after Google acquired or Aqua hired uh him and his group from Character AI.
00:35:50.320 --> 00:35:51.360
This is a pretty big one, right?
00:35:51.440 --> 00:35:57.519
So one of the original authors on the attention is all you need uh paper that led to the Transformer, right?
00:35:57.599 --> 00:36:05.920
So one of the OGs of AI leaving uh Google Gemini after they spent$2.7 billion to join OpenAI.
00:36:06.079 --> 00:36:15.599
Uh another big get for OpenAI this week, former White House AI advisor Dean Ball joined OpenAI to lead Frontier AI policy team.
00:36:15.679 --> 00:36:22.079
So some big hires for OpenAI after Anthropic landed uh Andre Carpathy a couple of weeks ago.
00:36:22.320 --> 00:36:33.199
Speaking of Anthropic, they released their Claude Code Artifacts, which allows teams to share live coding artifacts for collaborative and interactive sessions.
00:36:33.360 --> 00:36:35.760
So very similar to Codecs sites.
00:36:35.840 --> 00:36:38.079
Uh XAI, oh, already covered that one.
00:36:38.159 --> 00:36:40.480
Uh Copilot Cowork is generally available.
00:36:40.639 --> 00:36:43.360
Like I said, downside though is usage-based pricing.
00:36:43.440 --> 00:36:47.440
And last but not least, Google released the open knowledge format.
00:36:47.599 --> 00:36:54.639
So kind of an alternative, or what could in theory be an alternative uh for an uh agentic communication system.
00:36:54.800 --> 00:37:00.559
So instead of markdown files, you might have OKFs if a lot of people uh adapt this.
00:37:00.719 --> 00:37:01.840
So that's it.
00:37:02.000 --> 00:37:04.960
That is a wrap for the AI news that matter.
00:37:05.039 --> 00:37:11.039
So, y'all, I'm telling you, this is happening so fast, and the pace is only going to pick up.
00:37:11.199 --> 00:37:11.599
All right.
00:37:11.679 --> 00:37:18.400
So you probably shouldn't be spending like four or five hours every single day, right, taking all this work home for you.
00:37:18.480 --> 00:37:19.360
That's what I do.
00:37:19.519 --> 00:37:25.199
So make sure I try to keep these podcasts 30-ish minutes, uh, right, even faster if you listen on 2x.
00:37:25.280 --> 00:37:29.599
So make sure you subscribe to the podcast, but you really need to be reading our newsletter.
00:37:29.760 --> 00:37:32.880
So make sure you go do that at your everydayai.com.
00:37:32.960 --> 00:37:35.280
So thanks for tuning in and FYI.
00:37:35.440 --> 00:37:38.639
If you are new here, Mondays we do our AI news that matters.
00:37:38.800 --> 00:37:44.079
Wednesdays, we go hands-on with a kind of live demo with our AI at work on Wednesdays.
00:37:44.239 --> 00:37:49.039
Friday, we go over Friday features, which are new features that you can actually go use today.
00:37:49.119 --> 00:37:50.880
And I kind of give you the 101.
00:37:50.960 --> 00:37:53.760
And then Tuesdays and Thursdays, we rotate what those shows are.
00:37:53.840 --> 00:37:56.880
So if you knew if you're new here, that's kind of our weekly lineup.
00:37:56.960 --> 00:37:57.920
So thanks for tuning in.
00:37:58.079 --> 00:38:01.119
Hope to see you back tomorrow in every day for more everyday AI.
00:38:01.199 --> 00:38:02.239
Thanks, y'all.
00:00:02.000 --> 00:00:11.039
While the US government and Anthropic continued trying to make up and be friends again, the rest of the AI world didn't stop.
00:00:11.199 --> 00:00:23.359
And it was actually some smaller players that were making big splashes this week, potentially signaling more competition at the top for the big four of Anthropic, OpenAI, Google, and Microsoft.
00:00:23.519 --> 00:00:38.320
That's because open source may finally be having its ChatGPT moment, as a new Chinese model might actually be closing the gap between proprietary US models and open source or open weight options.
00:00:38.560 --> 00:00:41.520
Oh, and Midjourney made some news as well.
00:00:41.679 --> 00:00:45.679
Yeah, that mid journey that you probably haven't heard from in years.
00:00:45.920 --> 00:00:50.399
That's because they may be trying to disrupt the medical imaging space.
00:00:50.640 --> 00:00:55.359
Yeah, there's a lot going on if you don't follow every single day.
00:00:55.600 --> 00:00:58.240
And don't worry, I do that for you.
00:00:58.399 --> 00:01:00.320
So welcome to Everyday AI.
00:01:00.399 --> 00:01:07.280
And this is our Monday segment where we bring you only the most relevant news updates.
00:01:07.439 --> 00:01:11.920
I tell you what matters, what doesn't, and how you can use that information to grow your company and career.
00:01:12.000 --> 00:01:13.280
This is Everyday AI.
00:01:13.359 --> 00:01:20.159
Uh, so if you're new here, yeah, we do this every single day, daily, unedited, unscripted, live stream podcasts, and free daily newsletter.
00:01:20.239 --> 00:01:26.159
So uh if you are new, please make sure to subscribe to the show uh on Spotify or Apple Podcasts.
00:01:26.239 --> 00:01:26.799
I'd appreciate that.
00:01:26.879 --> 00:01:32.480
And then make sure you go to our website at your everydayai.com, sign up for the free daily newsletter each and every day.
00:01:32.560 --> 00:01:39.680
Uh, we recap the highlights from today's show as well as give you all the other AI news you need to know just in a quick digestible form.
00:01:39.920 --> 00:01:47.120
All right, uh, so without further ado, let's go over the biggest stories of the week that you might have missed.
00:01:47.280 --> 00:01:49.840
And I'm gonna break it down for you simply.
00:01:50.159 --> 00:02:02.000
All right, let's start at the top because yes, at least at the time of this recording, we still don't have access to Fable Five or Mythos V at whole.
00:02:02.079 --> 00:02:07.920
That's because the Trump White House and Anthropic are still trying to make things work.
00:02:08.159 --> 00:02:30.000
So, uh, actually a little bit of a controversial statement, maybe, uh, from President Trump as he told the Axios show on Friday that he considered Anthropic previously and or its CEO, Dario Amati, to be a national security threat, though he now believes the situation is improving.
00:02:30.159 --> 00:02:31.439
Yeah, he did say that.
00:02:31.520 --> 00:02:34.159
You might go want to listen to the interview yourself.
00:02:34.319 --> 00:02:46.800
We'll be linking it in today's newsletter, but he definitely said that he considered them a few weeks ago, whether he was talking about Anthropic as a whole or Dario Amati to be a national security threat, though it's now he says it's improving.
00:02:46.960 --> 00:02:47.599
So the U.S.
00:02:47.680 --> 00:03:05.520
government, if you haven't been following along, the US government recently imposed strict export controls on Anthropic, barring any country outside of the US and foreign nationals within the US from accessing uh Anthropic's most recent models called Fable V and Mythos V.
00:03:06.159 --> 00:03:24.319
So Anthropic has completely disabled in response to that, they've disabled public access uh to those models, Fable V, which is the version of Mythos V with more guardrails, uh so yeah, but it's been down and gone now for almost a week since that export control.
00:03:24.719 --> 00:03:34.560
Uh so the directive bars any foreign nationals, including uh foreign national anthropic employees, from accessing uh these frontier systems.
00:03:34.719 --> 00:03:47.680
But because anthropic can't instantly verify the exact citizens uh citizenship of every single user on its platform, it was forced to yank both those models offline globally to remain compliant.
00:03:47.840 --> 00:03:49.520
And that's been about a week.
00:03:49.680 --> 00:04:01.280
So, according to Axios in this new interview and story, the Trump administration's concerns were triggered by the Amazon report detailing the vulnerability in Anthropics technology.
00:04:01.439 --> 00:04:09.680
Uh, right, and a lot of reports said that was their CEO, Andy Jassy, and that led to the urgent discussions with the company leadership.
00:04:10.000 --> 00:04:25.120
Trump said he was initially alarmed by Anthropics' response, but was reassured after Amoti uh acted quickly and responsibly, calling him nice and smart after beating him at the G7 summit this past week.
00:04:25.279 --> 00:04:32.079
So the Pentagon has labeled Anthropic a supply chain risk, a designation typically reserved for foreign adversaries.
00:04:32.160 --> 00:04:33.759
And it was the first time a U.S.
00:04:33.839 --> 00:04:36.560
company had gotten that label uh from the U.S.
00:04:36.639 --> 00:04:37.120
government.
00:04:37.279 --> 00:04:50.160
So Trump did not rule out using emergency powers under the Defense Production Act if Anthropic does not comply with new standards, but stated he doesn't think that step will be necessary this time.
00:04:50.399 --> 00:05:00.800
Uh, Trump revealed that it was a competitor and part owner identified later as Amazon that alerted his administration about Anthropic's uh practice.
00:05:00.879 --> 00:05:03.680
So we went over this uh in detail, FYI.
00:05:03.920 --> 00:05:08.319
If you missed the story, let me just go ahead and pull the uh the episode number.
00:05:08.399 --> 00:05:10.560
That was from Thursday, episode 801.
00:05:10.720 --> 00:05:16.319
So if you want more on uh the background on those developments, you go listen to that episode 801.
00:05:16.399 --> 00:05:20.560
But at least for now, Mythos 5, still offline.
00:05:20.720 --> 00:05:25.839
Uh, but the latest here, uh, President Trump sat down with an interview with Axios.
00:05:25.920 --> 00:05:29.199
So, yeah, if you missed that, uh, definitely worth checking out.
00:05:29.600 --> 00:05:41.920
All right, our next story, and this seems to be an underlying theme this week on open source models, uh, apparently being considered at the highest levels.
00:05:42.160 --> 00:05:49.120
So, Microsoft is moving its co-pilot co-work AI assistant to a usage-based pricing model.
00:05:49.279 --> 00:05:51.439
So, uh kind of a couple pieces of news here.
00:05:51.519 --> 00:06:01.759
Number one, that Microsoft is going usage based on co-pilot cowork instead of just including it or going uh with you know a certain number that's included in a monthly subscription.
00:06:01.920 --> 00:06:02.639
That's number one.
00:06:02.800 --> 00:06:06.319
Number two is it's generally available, whereas before it was in preview.
00:06:06.480 --> 00:06:22.480
And number three, the most recent news, according to reports, is that they're actually looking at models like China's Deep Seek V4, a version that Microsoft may fine-tune uh just for co-pilot co-work to bring down costs.
00:06:23.199 --> 00:06:35.920
So, according to reports, the company is exploring a lower cost self-hosted AI model as an alternative to the more expensive anthropic models that are currently powering co-pilot co-work by default.
00:06:36.079 --> 00:06:56.240
And Axios did report that Microsoft is considering, among others, Deep uh Deep Seek V4, a Chinese developed open source model, which could dramatically reduce costs, potentially charging just about 87 cents per million tokens compared to Anthropic's$50 per million tokens.
00:06:56.399 --> 00:07:03.759
So, yeah, not the best at math, but that's like 18x cheaper, uh, or something like that.
00:07:03.920 --> 00:07:16.720
Uh, so Microsoft expects to announce its final model choice and deployment details within weeks with new uh with any new option to be hosted on Azure to keep customer data within Microsoft's cloud infrastructure.
00:07:16.879 --> 00:07:24.720
So the shift responds to the high compute cost that we've been hearing about over the last month or two uh of agentic AI tools.
00:07:24.800 --> 00:07:31.040
Uh, specifically, we've seen a lot of reports from enterprises uh, you know, getting claud bills that are pretty high, right?
00:07:31.120 --> 00:07:34.000
We saw that report uh a couple couple of weeks ago.
00:07:34.160 --> 00:07:45.519
We don't know if it was legitimate or not, but apparently an enterprise company had a uh nine-figure uh bill from Anthropic, which is a lot, uh, right?
00:07:45.680 --> 00:07:51.680
Hundreds of millions of dollars uh because they weren't keeping uh an eye on usage caps.
00:07:51.839 --> 00:08:06.000
So that's why open source models like Deep Seek V4 can cut costs, but they do require companies like Microsoft to handle additional hosting, optimization, and safety engineering, shifting some burdens away from commercial API providers.
00:08:06.160 --> 00:08:18.639
So there are obviously a ton of political and regulatory concerns as DeepSeek's Chinese origins have attracted scrutiny in Washington, raising questions about data security and compliance for U.S.
00:08:18.800 --> 00:08:19.680
companies.
00:08:20.000 --> 00:08:32.480
Uh, Microsoft has stated that any Deep Seek-based solution would be optional and subject to additional safety and compliance filters, aiming to address enterprise risk and regulatory requirements.
00:08:32.639 --> 00:08:48.799
Uh, so industry observers are watching for Microsoft's final decision, noting that total uh cost of ownership will depend not just on it per token prices, but also on model performance, reliability, and the need for robust monitoring and safety systems.
00:08:48.960 --> 00:09:01.039
So, this one it is kind of uh intriguing, I'll say that to say the least, uh, with Microsoft making this move and specifically, according to reports, looking at Deep Seek.
00:09:01.120 --> 00:09:02.639
So, why is that noteworthy?
00:09:02.799 --> 00:09:10.799
Well, uh, Microsoft obviously has good relationships and huge ownership stakes in both Anthropic and OpenAI.
00:09:10.879 --> 00:09:11.360
Uh, right.
00:09:11.519 --> 00:09:17.919
So uh traditionally, uh it's been the GBT models from OpenAI that have powered uh Copilot.
00:09:18.000 --> 00:09:25.120
They recently started offering some Claude models as well, but the Copilot Cowork was running Claude models by default.
00:09:25.279 --> 00:09:36.159
Uh, so co-work is kind of the original technology from Anthropic that Microsoft essentially had their own version that was powered under the hood by Anthropic's Copilot technology.
00:09:36.399 --> 00:09:58.240
So, what's interesting here with Microsoft looking at model providers like DeepSeq is well, OpenAI and Anthropic have launched some serious concerns in uh regulatory um steps that they've taken, uh, you know, going to the US government, accusing DeepSeq and other Chinese companies of distilling their models.
00:09:58.399 --> 00:10:00.480
So that's what's gonna make this really juicy.
00:10:00.639 --> 00:10:31.759
Uh, pretty, if I'm being honest, a pretty bold uh step here from Microsoft, even though it is just according to reports uh to do this, considering uh two of their biggest and most important partners in Open AI and Anthropic probably don't have the best view of Deep Seek for those exact reasons because they've shown uh right, Anthropic came out with a little bit more of a report how DeepSeq and other Chinese companies are essentially taking all of their hard work to make these open source models.
00:10:31.919 --> 00:10:35.519
All right, so our next AI story staying in the same vein.
00:10:35.679 --> 00:10:39.600
Yeah, we have a new king of the open hill.
00:10:39.759 --> 00:10:50.240
Uh that's Chinese startup ZAI is making a ton of headlines this week with their new open weights model GLM52.
00:10:50.720 --> 00:11:04.720
So ZAI has released that model, GLM5.2, a massive 753 billion parameter language model, now available with unrestricted MIT open source licensing.
00:11:04.879 --> 00:11:19.440
So GLM 5.2 is designed for long horizon autonomous coding and engineering tasks, featuring a highly stable 1 million token context window, allowing it to handle extremely long documents and complex workflows.
00:11:19.600 --> 00:11:23.440
So the model is immediately accessible for download on Hugging Face.
00:11:23.679 --> 00:11:33.360
But yeah, you kind of have to be an enterprise with a little bit of compute to be able to do that because it is a large model, unless you're going to run a highly quantized version of it.
00:11:33.519 --> 00:11:42.639
Uh, but you can also use it through Hugging Face, uh, through the ZAI API, and in over 20 third-party coding environments.
00:11:42.799 --> 00:11:50.879
Uh, so uh the new model, which is pretty interesting here, uh, it does look like it's very benchmaxed FYI, uh right.
00:11:50.960 --> 00:12:15.440
That's when a company maybe puts a little too much priority, uh priority on getting certain benchmarks, uh, but still it outperforms most other open source competitors, and it is even now challenging the top proprietary models on several industry benchmarks, including Suitebench Pro, which I think is thankfully on its way out of being an important model for software engineering.
00:12:15.600 --> 00:12:27.919
Yet still, uh, it is batting on some of these benchmarks, including SuiteBench Pro, Frontier Suite, uh, and others around GPT-5.5 and Anthropics Claude Opus 4.8.
00:12:28.000 --> 00:12:31.120
So with Fable 5 off the shelf, right?
00:12:31.279 --> 00:12:37.440
I mean, oh GPT-5.5 and uh Opus 4.8 are the undisputed leaders.
00:12:37.600 --> 00:12:53.840
So uh pretty significant, even if it is maybe a little bit oversaturated and a little bit bench maxed, uh, that an open source model, uh, I I won't say they've closed the gap, but they're starting to close the gap on uh frontier models.
00:12:53.919 --> 00:13:01.440
So there's obviously some use cases that you would absolutely not want to use uh GLM5.2, mainly because it's not multimodal.
00:13:01.519 --> 00:13:02.720
So it's text only.
00:13:02.879 --> 00:13:04.480
So that's a big downside, right?
00:13:04.559 --> 00:13:15.759
So for someone like me, I'm constantly uh doing multimodal inputs, you know, whether that's in Google's AI Studio, whether it's in uh, you know, Anthropic Claude, Codex, ChatGPT, etc.
00:13:16.000 --> 00:13:16.240
Right.
00:13:16.320 --> 00:13:19.919
So you can't have multimodal input or multimodal output.
00:13:20.000 --> 00:13:39.120
But with text only, uh, you know, GLM5.2 does seem to be a competitor, and it is technically state of the art, even beating uh in a couple benchmarks, I kid you not, Fable 5, Opus 4.8, and GPT-55 on some more niche uh kind of front-end coding design, some certain benchmarks out there.
00:13:39.279 --> 00:13:52.159
So uh pretty surprising uh considering that even the best open source models prior to this were generally, I would say, about four to eight months behind the frontier counterpart.
00:13:52.399 --> 00:13:58.960
So still GLM52 is far behind, mainly in terms of capabilities because it's not multimodal.
00:13:59.120 --> 00:14:06.799
But for text only, you know, at this point, it is only like two-ish, two to three months behind, uh, which is fairly impressive.
00:14:06.879 --> 00:14:18.240
And I think it does change the narrative, especially when uh, you know, kind of all of the rage in the second quarter in the summer so far has been shifting to token efficiency.
00:14:18.399 --> 00:14:45.919
I do think most companies, the biggest decision that they're gonna make in 2026, similarly, similarly to what Microsoft is doing, uh, you know, looking at switching over from some anthropic models uh to deep seek is well, most companies are gonna have to start looking at their uh their whether you're talking about token budgets, your AI spend, but you know, switching over uh from token maxing to token efficiency, which we did a story on that in our start here series.
00:14:46.000 --> 00:14:47.120
So make sure you go check that out.
00:14:47.200 --> 00:15:00.720
But essentially, uh, right, some of these agents run in loops nonstop, and as companies have kind of started to uh pluck away the uh you know, um them kind of subsidizing, right?
00:15:00.879 --> 00:15:14.159
Because you see different reports, you know, a$200 a month plan if you were paying API costs might cost actually, if you're paying for those tokens, anywhere from you know 10 to 15 to 20,000 uh, you know, for 200.
00:15:14.559 --> 00:15:22.960
So, you know, essentially these companies have been eating uh those costs just to get users in and happy and sticky and to you know get it rolled out into their organization.
00:15:23.039 --> 00:15:26.720
So a lot of companies are going to be have to be looking at token efficiency.
00:15:26.879 --> 00:15:30.240
So uh pretty big news, both with the last two stories.
00:15:30.399 --> 00:15:43.840
Microsoft looking at open alternatives to power its co-pilot cowork as they switch to usage base and with the new ZAI 5 uh GLM 5.2 model, definitely one worth keeping an eye on.
00:15:44.000 --> 00:15:57.360
All right, our next AI news story the CEOs of Anthropic and Google DeepMind uh and OpenAI met with G7 leaders and top US officials at the G7 Summit in France.
00:15:58.000 --> 00:16:07.519
So the main news out of here, at least when it comes to an artificial intelligence standpoint, was a push for a US-led coalition on artificial intelligence.
00:16:07.759 --> 00:16:23.440
So uh Dario Amati and Google DeepMind CEO Damas uh Demas Hasabas urged the creation of a US-led international coalition to set rules and standards for AI, citing urgent risks and the need for global cooperation.
00:16:23.679 --> 00:16:25.519
So, President Donald Trump, uh U.S.
00:16:25.600 --> 00:16:33.200
Secretary of State Marco Rubio, Treasury Secretary Scott Bessant, um, and Commerce Secretary Howard Lutnick represented the U.S.
00:16:33.279 --> 00:16:36.639
at the meeting, highlighting the high-level attention on AI governance.
00:16:36.879 --> 00:16:42.639
We also had Canadian Prime Minister uh Mark Carney, who expressed support for U.S.
00:16:42.720 --> 00:16:48.799
leadership in forming an AI standards coalition, signaling growing international alignment on the issue.
00:16:48.960 --> 00:16:51.679
So the call for action comes after the U.S.
00:16:51.759 --> 00:17:05.359
government imposed export controls on Anthropic's latest AI models, Fable V and Mythos V, which led the company to disable access due to the national security concerns that the government labeled.
00:17:05.519 --> 00:17:17.920
Uh, so emoti recommended that international cooperation should cover structured access to advanced AI models and trade of chips and critical components, specifically excluding China from these exchanges.
00:17:18.079 --> 00:17:29.839
So the tech leaders emphasized the urgent need to address AI risks in cybersecurity, bioterrorism, and intelligence, warning that unchecked AI could lead to major disasters if misused.
00:17:30.160 --> 00:17:44.400
OpenAI CEO Sam Altman advocated for a global forum to set standards for testing, risk analysis, and international collaboration, while OpenAI's Chris Lehan noted broad recognition that the US should take the lead.
00:17:44.640 --> 00:17:51.519
So the big development out of G7 Summit, well, number one is that AI was front and center, uh.
00:17:51.920 --> 00:17:59.920
So maybe it was just timing, circumstantial timing, uh, with the mythos uh pull down that this happened in the middle of it.
00:18:00.079 --> 00:18:23.200
So uh, you know, especially when it comes to the rise of Chinese open source models, uh, obviously, from an adversarial perspective, the G7 uh collectively wants to keep an eye on what China is doing in AI, how they're using this technology, because, like I've been saying for years, uh ultimately this will play out one way or another on the battlefield, right?
00:18:23.279 --> 00:18:47.920
Uh, this is, I think, why uh the G7 and world leaders uh are starting to pay more and more attention to AI, not just because of how it can be used um in cyber warfare, but also on the actual battlefield, which is uh actually goes back to the initial uh kind of feud between Anthropic and the US government for how anthropic did not want the military to use its models.
00:18:48.160 --> 00:19:00.799
So uh the other big takeaway here is obviously seemingly some international support uh for the US to take a lead in this G7 AI coalition, uh, which makes sense, right?
00:19:00.960 --> 00:19:07.920
A lot of times other world leaders, you know, might not want to, you know, step aside and let the US take a lead on something like this.
00:19:08.079 --> 00:19:17.599
But obviously, uh, with the uh the biggest and most powerful systems in the world residing in the US, it does make a lot of sense.
00:19:18.319 --> 00:19:18.720
All right.
00:19:18.880 --> 00:19:22.319
Here's one that maybe didn't make as much sense.
00:19:22.559 --> 00:19:24.319
That's mid-journey.
00:19:24.640 --> 00:19:29.119
They've got an interesting pivot as I take a sip here of my coffee.
00:19:30.880 --> 00:19:33.200
Uh love the headline here.
00:19:33.279 --> 00:19:35.279
Uh, I I believe this was from The Verge.
00:19:35.440 --> 00:19:41.039
They said Midjourney goes from generating cat images to full-body ultrasound scans.
00:19:41.200 --> 00:19:42.480
Yeah, that's where we're at.
00:19:43.039 --> 00:19:51.680
So if you missed it, pretty uh, I mean, this could either end up being one of the biggest pivots or one of just the most confusing.
00:19:51.920 --> 00:20:04.960
But Midjourney CEO David Holes has revealed the company's first hardware product, marking obviously a very significant shift from its well AI well-known AI image generator.
00:20:05.039 --> 00:20:09.119
Uh, so here's what they announced at their events late last week.
00:20:09.359 --> 00:20:22.559
So they announced the Mid Journey Scanner, which is a full-body ultrasound device designed to capture detailed internal images of muscle, fat, bone, and organs using a ring of sensors.
00:20:22.720 --> 00:20:33.680
So Holse claims the scanner aims for image quality comparable to an actual MRI, though the scans should only take about 60 seconds, according to the company.
00:20:33.839 --> 00:20:46.160
So this pivot, well, you might be like being like, okay, how did they literally go from, you know, one of the first AI image models, you know, generating silly images, you know, especially early on that didn't look very good, right?
00:20:46.240 --> 00:20:52.400
But Midjourney, I think, was uh, you know, the big player early on, before Dolly and all these other players.
00:20:52.480 --> 00:21:01.279
Uh, but before Midjourney, uh, their CEO, um, David Holes, actually had a relevant background.
00:21:01.359 --> 00:21:12.240
So he spent a decade as the CTO of Leap Motion, which is a company entirely focused on advanced hardware, optical spatial computing, and tracking human anatomy.
00:21:12.319 --> 00:21:21.359
Uh, and the other reason why this pivot maybe makes sense is well, because Mid Journey doesn't have any investors, so they can do what they want.
00:21:21.519 --> 00:21:29.519
So, presumably, aside from there's a lot of lawsuits out there that are still pending, so we'll see where they uh land or settle in those.
00:21:29.599 --> 00:21:39.680
But presumably, Mid Journey has a nice stack of cash, and they don't have you know, let alone companies like you know, SpaceX and Cerebrus that just went public.
00:21:39.839 --> 00:21:43.519
Obviously, anthropic and open AI will be going public uh fairly soon.
00:21:43.599 --> 00:21:47.039
So that changes what you can and can't do, and even in the time leading up.
00:21:47.119 --> 00:21:53.680
But Mid Journey, as a company that's essentially bootstrapped in uh at a time at least, they were printing money.
00:21:53.839 --> 00:21:55.920
Uh, they can make these kind of pivots.
00:21:56.000 --> 00:21:59.759
And with Holz's background, it maybe makes a lot more sense than people.
00:22:00.160 --> 00:22:02.640
People initially thought when they just saw the headlines.
00:22:02.799 --> 00:22:04.880
So let's talk a little bit more about the device.
00:22:05.039 --> 00:22:17.759
So it uses 40 butterfly ultrasound on chip imaging modules and two petaflops of processing power developed in network uh in partnership with the butterfly network.
00:22:17.839 --> 00:22:33.759
So the scanning process involves the subject stepping into a shallow pool of light and descending into water, where thousands of underwater sensors use ultrasonic waves to create 3D images of the person that is in the machine.
00:22:34.000 --> 00:22:36.160
So here's the other interesting part.
00:22:36.400 --> 00:22:43.279
Well, you're not gonna, well, maybe you will, but at least that's not the plan to find these things in a uh a doctor's office.
00:22:43.359 --> 00:22:58.880
That's because Midjourney First plans to open a spa in San Francisco's Union Square by the end of next year, featuring 10 scanners, a gym, saunas, cold plunges, and hot tub equipped scanning rooms.
00:22:59.119 --> 00:23:06.720
So yeah, not something that they're you know going to be rolling out to uh, you know, AI equipped doctor offices.
00:23:06.799 --> 00:23:10.640
It seems like they're gonna be opening spas with these scanners.
00:23:10.799 --> 00:23:16.720
So the company is currently focused on body composition maps, which do not require FDA clearance.
00:23:16.799 --> 00:23:17.759
That's the key there.
00:23:17.920 --> 00:23:22.559
But future medical applications would need regulatory approval.
00:23:22.720 --> 00:23:28.079
So users will be able to share their scan library with doctors and AI health tools.
00:23:28.160 --> 00:23:33.119
And Midjourney promised to prioritize data privacy as the launch approaches.
00:23:33.279 --> 00:23:46.559
So Holse envisions the scanner becoming a faster, safer alternative to an MRI without radiation or magnets in hopes for future FDA device classifications to enable broader data collection.
00:23:47.119 --> 00:23:53.680
So, yeah, a strange one here, but uh the more you think about it, it kind of makes sense.
00:23:53.920 --> 00:23:54.240
All right.
00:23:54.480 --> 00:24:00.160
MRIs are expensive, they're not easy to get into.
00:24:00.319 --> 00:24:04.480
So we'll see uh how this is initially received, and we'll see.
00:24:04.640 --> 00:24:18.480
I think the telling thing will be how this is received outside of San Francisco, uh, because let's be honest, there's a decent chance no one knows how much this is gonna cost, but this just seems like it's made for tech bros, right?
00:24:18.559 --> 00:24:19.200
Let's be honest.
00:24:19.279 --> 00:24:29.839
Let's it it it it sounds like it's made for people with you know who uh put you know use stacks of cash to, you know, as a booster seat, essentially.
00:24:30.000 --> 00:24:38.720
So I think people who have you know millions of dollars of disposable income who live in Silicon Valley might see something like this and say, oh, absolutely.
00:24:38.880 --> 00:24:44.480
Uh outside of San Francisco, I mean, we'll see how this catches on, right?
00:24:44.559 --> 00:24:45.680
I'm from Chicago.
00:24:45.839 --> 00:24:47.920
I don't necessarily see this catching on.
00:24:48.079 --> 00:24:49.680
Again, it depends on price.
00:24:49.839 --> 00:24:51.359
Yes, there's all these other things.
00:24:51.599 --> 00:24:55.279
The uh the spa, the plunge pool sounds great and all.
00:24:55.359 --> 00:25:12.319
Uh, but I don't know how many people would be interested in, you know, in their kind of thought on this is the more that you scan, the more scans you get, uh, the more likely you are to pick something up and detect a potential uh disease or abnormal uh abnormality sooner, which makes total sense.
00:25:12.400 --> 00:25:26.160
But I just don't know the amount of people who can afford something like that, or if there's an appetite outside of tight tech circles for people to be going, I don't know, monthly or even quarterly uh to get these scans.
00:25:26.240 --> 00:25:29.920
But regardless, hey, you got to swing for the fences to hit a home run.
00:25:30.000 --> 00:25:33.680
You're not gonna put runs on the board by bunting every single time at bat.
00:25:34.160 --> 00:25:34.640
All right.
00:25:34.799 --> 00:25:43.119
Speaking of swinging for the fences, this one did not get a lot of coverage, but I decided to make it a main story because I think it's really important.
00:25:43.359 --> 00:25:59.200
So Cursor announced a new, well, they didn't actually announce it, but they kind of previewed it that they're working on their own from scratch, 1.5 trillion parameter model, and they are releasing a GitHub competitor.
00:25:59.359 --> 00:26:03.680
So two big pieces from uh Cursor.
00:26:03.839 --> 00:26:14.480
So days after officially being acquired by the now public SpaceX, Cursor dropped some major announcements at their compile conference.
00:26:14.640 --> 00:26:30.400
So those are that they're launching a new made-from scratch AI model with over 1.5 trillion parameters, trained on scratch on more than 100,000 GPUs, making it potentially one of the largest and most powerful models to date.
00:26:30.640 --> 00:26:48.160
So that's what happens when you uh are now owned by SpaceX and you have access uh to all these uh you know compute factories, uh Colossus one and two, uh, that uh Groc slash XAI slash SpaceX, right?
00:26:48.240 --> 00:26:50.720
They're now just one thing they've invested heavily in.
00:26:50.880 --> 00:27:04.640
So uh cursor says this new model is designed to be a generally intelligent assistant moving beyond code generation to handle complex engineering tasks like planning, testing, and interacting with user interfaces.
00:27:04.880 --> 00:27:08.000
So, a couple of reasons why I think just the model is important.
00:27:08.160 --> 00:27:19.920
Well, number one, their last model that they up uh that they came out with that I believe was based on uh Kimmy, so it was just a fine-tuned or built on top of an open source model.
00:27:20.000 --> 00:27:23.759
So Composer 2.5 was actually a really good model.
00:27:23.920 --> 00:27:28.400
Uh, it was the first time that Cursor, I think, had a model that a lot of people were starting to consider.
00:27:28.559 --> 00:27:32.880
And as we start to shift into the super app phase, all right.
00:27:32.960 --> 00:27:40.480
So make sure you go listen to episode 799 where we talked about AI super apps, and I did mention Cursor is a player uh now.
00:27:40.799 --> 00:27:43.839
Um, you like you have to look at the cost efficiency.
00:27:44.000 --> 00:27:57.200
So, you know, essentially with Composer 2.5, you got about maybe 90% of the power that you would get from something like Opus uh 4.8 at like 10% of the cost.
00:27:57.440 --> 00:28:22.240
Uh so with cursor going from scratch and now having that relationship with XAI, Grok, SpaceX, uh, I do think we have to start paying a little bit of attention, a little bit more attention uh to cursor, especially because I do think on the harness side, uh it is 1A and 1B in terms of I do think still, uh Codex on the Harness side, 1A and Cursor 1B.
00:28:22.400 --> 00:28:27.359
They have a great uh harness in the um in the cursor platform.
00:28:27.920 --> 00:28:38.559
So uh cursor's CEO Michael Truel emphasized the scale and fresh training approach, allows for better control over model behavior, and supports a wider range of engineering workloads.
00:28:38.720 --> 00:28:57.839
The other big news is they did announce their origin feature, a new Git native code hosting platform built for AI agents, which aims to challenge Microsoft's GitHub by supporting thousands of automated code operations daily in featuring automated conflict uh resolution.
00:28:58.000 --> 00:29:17.039
So cursor's acquisition gives it access to SpaceX's vast computational resources, with analysts saying the combination of Cursor's reinforcement learning expertise and SpaceX's compute power could quickly make Cursor a top competitor to Claude Code, Codex, and even GitHub.
00:29:17.200 --> 00:29:30.640
Uh so the GitHub play is interesting because we saw reports about a year and a half ago that OpenAI was developing their own uh GitHub uh kind of um competitor, but it never came to fruition.
00:29:30.720 --> 00:29:34.240
So maybe that was one of those you know side quests that was tabled.
00:29:34.319 --> 00:29:38.880
Uh so cursor maybe could start putting itself into the conversation.
00:29:38.960 --> 00:29:45.599
We'll see if we start shifting from talking about the big four of anthropic Google, open AI, and Microsoft.
00:29:45.680 --> 00:29:50.319
And maybe Cursor might thrust itself up into that first tier.
00:29:50.640 --> 00:29:55.279
All right, our last big piece of AI news.
00:29:55.519 --> 00:30:02.240
Yeah, it seems GPT-5.6 from OpenAI could be days away.
00:30:02.480 --> 00:30:15.759
So, according to reports, leaks, sleuths online, all that stuff, open AI is preparing to release its new GPT-5.6 family of AI models as soon as next week.
00:30:15.920 --> 00:30:31.359
Uh, so according to reports and leaks, uh, you might see a couple different variations of the GPT 5.6 model, including the standard, a new mini, and pro versions that could debut together.
00:30:31.519 --> 00:30:46.720
Uh, early access to GPT-5 Pro has reportedly already reached some pro subscribers who report noticeably noticeably better understanding on technical prompts, uh, though certain web development issues remain unresolved.
00:30:46.799 --> 00:30:54.079
So, yeah, uh, if you've been on you know Twitter at all over the last like five days, you'll see a lot of people.
00:30:54.160 --> 00:30:59.759
And I need to do a better job of you know coming up with these uh use cases that I actually save.
00:31:00.000 --> 00:31:15.599
The thing is, I try to not take on as many of the visual use cases because I think there's plenty of people out there, and I don't think that for the most part, these are things aside from you know creating things like decks, PowerPoints, um, you know, Word documents, you know, business dashboards.
00:31:15.680 --> 00:31:18.799
I think a lot of those um are easy to compare.
00:31:18.960 --> 00:31:28.000
But I think you know, what a lot of people are doing is just like you know, 3GS, uh, you know, 3D worlds, uh, you know, coding games.
00:31:28.160 --> 00:31:40.400
I don't think for the majority of our audience that's uh uh something that people are excited about, but that's why I think we've seen all these reports that apparently OpenAI maybe testing, A B testing uh GVD5.
00:31:44.880 --> 00:31:50.799
You know, create an interactive world model that you can walk through, you know, just using 3GS or something like that.
00:31:50.960 --> 00:31:56.240
And people have those kind of saved, and all of a sudden it's wow, like three, four times better.
00:31:56.400 --> 00:31:59.039
Uh, so a couple more details.
00:31:59.200 --> 00:32:13.839
The new GPT-5.6 is rumored to push the context window to 1.5 million tokens up from 1 million tokens in GPT-5.5, allowing for significantly longer and more complex conversations and coding sessions.
00:32:14.000 --> 00:32:30.319
So, developers who think or believe that they have access to GPT-5.6 Pro say that it outperforms anthropic's mythos models in agentic coding tasks, raising the stakes in the competition for the AI-powered software development.
00:32:30.720 --> 00:32:35.359
Uh, so uh make sure to keep an eye on the newsletter uh this week.
00:32:35.440 --> 00:32:51.519
And obviously, if it does come out, you better believe we will be having a show because there's a good chance that uh maybe as soon as today, maybe by the time that you uh you know hear this podcast, that we might have Fable Five from Anthropic back.
00:32:51.839 --> 00:32:57.039
Uh again, according to reports, we could have GVD5.6 as soon as this week.
00:32:57.200 --> 00:33:01.359
And we did earlier uh from Google themselves at their I.O.
00:33:01.599 --> 00:33:03.279
conference last month.
00:33:03.359 --> 00:33:11.759
Uh, they did say that next month, which would be this month in June, they did say that they're going to be releasing their Gemini 3.5 Pro model.
00:33:12.079 --> 00:33:19.519
So, in theory, this week we could get access or get access again to three state-of-the-art models.
00:33:19.599 --> 00:33:21.119
So, uh pretty big week.
00:33:21.200 --> 00:33:32.400
Uh, buckle in, make sure you got your tokens ready or your API uh budgets in check because there could be a lot of new model access rolling down the pipeline.
00:33:32.559 --> 00:33:36.160
All right, so that is it for our main AI news stories.
00:33:36.319 --> 00:33:38.480
So let's quickly go over everything else.
00:33:38.640 --> 00:33:50.480
This is the what's new and what's next, just stories that didn't make our top cut, some rumors, uh, some leaks, uh, but let's go ahead and go through them quickly, bullet point style.
00:33:51.359 --> 00:33:52.480
All right, here we go.
00:33:52.640 --> 00:34:09.280
So Amazon, Nvidia, and AMD invested$310 million in Odyssey ML to advance real-world AI understanding, anthropic upgraded claw and design, adding custom design systems, code sync, and export features.
00:34:09.519 --> 00:34:10.800
That one really good.
00:34:10.960 --> 00:34:11.840
FYI.
00:34:12.000 --> 00:34:19.039
OpenAI is reportedly preparing to integrate the library feature from ChatGPT into Codex.
00:34:19.199 --> 00:34:20.000
I would like that.
00:34:20.079 --> 00:34:22.400
I spent a lot of time trying to find all those files.
00:34:22.480 --> 00:34:27.039
Uh, XAI released their Grok Imagine Video 1.5 model, looks pretty good.
00:34:27.280 --> 00:34:38.239
Adobe expanded their Firefly AI assistant to other uh tools in their stack, including Premiere, Illustrator, InDesign, and Frame IO with some new features.
00:34:38.400 --> 00:34:44.480
OpenAI released unified usage analytics and spend controls for ChatGPT Enterprise.
00:34:44.639 --> 00:34:53.519
Uh, some new leaks show that Anthropic is working on a schedules feature for its always-on upcoming Claude Conway agent.
00:34:53.760 --> 00:35:01.280
Google rolled out their Ask Ad Manager Beta, an AI agent for custom insights and troubleshooting for Google Ads.
00:35:01.840 --> 00:35:06.480
Uh, like I said, GBT5.6 and Gemini 3.5 are likely around the corner.
00:35:06.719 --> 00:35:12.960
Codex, this is the big one, released their record and replay workflow capture for Mac.
00:35:13.280 --> 00:35:15.119
Uh, so excluding some countries.
00:35:15.280 --> 00:35:18.719
And I do believe I'll have to check again at the uh the newsletter.
00:35:18.880 --> 00:35:25.119
Most of the times, either on Friday or Monday, I'll ask you guys what you want for our Wednesday show, which is a hands-on demo.
00:35:25.280 --> 00:35:31.280
And I do believe that the new Codex uh record and replay uh one, which I was surprised.
00:35:31.360 --> 00:35:39.440
I thought Claude Design was gonna win our poll, uh, but it seems like everyone wanted the new uh record and replay feature, which is really cool.
00:35:39.599 --> 00:35:50.239
Uh Google Gemini co-lead Noam Shazir left for Open AI after Google acquired or Aqua hired uh him and his group from Character AI.
00:35:50.320 --> 00:35:51.360
This is a pretty big one, right?
00:35:51.440 --> 00:35:57.519
So one of the original authors on the attention is all you need uh paper that led to the Transformer, right?
00:35:57.599 --> 00:36:05.920
So one of the OGs of AI leaving uh Google Gemini after they spent$2.7 billion to join OpenAI.
00:36:06.079 --> 00:36:15.599
Uh another big get for OpenAI this week, former White House AI advisor Dean Ball joined OpenAI to lead Frontier AI policy team.
00:36:15.679 --> 00:36:22.079
So some big hires for OpenAI after Anthropic landed uh Andre Carpathy a couple of weeks ago.
00:36:22.320 --> 00:36:33.199
Speaking of Anthropic, they released their Claude Code Artifacts, which allows teams to share live coding artifacts for collaborative and interactive sessions.
00:36:33.360 --> 00:36:35.760
So very similar to Codecs sites.
00:36:35.840 --> 00:36:38.079
Uh XAI, oh, already covered that one.
00:36:38.159 --> 00:36:40.480
Uh Copilot Cowork is generally available.
00:36:40.639 --> 00:36:43.360
Like I said, downside though is usage-based pricing.
00:36:43.440 --> 00:36:47.440
And last but not least, Google released the open knowledge format.
00:36:47.599 --> 00:36:54.639
So kind of an alternative, or what could in theory be an alternative uh for an uh agentic communication system.
00:36:54.800 --> 00:37:00.559
So instead of markdown files, you might have OKFs if a lot of people uh adapt this.
00:37:00.719 --> 00:37:01.840
So that's it.
00:37:02.000 --> 00:37:04.960
That is a wrap for the AI news that matter.
00:37:05.039 --> 00:37:11.039
So, y'all, I'm telling you, this is happening so fast, and the pace is only going to pick up.
00:37:11.199 --> 00:37:11.599
All right.
00:37:11.679 --> 00:37:18.400
So you probably shouldn't be spending like four or five hours every single day, right, taking all this work home for you.
00:37:18.480 --> 00:37:19.360
That's what I do.
00:37:19.519 --> 00:37:25.199
So make sure I try to keep these podcasts 30-ish minutes, uh, right, even faster if you listen on 2x.
00:37:25.280 --> 00:37:29.599
So make sure you subscribe to the podcast, but you really need to be reading our newsletter.
00:37:29.760 --> 00:37:32.880
So make sure you go do that at your everydayai.com.
00:37:32.960 --> 00:37:35.280
So thanks for tuning in and FYI.
00:37:35.440 --> 00:37:38.639
If you are new here, Mondays we do our AI news that matters.
00:37:38.800 --> 00:37:44.079
Wednesdays, we go hands-on with a kind of live demo with our AI at work on Wednesdays.
00:37:44.239 --> 00:37:49.039
Friday, we go over Friday features, which are new features that you can actually go use today.
00:37:49.119 --> 00:37:50.880
And I kind of give you the 101.
00:37:50.960 --> 00:37:53.760
And then Tuesdays and Thursdays, we rotate what those shows are.
00:37:53.840 --> 00:37:56.880
So if you knew if you're new here, that's kind of our weekly lineup.
00:37:56.960 --> 00:37:57.920
So thanks for tuning in.
00:37:58.079 --> 00:38:01.119
Hope to see you back tomorrow in every day for more everyday AI.
00:38:01.199 --> 00:38:02.239
Thanks, y'all.
