WEBVTT
00:00:02.080 --> 00:00:05.919
There was no Blockbuster AI story this past week.
00:00:06.240 --> 00:00:10.640
No landmark AI court case or chart breaking AI model.
00:00:10.880 --> 00:00:17.120
Although there were plenty of new updates, developments, and stories that you need to know for your company to keep pace.
00:00:17.359 --> 00:00:22.719
This was one of those rare quiet weeks in AI news and developments.
00:00:22.960 --> 00:00:25.199
You know what that means, though, right?
00:00:25.519 --> 00:00:30.000
We're likely in for a doozy of new news this week.
00:00:30.160 --> 00:00:33.679
Like a sharkshooter that never misses twice from the three.
00:00:33.759 --> 00:00:38.880
I'm guessing we're going to be in for a big week of AI updates this coming week.
00:00:39.039 --> 00:00:43.200
Regardless, we did get enough to keep up with the past seven days.
00:00:43.359 --> 00:00:58.960
I mean, we have the world's new best open model, some juicy AI lawsuits, and the Microsoft CEO calling out not just one of his biggest competitors by name, but also one of the companies that he's invested five billion dollars in.
00:00:59.119 --> 00:01:02.560
Yeah, still plenty you need to know to keep up.
00:01:02.640 --> 00:01:06.719
And that's what we're gonna do on today's AI News That Matters.
00:01:06.879 --> 00:01:07.680
What's going on?
00:01:07.840 --> 00:01:09.359
Welcome to Everyday AI.
00:01:09.439 --> 00:01:12.239
If you're new here, my name's Jordan Wilson, and we do this every day.
00:01:12.319 --> 00:01:20.799
It's your daily live stream podcast and free daily newsletter helping business leaders like you and me keep up with the AI developments, use that information to grow your company and career.
00:01:21.040 --> 00:01:24.079
So if that's what you're trying to do, sweet, me too.
00:01:24.239 --> 00:01:31.840
Uh, it starts here with the unedited, unscripted live stream podcast, but please make sure you go to our website at your everydayai.com.
00:01:32.000 --> 00:01:37.920
We're going to be keeping you up to date with the highlights from today's story and everything else that's happening in AI world.
00:01:38.159 --> 00:01:41.920
So let's just jump straight into it.
00:01:42.079 --> 00:01:58.000
First, Meta, after having an amazing July so far with their new Muse Spark 1.1 that we talked about last week, their new Muse uh image and video models, well, they got uh maybe a little bit of a black eye this week.
00:01:58.079 --> 00:02:14.080
That's because a group of 26 uh current and previous meta employees has filed a federal lawsuit claiming the company's use of AI in layoff decisions on fairly targeted workers on medical, parental, or family leave.
00:02:14.639 --> 00:02:28.960
So, according to the Associated Press, the lawsuit was filed in Oakland and involves 26 anonymous meta employees who say they were selected for layoffs after taking protected leave or requesting disability accommodations.
00:02:29.199 --> 00:02:37.599
So the employees are among the 8,000 workers or about 10% of Meta's workforce who were notified of layoffs in May.
00:02:37.759 --> 00:02:54.240
So the lawsuit alleges that Meta used internal AI systems, keystroke in activity monitoring, uh, in algo performance rankings and AI token usage dashboards uh dashboards to determine who would be laid off.
00:02:54.400 --> 00:03:06.319
So the plaintiffs claim that these AI-powered systems penalized employees who were on protective leave, as their output was automatically reduced, making them more likely to be selected for termination.
00:03:06.560 --> 00:03:14.800
Many of the affected workers were on pregnancy or parental leave, while others took medical or family caregiving lead.
00:03:14.960 --> 00:03:23.199
Eight are women who had been who had taken maternity or pregnancy related leaves, and four are men who took parental leave.
00:03:23.520 --> 00:03:34.800
So the lawsuit argues that meta systems failed to account for legally protected absences and did not pause for individualized review as required by law.
00:03:34.960 --> 00:03:46.159
The complaint cites violations of several state and federal laws, including the Family and Medical Leave Act, the Americans with Disabilities Act, and the Pregnancy Discrimination Act.
00:03:46.319 --> 00:03:57.919
So Meta responded saying that those claims lacked merit and are not based on facts, and stated that workforce decisions were made by people, not AI.
00:03:59.120 --> 00:04:13.199
So, not exactly the week that Meta wanted to follow up with after their hot start to July that saw their stock really just rebound from like a year of lackluster performance and losses.
00:04:13.439 --> 00:04:20.879
So uh Meta honestly had the best start, I would say, uh to July outside of Open AI, right?
00:04:20.959 --> 00:04:29.360
If you look at kind of the the big six of AI, uh right, which is the big four model makers, uh that's open AI, anthropic Google, Microsoft.
00:04:29.439 --> 00:04:33.040
And then you put in Meta and SpaceX into the equation, right?
00:04:33.120 --> 00:04:40.319
If I talk about the big six in AI, uh right, it's kind of the big four plus Meta and uh SpaceX.
00:04:40.720 --> 00:04:47.439
So Meta actually had one of the best Julies, you know, so far, uh, until this happened.
00:04:47.600 --> 00:04:48.879
So not a good look.
00:04:48.959 --> 00:04:51.920
Uh, this one's gonna be interesting to see how it plays off.
00:04:52.000 --> 00:05:00.079
And if nothing else, a good takeaway for business leaders here is you always have to make sure that your systems are talking to each other, right?
00:05:00.240 --> 00:05:04.480
If this is truly the case, this is absolutely horrible, right?
00:05:04.560 --> 00:05:22.560
If women that were on pregnancy leave uh were, you know, essentially um decided and AI decided that they were going to be up for termination because they weren't hitting their keystrokes or their, you know, token, you know, targets, which is absolutely asinine if this is true.
00:05:22.720 --> 00:05:23.600
I hope it's not.
00:05:23.680 --> 00:05:28.319
Uh, but regardless, when it does break one way or the other, we'll let you know in the newsletter.
00:05:28.399 --> 00:05:31.680
I'm guessing this will eventually just be settled out of court.
00:05:32.160 --> 00:05:32.800
All right.
00:05:32.959 --> 00:05:38.480
Uh, our next one, which could end up being a big joint agreement for U.S.
00:05:38.639 --> 00:05:41.839
tech companies, which you don't really see a lot.
00:05:42.079 --> 00:05:58.160
Uh, that's because the Google Deep Mine CEO is pushing for a unified AI watchdog after the anthropic ban has led to kind of this staggered and what I'm calling the permission slip AI reality that we're all living in.
00:05:58.399 --> 00:06:18.240
So, Google Deep Mine CEO Demis Hasabas is calling for a FINRA-style industry-funded watchdog to oversee advanced AI, uh, aiming to prevent the chaotic kind of government crackdowns like the one that halted Anthropics Fable 5 and Mythos 5 models in June.
00:06:18.399 --> 00:06:36.480
So, the Trump administration's abrupt ban on those vital models when they were citing national security risks triggered a tense two and a half week negotiation and forced last-minute changes, exposing the lack of a formal review process for powerful AI systems.
00:06:36.800 --> 00:06:43.360
So, this is what uh Demis's proposed body would require.
00:06:43.519 --> 00:06:56.240
So, it would require uh AI labs to submit their frontier models for pre-release vetting up to 30 days before launch, focusing on threats like cyber warfare, bioweapons, and nuclear misuse.
00:06:56.399 --> 00:07:06.000
So, the framework would also include emergency powers to slow or halt industry-wide AI development if a model is found to cross critical risk threshold.
00:07:06.399 --> 00:07:16.800
So, the reason why this is important because this is the first uh kind of initiative that's picked up steam across all of, well, not all of the big six, but I think five of them.
00:07:17.040 --> 00:07:32.240
That's because major tech leaders, including Microsoft's Satya Nadella, OpenAI's Sam Altman, Google's Sindar Pichai, and others have publicly endorsed the core principles of the proposal, marking a rare show of unity in the competitive sector.
00:07:32.399 --> 00:07:42.720
Uh, Inthropics, Dario Modi supports binding regulation, but has called for an even stricter government-led agency with authority to block unsafe model releases outright.
00:07:42.879 --> 00:07:54.480
So, Meta, led by Mark Zuckerberg, remains the notable holdout here, continuing to advocate for open source AI in drafting its own regulatory response.
00:07:54.639 --> 00:08:17.360
That one there is pretty uh, I don't want to give it a certain name, but yeah, you know, the fact that Meta cited open source AI being an advocate for, even though they are the company that has famously shifted away from open models in their previous uh llama series to now the Mews, which are proprietary closed source.
00:08:17.519 --> 00:08:22.800
So I don't really know why Meta's response was was, well, hey, we care more about open source.
00:08:22.959 --> 00:08:40.080
Anyways, all of the other major US labs have already signed voluntary agreements to submit their models for government safety testing, but the industry, the industry wants a predictable, standardized process instead of the ad hoc intervention that we've seen.
00:08:40.320 --> 00:08:50.000
So the White House is currently reviewing the proposal as a potential compromise to keep America competitive in AI while managing security risks.
00:08:50.159 --> 00:08:54.879
So if you didn't see this, we did share this in our newsletter last week.
00:08:55.120 --> 00:08:57.440
I think this is sorely needed.
00:08:57.519 --> 00:09:10.080
Uh, presumably, you would have probably liked this uh to come from the government and not necessarily a private company, but maybe that's what might make this stick, right?
00:09:10.240 --> 00:09:20.080
Similar to the FINRA style, um, you know, having a private to public uh sector agreement that says, hey, we're essentially the experts in AI.
00:09:20.399 --> 00:09:32.000
You probably don't know, government, uh, exactly what you should be doing when it comes to regulating our industry, which I absolutely agree with.
00:09:32.240 --> 00:09:46.799
Um that's not me saying that there's not smart people in AI in the government, there's obviously absolutely brilliant people uh, you know, who have both backgrounds um at these big six tech companies working in government now.
00:09:46.879 --> 00:09:55.440
But for the most part, when it comes to like the bleeding edge technology, I don't think the White House really has a clue.
00:09:55.519 --> 00:09:56.799
And that's not a political statement.
00:09:57.120 --> 00:09:58.480
That's just any government.
00:09:58.639 --> 00:10:04.720
I don't think any government would have the clue how to regulate the United States acceleration in AI right now.
00:10:04.799 --> 00:10:15.759
Uh right, but that's because you know, 60 days ago, you know, six months ago, the capabilities that we have today, no one probably could have predicted them, especially six months ago.
00:10:16.000 --> 00:10:24.799
So it is good uh that you know Google DeepMinds uh Demis Hossibus is kind of calling for this and taking uh the step forward.
00:10:24.960 --> 00:10:33.120
Uh Demis obviously has one of the deepest and um most respected backgrounds in AI in the entire world.
00:10:33.279 --> 00:10:36.720
Uh, so we'll see if Sir Demis can get it done.
00:10:37.039 --> 00:10:50.159
All right, uh moving from one uh kind of compromise, I guess, of US big tech companies uh to a leak that could maybe impact millions of people.
00:10:50.320 --> 00:10:57.360
That's because we just got a report from Bloomberg that says OpenAI is working on its first hardware device.
00:10:57.519 --> 00:11:07.840
So finally, we get some details on what this uh, you know, much reported on device, you know, we've been hearing about this thing for uh almost nine or ten months now.
00:11:08.000 --> 00:11:16.399
So according to Bloomberg Reports, it is a mobile smart speaker that integrates AI and syncs with Chat GPT.
00:11:16.960 --> 00:11:29.759
So, according to reports, the device is designed to be a screen-free device and it acts as a human-like AI companion for the home, offering proactive learning and a personalized service.
00:11:30.320 --> 00:11:47.679
So, unlike traditional smart speakers, OpenAI's model reportedly features mechanical elements that can move on their own, aiming to feel uh more like a living companion or one that can at least kind of wander maybe from room to room with you.
00:11:47.840 --> 00:11:55.039
So the speaker is intended to access a user's digital life, including emails to deliver highly tailored assistance.
00:11:55.279 --> 00:12:03.519
So, development is being led by former Apple engineers who played key roles in creating products like the iPhone and Mac.
00:12:03.840 --> 00:12:13.519
So the project comes amid legal friction as Apple recently sued OpenAI, accusing it of stealing trade secrets and warning of more revelations during discovery.
00:12:13.679 --> 00:12:16.320
OpenAI denied any wrongdoings on that.
00:12:16.559 --> 00:12:26.240
Sources told Bloomberg the new device veers significantly from anything Apple has on the market today and is unlikely to violate trade secrets.
00:12:26.480 --> 00:12:39.279
So OpenAI's push reflects the growing industry excitement about consumer AI hardware, with startups like Hark raising around$700 million to build personal intelligence devices.
00:12:39.519 --> 00:12:50.399
So the reported OpenAI devices form factor hasn't been revealed, highlighting how much investor interest exists before these products are even unveiled.
00:12:50.960 --> 00:12:55.679
So I'm all for anyone that's listening to the show for a long time.
00:12:55.840 --> 00:12:57.440
I don't really care about my data.
00:12:57.600 --> 00:12:58.559
I don't, right?
00:12:58.639 --> 00:13:09.039
If you want to know all the random AI stuff I've been searching, and uh if an open AI smart speaker wants to follow me around and help me with that, sweet, sounds good, right?
00:13:09.200 --> 00:13:15.919
Anything to get rid of uh Alexa and Siri sooner, uh the better.
00:13:16.080 --> 00:13:17.519
Like, please sign me up for this.
00:13:17.600 --> 00:13:19.120
Put me on a wait list, all right.
00:13:19.279 --> 00:13:23.200
Uh, next, you don't have to get on a wait list because this is out.
00:13:23.759 --> 00:13:30.399
This is the new king of the open, well, soon open source hill.
00:13:30.480 --> 00:13:37.519
All right, more on that here in a second, but we do have a new open source king in Kimmy K3.
00:13:38.159 --> 00:13:54.000
So Moonshot AI, a Chinese startup backed by Alibaba and Tencents, uh, introduced Kimi K3, a absolutely massively monster open source model with nearly three trillion parameters.
00:13:54.159 --> 00:13:57.600
Yes, three almost three trillion parameters.
00:13:57.759 --> 00:14:08.240
So Moonshots revealed Kimi 3 at the World AI conference in Shanghai, touting it as the world's first open source model in the 3 trillion parameter class.
00:14:08.399 --> 00:14:18.159
So Kimi K3 is not technically open source yet, but the company says they will release the weights on July 27th.
00:14:18.399 --> 00:14:30.720
So at 2.8 trillion parameters, Kimi K3 puts it in direct competition with leading American proprietary models from OpenAI and Anthropic.
00:14:31.279 --> 00:15:00.159
So the announcement comes just weeks after the US temporarily forced Anthropic to withdraw its its mythos um and Fable 5 models due to cybersecurity concern, highlighting the growing not just national security focus on advanced AI, but also the interest on potential open source AI projects that can live in the frontier or like frontier point five sector.
00:15:00.559 --> 00:15:01.120
Uh, right.
00:15:01.200 --> 00:15:12.960
So, and this is why it's actually important because third-party evaluations by artificial analysis and arena show Kimi K3 performing uh on par-ish, right?
00:15:13.200 --> 00:15:18.480
Uh with uh uh GPD 5.6 soul and fable five.
00:15:18.559 --> 00:15:21.919
So in some benchmarks, they're a little bit better, actually.
00:15:22.159 --> 00:15:35.759
But right now I would say that you have uh GPT 5.6 soul and fable, they are class 1A, and now well, you might have an open source model in Kimi K3, technically 1B.
00:15:35.919 --> 00:15:40.000
So at least right now, I wouldn't even say it's the second tier.
00:15:40.080 --> 00:15:48.720
It's you know, we have a tie for 1A, and yeah, we have a Chinese open source model that has closed the gap uh to be 1B.
00:15:49.120 --> 00:16:03.360
So Kimi K3's open source approach stands in contrast to obviously the closed proprietary systems of American firms, throwing a wrench in the AI strategy game plan for some larger enterprises.
00:16:03.519 --> 00:16:11.200
So the model itself is designed to operate with minimal human supervision, handling complex tasks like engineering and software development.
00:16:11.360 --> 00:16:24.000
The release immediately impacted global markets with shares in Chinese competitors ZAI in Minimax dropping sharply in Hong Kong after the announcement.
00:16:24.320 --> 00:16:26.879
So I'll cut it to you straight on this.
00:16:26.960 --> 00:16:33.759
Uh, because I think people, for the most part, when they look at open source models, they usually think of two things.
00:16:33.919 --> 00:16:38.399
They think, well, if I have a powerful enough computer, right?
00:16:38.480 --> 00:16:43.679
Uh, like as an example, I have a very beefy Mac Studio uh sitting here on my desktop.
00:16:43.840 --> 00:16:49.679
People think, oh, if I have very expensive hardware, I can run the best open source models.
00:16:50.000 --> 00:16:51.679
No, we are past that.
00:16:51.919 --> 00:16:59.840
Unless you literally have a rack of NVIDIA GPUs and a cooling uh center in your basement, you can't run this, right?
00:16:59.919 --> 00:17:02.639
This is for 0% of consumers out there.
00:17:02.879 --> 00:17:08.079
You can't run a 3 trillion parameter model on any consumer hardware.
00:17:08.240 --> 00:17:09.759
So that's number one.
00:17:10.000 --> 00:17:15.200
So, yes, this is really for those enterprises that have access to compute at scale.
00:17:15.359 --> 00:17:40.880
Number one, especially if you can host it, which is not a lot of companies, but if you do have a couple of racks, uh and maybe you have a large AI bill, you know, maybe you're paying uh eight, nine figures in a ice fan with, you know, a bunch of GPUs, which again, not a lot of companies, but for some companies, yes, in uh this week or sorry, in like 10 days or so, uh, on July 27th.
00:17:40.960 --> 00:17:49.920
So that's like in seven days, um, you'll technically be able to download this model and if you want to fine-tune it because it's open source.
00:17:50.000 --> 00:18:03.359
Um, so for a very, very, very, very few companies, this is actually could be fantastic news if you have uh the resources and uh the the human uh skill needed to make that happen.
00:18:03.599 --> 00:18:04.640
That's one thing people think.
00:18:04.799 --> 00:18:06.079
Oh, I can just download this and run it.
00:18:06.240 --> 00:18:06.799
No, you can't.
00:18:06.960 --> 00:18:09.599
Number two, they think it's extremely cheap.
00:18:10.240 --> 00:18:11.440
Also not the case, right?
00:18:11.519 --> 00:18:19.200
So if you're not using this, you know, downloading it and running it locally, people think, well, oh, open source models are super cheap.
00:18:19.519 --> 00:18:21.519
Previously, yeah, they were.
00:18:22.160 --> 00:18:24.079
Not really with this one, right?
00:18:24.240 --> 00:18:26.559
Um, we we we see this Kimi K3.
00:18:26.880 --> 00:18:35.759
Uh, we also saw uh that Quen will be releasing their next big multiple trillion parameter model uh here any minute or any hour.
00:18:36.000 --> 00:18:45.359
So I don't think this next tier of frontier-esque open models are gonna be open in the sense that we thought of them previously.
00:18:45.519 --> 00:18:46.079
Um, right.
00:18:46.160 --> 00:18:53.359
And the other thing is, there's no real guarantee that we keep getting these open source Chinese models for much longer.
00:18:53.599 --> 00:19:03.839
We saw reports, which we talked about on last week's AI News That Matters, that China may start uh restricting the export of their models, right?
00:19:04.000 --> 00:19:16.640
Uh, specifically not wanting uh maybe the US to get a hold of these, which is obviously ironic considering that these models have largely just been distilled from US models.
00:19:16.720 --> 00:19:19.200
Uh so regardless, there's a little bit of background.
00:19:19.359 --> 00:19:20.559
You can't go download this.
00:19:20.640 --> 00:19:24.880
And if you're paying API prices, it's not like they're super cheap.
00:19:24.960 --> 00:19:31.680
Uh, right, they're actually in many cases more expensive than GPT 5.6 uh Seoul.
00:19:31.839 --> 00:19:41.519
Uh, you know, they're probably a little cheaper uh than Fable 5 and Mythos 5, since those are the most expensive models on the market when it comes to price per performance.
00:19:41.839 --> 00:19:42.480
All right.
00:19:42.960 --> 00:19:45.920
More model news, but this is not we get one.
00:19:46.000 --> 00:19:47.839
This is just we're waiting longer.
00:19:48.000 --> 00:20:04.880
So Alphabet, parent company of Google, its shares fell 4% Thursday after news broke that the highly anticipated Google Gemini 3.5 Pro, yeah, months behind schedule, not coming out anytime soon.
00:20:05.039 --> 00:20:17.279
So, according to a new Bloomberg report, the delay is due to Google's efforts to improve the model's coding performance, which currently falls short according to their own internal expectations.
00:20:17.599 --> 00:20:31.039
So the setback comes as competitors like OpenAI, Anthropic, and uh Meta have all launched AI models recently that outperform Google's current offerings in software code generation.
00:20:31.200 --> 00:20:39.920
So Alphabet first teased Gemini 3.5 Pro in May after releasing Gemini 3.5 Flash.
00:20:40.160 --> 00:20:41.920
So in May at their I.O.
00:20:42.160 --> 00:20:45.920
conference, they said that the model would be coming out in June.
00:20:46.160 --> 00:20:50.960
Obviously, June has already passed and we're near the end of July.
00:20:51.119 --> 00:20:58.480
And now the latest report says Google's next model in Gemini 3.5 Pro could still be months away.
00:20:58.640 --> 00:21:14.799
So an alphabet spokesperson told CNBC that the company is shipping quickly across a wide range of models and continues to test Gemini 3.5 Pro and other upgrades with partners and the US government.
00:21:15.759 --> 00:21:30.880
Uh so yeah, if you were, you know, holding your breath or if you were you know trying to decide between GPT-5.6 soul and fable five, and you're like, wait, let's just wait until you know Google Gemini 3.5 Pro because it was supposed to come in June.
00:21:31.039 --> 00:21:32.319
Surely it's coming in July.
00:21:32.480 --> 00:21:33.279
Well, no.
00:21:33.519 --> 00:21:42.559
And if this report holds to be true, we might be heading into the fall without even Gemini 3.5 Pro.
00:21:42.880 --> 00:21:49.440
And at that point, I mean, presumably, we're gonna be getting a Fable 5.1 soon.
00:21:49.519 --> 00:21:54.319
Uh, we may be getting a GPT uh six around the horizon.
00:21:54.480 --> 00:21:55.200
We'll see.
00:21:55.359 --> 00:21:59.519
So, you know, it's it's gonna um be worth following.
00:22:00.160 --> 00:22:11.839
To see just how big of a jump this 3.5 Pro could be, uh, or maybe I don't know, uh, if Google comes out with another flash model in the interim.
00:22:12.000 --> 00:22:18.480
Because if they're still working on the same pre-train that they were previously, that can might be concerning.
00:22:18.640 --> 00:22:40.240
And I mean, we'll see if for the first time in a long while, at least two plus years, if Google is not able to compete on the model front, which to me, I would be shocked because I've always been a big believer in that Google at any point, whenever they want, can drop the world's most powerful model.
00:22:40.640 --> 00:22:50.160
So I don't know, maybe when I said that a year or so ago, maybe I was biting off more than I could chew and believing in uh the big G a little bit too much.
00:22:50.400 --> 00:22:53.599
So uh yeah, we'll we'll see if that holds true.
00:22:53.759 --> 00:22:55.519
I still think they'll be able to.
00:22:55.599 --> 00:23:01.359
Uh you know, Google it's they have their own, right?
00:23:01.599 --> 00:23:20.559
They have their own uh the you know GPUs, TPUs, they can they can train these things quickly, uh, without having to, you know, pay their competitors billions of dollars, like Anthropic is having to do right now, uh, for inference and to train these models.
00:23:20.799 --> 00:23:22.160
So, all right.
00:23:22.319 --> 00:23:32.720
Uh speaking of big tech companies, we're moving from Google to Microsoft for our last big AI news story of the week.
00:23:33.039 --> 00:23:35.119
That is the juicy one.
00:23:35.200 --> 00:23:35.839
Here we go.
00:23:36.000 --> 00:23:58.079
So, Microsoft CEO Satya Nadella recently, according to reports, started to kind of seemingly shift his public stance on mainstream AI, moving from an aggressive investor in closed source frontier labs to a vocal critic of their current trajectory.
00:23:58.240 --> 00:24:05.839
So the shift began with Nadella's leaked criticism of Anthropics Claude Fable 5.
00:24:06.079 --> 00:24:23.519
So, according to reports, Nadella slammed the flagship model as being too quote unquote editorially controlled, arguing that its rigid guardrails and refusal protocols don't make sense for a creative utility.
00:24:23.759 --> 00:24:40.000
So, reports also said that Nadella questioned Anthropics limits on user requests to its high-end fable AI model, calling the editorial controls on creation tools nonsensical during remarks to Microsoft's co-pilot engineers.
00:24:40.160 --> 00:24:54.240
So, Anthropic's Fable model sometimes rejects users' queries or responds with older versions of the model, with essentially falling back to Opus 4.8, leading to a lot of frustration and criticism on social media, right?
00:24:54.400 --> 00:25:07.920
You've seen many examples, and I've tried some myself, even asking Fable 5 about certain types of cancers, or you know, hey, what are the biggest uh you know, in certain medical fields, what are the biggest unanswered questions?
00:25:08.079 --> 00:25:16.319
And it will say essentially, you can't ask this, and it's gonna roll you back uh to an older model like Opus 4.8.
00:25:16.559 --> 00:25:26.960
So this is extremely noteworthy as Satya Nadella is uh, at least according to these reports, publicly criticizing Anthropic.
00:25:27.359 --> 00:25:44.559
Um uh that's because Microsoft has invested five billion dollars in Anthropic, uh, with Anthropic agreeing to also spend$30 billion in future years on Microsoft's Azure cloud infrastructure.
00:25:44.880 --> 00:26:08.799
So this is uh right, you normally wouldn't hear shots fired like this, uh, you know, slamming a company in their flagship model, especially when said company is well gonna be reportedly paying you$30 billion over the next few years, and you have a large ownership in that company.
00:26:09.200 --> 00:26:16.880
But that wasn't the only uh kind of shot at anthropic that Nadella took.
00:26:17.119 --> 00:26:20.480
That one was again, according to reports, direct.
00:26:20.720 --> 00:26:26.400
The second one, a little indirect, but most industry uh insiders are connecting the dots.
00:26:26.640 --> 00:26:44.079
So, in a seemingly somewhat related long-form online post last week, Nadella coined the term reverse information paradox to explain that businesses essentially just pay for AI twice, they pay for it once with actual money, right?
00:26:44.240 --> 00:26:48.079
So they're buying subscriptions for their employees or paying for tokens.
00:26:48.240 --> 00:26:51.599
And then he said they're actually paying for it a second time.
00:26:51.759 --> 00:27:04.240
And the second time is more dangerous because they're giving away their proprietary knowledge and institutional know-how that they're essentially forced to hand over to make the model useful.
00:27:04.880 --> 00:27:18.720
So in his posts, Nadella didn't name names on who does this, but well, everyone's looking at the fine print here and being like, wait, the only company that actually forces you to do that right now is anthropic.
00:27:18.880 --> 00:27:29.599
Because if you're a at least for paying business customers, at least, uh, Anthropic is the only one that has a requirement that they're going to retain your data and there's nothing you can do about it, right?
00:27:29.759 --> 00:27:46.480
So with the new Mythos 5 and Fable 5, uh Anthropic essentially throws zero uh data retention or uh ZDR, you know, Anthropic's just throwing zero data retention out the window.
00:27:46.640 --> 00:27:57.599
Uh so again, although Nadella didn't mention anthropic by name, everyone's you know reading his posts and being like, okay, this is also definitely about Fable.
00:27:58.000 --> 00:28:04.880
So for all you people out there that you know sometimes say, Oh, Jordan's being so hard on Anthropic, no, I'm not.
00:28:05.039 --> 00:28:05.519
My gosh.
00:28:05.680 --> 00:28:20.640
Uh, like the fact that Anthropic still has that required uh data retention, again, I don't have five billion dollars invested in anthropic, and anthropic hasn't agreed to pay me 30 billion of dollars over the next couple of years.
00:28:20.799 --> 00:28:30.000
So if Saudi Nadella is number one, again, according to reports, calling them out by name in one uh you know, breath.
00:28:30.079 --> 00:28:38.720
And then in the next breath, he's essentially describing them because that's the only company that truly fits that description for enterprise customers.
00:28:38.960 --> 00:28:42.640
That's not a good sign for anthropic, just saying that.
00:28:43.359 --> 00:28:56.480
Uh, so yeah, yeah, many attributed that second half of the statement to Anthropic's Mythos 5 and Fable 5 data policy, which forces that data retention even for enterprise customers.
00:28:57.039 --> 00:29:04.079
All right, that's a wrap on our big stories, but let's quickly go into the what's new and what's next.
00:29:04.240 --> 00:29:09.839
So these are some still important happenings, uh, some maybe some leaks in here.
00:29:09.920 --> 00:29:14.319
Uh, but let's just quickly go into our what's new and what's next.
00:29:14.480 --> 00:29:15.119
Rapid fire.
00:29:15.200 --> 00:29:15.680
Here we go.
00:29:15.839 --> 00:29:22.960
So the Trump administration launched Gold Eagle Hub to accelerate cybersecurity patching for critical infrastructure.
00:29:23.200 --> 00:29:29.440
Meta reportedly entered talks to lease up to 10 billion dollars of compute to Anthropic.
00:29:29.599 --> 00:29:34.559
Google AI mode added actions through Canva, Instacart, and YouTube Music.
00:29:34.720 --> 00:29:36.880
Yeah, if you didn't check that out, uh, pretty cool.
00:29:36.960 --> 00:29:40.079
If you have a paid personal Gmail account, check that out.
00:29:40.240 --> 00:29:45.519
Uh, next, I referenced this earlier, but Quen 3.8 is launching soon.
00:29:45.680 --> 00:29:57.599
And the company says it is a 2.4 trillion parameter model, and they are reporting benchmarks near Fable 5 GPT 5.6 soul, Kimi K3 levels.
00:29:57.759 --> 00:30:04.480
All right, thinking Machines released a 975 billion parameter open weight inkling AI model.
00:30:04.720 --> 00:30:11.680
Uh, New York halted any data center construction more than 50 mag um MW.
00:30:12.079 --> 00:30:13.440
Uh, is that mega?
00:30:13.599 --> 00:30:15.359
I always forget that what what that is.
00:30:15.519 --> 00:30:18.720
MW is is mega mega hertz.
00:30:19.039 --> 00:30:19.680
Is that right?
00:30:19.839 --> 00:30:21.359
Mega, megawatt.
00:30:21.519 --> 00:30:22.480
My gosh, y'all.
00:30:22.640 --> 00:30:26.160
Sometimes I read too much AI news and I just forget things.
00:30:26.319 --> 00:30:26.640
All right.
00:30:26.799 --> 00:30:34.960
Anyways, New York halted uh construction for one year on any of those uh new data centers.
00:30:35.119 --> 00:30:44.079
Uh, next, more than 200 experts, including 15 Nobel laureates, called for urgent preparation for AI's economic impact.
00:30:44.319 --> 00:30:52.640
Anthropic permanently extended Fable 5 to max users, but yeah, it's only at about one-third of the original limit.
00:30:52.720 --> 00:30:55.920
So, yeah, uh, I talked about this on last week's show.
00:30:56.000 --> 00:31:03.599
I said there's no way Anthropic is actually taking away Fable 5 because they would lose most of their subscription revenue.
00:31:03.680 --> 00:31:10.160
Uh, so unfortunately for most people, it is only if you're on that$100 or$200 a month paid plan.
00:31:10.400 --> 00:31:18.799
Google renamed Notebook LM to Gemini Notebook and they added cloud code execution uh for paid users.
00:31:18.880 --> 00:31:23.759
I think that's only out uh to ultra users, but it will be coming out to pro users soon.
00:31:23.920 --> 00:31:28.640
So, yeah, let me know if we should do a Gemini Notebook uh episode when that drops.
00:31:28.720 --> 00:31:31.680
I'm gonna accidentally still call it Notebook LM all the time.
00:31:31.839 --> 00:31:41.200
Uh next, a hacker leaked Sunos secret data, uh, reportedly showing that millions of songs were scraped from YouTube and Deezer.
00:31:42.160 --> 00:31:42.799
All right.
00:31:43.119 --> 00:31:50.720
Uh, next, OpenAI released GPT Red, an automated system to detect prompt injection vulnerability in AI agents.
00:31:50.880 --> 00:31:55.599
Uh, Amazon Bedrock added support for OpenAI's new models, GPT5.6.
00:31:56.079 --> 00:32:09.680
Uh, Apple is reportedly testing Prism L, uh Prism L's uh AI model compression, which would essentially allow them to run 27 billion parameter models on iPhones, which would be pretty sweet.
00:32:09.839 --> 00:32:13.599
Please just fix the Siri Apple.
00:32:13.680 --> 00:32:15.039
If you're listening, thanks.
00:32:15.200 --> 00:32:22.400
Uh Anthropic added an in-app browser to Claude Code on desktop, enabling site in doc interaction.
00:32:22.480 --> 00:32:23.440
So yeah, I tried it out.
00:32:23.519 --> 00:32:24.079
It's pretty good.
00:32:24.240 --> 00:32:26.880
It's not quite as good as Codex's, but it does the job.
00:32:26.960 --> 00:32:32.640
Uh, next, Perplexity rolled out self-improving memory, faster models, and one-click website publishing.
00:32:32.799 --> 00:32:44.480
ChatGPT rolled out their universal search uh that helps you more easily find projects, images, documents, chats from one sidebar, a much richer and better search experience.
00:32:44.640 --> 00:32:57.359
Uh, email program superhuman released their new auto draft feature that writes email replies in your voice uh using better models, and then you can go send them from your drafts.
00:32:57.519 --> 00:33:04.880
Next, Spotify released a new feature where you can talk to Spotify by voice or text to control music and see listening history.
00:33:05.039 --> 00:33:12.960
Uh Gemini rolled out personal avatars to allow you to generate AI videos starring your own face and voice.
00:33:13.279 --> 00:33:19.920
And then last but not least, uh ChatGPT work, which I don't know if it's called just ChatGPT now.
00:33:20.079 --> 00:33:24.400
I noticed that in the dropdown that used to say ChatGPT work and codex.
00:33:24.559 --> 00:33:26.480
Now it just says ChatGPT and Codex.
00:33:26.640 --> 00:33:33.119
So I'm not sure if we're still calling it ChatGPT Work or not, but regardless, it actually makes it easier.
00:33:33.279 --> 00:33:41.119
That's funny, one of my uh AIs just uh shouted out at me, but it actually makes it much easier uh to show your conversation history.
00:33:41.279 --> 00:33:48.720
So essentially they brought your chats from Chat GPT in to a place that made more sense inside of the desktop app.
00:33:48.799 --> 00:33:51.839
Whereas before, you know, users were a little bit confused.
00:33:52.000 --> 00:33:55.440
Also, it brings your projects in, which a lot of users were complaining about.
00:33:55.599 --> 00:34:03.680
So essentially, if you tried out Chat GPT work when it came out uh about 10 days ago and were like, wait, some of this isn't adding up.
00:34:03.759 --> 00:34:05.200
Where's all my chat GPT stuff?
00:34:05.279 --> 00:34:06.319
Well, now it's all there.
00:34:06.480 --> 00:34:08.400
All right, so I hope this was helpful.
00:34:08.559 --> 00:34:17.840
All of the AI news that matters for this week, but like I said to start the show, I can almost guarantee we're gonna get a ton of new releases, right?
00:34:18.000 --> 00:34:27.360
Word on the street is we'll probably be getting an Opus V from Anthropic that could be near that GPT 5.6 or uh mythos level.
00:34:27.519 --> 00:34:31.440
Uh, you know, and who knows, it is still hot AI summer.
00:34:31.599 --> 00:34:37.360
And I think that once the steam starts to pick up, it is going to be a hot one.
00:34:37.599 --> 00:34:40.159
So uh I get it, it is hard.
00:34:40.239 --> 00:34:47.599
Uh, if you are a busy business leader and you're trying to implement AI in your company and you're like, how do I keep up?
00:34:47.760 --> 00:34:53.760
Well, it starts here with the daily podcast, but make sure you also go to our website at your everydayai.com.
00:34:53.840 --> 00:34:54.480
Thank you for tuning in.
00:34:54.639 --> 00:34:57.599
Hope to see you back tomorrow and every day for more everyday AI.
00:34:57.679 --> 00:34:58.320
Thanks, y'all.
00:00:02.080 --> 00:00:05.919
There was no Blockbuster AI story this past week.
00:00:06.240 --> 00:00:10.640
No landmark AI court case or chart breaking AI model.
00:00:10.880 --> 00:00:17.120
Although there were plenty of new updates, developments, and stories that you need to know for your company to keep pace.
00:00:17.359 --> 00:00:22.719
This was one of those rare quiet weeks in AI news and developments.
00:00:22.960 --> 00:00:25.199
You know what that means, though, right?
00:00:25.519 --> 00:00:30.000
We're likely in for a doozy of new news this week.
00:00:30.160 --> 00:00:33.679
Like a sharkshooter that never misses twice from the three.
00:00:33.759 --> 00:00:38.880
I'm guessing we're going to be in for a big week of AI updates this coming week.
00:00:39.039 --> 00:00:43.200
Regardless, we did get enough to keep up with the past seven days.
00:00:43.359 --> 00:00:58.960
I mean, we have the world's new best open model, some juicy AI lawsuits, and the Microsoft CEO calling out not just one of his biggest competitors by name, but also one of the companies that he's invested five billion dollars in.
00:00:59.119 --> 00:01:02.560
Yeah, still plenty you need to know to keep up.
00:01:02.640 --> 00:01:06.719
And that's what we're gonna do on today's AI News That Matters.
00:01:06.879 --> 00:01:07.680
What's going on?
00:01:07.840 --> 00:01:09.359
Welcome to Everyday AI.
00:01:09.439 --> 00:01:12.239
If you're new here, my name's Jordan Wilson, and we do this every day.
00:01:12.319 --> 00:01:20.799
It's your daily live stream podcast and free daily newsletter helping business leaders like you and me keep up with the AI developments, use that information to grow your company and career.
00:01:21.040 --> 00:01:24.079
So if that's what you're trying to do, sweet, me too.
00:01:24.239 --> 00:01:31.840
Uh, it starts here with the unedited, unscripted live stream podcast, but please make sure you go to our website at your everydayai.com.
00:01:32.000 --> 00:01:37.920
We're going to be keeping you up to date with the highlights from today's story and everything else that's happening in AI world.
00:01:38.159 --> 00:01:41.920
So let's just jump straight into it.
00:01:42.079 --> 00:01:58.000
First, Meta, after having an amazing July so far with their new Muse Spark 1.1 that we talked about last week, their new Muse uh image and video models, well, they got uh maybe a little bit of a black eye this week.
00:01:58.079 --> 00:02:14.080
That's because a group of 26 uh current and previous meta employees has filed a federal lawsuit claiming the company's use of AI in layoff decisions on fairly targeted workers on medical, parental, or family leave.
00:02:14.639 --> 00:02:28.960
So, according to the Associated Press, the lawsuit was filed in Oakland and involves 26 anonymous meta employees who say they were selected for layoffs after taking protected leave or requesting disability accommodations.
00:02:29.199 --> 00:02:37.599
So the employees are among the 8,000 workers or about 10% of Meta's workforce who were notified of layoffs in May.
00:02:37.759 --> 00:02:54.240
So the lawsuit alleges that Meta used internal AI systems, keystroke in activity monitoring, uh, in algo performance rankings and AI token usage dashboards uh dashboards to determine who would be laid off.
00:02:54.400 --> 00:03:06.319
So the plaintiffs claim that these AI-powered systems penalized employees who were on protective leave, as their output was automatically reduced, making them more likely to be selected for termination.
00:03:06.560 --> 00:03:14.800
Many of the affected workers were on pregnancy or parental leave, while others took medical or family caregiving lead.
00:03:14.960 --> 00:03:23.199
Eight are women who had been who had taken maternity or pregnancy related leaves, and four are men who took parental leave.
00:03:23.520 --> 00:03:34.800
So the lawsuit argues that meta systems failed to account for legally protected absences and did not pause for individualized review as required by law.
00:03:34.960 --> 00:03:46.159
The complaint cites violations of several state and federal laws, including the Family and Medical Leave Act, the Americans with Disabilities Act, and the Pregnancy Discrimination Act.
00:03:46.319 --> 00:03:57.919
So Meta responded saying that those claims lacked merit and are not based on facts, and stated that workforce decisions were made by people, not AI.
00:03:59.120 --> 00:04:13.199
So, not exactly the week that Meta wanted to follow up with after their hot start to July that saw their stock really just rebound from like a year of lackluster performance and losses.
00:04:13.439 --> 00:04:20.879
So uh Meta honestly had the best start, I would say, uh to July outside of Open AI, right?
00:04:20.959 --> 00:04:29.360
If you look at kind of the the big six of AI, uh right, which is the big four model makers, uh that's open AI, anthropic Google, Microsoft.
00:04:29.439 --> 00:04:33.040
And then you put in Meta and SpaceX into the equation, right?
00:04:33.120 --> 00:04:40.319
If I talk about the big six in AI, uh right, it's kind of the big four plus Meta and uh SpaceX.
00:04:40.720 --> 00:04:47.439
So Meta actually had one of the best Julies, you know, so far, uh, until this happened.
00:04:47.600 --> 00:04:48.879
So not a good look.
00:04:48.959 --> 00:04:51.920
Uh, this one's gonna be interesting to see how it plays off.
00:04:52.000 --> 00:05:00.079
And if nothing else, a good takeaway for business leaders here is you always have to make sure that your systems are talking to each other, right?
00:05:00.240 --> 00:05:04.480
If this is truly the case, this is absolutely horrible, right?
00:05:04.560 --> 00:05:22.560
If women that were on pregnancy leave uh were, you know, essentially um decided and AI decided that they were going to be up for termination because they weren't hitting their keystrokes or their, you know, token, you know, targets, which is absolutely asinine if this is true.
00:05:22.720 --> 00:05:23.600
I hope it's not.
00:05:23.680 --> 00:05:28.319
Uh, but regardless, when it does break one way or the other, we'll let you know in the newsletter.
00:05:28.399 --> 00:05:31.680
I'm guessing this will eventually just be settled out of court.
00:05:32.160 --> 00:05:32.800
All right.
00:05:32.959 --> 00:05:38.480
Uh, our next one, which could end up being a big joint agreement for U.S.
00:05:38.639 --> 00:05:41.839
tech companies, which you don't really see a lot.
00:05:42.079 --> 00:05:58.160
Uh, that's because the Google Deep Mine CEO is pushing for a unified AI watchdog after the anthropic ban has led to kind of this staggered and what I'm calling the permission slip AI reality that we're all living in.
00:05:58.399 --> 00:06:18.240
So, Google Deep Mine CEO Demis Hasabas is calling for a FINRA-style industry-funded watchdog to oversee advanced AI, uh, aiming to prevent the chaotic kind of government crackdowns like the one that halted Anthropics Fable 5 and Mythos 5 models in June.
00:06:18.399 --> 00:06:36.480
So, the Trump administration's abrupt ban on those vital models when they were citing national security risks triggered a tense two and a half week negotiation and forced last-minute changes, exposing the lack of a formal review process for powerful AI systems.
00:06:36.800 --> 00:06:43.360
So, this is what uh Demis's proposed body would require.
00:06:43.519 --> 00:06:56.240
So, it would require uh AI labs to submit their frontier models for pre-release vetting up to 30 days before launch, focusing on threats like cyber warfare, bioweapons, and nuclear misuse.
00:06:56.399 --> 00:07:06.000
So, the framework would also include emergency powers to slow or halt industry-wide AI development if a model is found to cross critical risk threshold.
00:07:06.399 --> 00:07:16.800
So, the reason why this is important because this is the first uh kind of initiative that's picked up steam across all of, well, not all of the big six, but I think five of them.
00:07:17.040 --> 00:07:32.240
That's because major tech leaders, including Microsoft's Satya Nadella, OpenAI's Sam Altman, Google's Sindar Pichai, and others have publicly endorsed the core principles of the proposal, marking a rare show of unity in the competitive sector.
00:07:32.399 --> 00:07:42.720
Uh, Inthropics, Dario Modi supports binding regulation, but has called for an even stricter government-led agency with authority to block unsafe model releases outright.
00:07:42.879 --> 00:07:54.480
So, Meta, led by Mark Zuckerberg, remains the notable holdout here, continuing to advocate for open source AI in drafting its own regulatory response.
00:07:54.639 --> 00:08:17.360
That one there is pretty uh, I don't want to give it a certain name, but yeah, you know, the fact that Meta cited open source AI being an advocate for, even though they are the company that has famously shifted away from open models in their previous uh llama series to now the Mews, which are proprietary closed source.
00:08:17.519 --> 00:08:22.800
So I don't really know why Meta's response was was, well, hey, we care more about open source.
00:08:22.959 --> 00:08:40.080
Anyways, all of the other major US labs have already signed voluntary agreements to submit their models for government safety testing, but the industry, the industry wants a predictable, standardized process instead of the ad hoc intervention that we've seen.
00:08:40.320 --> 00:08:50.000
So the White House is currently reviewing the proposal as a potential compromise to keep America competitive in AI while managing security risks.
00:08:50.159 --> 00:08:54.879
So if you didn't see this, we did share this in our newsletter last week.
00:08:55.120 --> 00:08:57.440
I think this is sorely needed.
00:08:57.519 --> 00:09:10.080
Uh, presumably, you would have probably liked this uh to come from the government and not necessarily a private company, but maybe that's what might make this stick, right?
00:09:10.240 --> 00:09:20.080
Similar to the FINRA style, um, you know, having a private to public uh sector agreement that says, hey, we're essentially the experts in AI.
00:09:20.399 --> 00:09:32.000
You probably don't know, government, uh, exactly what you should be doing when it comes to regulating our industry, which I absolutely agree with.
00:09:32.240 --> 00:09:46.799
Um that's not me saying that there's not smart people in AI in the government, there's obviously absolutely brilliant people uh, you know, who have both backgrounds um at these big six tech companies working in government now.
00:09:46.879 --> 00:09:55.440
But for the most part, when it comes to like the bleeding edge technology, I don't think the White House really has a clue.
00:09:55.519 --> 00:09:56.799
And that's not a political statement.
00:09:57.120 --> 00:09:58.480
That's just any government.
00:09:58.639 --> 00:10:04.720
I don't think any government would have the clue how to regulate the United States acceleration in AI right now.
00:10:04.799 --> 00:10:15.759
Uh right, but that's because you know, 60 days ago, you know, six months ago, the capabilities that we have today, no one probably could have predicted them, especially six months ago.
00:10:16.000 --> 00:10:24.799
So it is good uh that you know Google DeepMinds uh Demis Hossibus is kind of calling for this and taking uh the step forward.
00:10:24.960 --> 00:10:33.120
Uh Demis obviously has one of the deepest and um most respected backgrounds in AI in the entire world.
00:10:33.279 --> 00:10:36.720
Uh, so we'll see if Sir Demis can get it done.
00:10:37.039 --> 00:10:50.159
All right, uh moving from one uh kind of compromise, I guess, of US big tech companies uh to a leak that could maybe impact millions of people.
00:10:50.320 --> 00:10:57.360
That's because we just got a report from Bloomberg that says OpenAI is working on its first hardware device.
00:10:57.519 --> 00:11:07.840
So finally, we get some details on what this uh, you know, much reported on device, you know, we've been hearing about this thing for uh almost nine or ten months now.
00:11:08.000 --> 00:11:16.399
So according to Bloomberg Reports, it is a mobile smart speaker that integrates AI and syncs with Chat GPT.
00:11:16.960 --> 00:11:29.759
So, according to reports, the device is designed to be a screen-free device and it acts as a human-like AI companion for the home, offering proactive learning and a personalized service.
00:11:30.320 --> 00:11:47.679
So, unlike traditional smart speakers, OpenAI's model reportedly features mechanical elements that can move on their own, aiming to feel uh more like a living companion or one that can at least kind of wander maybe from room to room with you.
00:11:47.840 --> 00:11:55.039
So the speaker is intended to access a user's digital life, including emails to deliver highly tailored assistance.
00:11:55.279 --> 00:12:03.519
So, development is being led by former Apple engineers who played key roles in creating products like the iPhone and Mac.
00:12:03.840 --> 00:12:13.519
So the project comes amid legal friction as Apple recently sued OpenAI, accusing it of stealing trade secrets and warning of more revelations during discovery.
00:12:13.679 --> 00:12:16.320
OpenAI denied any wrongdoings on that.
00:12:16.559 --> 00:12:26.240
Sources told Bloomberg the new device veers significantly from anything Apple has on the market today and is unlikely to violate trade secrets.
00:12:26.480 --> 00:12:39.279
So OpenAI's push reflects the growing industry excitement about consumer AI hardware, with startups like Hark raising around$700 million to build personal intelligence devices.
00:12:39.519 --> 00:12:50.399
So the reported OpenAI devices form factor hasn't been revealed, highlighting how much investor interest exists before these products are even unveiled.
00:12:50.960 --> 00:12:55.679
So I'm all for anyone that's listening to the show for a long time.
00:12:55.840 --> 00:12:57.440
I don't really care about my data.
00:12:57.600 --> 00:12:58.559
I don't, right?
00:12:58.639 --> 00:13:09.039
If you want to know all the random AI stuff I've been searching, and uh if an open AI smart speaker wants to follow me around and help me with that, sweet, sounds good, right?
00:13:09.200 --> 00:13:15.919
Anything to get rid of uh Alexa and Siri sooner, uh the better.
00:13:16.080 --> 00:13:17.519
Like, please sign me up for this.
00:13:17.600 --> 00:13:19.120
Put me on a wait list, all right.
00:13:19.279 --> 00:13:23.200
Uh, next, you don't have to get on a wait list because this is out.
00:13:23.759 --> 00:13:30.399
This is the new king of the open, well, soon open source hill.
00:13:30.480 --> 00:13:37.519
All right, more on that here in a second, but we do have a new open source king in Kimmy K3.
00:13:38.159 --> 00:13:54.000
So Moonshot AI, a Chinese startup backed by Alibaba and Tencents, uh, introduced Kimi K3, a absolutely massively monster open source model with nearly three trillion parameters.
00:13:54.159 --> 00:13:57.600
Yes, three almost three trillion parameters.
00:13:57.759 --> 00:14:08.240
So Moonshots revealed Kimi 3 at the World AI conference in Shanghai, touting it as the world's first open source model in the 3 trillion parameter class.
00:14:08.399 --> 00:14:18.159
So Kimi K3 is not technically open source yet, but the company says they will release the weights on July 27th.
00:14:18.399 --> 00:14:30.720
So at 2.8 trillion parameters, Kimi K3 puts it in direct competition with leading American proprietary models from OpenAI and Anthropic.
00:14:31.279 --> 00:15:00.159
So the announcement comes just weeks after the US temporarily forced Anthropic to withdraw its its mythos um and Fable 5 models due to cybersecurity concern, highlighting the growing not just national security focus on advanced AI, but also the interest on potential open source AI projects that can live in the frontier or like frontier point five sector.
00:15:00.559 --> 00:15:01.120
Uh, right.
00:15:01.200 --> 00:15:12.960
So, and this is why it's actually important because third-party evaluations by artificial analysis and arena show Kimi K3 performing uh on par-ish, right?
00:15:13.200 --> 00:15:18.480
Uh with uh uh GPD 5.6 soul and fable five.
00:15:18.559 --> 00:15:21.919
So in some benchmarks, they're a little bit better, actually.
00:15:22.159 --> 00:15:35.759
But right now I would say that you have uh GPT 5.6 soul and fable, they are class 1A, and now well, you might have an open source model in Kimi K3, technically 1B.
00:15:35.919 --> 00:15:40.000
So at least right now, I wouldn't even say it's the second tier.
00:15:40.080 --> 00:15:48.720
It's you know, we have a tie for 1A, and yeah, we have a Chinese open source model that has closed the gap uh to be 1B.
00:15:49.120 --> 00:16:03.360
So Kimi K3's open source approach stands in contrast to obviously the closed proprietary systems of American firms, throwing a wrench in the AI strategy game plan for some larger enterprises.
00:16:03.519 --> 00:16:11.200
So the model itself is designed to operate with minimal human supervision, handling complex tasks like engineering and software development.
00:16:11.360 --> 00:16:24.000
The release immediately impacted global markets with shares in Chinese competitors ZAI in Minimax dropping sharply in Hong Kong after the announcement.
00:16:24.320 --> 00:16:26.879
So I'll cut it to you straight on this.
00:16:26.960 --> 00:16:33.759
Uh, because I think people, for the most part, when they look at open source models, they usually think of two things.
00:16:33.919 --> 00:16:38.399
They think, well, if I have a powerful enough computer, right?
00:16:38.480 --> 00:16:43.679
Uh, like as an example, I have a very beefy Mac Studio uh sitting here on my desktop.
00:16:43.840 --> 00:16:49.679
People think, oh, if I have very expensive hardware, I can run the best open source models.
00:16:50.000 --> 00:16:51.679
No, we are past that.
00:16:51.919 --> 00:16:59.840
Unless you literally have a rack of NVIDIA GPUs and a cooling uh center in your basement, you can't run this, right?
00:16:59.919 --> 00:17:02.639
This is for 0% of consumers out there.
00:17:02.879 --> 00:17:08.079
You can't run a 3 trillion parameter model on any consumer hardware.
00:17:08.240 --> 00:17:09.759
So that's number one.
00:17:10.000 --> 00:17:15.200
So, yes, this is really for those enterprises that have access to compute at scale.
00:17:15.359 --> 00:17:40.880
Number one, especially if you can host it, which is not a lot of companies, but if you do have a couple of racks, uh and maybe you have a large AI bill, you know, maybe you're paying uh eight, nine figures in a ice fan with, you know, a bunch of GPUs, which again, not a lot of companies, but for some companies, yes, in uh this week or sorry, in like 10 days or so, uh, on July 27th.
00:17:40.960 --> 00:17:49.920
So that's like in seven days, um, you'll technically be able to download this model and if you want to fine-tune it because it's open source.
00:17:50.000 --> 00:18:03.359
Um, so for a very, very, very, very few companies, this is actually could be fantastic news if you have uh the resources and uh the the human uh skill needed to make that happen.
00:18:03.599 --> 00:18:04.640
That's one thing people think.
00:18:04.799 --> 00:18:06.079
Oh, I can just download this and run it.
00:18:06.240 --> 00:18:06.799
No, you can't.
00:18:06.960 --> 00:18:09.599
Number two, they think it's extremely cheap.
00:18:10.240 --> 00:18:11.440
Also not the case, right?
00:18:11.519 --> 00:18:19.200
So if you're not using this, you know, downloading it and running it locally, people think, well, oh, open source models are super cheap.
00:18:19.519 --> 00:18:21.519
Previously, yeah, they were.
00:18:22.160 --> 00:18:24.079
Not really with this one, right?
00:18:24.240 --> 00:18:26.559
Um, we we we see this Kimi K3.
00:18:26.880 --> 00:18:35.759
Uh, we also saw uh that Quen will be releasing their next big multiple trillion parameter model uh here any minute or any hour.
00:18:36.000 --> 00:18:45.359
So I don't think this next tier of frontier-esque open models are gonna be open in the sense that we thought of them previously.
00:18:45.519 --> 00:18:46.079
Um, right.
00:18:46.160 --> 00:18:53.359
And the other thing is, there's no real guarantee that we keep getting these open source Chinese models for much longer.
00:18:53.599 --> 00:19:03.839
We saw reports, which we talked about on last week's AI News That Matters, that China may start uh restricting the export of their models, right?
00:19:04.000 --> 00:19:16.640
Uh, specifically not wanting uh maybe the US to get a hold of these, which is obviously ironic considering that these models have largely just been distilled from US models.
00:19:16.720 --> 00:19:19.200
Uh so regardless, there's a little bit of background.
00:19:19.359 --> 00:19:20.559
You can't go download this.
00:19:20.640 --> 00:19:24.880
And if you're paying API prices, it's not like they're super cheap.
00:19:24.960 --> 00:19:31.680
Uh, right, they're actually in many cases more expensive than GPT 5.6 uh Seoul.
00:19:31.839 --> 00:19:41.519
Uh, you know, they're probably a little cheaper uh than Fable 5 and Mythos 5, since those are the most expensive models on the market when it comes to price per performance.
00:19:41.839 --> 00:19:42.480
All right.
00:19:42.960 --> 00:19:45.920
More model news, but this is not we get one.
00:19:46.000 --> 00:19:47.839
This is just we're waiting longer.
00:19:48.000 --> 00:20:04.880
So Alphabet, parent company of Google, its shares fell 4% Thursday after news broke that the highly anticipated Google Gemini 3.5 Pro, yeah, months behind schedule, not coming out anytime soon.
00:20:05.039 --> 00:20:17.279
So, according to a new Bloomberg report, the delay is due to Google's efforts to improve the model's coding performance, which currently falls short according to their own internal expectations.
00:20:17.599 --> 00:20:31.039
So the setback comes as competitors like OpenAI, Anthropic, and uh Meta have all launched AI models recently that outperform Google's current offerings in software code generation.
00:20:31.200 --> 00:20:39.920
So Alphabet first teased Gemini 3.5 Pro in May after releasing Gemini 3.5 Flash.
00:20:40.160 --> 00:20:41.920
So in May at their I.O.
00:20:42.160 --> 00:20:45.920
conference, they said that the model would be coming out in June.
00:20:46.160 --> 00:20:50.960
Obviously, June has already passed and we're near the end of July.
00:20:51.119 --> 00:20:58.480
And now the latest report says Google's next model in Gemini 3.5 Pro could still be months away.
00:20:58.640 --> 00:21:14.799
So an alphabet spokesperson told CNBC that the company is shipping quickly across a wide range of models and continues to test Gemini 3.5 Pro and other upgrades with partners and the US government.
00:21:15.759 --> 00:21:30.880
Uh so yeah, if you were, you know, holding your breath or if you were you know trying to decide between GPT-5.6 soul and fable five, and you're like, wait, let's just wait until you know Google Gemini 3.5 Pro because it was supposed to come in June.
00:21:31.039 --> 00:21:32.319
Surely it's coming in July.
00:21:32.480 --> 00:21:33.279
Well, no.
00:21:33.519 --> 00:21:42.559
And if this report holds to be true, we might be heading into the fall without even Gemini 3.5 Pro.
00:21:42.880 --> 00:21:49.440
And at that point, I mean, presumably, we're gonna be getting a Fable 5.1 soon.
00:21:49.519 --> 00:21:54.319
Uh, we may be getting a GPT uh six around the horizon.
00:21:54.480 --> 00:21:55.200
We'll see.
00:21:55.359 --> 00:21:59.519
So, you know, it's it's gonna um be worth following.
00:22:00.160 --> 00:22:11.839
To see just how big of a jump this 3.5 Pro could be, uh, or maybe I don't know, uh, if Google comes out with another flash model in the interim.
00:22:12.000 --> 00:22:18.480
Because if they're still working on the same pre-train that they were previously, that can might be concerning.
00:22:18.640 --> 00:22:40.240
And I mean, we'll see if for the first time in a long while, at least two plus years, if Google is not able to compete on the model front, which to me, I would be shocked because I've always been a big believer in that Google at any point, whenever they want, can drop the world's most powerful model.
00:22:40.640 --> 00:22:50.160
So I don't know, maybe when I said that a year or so ago, maybe I was biting off more than I could chew and believing in uh the big G a little bit too much.
00:22:50.400 --> 00:22:53.599
So uh yeah, we'll we'll see if that holds true.
00:22:53.759 --> 00:22:55.519
I still think they'll be able to.
00:22:55.599 --> 00:23:01.359
Uh you know, Google it's they have their own, right?
00:23:01.599 --> 00:23:20.559
They have their own uh the you know GPUs, TPUs, they can they can train these things quickly, uh, without having to, you know, pay their competitors billions of dollars, like Anthropic is having to do right now, uh, for inference and to train these models.
00:23:20.799 --> 00:23:22.160
So, all right.
00:23:22.319 --> 00:23:32.720
Uh speaking of big tech companies, we're moving from Google to Microsoft for our last big AI news story of the week.
00:23:33.039 --> 00:23:35.119
That is the juicy one.
00:23:35.200 --> 00:23:35.839
Here we go.
00:23:36.000 --> 00:23:58.079
So, Microsoft CEO Satya Nadella recently, according to reports, started to kind of seemingly shift his public stance on mainstream AI, moving from an aggressive investor in closed source frontier labs to a vocal critic of their current trajectory.
00:23:58.240 --> 00:24:05.839
So the shift began with Nadella's leaked criticism of Anthropics Claude Fable 5.
00:24:06.079 --> 00:24:23.519
So, according to reports, Nadella slammed the flagship model as being too quote unquote editorially controlled, arguing that its rigid guardrails and refusal protocols don't make sense for a creative utility.
00:24:23.759 --> 00:24:40.000
So, reports also said that Nadella questioned Anthropics limits on user requests to its high-end fable AI model, calling the editorial controls on creation tools nonsensical during remarks to Microsoft's co-pilot engineers.
00:24:40.160 --> 00:24:54.240
So, Anthropic's Fable model sometimes rejects users' queries or responds with older versions of the model, with essentially falling back to Opus 4.8, leading to a lot of frustration and criticism on social media, right?
00:24:54.400 --> 00:25:07.920
You've seen many examples, and I've tried some myself, even asking Fable 5 about certain types of cancers, or you know, hey, what are the biggest uh you know, in certain medical fields, what are the biggest unanswered questions?
00:25:08.079 --> 00:25:16.319
And it will say essentially, you can't ask this, and it's gonna roll you back uh to an older model like Opus 4.8.
00:25:16.559 --> 00:25:26.960
So this is extremely noteworthy as Satya Nadella is uh, at least according to these reports, publicly criticizing Anthropic.
00:25:27.359 --> 00:25:44.559
Um uh that's because Microsoft has invested five billion dollars in Anthropic, uh, with Anthropic agreeing to also spend$30 billion in future years on Microsoft's Azure cloud infrastructure.
00:25:44.880 --> 00:26:08.799
So this is uh right, you normally wouldn't hear shots fired like this, uh, you know, slamming a company in their flagship model, especially when said company is well gonna be reportedly paying you$30 billion over the next few years, and you have a large ownership in that company.
00:26:09.200 --> 00:26:16.880
But that wasn't the only uh kind of shot at anthropic that Nadella took.
00:26:17.119 --> 00:26:20.480
That one was again, according to reports, direct.
00:26:20.720 --> 00:26:26.400
The second one, a little indirect, but most industry uh insiders are connecting the dots.
00:26:26.640 --> 00:26:44.079
So, in a seemingly somewhat related long-form online post last week, Nadella coined the term reverse information paradox to explain that businesses essentially just pay for AI twice, they pay for it once with actual money, right?
00:26:44.240 --> 00:26:48.079
So they're buying subscriptions for their employees or paying for tokens.
00:26:48.240 --> 00:26:51.599
And then he said they're actually paying for it a second time.
00:26:51.759 --> 00:27:04.240
And the second time is more dangerous because they're giving away their proprietary knowledge and institutional know-how that they're essentially forced to hand over to make the model useful.
00:27:04.880 --> 00:27:18.720
So in his posts, Nadella didn't name names on who does this, but well, everyone's looking at the fine print here and being like, wait, the only company that actually forces you to do that right now is anthropic.
00:27:18.880 --> 00:27:29.599
Because if you're a at least for paying business customers, at least, uh, Anthropic is the only one that has a requirement that they're going to retain your data and there's nothing you can do about it, right?
00:27:29.759 --> 00:27:46.480
So with the new Mythos 5 and Fable 5, uh Anthropic essentially throws zero uh data retention or uh ZDR, you know, Anthropic's just throwing zero data retention out the window.
00:27:46.640 --> 00:27:57.599
Uh so again, although Nadella didn't mention anthropic by name, everyone's you know reading his posts and being like, okay, this is also definitely about Fable.
00:27:58.000 --> 00:28:04.880
So for all you people out there that you know sometimes say, Oh, Jordan's being so hard on Anthropic, no, I'm not.
00:28:05.039 --> 00:28:05.519
My gosh.
00:28:05.680 --> 00:28:20.640
Uh, like the fact that Anthropic still has that required uh data retention, again, I don't have five billion dollars invested in anthropic, and anthropic hasn't agreed to pay me 30 billion of dollars over the next couple of years.
00:28:20.799 --> 00:28:30.000
So if Saudi Nadella is number one, again, according to reports, calling them out by name in one uh you know, breath.
00:28:30.079 --> 00:28:38.720
And then in the next breath, he's essentially describing them because that's the only company that truly fits that description for enterprise customers.
00:28:38.960 --> 00:28:42.640
That's not a good sign for anthropic, just saying that.
00:28:43.359 --> 00:28:56.480
Uh, so yeah, yeah, many attributed that second half of the statement to Anthropic's Mythos 5 and Fable 5 data policy, which forces that data retention even for enterprise customers.
00:28:57.039 --> 00:29:04.079
All right, that's a wrap on our big stories, but let's quickly go into the what's new and what's next.
00:29:04.240 --> 00:29:09.839
So these are some still important happenings, uh, some maybe some leaks in here.
00:29:09.920 --> 00:29:14.319
Uh, but let's just quickly go into our what's new and what's next.
00:29:14.480 --> 00:29:15.119
Rapid fire.
00:29:15.200 --> 00:29:15.680
Here we go.
00:29:15.839 --> 00:29:22.960
So the Trump administration launched Gold Eagle Hub to accelerate cybersecurity patching for critical infrastructure.
00:29:23.200 --> 00:29:29.440
Meta reportedly entered talks to lease up to 10 billion dollars of compute to Anthropic.
00:29:29.599 --> 00:29:34.559
Google AI mode added actions through Canva, Instacart, and YouTube Music.
00:29:34.720 --> 00:29:36.880
Yeah, if you didn't check that out, uh, pretty cool.
00:29:36.960 --> 00:29:40.079
If you have a paid personal Gmail account, check that out.
00:29:40.240 --> 00:29:45.519
Uh, next, I referenced this earlier, but Quen 3.8 is launching soon.
00:29:45.680 --> 00:29:57.599
And the company says it is a 2.4 trillion parameter model, and they are reporting benchmarks near Fable 5 GPT 5.6 soul, Kimi K3 levels.
00:29:57.759 --> 00:30:04.480
All right, thinking Machines released a 975 billion parameter open weight inkling AI model.
00:30:04.720 --> 00:30:11.680
Uh, New York halted any data center construction more than 50 mag um MW.
00:30:12.079 --> 00:30:13.440
Uh, is that mega?
00:30:13.599 --> 00:30:15.359
I always forget that what what that is.
00:30:15.519 --> 00:30:18.720
MW is is mega mega hertz.
00:30:19.039 --> 00:30:19.680
Is that right?
00:30:19.839 --> 00:30:21.359
Mega, megawatt.
00:30:21.519 --> 00:30:22.480
My gosh, y'all.
00:30:22.640 --> 00:30:26.160
Sometimes I read too much AI news and I just forget things.
00:30:26.319 --> 00:30:26.640
All right.
00:30:26.799 --> 00:30:34.960
Anyways, New York halted uh construction for one year on any of those uh new data centers.
00:30:35.119 --> 00:30:44.079
Uh, next, more than 200 experts, including 15 Nobel laureates, called for urgent preparation for AI's economic impact.
00:30:44.319 --> 00:30:52.640
Anthropic permanently extended Fable 5 to max users, but yeah, it's only at about one-third of the original limit.
00:30:52.720 --> 00:30:55.920
So, yeah, uh, I talked about this on last week's show.
00:30:56.000 --> 00:31:03.599
I said there's no way Anthropic is actually taking away Fable 5 because they would lose most of their subscription revenue.
00:31:03.680 --> 00:31:10.160
Uh, so unfortunately for most people, it is only if you're on that$100 or$200 a month paid plan.
00:31:10.400 --> 00:31:18.799
Google renamed Notebook LM to Gemini Notebook and they added cloud code execution uh for paid users.
00:31:18.880 --> 00:31:23.759
I think that's only out uh to ultra users, but it will be coming out to pro users soon.
00:31:23.920 --> 00:31:28.640
So, yeah, let me know if we should do a Gemini Notebook uh episode when that drops.
00:31:28.720 --> 00:31:31.680
I'm gonna accidentally still call it Notebook LM all the time.
00:31:31.839 --> 00:31:41.200
Uh next, a hacker leaked Sunos secret data, uh, reportedly showing that millions of songs were scraped from YouTube and Deezer.
00:31:42.160 --> 00:31:42.799
All right.
00:31:43.119 --> 00:31:50.720
Uh, next, OpenAI released GPT Red, an automated system to detect prompt injection vulnerability in AI agents.
00:31:50.880 --> 00:31:55.599
Uh, Amazon Bedrock added support for OpenAI's new models, GPT5.6.
00:31:56.079 --> 00:32:09.680
Uh, Apple is reportedly testing Prism L, uh Prism L's uh AI model compression, which would essentially allow them to run 27 billion parameter models on iPhones, which would be pretty sweet.
00:32:09.839 --> 00:32:13.599
Please just fix the Siri Apple.
00:32:13.680 --> 00:32:15.039
If you're listening, thanks.
00:32:15.200 --> 00:32:22.400
Uh Anthropic added an in-app browser to Claude Code on desktop, enabling site in doc interaction.
00:32:22.480 --> 00:32:23.440
So yeah, I tried it out.
00:32:23.519 --> 00:32:24.079
It's pretty good.
00:32:24.240 --> 00:32:26.880
It's not quite as good as Codex's, but it does the job.
00:32:26.960 --> 00:32:32.640
Uh, next, Perplexity rolled out self-improving memory, faster models, and one-click website publishing.
00:32:32.799 --> 00:32:44.480
ChatGPT rolled out their universal search uh that helps you more easily find projects, images, documents, chats from one sidebar, a much richer and better search experience.
00:32:44.640 --> 00:32:57.359
Uh, email program superhuman released their new auto draft feature that writes email replies in your voice uh using better models, and then you can go send them from your drafts.
00:32:57.519 --> 00:33:04.880
Next, Spotify released a new feature where you can talk to Spotify by voice or text to control music and see listening history.
00:33:05.039 --> 00:33:12.960
Uh Gemini rolled out personal avatars to allow you to generate AI videos starring your own face and voice.
00:33:13.279 --> 00:33:19.920
And then last but not least, uh ChatGPT work, which I don't know if it's called just ChatGPT now.
00:33:20.079 --> 00:33:24.400
I noticed that in the dropdown that used to say ChatGPT work and codex.
00:33:24.559 --> 00:33:26.480
Now it just says ChatGPT and Codex.
00:33:26.640 --> 00:33:33.119
So I'm not sure if we're still calling it ChatGPT Work or not, but regardless, it actually makes it easier.
00:33:33.279 --> 00:33:41.119
That's funny, one of my uh AIs just uh shouted out at me, but it actually makes it much easier uh to show your conversation history.
00:33:41.279 --> 00:33:48.720
So essentially they brought your chats from Chat GPT in to a place that made more sense inside of the desktop app.
00:33:48.799 --> 00:33:51.839
Whereas before, you know, users were a little bit confused.
00:33:52.000 --> 00:33:55.440
Also, it brings your projects in, which a lot of users were complaining about.
00:33:55.599 --> 00:34:03.680
So essentially, if you tried out Chat GPT work when it came out uh about 10 days ago and were like, wait, some of this isn't adding up.
00:34:03.759 --> 00:34:05.200
Where's all my chat GPT stuff?
00:34:05.279 --> 00:34:06.319
Well, now it's all there.
00:34:06.480 --> 00:34:08.400
All right, so I hope this was helpful.
00:34:08.559 --> 00:34:17.840
All of the AI news that matters for this week, but like I said to start the show, I can almost guarantee we're gonna get a ton of new releases, right?
00:34:18.000 --> 00:34:27.360
Word on the street is we'll probably be getting an Opus V from Anthropic that could be near that GPT 5.6 or uh mythos level.
00:34:27.519 --> 00:34:31.440
Uh, you know, and who knows, it is still hot AI summer.
00:34:31.599 --> 00:34:37.360
And I think that once the steam starts to pick up, it is going to be a hot one.
00:34:37.599 --> 00:34:40.159
So uh I get it, it is hard.
00:34:40.239 --> 00:34:47.599
Uh, if you are a busy business leader and you're trying to implement AI in your company and you're like, how do I keep up?
00:34:47.760 --> 00:34:53.760
Well, it starts here with the daily podcast, but make sure you also go to our website at your everydayai.com.
00:34:53.840 --> 00:34:54.480
Thank you for tuning in.
00:34:54.639 --> 00:34:57.599
Hope to see you back tomorrow and every day for more everyday AI.
00:34:57.679 --> 00:34:58.320
Thanks, y'all.
