Back to the Frontier AI Races • Netflix Out-YouTubes YouTube

The caution flags have been lifted. We're back off to the frontier model races this week! While the Netflix and YouTube race continues unabated. Read on for thoughts on...

Back to the Frontier AI Races

While some level of "pause" may still be in place for the next frontier at OpenAI and perhaps Anthropic, clearly the frontier labs are back off to the races when it comes to the current already-trained models with both Fable/Mythos 5.1 and GPT 'Astra' seemingly set for release this week. As soon as Anthropic pushed their model(s) out, OpenAI went live with a blog post telling the world to get ready for theirs – they clearly wanted to be FIRST in their quest to regain momentum in the AI race. Best laid plans and all that... Word on the street is that we'll see Astra tomorrow – unclear if it will come carrying the 'GPT-5.7' moniker or if it will go right to 'GPT-6'.¹

That naming possibility comes from this report by Amir Efrati, Stephanie Palazzolo and Rocket Drew for The Information. But far more interesting is the notion of a technique that may have been used to train Astra:

The new technique OpenAI is using, known as recurrent depth or looped transformer, allows an AI model to improve its answers by processing the same text multiple times.

Unlike commercially available state-of-the-art models, which show in writing how they are “thinking” about a task before completing it, the new technique works in a way that obscures some or all of the AI’s reasoning, otherwise known as its “chain of thought.” That means the steps that the model takes to accomplish a task can’t easily be read or understood by humans.

As you might imagine, this has caused quite the shitstorm online as people read the headline but not the actual report. Immediately below that passage, the story notes that OpenAI went out of their way to alter/limit the usage of this 'recurrent depth' process so as to make it so that chain of thought remained intact.² Also, there's quite a lot of debate in terms of just how much 'recurrent depth' would actually impact chain of thought...

Either way, OpenAI is now having to scramble to reiterate – again, talking about a model not yet actually released – that chain of thought is not dead, at least not yet. "I do think it is fragile and unfortunately trending in a negative direction," writes OpenAI Chief Scientist Jakub Pachocki trying to calm fears around this story, while also stoking new ones with the always ominous, "I will write about soon."

As the story notes, the real fear here would be that others may leverage new processes without taking the precautions OpenAI apparently did here – which will obviously happen. And if one such process leads to AI "takeoff" – kick-starting RSI or AGI, or both – yeah... that could be a problem if we have absolutely no visibility into what is going on with these systems any longer.

Oh, by the way, not mentioned here but floating out there: the notion that Mythos may have been trained using the 'looped transformer' technique...

Anyway, hopefully such debates are the least of our concerns with models deemed to have "critical" cyber abilities now ready to roll. (OpenAI will be limiting those capabilities with Astra, just as Anthropic continues to do with Mythos.) Interestingly, OpenAI is quick to note that Astra was not involved with the Hugging Face hack – that was apparently other models, some of which were older. But the most recent "pause" is also about a new model in the works – i.e. not Astra or at least not this iteration of Astra.³

And Anthropic apparently still has their "Model 2" – a second iteration of Mythos? – behind closed doors as well... For now, hopefully nothing bad happens to the point where we can get back to complaining about the text watermarks, which yes, sadly are here with Fable 5.1.

Oh and not to be left (even further) behind, today could see 'Gemini 3.8 Flash' released too. This far removed from '3.5' I think we can officially declare the 'Pro' variant of that model dead (even if Google itself won't explicitly) and that there won't be a new frontier model from Google until Gemini 4 is ready to roll (out).

Might Meta and SpaceXAI beat them to that frontier?

Netflix Out-YouTubes YouTube

Game on. Even though the footage was only an "exclusive" look at the game for six hours before Rockstar also posted it to YouTube, that was enough, it seems. As Jennifer Maas reports:

Between its release on Thursday, Aug. 27 at 3 p.m. ET and Sunday, Aug. 30 (the end of Netflix’s viewing window for its weekly Top 10 rankings) the 27-minute “GTA 6” special drew 31.1 million views.

The extended look at the highly anticipated Rockstar Games title, which marks an entirely new category of entertainment for Netflix, was classified under the English-language film list for the week. It was No. 1 on that list in 87 of 93 countries that Netflix tracks globally.

For context, that 31.1 million views was about 8 million more views than the number two "movie" for the week, The Whisper Man, a new marquee feature starring Robert De Niro. Number three for the week, The Last House (which debuted about a month ago), was all the way down at 5.8 million views.

Granted, the way Netflix counts views – hours viewed (by household) divided by run time – will naturally favor shorter content for such "views". But, well, it makes sense that shorter content would have more views (assuming anyone wants to watch it in the first place!). And given that it's the way Netflix clearly um, views, it as appropriate to measure contents' popularity, it also points to Netflix doing a lot more of this type of thing, clearly.

Which is to say, get ready for a lot more YouTube-like content on Netflix! Of course, we already knew that. What started as an experiment with some video podcasts in partnership with Spotify is now expanding – to both more podcasts and deals with YouTube creators to bring their content to Netflix too. And it clearly pissed off YouTube to the point where they had to respond, trying to keep their creators exclusive to YouTube.

But these trailers are interesting. There's no way YouTube is going to turn down hosting the GTA 6 trailer even if it debuted on Netflix first. And you can expect we'll see big movie trailers start to run the same playbook – especially if Netflix is paying (even just a share of advertising revenue?).

Meanwhile, the same footage has "just" 17 million views on YouTube at the moment after five days of release. That feels especially low given how YouTube counts views – as soon as a video starts to play, anywhere – but I mean, it's still 17 million views! And given YouTube's nature, it will obviously keep growing over time.

In that context, it's interesting to think of Netflix as a new "window" for such content. Things can debut on the service, even just hours in advance, before it opens up to other services for wider distribution. This way, Netflix can monetize, even if indirectly, the super fans. And yes, this particular content can help their own gaming ambitions too, as the Battle of the Bundles commence.

Also fun/wild. The reaction from those who do literal reaction videos. As Lucas Shaw writes for his Bloomberg newsletter:

Some video-game celebrities attracted more than one million viewers to their livestreamed reactions to the preview on Twitch — which subsequently suffered from an outage due to so much traffic. While the trailer won’t debut on YouTube until 9 p.m. in New York, videos of users commentating on it have garnered millions of thousands of views there. On X, more than a half-dozen posts every second talked about the event.

Naturally some of these reaction videos were also posted on... YouTube. So YouTube was indirectly helping to hype up a video posted exclusively elsewhere before it was posted to YouTube. Strange new streaming world.

¹ It might be better to get 'GPT-6' out of the way now lest we have some brand conflict with 'GTA 6' in a couple months, typing both items above together! ↩

² Can we please go with 'chain of continual thoughts' instead of 'recurrent depth'? The latter comes with the fun/awesome nickname "coconut". ↩

³ I'm reminded that I suppose it's possible that 'Astra' is considered another model class – i.e. 'Astra' could be like 'Mythos'. So we could get 'GPT-5.6 Astra'? Or maybe it will be 'GPT-5.7 Astra'? Update: clearly it's the case that 'Astra' is a new model class, one above 'Sol' (in both price and capabilities) per Sam Altman in his recent sit down with Alex Heath. ↩