Release cadence, Polymarket odds, and rumours of a real-time Astra

All right, you heard it here first, folks. And James Cameron, if you're listening, it's your next movie.

A few quick glances at what's coming in the near future. You know, I've mentioned this before. We've seen the release of new models going from every few months now to every 5 days. Uh, some predictions from Polymarket: uh, the next Grok model, 4.7 — again we're waiting for Grok 5, uh, when it comes out — but Grok 4.7 expected in the next week or two. Uh, so super cool.

And then uh GPT-6.1 released by when? Uh, again, this is going to be sort of leapfrogging each other. Uh, here it is, their prediction of, you know, 48% by end of October and 85% by the end of September. And then finally uh Anthropic with Fable 5.2 — two very similar numbers, again, uh, you know, 43% by October 31st and 86% by — no, uh — December 31st. And again, this is where the frontier labs are sort of playing chicken with each other, waiting for the model to come out and then the next day releasing their model.

Any thoughts on this, Alex, on model — just dwelling in particular on what the rumor mill on social media are alleging regarding future versions of Astra?

The pretty consistent rumor mill message is that sometime later this year we'll see — and we've started to see this a bit with Astra winning Pokemon, winning Portal, winning all of these other interactive, visually intensive, visually reasoning intensive games — the rumor mill is furiously alleging that sometime before the end of this year, we're going to see a future version of Astra that offers honest to goodness real time control, including third party extrapolations, that if you take Astra, you extrapolate, say, robotic control — maybe we'll talk about that in a bit — we're going to see real time generally intelligent embodiment, either controlling video games, which is again ironically where the whole industry of modern RL started, with the Google DeepMind folks trying to win at computer games using [unclear] and GPUs being developed for video games in the first place.

Well, the modern use of GPUs started with ImageNet and the ILSVRC for computer vision. Uh, and then we started to see out of the DeepMind folks in particular deep reinforcement learning for winning games like that. That's where RL started, and then we went off on this detour of LLMs that were just self-supervised objectives. But it seems, as alleged, in the next few months we're supposed to see a next version of Astra that will be general purpose, interactive, real-time, game-winning capable AI.