Alexander Wissner-Gross

"It's maybe close to recent Sonnet 5.5 or Opus 5.5 if you ignore cost."

"Claude is now leading 26% of model research and development internally as of August, up from 3% in April. If you just extrapolate that one trend sigmoidally, you find that approximately in the next 3 to 12 months, depending on uncertainty, AI is just completely leading all of its own R&D."

"It had to copy Anthropic and catch up via Codex to Anthropic's Claude and Claude for code. So it's running a little bit late."

"Users of Claude can get Claude where and when they want on the infra they want. That's powerful. And you see now OpenAI copying Anthropic and doing that."

"The recent launch of Opus 5 is a step in the right direction to seeing a convergence between frontier models on the one hand and videogen or world models on the other, because Opus 5 seemed almost mildly bench-maxed towards front-end development and the loop between visuals and code."

"Fable was almost certainly pre-trained off of a common corpus and probably post-trained off of a good deal of the same synthetic corpus as earlier models like Opus 4.8. So the signatures would be reasonably expected to rhyme if K3 were being post-trained off of Opus 4.8."

"If you try to ask Fable 5 any sort of question about biology, it'll autorevert to Opus 4.8. So adding more buffer is the go-to strategy right now on the jailbreak or prompt injection side."

"Opus 4.8, which has been out for what in these singularity times passes for an eternity, is better. It's superior on a cost-performance basis."

"Folks I know are actually getting real performance gains out of running GLM 5.2 locally instead of say Opus 4.8 or GPT 5.5. This is just hugely liberating for anyone who wants near the level of Opus 4.8 performance that they can run locally and control locally."

"I would expect lots of new targets, lots of new approaches to emerge from both the AlphaFold 3 style of protein folding problem and structural biology being solved as well as virtual cell models."

"Quantum computing was supposed to give us protein folding. Turns out the protein folding problem was solved by arguably AlphaFold 3 purely classically without use of any quantum computing."

"Neuralink announcing the successful pre-training of the first foundation model off of recordings at scale from one human, soon to be many humans."

"On the BCI side, Science and Neuralink, Blind Sight — two most prominent examples of BCI augmented vision."

"Neuralink, which also has its own approach for curing blindness called Blindsight, seems to primarily be focused on direct brain stimulation, direct brain intervention."

"I think we'll see to earlier discussion borg-anisms, so collective human intelligences, whether it's via Neuralink or some other format."

"Anyone who's ever used Claude Code understands Claude Code can use all of the tools available in your local desktop environment. And that was a pretty big leap compared to what was available on the market prior to Claude Code."

"Not only the attraction to pre-ChatGPT books because maybe they contained fewer generative artifacts, but also rumors that in newer books authors are attempting to defend themselves with poisoning attacks."

"A recognizable transformer-like architecture can almost match GPT-5.5 max on the task-cost frontier."

"I am very very excited to see 5.6 come out. We've seen improvements on a number of biology benchmarks. We haven't seen the full suite of benchmark results on 5.6 yet."

"Chat GPT also is going through the same brain transplant of making Codex essentially a model or class of models together with scaffolding that were optimized for code generation and recursive self-improvement — that's becoming the new mainline ChatGPT."

"Folks I know are actually getting real performance gains out of running GLM 5.2 locally instead of say Opus 4.8 or GPT 5.5."

"If China hadn't forced the issue by open sourcing, suppose there was just OpenAI, Gemini, Anthropic, XAI, all four US companies."

"A more conventional cadence would be quarterly or annually. In Gemini — Google DeepMind's case — it's certainly on an annual basis."

"Right now we are in a arguably a duopoly between OpenAI and Anthropic. They're just running away with the race with Chinese open-weight models a few months on their heels. And I want there to be a competitive frontier. And I think Grok is one of the possible competitors along with Gemini."

"Google DeepMind has released Gemini Omni which will generate at best 10 to 15 second clips, but they've basically abandoned long-form video generation."

"What do we have? We have Gemini Omni, which is still limited to about 10 seconds."

"Making Codex essentially a model or class of models together with scaffolding that were optimized for code generation and recursive self-improvement — that's becoming the new mainline ChatGPT."

"Codex is a wonderful product, and Codex revenue, according to OpenAI's reports, is skyrocketing. If I were Sam and I were OpenAI, I would be asking the question: how long until Codex can be fully brain-swapped in with all of ChatGPT and powering the revenue engine that I need to motivate the trillion-dollar-plus IPO."

"Sam is just putting a happy face on the need to grow Codex revenue even more before OpenAI is ready to go public. There's a deeper take — if IPOs are being delayed because large private companies don't need external capital because they're achieving recursive self-improvement, this marks potentially the beginning of a decoupling between technology and capital."

"I view Grok 4.6 as essentially the next version of Cursor. XAI, SpaceX, and Elon have been quite public about how 4.6 leaned heavily on post-training thanks to the Cursor acquisition, which I think is still in the process of being completed. Their licensing of all of the reasoning trace data from Cursor as essentially pulling a westernized version of what the Chinese frontier labs were doing — siphoning off reasoning traces from lots of people historically interacting via Cursor with Claude and Claude's competitors."

"He basically had to gut his foundation model team and acquire Cursor with the IPO riches from SpaceX."

"The Grok 4.5 that we saw has finally touched the cost-per-task optimal frontier. It's a Grok that's basically merged in and apparently on its way to becoming Cursor's model, but rebranded as Grok."

"I see the beginnings with Grok 4.5 in combination with Cursor. Right now the Cursor acquisition is taking the form of post-training of models that he already had. In the near future, Elon's promised that Cursor is going to be part of the pre-training recipe. If he can make that work, I think Grok has a fighting chance to join the frontier."

"Grok infamously doing it. El admitted it and then also purchased Cursor which had been fine-tuning off of traces on top of Claude. So this is everyone else doing this."

"SpaceX just announced a definitive agreement to consummate its purchase of AnySphere, the corporate owner of Cursor, for 60 billion plus. Cursor is the new Grok for the time being."