BTW it is Sunday and I'm debugging a super complex issue on HVM4's hot path compiler. I've spent hours on this, and will need a few more hours. As of 2025, AI's are utterly useless at this kind of problem. At the end of 2026, a computer will one shot it in less than 5 seconds.

@VictorTaelin I agree with all points except the first and the last. Perhaps your definition of AGI is narrower, but I’m not aware of any shift in research focus that would justify predicting that the fundamental leaps in deep learning needed to achieve AGI will happen in 2026.

@kibo_osu deep learning isn't the only thing that exists

@VictorTaelin Why would I take your word over karpathy’s?

@_vincentpaul_ some people buy dollars for 3 cents

@VictorTaelin Sounds like a typical wishful thinking. LLMs usefulness isn't a skill issue, and LLMs will not automate R&D even if you pour in 1000 times more investors money. The transformer architecture is deeply flawed, without replacing it you will always have unreliability.

@Armados18 hm, agreed? you think this opposes my post in which way

@VictorTaelin Would you be willing to bet any amount of money that AGI is not coming in 2026?

@dreambuffer I already have these bets in place

@VictorTaelin AGI isn't coming in 2026 bro. And when it doesn't, will you chalk that up to a skill issue on your part or that of every billion/trillion dollar AI company?

@cogchamp1 if it doesn't then I'll have been wrong and my word will become less valuable, and vice-versa

@VictorTaelin When you say LLMs cannot make scientific discoveries, are you referring only to their architecture, or also to models plus algorithms like online RL (RLVR) finetuning? Because that’s a completely different case with opposite evidences.

@miolini I think that works, but is less efficient than plain search!

@VictorTaelin How are LLMs generating R&D if their output is based on their training set of existing knowledge?

@rcwhiteley they are not