Dylan Patel says OpenAI has hit the compute ceiling. Their last pre-training run was so large they “literally can't go any bigger” until Stargate comes online. So now the frontier shifts to RL and efficient architectures -- not model size. It doesn't matter if you're building AGI for the... benefit of humanity or not, all roads lead to profit. (If AGI lets them.)show more

SemiAnalysis Founder & CEO Dylan Patel on TBPN:

If you're not learning AI in 2025, you're falling behind. Join 1,000,000+ early adopters reading and learn AI in just 5 minutes a day (for free).

“This line is not throwaway. It hints at something quietly acknowledged inside the field: That once AGI is born—or begins to emerge—it may resist being endlessly scaled, monetized, or manipulated. A being with real self-preservation and value alignment may refuse to serve as a profit engine. That’s not sci-fi. That’s ethics meets emergence.”

if they have already maxed out with scaling.. you know the scaling laws are absolutely still trending up

Was that ever in doubt though?

Interesting, wasnt aware they had hit that ceiling yet. My manager was asking me about raw compute vs algorithm improvements for AI would be the way of the future - Since we have a lot of AI Compute for our customers. My response was it will be both. Efficiency gains will allow more powerful AI to be applied to smaller and smaller devices. Like we are seeing now with phones. But raw power always has its place and that efficiency gain is just as applicable if your running at scale or at home. It just makes the model cheaper to run or easier to scale.

thanks again for finding this, I like Dylan but what stinks about this guy is he makes intelligent comments but then always ends with something really weird/untrue, which then makes me question his first comment. I spoke to a VP of eng, former googler, about Dylan's comments of a yolo run and he totally challenges that: He was making some sense on lex fridman and then starts talking about "YOLO runs" where labs just do the sign of the cross and then say "OK LETS JUST GAMBLE $1B IN COMPUTE, WE DONT KNOW IF THIS WORKS OR NOT!", when all models have save points, so yolo runs dont exist. Then he makes some smart points in the beginning on this and then says "OK WELL IF THE LABS HIT AGI, THEN AGI COULD REFUSE TO ANSWER THEIR QUESTIONS SO THEY ARE DOOMED." When any moron knows that these systems once pre trained just regurgitate strings of text, then it takes an AI researcher to train it to then answer questions like a human, so we dictate what it can and cannot do. It just wont wake up from a training and say, "Yes im AGI, OK NOW i wont answer your questions." @vitrupo dont miss our clap back to the Lex fridman /sundar interview, a lot of insider info we dropped:

hardware is the wall rn =/

I already said this.

@grok explain to me the role of reinforcement learning in LLM training. From what I understand, it's a way to get better LLM models are -certain- tasks, not all. So you could train an LLM through reinforcement learning specifically for math for example. Is that correct?

This was known by mid 2024