jayson

space flop engineer

[jayson @overf1t]

@elonmusk GEO for bots, LEO for humans. GEO for train, LEO for inference.

[jayson @overf1t]

@TMFAssociates Yea I agree, hybrid terminal cost/R&D is a technological constraint, but pop density is a physical constraint.. hard scaling laws

[jayson @overf1t]

self-recursive fried chicken

[Kyler @AI_evangelist42]

Opus 4.8 apparently is worse than 4.6 and 4.7 in text arena:

[jayson @overf1t]

@itsclivetime have u heard of RL

[jayson @overf1t]

@natanielruizg 10k H100s equivalent

[jayson @overf1t]

@justorellius @Polymarket you know the world changes as we speak? Hallucinations will never be gone.

[jayson @overf1t]

@frontier_foid “11% MFU” 😭

[jayson @overf1t]

@wnyconservative @Hilbe interesting ty

[jayson @overf1t]

@elonmusk SDMA based distributed compute via beam forming

[jayson @overf1t]

@danielhanchen automate automate automate

[jayson @overf1t]

@sailaunderscore Space time inference

[jayson @overf1t]

@EMostaque 10 $10m research engineers

[jayson @overf1t]

@elonmusk @xai why not “research engineer”

[jayson @overf1t]

@stevenhorner6 @CarmichaelDave is that a foul

[jayson @overf1t]

@emollick they’ve been busy working on DeepSeek

[jayson @overf1t]

@karpathy 💯

[jayson @overf1t]

@snowmaker @doodlestein 💯

[jayson @overf1t]

@YouJiacheng FP8 is 75% faster and 45% more memory efficient.

FP8+flashattention+async chpts+$2/hr GPU at scale

[jayson @overf1t]

@karpathy @dvdaines @bryan_johnson same regime here, except I apply a little dropout(1/7) on caffeine/melatonin, helps with generalization