f(x) @Machine1235

Author of GRPO-Zero library, PhD, AlphaZero in JAX, text-to-speech models, mechanistic interpretability, exploring neural nets' internal circuits... github.com/ntt123 Singapore Joined February 2012

Tweets

529
Followers

236
Following

657
Likes

2K

f(x) @Machine1235

3 days ago

It was unclear to me at first why cute layout with simple inner product between logical coordinates and strides is powerful. The real power is in nested layouts that allow us to specify linear relationships in high dimensions, then project back to low dimensions, which leads to…

0 0 0 29 0

f(x) @Machine1235

a week ago

Humanity's most important job in this century is to guide AGI as it grows, like parents guiding a child. We probably won't stop developing AGI because the potential benefits are too great. Trying to prove mathematically that AGI is safe won't really work, since reality is too…

0 0 0 52 0

f(x) @Machine1235

a week ago

Set out to learn GPU programming; ended up learning Category theory 😅

Charles 🎉 Frye @charles_irl

a week ago

Set out to learn GPU programming; ended up learning Category theory 😅

5 10 191 14K 108

Download Image

0 0 1 94 0

f(x) @Machine1235

2 weeks ago

👀

Sholto Douglas @_sholtodouglas

2 weeks ago

👀

30 5 438 57K 29

0 0 0 93 0

f(x) @Machine1235

2 weeks ago

It was never make sense to me why a lot of people accused model provider to fallback to low bit precision when the load is high, because it would completely erase users trust (likely to be illegal as well). I think there are a few factors that cause users see degradation in…

Claude @claudeai

2 weeks ago

29 20 688 259K 51

0 0 1 107 0

f(x) @Machine1235

2 weeks ago

We usually think of matrix multiplication as a collection of independent inner products between rows of matrix A and columns of matrix B. But this is misleading in terms of GPU programming, matrix multiplication should be considered as sum of outer products between columns of…

0 0 0 32 0

f(x) @Machine1235

3 months ago

👀

Elon Musk @elonmusk

a year ago

👀

158 308 3K 467K 200

0 0 0 50 0

f(x) @Machine1235

4 months ago

LLM + Tool Use + CoT Reasoning = Turing-complete computer.

0 0 1 71 0

f(x) @Machine1235

4 months ago

The problem with social networks: horrible takes get the same megaphone as brilliant insights.

Dmitrii Kovanikov @ChShersh

4 months ago

The problem with social networks: horrible takes get the same megaphone as brilliant insights.

74 69 2K 218K 241

0 0 0 79 0

f(x) @Machine1235

4 months ago

I wonder if DeepMind purposely trained Veo 3 with captions turned on. I can imagine it being helpful for (for example) aligning speech with video frames.🤔

Google DeepMind @GoogleDeepMind

4 months ago

I wonder if DeepMind purposely trained Veo 3 with captions turned on. I can imagine it being helpful for (for example) aligning speech with video frames.🤔

657 1K 8K 1.5M 3K

Download Video

0 0 1 111 0

f(x) @Machine1235

4 months ago

This is a very... entertaining exchange. @EricRWeinstein and @seanmcarroll go at it over theories about the universe, and yet it gets deeply personal.

Piers Morgan Uncensored @PiersUncensored

4 months ago

This is a very... entertaining exchange. @EricRWeinstein and @seanmcarroll go at it over theories about the universe, and yet it gets deeply personal.