In this video, I'll be deriving and coding Flash Attention from scratch. No prior knowledge of CUDA or Triton is required. Link to the video: All the code will be written in Python with Triton, but no prior knowledge of Triton is required. I'll also.

There’s also a bit of ERPO 🤌🏽.

New alignment technique dropped: RLEF.

RT @MistralAI: Introducing Devstral Small and Medium 2507! This latest update offers improved performance and cost efficiency, perfectly su….

Recently asked ChatGPT to generate a picture of me and it added some private info about me in the picture that were not requested, taken from memories (e.g. my street name). Sometimes they're small details and you risk leaking private info.

How it started vs how it’s going

Looking for top AI Scientists & Research Engineers (incl. interns!) for my team:. 💻 PyTorch / JAX.🧠 LLM / robotics / multimodal / diffusion.📚 Strong research or project track record.🎓 PhD / MSc or exceptional background. Apply now:

It’s sad to admit but Soham Parekh can make millions selling courses on interviewing skills. It’s a @cluely market: cheating pays. Apparently a lot.

TLDR: solve problems.

pip is to uv what Edge is to Chrome.

Learning is easy. Unlearning all the things that prevent you from learning is hard.

Talking to LLMs is all fun until you need to talk to customer support.

Every time you ask ChatGPT to rewrite the answer using a higher quality model (for example 4o -> o3), you're implicitly generating a signal they'll use to improve their weaker model on your prompt.

UK government launches 54 million £ fund (over five years) to attract “global talent”. 54 million £. That’s half the signing bonus Meta is offering to get OpenAI talent.

Wife is 10,000km away experimenting with diffusion language models. Somehow the model keeps generating “import Umar”. No matter how ill-conditioned the loss landscape, true love always finds a way to converge.

Started practicing QiGong outside and felt shy at first. Thought people were judging me - turns out, it was just me. People are actually just curious. With few days of persistence, the inner voice faded away. Now it feels like flying. Most things in life work the same way.

What’s the equivalent of Hulk’s pants in deep learning? . GPT4.5:

You know the prompts. The issue is you’re generating too few tokens. Generate more output. It comes with prefix caching: that’s how you compound learning. KV-Cache decays over time. Make sure to prefill with quality tokens, not TikTok shorts.

Back pain since many years. Nothing worked like QiGong. I practice it every morning since months, outside in the open air. Let people stare at you, they’re just curious. This is the only channel you need to get started:

Who's building cluely for gf/wife questions?.

99% of people: . 1 - Start something complex.2 - Give up because afraid of not making it.3 - Go for something simple.4 - Give up because boring and afraid not benefiting career. Repeat forever.