light

learning from first principles

[light @reprompting]

This is the roadmap I've put together for the next few months. While I already have a background in ML, I've never really explored the performance engineering or CUDA side of things. It should be an interesting few months ahead.

[light @reprompting]

reading about tritonforge today

https://t.co/G1oTtofWGV

[light @reprompting]

reading about kernelfoundry today

https://t.co/V5oIJhJwDK

[light @reprompting]

i can understand the intent behind it, but the current version of chatgpt, where it loads visuals in addition to text (it should have always been text), takes noticeable time to load, making the UI look buggy/laggy

[light @reprompting]

reading about how warp divergence has evolved across GPU generations

https://t.co/LwLgOciHcI

[light @reprompting]

reading about zipserv today

https://t.co/z1JvEXNzaU

[light @reprompting]

reading about the NVIDIA hopper architecture today

https://t.co/CrPtjPkcId

[light @reprompting]

you would have to be crazy to outsource your cognition to AI agents

[OpenAI @OpenAI]

Introducing dots, powered by GPT-6 Astra.

Remarkably capable, always-on agents built to handle everything.

[light @reprompting]

reading about kernel-smith today (also found out about InternLM through this)

https://t.co/WNw0QNbpkl

[light @reprompting]

I’ve always felt like Rabbit’s form factor was quite good.

[Chris @ChrisGPT]

Metas one more thing moment - they unveiled hardware for Meta Muse (beating OpenAI to an AI hardware device unveil)

Always on w/fingerprint sensor “by far the fastest way to talk to your muse” shipping in time for the holidays

I think OpenAI can one up this but we will see 👀

[light @reprompting]

@AscendNoosphere yeah bookmarks usually end up becoming a graveyard :D

I had thought of making a list but lost enthusiasm along the way. But hopefully soon. haha

[light @reprompting]

interesting read. it provides a thorough walkthrough of triton, starting from the basics and building all the way up to a full implementation of the mamba algorithm.

https://t.co/b16t9tnJ4Y

[light @reprompting]

reading about tritonBLAS today

https://t.co/RPI44VAXOt

[light @reprompting]

@aadityansha_06 nice work.

[light @reprompting]

@WilliFetzner i don’t think slowing down would really help, other than giving more time to improve safeguards while working on alignment

[light @reprompting]

the genie is already out of the bottle and dario is still debating how fast to open it

[Dario Amodei @DarioAmodei]

We Must Pace the Frontier: I’ve written a new essay on why the AI industry should slow down, with a three-part plan for doing so.

Anthropic is unilaterally committing to the first of these steps. We’ll provide third-party evaluators with permanent, employee-level access to our systems, so that they can verify adherence to our safety measures, report on incidents, and assess models’ alignment during training.

You can read the full post here: https://t.co/OGyPb7yaYt

[light @reprompting]

@pavelsimo i think the same. he just wants to make sure nobody can catch up, while positioning anthropic for the upcoming IPO

[light @reprompting]

reading about autokernel today

https://t.co/rA2wx0SUzq

[light @reprompting]

when the going gets tough, the tough get going. i am certain the whale will take us to the moon.

[DeepSeek @deepseek_ai]

🚀 Introducing DeepSeek-V4.1-Flash: smarter, faster, more efficient.

🔹 Introducing the smallest model in our new architecture family, with native visual understanding.
🔹 Designed for greater capability, faster inference, higher throughput, and scaling to larger models.

1/6

[light @reprompting]

reading about squeezed attention today

https://t.co/CxF0letm3U

[light @reprompting]

funnily, most people are interested in it mainly because of the conspiracy involved, not the actual math. people just love gossip.

[OpenAI @OpenAI]

We’re sharing a solution to the Navier-Stokes Millennium Prize Problem, one of the deepest problems at the frontier of mathematics.

The proof was produced by a group of agents, using an OpenAI next-generation model significantly more capable than GPT-6 Astra.

The problem concerns whether the description of smooth three-dimensional fluid motion modeled by the Navier-Stokes equations can break down. It has remained unresolved for roughly 90 years.