Joe retweeted
DeepSeek is the actual Open AI company 💀
🚀 Introducing DeepInfer, our in-house inference framework, now fully open source!
Powering DeepSeek-V4.1-Flash at scale with high throughput and low latency. We’re sharing our production stack to help the community lower compute costs, reduce energy use, and make AI more affordable.
🔗 github.com/deepseek-ai/DeepI…
1/4
Joe retweeted
You should post here with your real voice.
The more you think on what you have to write or use AI is a bad bad decision
Joe retweeted
There are many bad things about the way AI companies have been treating math, but by far the best thing has been how socially acceptable it has become to admit that you are interested in a famous open problem outside of your expertise
Joe retweeted
A video I watched of @stephen_wolfram yesterday he explained this more specifically
It isn’t all code!
And LLMs are good at being an intermediary between humans and computation (a very quick synopsis of Stephen Wolfram)
My completely uneducated guess (intuition?) about mechanistic interpretability is that it's all code. These synthetic emotions: code. Reasoning: code. It’s all code. And if it's code, that means we have decades of knowledge, intuition, strategies, just things that will let us understand it and then modify it. And now that we think that we can compile back and forth from neural networks and, let's call it, “human-readable” code, we will be able to “solve” alignment the same way that we figure out and control determinism in code.
Going to follow this intuition for a bit and see where it takes me (a complete noob with no preconceived notions of what is/isn’t possible)
Joe retweeted
In the next version of fframes you will be able to set various audio effects e.g. compressor, normalizer, noise reduction, and even voice isolation straight from the code.
no need for additional scripts, everything stays in the code definition for video
rust is beautiful:
Joe retweeted
Thinking of buying M1 Airs with a broken screen, 200/250€ as there will be a compute server crunch and these can be used locally with 15W (and are like $4k of GitHub Actions use per M1 per month) but will use them for running FEM/bio/chem/Eng stuff for Anabasis tests
Since I can't win the @Cloudflare GitHub contest due to where I live, I do want to share the idea and maybe spark some thought in people
jovidecroock.com/blog/gittif…
In the short term, there should be more math funding than ever. In the limit, you become a guy who need only ask the questions. You decide how compute is allocated by determining which questions to ask. There will always be a cost. There will always be finite compute.
To allocate it best and ask the right questions, you need to know the math. Therefore there will always a role for conventional-ish human mathematics, at least it will always be useful to have human input on that.
Joe retweeted
The fact is: no user or dev has a moat anymore and these tech bros don’t give a fuck about you



























