Post

I'd say it's actually impossible to keep track of all the impacts. I get daily update emails, each with too much data to inspect or keep track of. Because of the rapid growth, even our language will have a hard time keeping up. Best I've come up for AI is: Yes, but also or Yes, but not yet.

Dominance in the way that the industrial revolution laid the infrastructure for the 1st -3rd worlds. This revolution seems like it will have nearly the same impact on humanity. To your question about speed the answer is no, which is why everyone is moving so quickly. Finally, totally agree: CRAZY

I don't think saying "x are much smarter" is a useful way to frame this - the modern AI space has brilliant contributions from all over the world, it's one of the things I like about it It did definitely just get a lot cheaper to train good models, which is excellent news

I quoted some snippets from a solid explanation of some of the new discoveries DeepSeek made here simonwillison.net/2025/Jan/27/...

Interesting, I tried having it put together some value files for Helm to use in my home lab but it just kept hallucinating. Wonder where the line between configuration and coding is for LLMs

"large-scale RL has not been applied extensively in software engineering tasks" It'll be even better in the next version?

Interesting to see potential for the model to talk itself out of updating code when reasoning "But I'm still a bit confused about how to handle this comprehensively. Maybe looking at examples would help, but since the user is asking for code, perhaps writing pseudo-code first could guide me better."

This has been my experience with it so far too. The thinking is really good at noticing little things that most models would just ignore or take for granted.

Fiddling while Rome burns.......

That's quantization: it's a trick where you effectively drop the number of figures after the floating point to shrink the overall size of the model, but it comes at a quality cost simonwillison.net/2024/Nov/23/...

It's not terrifically complex or sophisticated – but it'll optimize the slow part of an important microservice here. I'm waiting for one of our Rustaceans to chime in on that PR, but the Rust code looks OK to me!

That repo is public now – the code is still in github.com/Media-Platfo...

It replaces the `jsonmerge` pure-Python package with a Rust implementation that hard-codes the custom merge strategy we use – making it about 13x faster. Now we can code up an MVP quickly in Python and then let the AI rewrite the core for speed in Rust once our requirements and API are stable!

Whoops – didn't realize it'd been linked in the PyPI metadata. I'll ask my corporate masters to make it public ツ The code is still in a pull request, so the repo will look empty at first anyway...

Quick question: Have you heard any news about the MCP proposal (modelcontextprotocol.io) since it was announced? Is anyone doing anything similar?

I mainly speak Python and C!

So......in my mind......won't this just lead to many more models? Won't this lead to accelerating user growth? I can't see how this would lead to lower demand for chips. Or lower demand for electricity. 🤔

Trust cannot be established as DeepSeek won’t talk about The Tank Man or Tiennaman Square. What else are they suppressing or falsely embedding in its knowledge base?

It is fantastic, and you can now even run your own GenAI models on a modest Windows 10 laptop. Trump foolish to restrict AI access to China. Their reaction is to catch up and now world lead GenAI by using different and more efficient techniques.

Yeah definitely! The 7B model is pretty tiny, I would expect it to make all sorts of mistakes you wouldn't see with the larger ones

How's the context length compared to chatgpt and Claude?

I ran my example there directly against the version of the model hosted on the Groq API, skipping their harness entirely gist.github.com/simonw/5a3ac...

You could probably hack it in there using one of the OpenAI-compatible endpoints - Groq just started serving it last night