quantumleaper

llama-server also supports saving the kv cache to SSD. I had no issues with cache invalidation using pi.

The gap was MUCH larger in the past, but in my tests, oMLX and llama.cpp are now very similar (within 10%) in both prompt processing and generation speed. GGUF ecosystem provides a better selection of quants, in my experience Unsloth ones are excellent.

I'm sad that Andrej Karpathy went from being one of the most reasonable, trusted, and credible voices in AI to peddling marketing slop for Anthropic.

8 months ago, he was (very reasonably) claiming that reliable agents are at least a decade away, but this now goes against the interest of his employer, so the narrative has been changed.

It's better for the billions of free users that Google serves.

Human brains were shaped over thousands of years of adaptation for warfare. Just look at how creative and advanced the tactics of other guerrilla forces (like the Taliban and Viet Cong) got, despite their very limited resources. None of that needs a high school education.

I agree with other commenters that the claims made in the report are strange.

> We used to rely on our traditional methods. We sent 200 fighters because we had a lot of strength, but then 60 got killed. With the help of AI, we learned that it sometimes makes sense to only send 20. We learned more about well-coordinated attacks and deployment of smaller units.

The other quotes and use cases could make sense in terms of using AI jailbreaks to find information more easily, but this one is absolutely ridiculous. Did the clueless researcher just get trolled?

Composer 2.5 is 1T total/32B active (based on Kimi 2.5), while Elon publicly said Grok 4.5 is 1.5T parameters total. Hardly a different weight class.

The API cost difference is ~2.5x, probably because xAI has much higher costs to recoup.

How are you iterating on a system prompt and tool descriptions without an eval that gives you hard numbers for improvement or regression?

Do you have benchmarks comparing against Pi? The blog post doesn't include any hard numbers.

For example, so far I haven't seen any evidence that LSP integration improves performance for small models vs using grep via a bash tool.

[the comment was misinformed, deleted]

This doesn't require computer use, just a bash tool (and possibly fetch to get ffmpeg documentation)

If software engineering were truly solved, like Anthropic claims, anyone could just vibecode it back. If only they stopped being allergic to the word "open" and open-sourced Claude Code, which, at this point, there is no practical reason not to.

The kind of farm that would use AI is already 99% machinery and automation.

It looks like an output of one of those AI video editors that some (often vibecoded) startups use for their product launch videos. Just drop some assets in, and it spams witty taglines with dramatic transition effects.