Leopold has the opportunity to do the funniest thing rn (tell us what Ilya saw)

Eval numbers for the new GPT-4 Turbo are out

currently zooming in 100x to count pixels to find out which lines are going up or down b/c no labels

twitter.com/i/spaces/1OyKA…

The moat of software AI agents is not the thin wrapper layer (Devin, SWE-Agent), but the underlying LLM. Instead of benchmarking the wrapper, I think SWE-Bench is excellent for evaluating coding LLMs instead:

Hold the agent layer fixed and vary only the LLM backend. Provide all…

This is the first music generation AI that has music quality high enough I could actually listen to it. Incredible.

the big labs are gonna keep one upping each other until someone casually ships AGI with a name like 'gpt-5-ultra-2025-preview' or 'claude-5-symphony-20241019'

Ok so from really early tests the new gpt4 definitely feels better at coding

Less lazy, more willing to write code. Was able to give it a few file, and it wrote perfect code (very uncommon before)

Might be switching away from opus.(gpt4 is cheaper & works better with cursor)

Majorly improved, but not even worth a blog post?

So just to recap this 🔥 Tuesday:

- Google refreshes 1.5 Pro, with sound understanding
- Google releases Gemma Code 3 models
- OpenAI gave us GA for vision + updated model
and now
Mistral AI doing torrents again, with 170B MoE

This is a day after cohere beat GPT-4 🚀

gpt4 update quite notable

rly miss ilya 💔

does anyone have any examples of the new gpt-4-turbo model and how it's better?

supposedly better at reasoning/math but haven't had a chance to test this yet