7b? Well even though it’s not all you need, you have my attention :)

Oh yes. When I say that data quality is the new scaling law, it’s not quality as 'one metric' but provenance, depth, correctness, diversity, reasoning-rich… And obviously it’s hard.

Taking advantage of the French holidays for nice things to come.

At this point, I’m not sure if we are benchmarking models or masculine insecurities.

For everyone interested in quality data, we may well have a big announcement next week.

Data quality is the new scaling law, but not sure OpenAI got the memo: 3 million calls on a day on the self-avowed 'world's lamest content farm' (!)

🫡

For an open everything model, definitely looking forward to train stuff on 78rpm phonographs.

Soon pitching MoE's eleven to Netflix: extraordinary team of 11 people for the ultimate LLM heist.

Unpopular opinion: big models are great but the most valuable open LLM release will be sub-7B well-optimized models.