Clear negative correlation between accuracy and reasoning gap. This goes directly against the hypothesis that larger models are more contaminated.

Best news for largest language models in a long time!

WTF is going on with Mistral Large 5 shots without CoT?

Next week I'll run a model on all the conversations of the week to estimate (usefulness, time saved or lost in minutes) so that I can compute # of humans saved / week by Dust users :)

Semantic search is powerful but bad at quantitative questions (by construction).

To circumvent that, we built Table Queries📓

Any structured data in your company (Google Sheets, Notion DBs, CSVs...) gets turned in to JIT in-memory sqlite DBs that models can query using SQL👨‍🏫

People ask me (often with a weird look):

'But dude, why did you leave OpenAI ??'

This is why 👇

We made two hard bets with Dust:

- An horizontal platform with access to all the SaaS relied on by our users (Notion, Github, Slack, Drive, Intercom, ...)
- Not one Assistant, but many Assistants specialized on specific tasks.
- Capability to do semantic rertieval but also…

Anybody tried to make models play chess against one another in standard algebraic notation?

We know models are quite good at it. But who wins?

Mistral-Large vs Claude 2 vs Gemini 1.5 vs GPT-4