emaadm

> We are already near the limits of what we can do

Hard disagree. If I had a million Claudes worth of compute I'd be livestreaming my entire reality feed to a local server 24/7 and having it organize my observations and thoughts, synthesize new ideas, implement prototypes and discard infeasible ones while I sleep. If you're in the business of knowledge creation, a million Claudes isn't enough. Text is an easy modality, I want foundation models that operate on text, images, audio, video, streaming point clouds, ...

> Thinking Machines released a product that helps developers adjust A.I. models to better perform specific tasks. Offered to businesses and software developers over the internet, the product was similar to technologies from Google, Amazon and Microsoft.

I cannot take this article seriously.

Stopping Agents [1] - LLM agents that stop conversations early and save time.

Because they're trained using imitation learning instead of RL, they're scalable and easy to deploy with your own data (also open-source!).

Mainly targeted at and tested on quickly disqualifying prospects in sales calls, but can be applied more broadly.

[1] https://stoppingagents.com

OpenAI has had a fine-tuning API since GPT-3.5, and a reinforcement fine-tuning API since last year.

There's some interesting work in NeurIPS this year on fused kernels for MoE too: https://flash-moe.github.io/