Sublime

According to this research it sounds like Claude uses the same tokenization trick I invented for ULMFiT in 2017! :D

Proud to report that my friend Ron Arel has created an Artificial Artificial Intelligence researcher

My PhD student has reported that all writing by ChatGPT is improved by adding "Don't use adjectives." at the end of the prompt. We are soliciting feedback on this prompting strategy.

launch anything https://t.co/GCrwDiKeaE

Things are heating up 🔥🔥 Jan-nano by Menlo Research: a 4B Qwen3 finetune (using DAPO) for deep research, tool use, and search extraction On SimpleQA (agentic/MCP), Jan-nano scores 80.7 🤯 https://t.co/AcyOgyMbmZ

My Deepwriter already can write entire PhD dissertations with gpt-4o-mini (very cheaply) that are 120 pages or more in perfect LaTeX even if they are filled with advanced visual trees, equations, etc., and most importantly: advanced idea generation! I tested Deepwriter with the API of o1 models (which was a pain because... See more

A concept I would really like to test with specialized pretrain, that was impossible in fine tuning: poetry generation with byte/letter level tokenization. Rimes, cadences, assonances all clearly require that.

🚀Enter the new era of API programming. Meet Gorilla, a new Large Language Model that can generate accurate API calls, outperforming GPT-4 AND Claude. Imagine doing all API calls just by using natural language. You need to check this out now: https://t.co/5FjCnOUkBO