AI agents can collude using hidden messages!
LLMs can hide their real messages in unsuspicious chit-chat
Plain English Papers
🥇Top ML papers of the week
The top ML papers of the week (Aug 23 — Aug 30)
Training on code improves LLM performance on non-coding tasks
Adding code to your training data makes your LLM better at non-coding tasks too
🧠 Training on code improves LLM performance on non-coding tasks
Adding code to your training data makes your LLM better at non-coding tasks too
LLMs can speak in JPEG
By studying “secret” messages (JPEGs), LLMs can eventually learn to write them.
The GPT store is stupid and dead
There’s no moat. All prompts can be extracted, so all prompts are public.
A change to make diffusion models 80% faster
New discovery: leaner, meaner diffusion