🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

Do you still remember this image? The definition of “reasoning” has been significantly changed in the past two years.

Remember reinforcement fine-tuning? We’ve been working away at it since last December, and it’s available today with OpenAI o4-mini! RFT uses chain-of-thought reasoning and task-specific grading to improve model performance—especially useful for complex domains. Take

You can now connect GitHub repos to deep research in ChatGPT. 🐙 Ask a question and the deep research agent will read and search the repo’s source code and PRs, returning a detailed report with citations. Hit deep research → GitHub to get started.

🧵1/ Everyone says toxic data = bad models. But what if more toxic data could help us build less toxic models? Our new paper explores this paradox. Here’s what we found 👇

House Energy and Commerce reconciliation text has language preempting all state AI regulations: "no state or political subdivision may enforce any law or regulation regulating artificial intelligence models, artificial intelligence systems, or automated decision systems during

So apparently someone from Anthropic submitted a report to the court... with a hallucinated citation by Claude.

Princeton engineers have identified a universal weakness in AI chatbots that allows users to bypass safety guardrails and elicit directions for malicious uses, from creating nerve gas to hacking government databases. bit.ly/3SzRto7

There are so many hallucinated citations in court nowadays, that I'm starting to put together a tracker. Check it out and feel free to send ones that I've missed along. New tabs coming for more categories of AI+Law cases!

Introducing a research preview of Codex in ChatGPT openai.com/live

I did have the “feeling-the-agi” moment when using Codex. It works well even in the gigantic OpenAI repo.

I asked Codex to convert a legacy project from Python 2.7 to 3.11 and from Django 1.x to 5.0 It literally took 12 minutes If you know, that’s usually weeks of pain This is actually insane

😱

Sam & Jony introduce io

🚨 New work: We rethink how we finetune safer LLMs — not by filtering after the generation, but by tracking safety risk token by token during training. We repurpose guardrail models like 🛡️ Llama Guard and Granite Guardian to score evolving risk across each response 📉 — giving

🚀 Excited to share the most inspiring work I’ve been part of this year: "Learning to Reason without External Rewards" TL;DR: We show that LLMs can learn complex reasoning without access to ground-truth answers, simply by optimizing their own internal sense of confidence. 1/n

Your LLM Guard Model is secretly a reliable LLM-finetuning-guardrail! IBM Granite Guardian and LLAMA Guard are particularly suited to tracking harmful levels of fine-tuning data at the token level and making training adjustments during fine-tuning Paper: arxiv.org/abs/2505.17196

🤯 We cracked RLVR with... Random Rewards?! Training Qwen2.5-Math-7B with our Spurious Rewards improved MATH-500 by: - Random rewards: +21% - Incorrect rewards: +25% - (FYI) Ground-truth rewards: + 28.8% How could this even work⁉️ Here's why: 🧵 Blogpost: tinyurl.com/spurious-rewar…

The next ~1-4 years will be taking the 2017-2020 years of Deep RL and scaling up: exploration, generalization, long-horizon tasks, credit assignment, continual learning, multi-agent interaction! Lots of cool work to be done! 🎮🤖 But we shouldn't forget big lessons from back

Our tracker of “hallucinated” or nonexistent citations in real-world legal contexts has reached over 140 cases across the world. There’s a notable spike in the last 6 months. 📈📈📈

So far, RL in LLMs has been "RL as a distillation method". - RL helped us distill great verifiers (e.g., code) in the model. - When models are better at verification than generation, we used RL to distill those abilities back to the model. That's about to change with agents!