Physical orders from Manning will ship in 1-2 weeks, and Amazon a week or so after. Thanks for your support! Website: rlhfbook.com Buy on Amazon: amzn.to/4cwCDJQ Buy on Manning: hubs.la/Q03TsMBq0 Course landing page: rlhfbook.com/course Course Playlist: www.youtube.com/watch?v=jQPi...

RLHF Book: Reinforcement Learning from Human Feedback and LLM Post-Training

A free online book and course on RLHF, preference tuning, reward models, RLVR, and post-training language models.

rlhfbook.com

How distillation is used today and what performance uplift it gives to open models (a rant) natolambert.substack.com/p/how-distil...

How distillation is used today and what performance uplift it gives to open models

This is in response to Ben Thompson’s recent piece where he said distillation is happening during RL and becoming more important to model performance.

natolambert.substack.com

Rght now American companies need Chinese models to secure their cyber infra due to guardrails on closed models. But if a Chinese model in training had infiltrated a prominent American tech company, it very likely could've been the cause of policy banning future Chinese models.

Incredibly proud of the Sakana AI team. We have developed an orchestration model right here out of Japan that achieves state-of-the-art performance on real-world cybersecurity benchmarks! 🎌 sakana.ai/fugu-cyber-r...

Sakana AI@sakanaai.bsky.social · 2w ago

Introducing Fugu-Cyber: an update to our Fugu orchestration model. It achieves state-of-the-art performance on real-world security benchmarks, matching cyber-focused frontier models like GPT-5.5-Cyber and Mythos Preview. Release Blog: sakana.ai/fugu-cyber-r... 🐡

Wellington, Vanguard, Blackstone launch funds targeting wealthy investors reut.rs/3T5fKGm

Wellington, Vanguard, Blackstone launch funds targeting wealthy investors

Wellington Management, Vanguard and Blackstone are launching two funds that will offer investments in both public and private markets for a growing ​cohort of wealthy individuals, the companies said on Wednesday.

reut.rs

The goal with my rlhf book is to make the "home on the internet" for the next generation learning post-training. That's why I'm doing all formats (lectures, code, book, discord, model completions... & ofc blog of interconnects). A hub is more lasting than non-fiction writing. Join! rlhfbook.com

I also made a page with extra resources (talks and books I recommend): rlhfbook.com/course#extra...

RLHF & Post-Training Course by Nathan Lambert

Free course lectures on RLHF, reward models, preference tuning, RLVR, and modern LLM post-training.

rlhfbook.com

On-policy distillation is on track to be a lasting method in post-training. The list of areas would be: Instruction tuning (SFT/IFT) RLHF Direct Preference Optimization (DPO et al) RLVR On-policy Distillation (OPD) New classes of methods are rare! Excited to play.

Check out our paper, technical blog post, and open-source code. Code: github.com/SakanaAI/spa... Blog: pub.sakana.ai/sparser-fast... Paper: arxiv.org/abs/2603.23198

GitHub - SakanaAI/sparser-faster-llms: Cuda kernels for leveraging LLM sparsity to improve throughput and decrease the memory requirements during inference and training.

Cuda kernels for leveraging LLM sparsity to improve throughput and decrease the memory requirements during inference and training. - SakanaAI/sparser-faster-llms

github.com

Great telling of the sights when visiting China’s AI and robotics companies (the same trip I was on!). open.substack.com/pub/ailibrar...

We Spent 10 Days Touring Chinese AI Labs. Here’s What We Saw.

Sleeping Cots, Robot Pharmacies, and the Race for AGI

open.substack.com

Reproducing all of Jürgen Schmidhuber’s papers (1990-2025) using an AI coding assistant. Cool project by Yaroslav! It even reproduced the “World Models” paper by me and Schmidhuber (2018) using a toy environment, with a full VAE + RNN world model implementation. Project: github.com/cybertronai/...

Visiting most of the leading Chinese AI labs, I'm struck by a culture that's extremely well suited to building LLMs with fewer resources, but one happening in a very different ecosystem, more companies at play, almost no data industry, etc. Full report: www.interconnects.ai/p/notes-from...

Notes from inside China's AI labs

Lessons from my trip to talk to most of the leading AI labs in China.

interconnects.ai

Excited to launch the accompanying free RLHF Course for my book. To kick it off, I've released: - Welcome video - Lecture 1: Overview of RLHF & Post-training - Lecture 2: IFT, Reward Models, Rejection Sampling - Lecture 3: RL Math - Lecture 4: RL Implementation Landing page: rlhfbook.com/course

Course | RLHF Book by Nathan Lambert

Course lectures and talks on RLHF and post-training.

rlhfbook.com

Most obvious immediate impressions on coming back to the US from China. 1. The cars here are so lame. So many cool EVs in China, feels like I went 2 decades back in time. 2. The coffee here is so much better. First real coffee in almost 2 weeks changes you.

Mac Mini M4 is becoming the default OpenClaw box — unified memory means you can run 24B+ models without a GPU. Full guide on running LLMs on M-series: insiderllm.com/guides/running-llms-mac-m-series/ #OpenClaw #LocalAI #AppleSilicon

Clouds rush to deliver OpenClaw-as-a-service offerings

Clouds rush to deliver OpenClaw-as-a-service offerings

As analyst house Gartner declares AI tool ‘comes with unacceptable cybersecurity risk’ and urges admins to snuff it out If you’re brave enough to want to run the demonstrably insecure AI assistant OpenClaw, several clouds have already started offering it as a service.…

dlvr.it

Here’s this week’s Better Offline, I’m joined by @davidgerard.co.uk for Hater Season to talk about the stupidity of Openclaw, how seemingly smart people keep being “one-shotted” by AI, and why we’ve been headed toward a calamity in the markets since 2008. podcasts.apple.com/us/podcast/b...

Hater Season: Openclaw with David Gerard

Podcast Episode · Better Offline · 02/04/2026 · 38m

podcasts.apple.com

Notion’s new AI Agents will basically do your job for you

Notion’s new AI Agents will basically do your job for you

The AI Agents in Notion 3.0 build pages and databases for you, and more. 

buff.ly

A trio of systemic weaknesses within AI agents leave them open to abuse. Experts worry it is only a matter of time before millions of dollars are stolen because of them

Why AI systems might never be secure

A “lethal trifecta” of conditions opens them to abuse

econ.st

I'm back in Computer Weekly with this feature examining the pros and cons of AI agents for cybersecurity... “If I’m a criminal and I know you’re using an AI agent... for me, that’s a way into the network to deploy ransomware” www.computerweekly.com/feature/Are-... via @computerweekly.bsky.social

Are AI agents a blessing or a curse for cyber security? | Computer Weekly

Agentic AI is touted as a helpful tool for managing tasks, and cyber criminals are already taking advantage. Should information security teams look to AI agents to keep up?

computerweekly.com

A wave of startups are creating RL environments to help AI labs train agents. It might be Silicon Valley’s next craze in the making.

Silicon Valley bets big on 'environments' to train AI agents | TechCrunch

A wave of startups are creating RL environments to help AI labs train agents. It might be Silicon Valley’s next craze in the making.

techcrunch.com

🚨Our NeurIPS 2025 competition Mouse vs. AI is LIVE! We combine a visual navigation task + large-scale mouse neural data to test what makes visual RL agents robust and brain-like. Top teams: featured at NeurIPS + co-author our summary paper. Join the challenge! Whitepaper: arxiv.org/abs/2509.14446

Mouse vs. AI: A Neuroethological Benchmark for Visual Robustness and Neural Alignment

Visual robustness under real-world conditions remains a critical bottleneck for modern reinforcement learning agents. In contrast, biological systems such as mice show remarkable resilience to environ...

arxiv.org

NEW: Weekend News Agents No-one can be certain how transformational AI will be:to the economy, to human psychology, to the human experience. I am absolutely certain our politicians haven’t thought about it enough or begun to properly grapple with the economic, regulatory and political implications

Figma made its design tools more accessible to AI agents

Figma made its design tools more accessible to AI agents

The next phase for Figma’s MCP server.

buff.ly

Microsoft is filling Teams with AI agents. Dedicated Copilot agents will join meetings and answer questions in channels, as part of Microsoft’s latest AI push www.theverge.com/news/781278/...

Microsoft is filling Teams with AI agents

A chatbot for every channel.

theverge.com

Read “Privacy, Cybersecurity, and Generative AI: Navigating the Digital Frontier“ by Vikram Jha on Medium: medium.com/agentic-ai-s...

Privacy, Cybersecurity, and Generative AI: Navigating the Digital Frontier

Stop Hyping Generative AI: We Need to Talk About Privacy and Security

medium.com

Fed up with Meta? Avoiding Instagram or Facebook isn’t enough to stop Meta from harvesting and profiting from your private information. Here’s how to limit Meta’s ability to monetize your personal data.

Mad at Meta? Don't Let Them Collect and Monetize Your Personal Data

If you’re fed up with Meta right now, you’re not alone. Meta tracks you across millions of websites and apps and its business model relies on your data. If you want to limit Meta’s ability to collect ...

eff.org

Microsoft’s Magnetic-One multi-agent framework, capable of interacting with file systems, web browsers, and terminals for coding, could supercharge Power Automate Desktop in the future. Just a guess! 🤔 #AgenticAI #AgenticAutomation #GenerativeAI #LLM #PowerAutomate #Microsoft #MSBuild

Magentic-One: A Generalist Multi-Agent System for Solving Complex Tasks - Microsoft Research

By Adam Fourney, Principal Researcher; Gagan Bansal, Senior Researcher; Hussein Mozannar, Senior Researcher; Victor Dibia, Principal Research Software Engineer; Saleema Amershi, Partner Research Manag...

microsoft.com

How agentic AI handles the speed and volume of modern threats www.helpnetsecurity.com/2024/10/31/l... #Infosec #Security #Cybersecurity #CeptBiro #AgenticAI #ModernThreats

How agentic AI handles the speed and volume of modern threats - Help Net Security

Agentic AI can actually enhance privacy protections because it doesn’t require a human to review sensitive data.

helpnetsecurity.com

Chain of Thought (CoT) reasoning in multi-agentic #AI helps different AI agents solve problems step-by-step by sharing their thoughts and progress, similar to humans discussing and breaking down complex issues arxiv.org/abs/2201.11903 #AgenticAI #GenAI

Chain-of-Thought Prompting Elicits Reasoning in Large Language Models

We explore how generating a chain of thought -- a series of intermediate reasoning steps -- significantly improves the ability of large language models to perform complex reasoning. In particular, we ...

arxiv.org