🚨Job Alert 🚨: My team at FAIR / AI at Meta (AI Research Agent) is now looking for additional contracting power! This time we are looking for ml-ops talent. If you are interested and based in the UK, please fill in the following form. Know someone? fwd pls🙏 docs.google.com/forms/d/e/1FAI…

Technology = power. AI is reshaping power — fast. Today’s AI doesn’t just assist decisions; it makes them. Governments use it for surveillance, prediction, and control — often with no oversight. Our new paper proposes some ML safeguards to resist AI-enabled authoritarianism:

We need a name for when you want to do something, pick up your phone to do it and then forget what you wanted to do.

Language models can be prompted or trained to verbalize reasoning steps using Chain of Thought (CoT). Despite prior work showing such reasoning can be unfaithful, we find that around 25% of recent papers that use cot still mistakenly claim CoT as an interpretability technique.

New paper on #CoT and #Explanations! There are two types of AI Safety researchers interpreting this: 1. “LLM outputs are not aligned with their internal representations.” 2. “#AI is so powerful that it is aiming to lie to us to pursue internal goals.”

🤩We are hiring a postdoctoral researcher to work on AI safety and multi-agent safety. This is a fully funded position in collaboration with Toyota Motor Europe (TME). More info and apply here: tinyurl.com/5c9jjcdy Closing 5th September Adel Bibi Engineering Science, Oxford

Come join us; I am hiring a postdoc position to work on agentic safety/security. You will be working with many amazing folks here in Oxford, in the industry, and in government.

Curious to know how consistent this is. Navigating this on-the-fly environment without any traces of historical past generations ought to have memory limitations with infinite loops, inconsistent U-turns, etc.

Super excited about this event! I will give an updated version of my talk on the Simulation Hypothesis - i.e. Machine Learning in the upcoming era of extremely fast computers. How can we do science that stands the test of time when compute capacity is accelerating?

It’s not every day that ones work is featured by the Scientific American Scientific American. Great work led by Lukas Aichberger and congrats to all co-authors. scientificamerican.com/article/hackin…

The New Frontier of AI Hacking—Could Online Images Hijack Your Computer? philip Adel Bibi & Alasdair Paren spoke to Scientificamerican about AI Agents. You can read the article here: tinyurl.com/ewpmtffe

I met Samuele Marro on a trip to London last year. We got coffee near DeepMind and talked about multi-agent systems, philosophy, and resisting centralized control. That talk led to today’s launch of Institute for Decentralized AI—a new Cosmos Institute project to advance protocols, standards, and

I have one slot to fully fund a visiting researcher position in #Oxford to work with me on the security and anomaly detection of #decentralized #multi #agentic #systems. You will also be working with amazing folks here in Oxford too! If this is of interest, please apply now!

Meet our next ML in PL Conference 2025 speaker Adel Bibi (Adel Bibi)! Adel Bibi is a senior researcher in machine learning and computer vision at the Department of Engineering Science of the University of Oxford, a Research Fellow (JRF) at Kellogg College, and a member of the

How can we efficiently monitor LLMs for safety? Strong monitors waste compute on easy inputs, but lightweight probes risk missing harms ⚠️ 𝙏𝙧𝙪𝙣𝙘𝙖𝙩𝙚𝙙 𝙥𝙤𝙡𝙮𝙣𝙤𝙢𝙞𝙖𝙡 𝙘𝙡𝙖𝙨𝙨𝙞𝙛𝙞𝙚𝙧𝙨 (TPCs) address this by generalizing linear probes for dynamic monitoring! 💫

Amazing new work by James Oldfield on inference scaling on monitors using polynomial approximations. If something’s clearly harmful, why waste expensive monitors? A simple polynomial approximation can flag it with its early monomial computations.

Looking forward to speaking in Poland next week!

We have 3/3 papers accepted at #NeurIPS this year — two on safety and security (lots of exciting new jailbreaks of agents coming soon 😉), and one on the science of benchmarking agentic systems (we’ll post details on all papers soon!).

I have one open position on multi-agent security — the deadline is near! 🚨

🚨How do we improve long-horizon reasoning capabilities by scaling RL with only existing data? Introducing our new paper: "h1: Bootstrapping LLMs to Reason over Longer Horizons via Reinforcement Learning"🫡 > RL on existing datasets saturates very quickly > Reasoning over