Herbie Bradley
@herbiebradley
a generalist agent | AI research & governance | @CambridgeMLG | formerly @AISecurityInst @AiEleuther
im helping organise this, come hang out :)
We are pleased to announce the Technical AI Governance (TAIG) Forum, a free two-day conference in the heart of London on August 16th - 17th, 2025, dedicated to advancing research and collaboration in the field of technical AI governance.
CEO of Aman: “It’s wartime. I’m going wartime CEO mode. I’ll be sleeping at work 7 days a week going forward.”
We already are. Just through a fairly narrow communication bottleneck.
what if I told you that OpenAI, Google, Anthropic and xAI will all be working together in a few years
a timely new report from @BritishProgress ! I think highly-capable AI agents have the potential to transform the way militaries are run over the next decade, so there is huge scope for impactful defense + AI work that ATI can tackle by narrowing its focus
🚨 NEW REPORT 🚨 The Alan Turing Institute has visibly struggled, and is now at a critical juncture. The Secretary of State has intervened. What happens next? A new report for @britishprogress by me and @jujulemons sets out a plan: 🧵
"writing is thinking" yes, but scientific writing is often not; it is writing *for communication* rather than as a tool for thought idk why this article ignores the difference
Yes. Writing is not a second thing that happens after thinking. The act of writing is an act of thinking. Writing *is* thinking. Students, academics, and anyone else who outsources their writing to LLMs will find their screens full of words and their minds emptied of thought.
reposting my operator thread from release day; I think it also holds up well for the implications of Agent Mode
*On Operator*. AI efforts are shifting to the widespread deployment of agents, not chatbots—with startling implications for the economy and geopolitics. The U.S. is likely to possess a strong *economic advantage*, due to its compute advantage. 1/N x.com/OpenAI/status/…
What's missing in the AI safety literature is a cost-benefit framework for evaluating when we should do more AI safety work vs. proceed with AI development. Indeed, one often finds an implicit assumption that we should ~always do more safety work, as if there are no tradeoffs.
Insurance is an underrated way to unlock secure AI progress. Insurers are incentivized to truthfully quantify and track risks: if they overstate risks, they get outcompeted; if they understate risks, their payouts bankrupt them. 1/9
Claude's Notion access tool is huge for me since I keep all my brainstorming there very convenient to name some notion pages as context for a query
classic VC trope: "no-one has figured out X well yet, someone should get on that" meanwhile there are 3 bay area startups and one solo digital nomad hacker in the space, but none of them are high profile on x dot com
My response to AI 2027: vitalik.eth.limo/general/2025/0… The AI 2027 post is high quality, I encourage people to read it at ai-2027.com I argue a misaligned AI will not be able to win nearly as easily as the AI 2027 scenario assumes, because it greatly underrates our…
Tired: force UK pension funds to invest in domestic firms Wired: force UK pension funds to invest in AGI labs
Some of these seem pretty concerning
As AI agents near real-world use, how do we know what they can actually do? Reliable benchmarks are critical but agentic benchmarks are broken! Example: WebArena marks "45+8 minutes" on a duration calculation task as correct (real answer: "63 minutes"). Other benchmarks…
overheard "he just read the entire book? like, rawdogged it? that's crazy"