I am very excited that AISI is announcing over £15M in funding for AI alignment and control, in partnership with other governments, industry, VCs, and philanthropists! . Here is a 🧵 about why it is important to bring more independent ideas and expertise into this space.

From near the end of Sleepwalkers, by Christopher Clark, as World War I starts.

What do you want to know?.

The correct mathematical definition is the one that makes the most intermediate lemmas happen to be true, along the way to the result you care about.

RT @ben_j_todd: The UK govt is now one of the bigger funders of AI alignment research. A new $20m fund was just announced. .

RT @alxndrdavies: I’ve been wanting to see this result for a while! We find evidence that narrow filtering of pre-training data can degrade….

RT @StephenLCasper: 🧵 New paper from @AISecurityInst x @AiEleuther that I led with Kyle O’Brien:. Open-weight LLM safety is both important….

RT @michael_nielsen: I'm quite bothered by the last post. It's written in social media advice speak :-(. The underlying point is, I think,….

RT @AISecurityInst: The more capable models become, the harder it is to safely evaluate them - but it’s essential that AI agents are tested….

RT @alxndrdavies: We at @AISecurityInst worked with @OpenAI to test GPT-5's safeguards. We identified multiple jailbreaks, including a univ….

This is the key point about infinities: the world has a lot of big numbers, and it is often way easier to do mathematics if we approximate those big numbers with infinities of one form or another.

I often see the implicit assumption that if we automate AI safety research, it’s enough to do so much of it that we generate a right answer. But this is not correct: we also need to know that all the other, cheaper, wrong answers are wrong.

RT @danielmurfet: Neural networks are grown, not programmed. What does that growth process look like? Like this!. This is a small language….

RT @activewarp: get in loser, we’re ranking our favorite fp4 representable values

Thank you, Yoshua! Very excited to get to work together in this new capacity!.

RT @matt_levine: This newsletter has been an embarrassment internally, but its failure isn’t a big concern.

RT @juddrosenblatt: Excellent point from @geoffreyirving :. cognitive science, theory, and clever ML experiments are "massively, systematic….

RT @AmmannNora: Very excited to see this come out, and to be able to support! . Beyond the funding itself, the RfP itself is a valuable res….

A final note: I am confident that alignment is solvable, but this does not imply that we'll find the solution in time, nor that technical alignment alone is sufficient for safety. We should try! If we don't succeed, something else that follows us will succeed instead.

Therefore, a core goal of the Alignment Project is bringing in researchers with relevant expertise who aren't yet working on alignment, both via funds and by helping to map the research space. Our research agenda aims to help here, and we would love to talk to interested people!.

There is a lot of risk in trying only one approach to simulating the future! There are only so many SOTA LLM-focused safety research areas. They might succeed, or they might all fail in correlated ways. Spreading our bets helps!.