an update: I’ve left AISI to focus on independent writing / advocacy for the next few months
it increasingly feels like The Big AI Thing is getting close, and I wanted the freedom to comment on that. I’ll be aiming to post ~weekly on my blog: https://t.co/prndz1yAa9
Incubator Week is back!
We've been quietly recruiting for it over the past few days and are now launching publicly.
If you have been thinking about what needs to be built to make AI go well, apply by May 26th.
We'll help you find the right idea and co-founder.
You will get hands on experience with classic mechanistic interpretability methods and build strong intuitions for how AI models can represent and transform information.
Prizes: $1,000 / $750 / $500 / $250 honourable mentions
Submissions close June 12: https://t.co/euFmPt60Ip
The linear representation hypothesis says neural networks encode concepts as directions in activation space.
We trained a small model where 7 of 8 features behave this way. The 8th doesn't.
$2,500+ in prizes to whoever can tell us how it's actually encoded. Bonus points if you can train a model with an even weirder representation.
Link in thread 🧵
The features are simple text properties: is-a-question, mentions-a-food, contains-a-person's-name, etc. Your 3 tasks are:
1. Identify which feature is not represented linearly
2. Characterise its geometric structure
3. (Bonus) Train your own model with an even weirder feature encoding
GPT-5.5 just scored higher than every PhD virologist on wet-lab troubleshooting.
Want to help build the defenses before the open weight version of it ships in a few months?
Come to our AI x Bio hackathon this weekend in London. Travel and hotel covered. $9k+ in prizes. Link 👇
Want to fast-track your technical AI safety career? Apply to our course by 26th April!
Our alumni work at Anthropic, Google DeepMind, UK AISI, Apollo, Redwood, MATS and more. We have given top talent introductions to top orgs over $500k in grants to build portfolios and build impactful careers.
The course is facilitated by experts to help you navigate the safety landscape and transition into impactful safety work and it is widely recognised by major AI safety orgs and fellowships. You can either do the intensive version over 6 days or part-time over 6 weeks to fit around your job!
🔗https://t.co/fvwMPy9YiL
Over the last few months we've given out $50,000+ in rapid grants to 77 people, funding research, fieldbuilding, early-stage projects and lots of career acceleration.
Now we're looking to triple that. Rapid grants go up to $10k with decisions made in under a week.
More in 🧵
We are hosting an AGI teach-in with @willsaunter, co-founder of @BlueDotImpact, next Tuesday lunchtime in Westminster.
We have a few places remaining, so if you want to be walked through all the need-to-knows of this new technology sign up here:
https://t.co/gnanMG82Tu
Mia Hopman just joined @apolloaievals as a Member of Technical Staff.
When she applied to our Alignment course in 2024, she was a data scientist at a medical device company. She wrote in her application: "I believe my skills could have a greater impact in the AI safety space, as opposed to applied ML in the medical device area."
The course was her first organized step into AI safety. She quit her job after completing it and went all in. Next came Cambridge AI Safety Hub, Tutke, LASR Labs. And now Apollo.
Both roles are SF-based (we sponsor US visas), in-person, with serious autonomy and unlimited PTO. If you're high-agency, care about the mission, and have a track record of building things that work - or if you know someone who is - reach out or directly apply here: https://t.co/yOJcsNg1bX
Community Lead - Turn course participants into lifelong collaborators. The people who take our courses could shape how AI plays out for humanity, but a course alone isn't enough. You'll make sure they meet the right people, stay engaged, and leave fired up to build. $110-150k.