Cam Tice

I'm extremely excited to bring people on the team who live for making the future of humanity brighter. If you know anyone who you think would be a good fit (or you think you would be), DM me!

"Calling for prudence, rigorous evaluation and even, at times, a slower pace in adopting AI does not mean opposing progress; instead, it is an exercise of responsible care for the human family." - @Pontifex I'm happy this morning to have a new teammate

Great paper on continual learning by @LPacchiardi . We should be far more prepared for a world where agents learn extensively from their environments. This work seems like a solid starting point and was very helpful for conceptualizing different types of CL systems.

(3) METR (and other independent orgs, as well as safety/security teams at labs) feel woefully under-resourced compared to the scale and pace of AI development - we’re struggling to build benchmarks fast enough, keep ahead of latest capability developments, read and respond to all the safety-related claims that AI developers are making, run all the evaluations and assessments that companies + governments are asking us to, plus develop the science needed to assess risks from increasingly capable AIs.

I like Geodesic a lot. The founders are highly motivated by catastrophic risks, low ego, and particularly "for the mission". Spaces that "feel RSI" are rare, so when I moved back to Cambridge, I ran weekly "ASI futurism" dinners - half of the regulars now work at Geodesic.

Geodesic is a new AI safety org i’m particularly excited about: they do awesome neglected work trying to figure out how to shape alignment priors for frontier AI. People should consider applying!

This is very important work and the Geodesic team is cracked + extremely nice! Apply apply

Come help us solve alignment 👍

Anthropic does a lot of alignment science. This is good, but I also wonder how much the conclusions they draw rely on baked-in assumptions about how they did pretraining, midtraining, etc. Better open-source alignment needed as a contrast point

Excited to see @AnthropicAI landing on improving pretraining priors as a central alignment intervention. In line with our findings in Alignment Pretraining, they find that a small amount of positive AI discourse in pretraining substantially reduces misalignment.

Excited to see our work featured in Isambard-AI's case study series; kind of compute-intensive research Geodesic does wouldn't happen without infrastructure like this. Thanks @BeestonMedia and BriCS for the chance to talk about our research!

Full system card can be found here: deploymentsafety.openai.com/gpt-5-5/gpt-5-…

I was disappointed in the rigor of this system card. It seems obvious that the safety team had extremely limited time and resources to put the report together. I hope this changes as we continue to push the frontier of intelligence and misalignment becomes harder to spot.

@OpenAI has a beautiful sweet of agentic misalignment benchmarks developed with @apolloaievals. I'm saddened to see that these were not included in the GPT-5.5 system card. This is especially frustrating given a meaningful increase in misalignment compared to GPT-5.4.

“WTBU” is one of the most useful communication technologies I know of. It stands for “Watch Team Back-Up”, I believe it originated to reduce mistakes on nuclear submarines. You prepend it to a message to someone where you’re pointing out something that might be obvious, but you want to check/confirm that they’re on top of it. For example, you might say “WTBU: you checked that we’re cleared to share this info with XYZ person?” It takes the pressure/ego out of the message, by letting you communicate something to the effect of “I’m not saying this because I think you’re incompetent/dumb, so ignore this if it isn’t relevant/useful—I just really want to make sure we don’t mess up/make silly mistakes, and smart/competent people can make silly mistakes!”—except once you’ve coordinated on using the letters WTBU to communicate that, you can just say “WTBU:”. This lets you now check basic/obvious things with coworkers much more easily and with less ego/emotion, which makes it much easier to catch mistakes in advance. Worth considering adopting into your org as a standard communication practice!

Opus 4.7 has a new tokenizer. This means it's also a new base model. Glory days of pretraining still very much going.

I'd recommend giving AI 2027 another pass if you're trying to get a better understanding of how the coming years play out. On the authors own guidance, reading it as AI 2029 is probably closer to the truth. But there is lots of danger in the tails :/ ai-2027.com

4. AI 2027 is a year out, and the AI progress predicted is not far off. @eli_lifland and crew made concrete predictions that were easy to falsify. Predicted AI revenue is ahead of schedule (predicted 26 Billion by May 2026) and their 80% METR time-horizon is basically tracking.

You should take some time to think about what short-timelines to superintelligence means for your career/research agenda/family. My main takeaways from the Anthropic release I presented at the Meridian office yesterday: