🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

Excited to have this research agenda out! Now here's my sarcastic/opinionated version 🧵

a fun prompt for introspection is "what contribution to societal flourishing do you most crave to be recognized for" and mine is definitely my list of synonym-based puns Unparalleled Misalignments

I like using o3 for quick and dirty automated data analysis. Eg give it a CSV, plot things, look at trends, brainstorm what to look into. I usually just use the default chatGPT interface for this. Are there better ways to do this, or easy ways to do this with other models?

Nate Soares and I are publishing a traditional book: _If Anyone Builds It, Everyone Dies: Why Superhuman AI Would Kill Us All_. Coming in Sep 2025. You should probably read it! Given that, we'd like you to preorder it! Nowish!

AlphaEvolve uses Gemini 2.0 Flash and Pro, I'd guess it would be substantially more capable today if they subbed in Gemini 2.5 models.

I'm vegan but I am fairly pro people eating lobster (assuming you don't boil it alive). Lobsters have about the same number of neurons as fruit flies, despite weighing 500,000x more. (Data from this plot is from o3 and might be wrong lol)

Casey, not to be mistaken for U-Casey

I think it’s under discussed that many plans for navigating the transition to ASI (at least implicitly) likely involve pausing AI capabilities progress for an extended period of time.

Yoshua Bengio is a continuing inspiration in this dire time with AI safety, and a reminder of why we do what we do - out of a sincere, and genuine love for humanity.

Weak prediction that Claude 4 will be closer to the AI 2027 extrapolation than the METR trend (including the METR trend updated for o3).

🤯 MIND-BLOWN! A new paper just SHATTERED everything we thought we knew about AI reasoning! This is paradigm-shifting. A MUST-READ. Full breakdown below 👇 🧵 1/23

mfw I prompt both gemini and claude to compare

Maybe AI safety orgs should have a "Carthago delenda est" passage that they add to the end of all their outputs, saying "To be clear, we think that AI development poses a double digit percent chance of literally killing everyone; this should be considered crazy and unacceptable".

Pack it up, the trends are clear, AI only gets less capable from now onwards.

Successful sabotage rates were “low”, that is Claude Opus 4 manages to sabotage and evade detection on “only” 30% of tasks. Very cool work, but this doesn’t seem that low to me!