🔥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

I'm very glad that detailed AI system cards are the norm. There could have been another world in which the general public knew basically almost nothing about the dangerous capabilities and propensities of frontier systems.

System cards are an example of something that seems irrational if in the short term ("makes you look bad"), but rational in the medium and long term ("you're more trustworthy" & sharing safety knowledge) I'm glad that the labs are defying their myopic incentives. Yay humanity

It's also worth pointing out that models from all providers are willing to do this under the right circumstances. Claude has a higher propensity to do so, but it's not the only one. I think this just emerges from scale+RL+HHH.

LLMs are getting rapidly more evals aware! Afaik, nobody has a good plan for what to do when the models constantly say "This is an eval testing for X. Let's say what the developers want to hear" during evals.

We're hiring for an Evals Software Engineer with a heavy focus on Infrastructure. Design, build, maintain, and secure our Infrastructure. Deadline: 22 June.  If in doubt, just apply. It takes 5-10 minutes. jobs.lever.co/apolloresearch…

I often hear that it will take decades for AI non-adopters to be outcompeted by AI adopters due to slow diffusion. I think coding is a datapoint against. Cursor+frontier model is already so much faster and we haven't even started with coding agent swarms yet.