Jamie Bernardi
@The_JBernardi
Doing AI policy research, ex-Bluedot Impact. Climber, guitarist and sporadic musician. he/him.
🚀 New blog! Achieving AI Resilience: Exploring AI safety through a lens of adaptation & societal resilience. Advanced AI is diffusing fast. How will societal systems keep pace? Will def/acc work? Why can't we rely on AI safeguards? Explore these ideas with me as they develop 👇
The AI Action Plan is out. Immediate reactions in this thread:
I'm wondering if we'll see the same phenomenon of demotivation amongst whitehat hackers--an activity often pursued for the joy and prowess of discovering something nobody else did. Many interesting jobs could move from being creativity-based to verification-based
the openai IMO news hit me pretty heavy this weekend i'm still in the acute phase of the impact, i think i consider myself a professional mathematician (a characterization some actual professional mathematicians might take issue with, but my party my rules) and i don't think i…
Spearphishing PSA—looks like there's a concerted attack on AI safety/governance folks going around. Be wary of calendar links via DM, and *never* give a 2-factor auth code over the phone. I almost got caught by this—got a phone call last week, but figured out it was sus. 🧵
And @ajeya_cotra's account has been hacked by the same folks - if you get messages from her asking to schedule a meeting, be very wary! She says she will never reach out to a potential grantee by Twitter, always email.
🚨 We're hiring at @BlueDotImpact to build the AI governance pipeline. Your mission: Figure out what governance is needed to make AI go well, then build the workforce to make it happen. Imo, this is a top 1% role if you want influence on AI's trajectory 🧵
Another fascinating wrinkle in the unfolding story of LLM chain-of-thought faithfulness…task complexity seems to matter. When the task is hard enough, the model *needs* the CoT to be faithful in order to succeed:
Is CoT monitoring a lost cause due to unfaithfulness? 🤔 We say no. The key is the complexity of the bad behavior. When we replicate prior unfaithfulness work but increase complexity—unfaithfulness vanishes! Our finding: "When Chain of Thought is Necessary, Language Models…
I'm pretty confident we won't have AGI/country of geniuses in a datacenter within 2 years. I like ai-2027 as a piece of futures work, but I think too many people are treating it as a mainline scenario, rather than unlikely-but-not-impossible. I think this is resulting in too…
Important work. Non-Claude models seem to refuse reasoning about alignment faking, and have less intrinsic tendency for goal-guarding. Observing this diff is a step towards better aligning AI. I'm in awe that 2025 is seeing alignment become an increasingly empirical discipline!
it is easier than ever to make a resume and even custom cover letters and flood 1000 jobs with it, so a lot of jobs are (even more than always) absolutely drowning in resumes. I suspect this is well on its way to entirely killing off the 'send in an application' style of hiring.
So this is actually not a joke. Sam literally crashed the interview and started attacking Kevin for the NYT lawsuit against OpenAI. Can you imagine if the CEO of Coca Cola (also valued $300B) did something like this? Wild timeline we're in. Video link below
We'll discuss it on the show, but since I'm getting asked about it: no, Sam Altman barging onto the Hard Fork Live stage 10 minutes ahead of schedule to harangue us about the NYT's lawsuit against OpenAI was not planned. Video coming shortly!
We'll discuss it on the show, but since I'm getting asked about it: no, Sam Altman barging onto the Hard Fork Live stage 10 minutes ahead of schedule to harangue us about the NYT's lawsuit against OpenAI was not planned. Video coming shortly!