we are in end game.

My guest essay in @nytimes today. I make 5 key points: 1. There’s little clear evidence of AI eliminating jobs at scale yet. But waiting to see is risky. Pittsburgh’s steel towns saw early signs with mini-mills before the losses showed up. Service capitals like London and New

I keep seeing this paper (and the others like it) being misinterpreted. This is a measure of AI applicability, not job loss. That might mean the jobs are at risk, but it also can be the jobs that will be most benefitted by using AI, or those that will be most transformed by it.

Sectors that are more exposed to LLMs experienced significant increases in wages and employment following ChatGPT’s release.

We're introducing GPT-5. The evals are SOTA, but the real story is usefulness. It helps with what people care about-- shipping code, creative writing, and navigating health info-- with more steadiness and less friction. We also cut hallucinations. It's better calibrated, says "I

Today we release gpt-oss-120b and gpt-oss-20b—two open-weight LLMs that deliver strong performance and agentic tool use. Before release, we ran a first of its kind safety analysis where we fine-tuned the models to intentionally maximize their bio and cyber capabilities 🧵

this kind of robust research -- really focusing on surveys as an eval (SaaE) for impact -- is so badly needed! we can only glean so much from a chat log or even RCT

ai’s “economic impacts” aren’t just productivity gains or layoffs. fear of being replaced + algorithmic bossing ==> despair. and despair begets inefficiency! if you're worried about the job market do you buy a home? change your major? have kids?

cRaZy PaPeR! yet more evidence young workers are feeling despair. why? - social media (“compare-and-despair” effect) - underemployment blues - pandemic burnout & rethinking the grind and turns out they don't like surveillance either!

if you think it's important to ~at least explore~ policies that respond as disruption happens -- automatic stabilizers, targeted support for displaced workers -- you need data! bls gives us the who, doing what, where, for how long of (un)employment.

to be clear, it doesn't mean disruptions don't happen, it means that when they happen policymakers have better data to underpin their response flying blind, great title for a romcom

the bls isn't just some political scoreboard. it has repeatedly served as a leading indicator for big labor market changes: - the China shock to American manufacturing - the "jobless recoveries" after 2008 - the rise of gig work - COVID!

if you care about ai safety you should care about bls!

big big kudos to @kiranTomlinson @soniajaffe Will Wang, @scottjcounts @ssuri

big fan of this work from the microsoft team! what would be great next: * more from more labs on the where/when/who of AI use * experiments to validate how AI use translates to productivity gains * policy proposals for using data from labs as leading indicators of impact

meanwhile, i continue to get a lot of angry messages from mathematicians

imho, lots of space for alpha in finding imo gold

love this paper btw! if you're an economist trying to decide what to work on in ai, there's a whole bunch of work on evals for interactions that augment, uplift, and assist to pick up

in a fun coincidence, i had emergency surgery in nairobi a few weeks ago (i'm fine now!), the team's work gave me so much more confidence navigating a foreign health setting and advocating for myself there :)

great work from karan and the team! more in-context studies, more!!