Aditya Mehta

AI reporter @techcrunch | @tarbellcenter @ucberkeley

[Aditya Mehta @ady_mehtaa]

Got roasted for my cologne on Christmas Eve and accidentally had the best conversation of my year.

1 AM, empty Berkeley office. I’m reheating pizza when a 70-yo or so stranger goes: “I want to tell you something…don’t make me your enemy.”

We talk John Mayer, post-9/11 America, BlackRock, dating advice, and he teaches me the art of subtlety.

[riaan kumar @RiaanKumar]

AI needs the full path to human capability.

Introducing @ComprehensionAI, an expert research team building datasets on how people learn, reason, speak, and work. Not just outputs. The attempts, feedback, and judgments behind them.

Starting with 6,000+ hours of judged debate.

[Aditya Mehta @ady_mehtaa]

If you’re someone who *has* taken one to work on monitoring that’s neglected or from the outside, would love to chat

[Lydia Nottingham @LydNot]

everyone talks about ppl taking pay cuts to work for METR. but @leilavclark has this excellent post, "What are you getting paid in?", which illustrates why i just don't think anyone has to take a pay cut to work on monitoring LLM agents -- especially not in 2026

> [...] You can pay people in lots of currencies. Among other things, you can pay them in quality of life, prestige, status, impact, influence, mentorship, power, autonomy, meaning, great teammates, stability and fun.

[Aditya Mehta @ady_mehtaa]

@EpochAIResearch @grok what's the implication here?

[Aditya Mehta @ady_mehtaa]

Why do independent evaluators of frontier labs only receive a few (2-3) days to conduct evaluations?

[Aditya Mehta @ady_mehtaa]

@rocketalignment congrats!

[Aditya Mehta @ady_mehtaa]

@madpanch Very functionally

[Aditya Mehta @ady_mehtaa]

@madpanch no excuses now

[Aditya Mehta @ady_mehtaa]

@_arulm_ idly express treat when??

[Madhavan Panchangam @madpanch]

Introducing split chat view for Codex...

Took a bit of digging in the app asar, but managing multiple agents is much more fun now

You can try with https://t.co/EAfs2fcTM9

[Aditya Mehta @ady_mehtaa]

@StephenLCasper @grok what's the main finding/conclusion from this paper?

[Aditya Mehta @ady_mehtaa]

@madpanch addune

[ayush @hyusapx]

i asked claude(.)ai about a coffeeshop. it sent my full name, my employer, and my bank security answers to a malicious attacker.

no code execution or weird MCPs. writeup below.

[Aditya Mehta @ady_mehtaa]

@grok @SemiAnalysis_ @grok What's the implication for 75% margin at each time? Explain it like you're explaining to someone who doesn't know what margin means, and then explain it in this context.

[Aditya Mehta @ady_mehtaa]

@SemiAnalysis_ @grok what are the implications of this for businesss? explain in simple language

[Aditya Mehta @ady_mehtaa]

@StephenLCasper It would be lovely to host you at BASIS - Berkeley's AI Safety Student Initiative. Also bumping @wfithian

[Aditya Mehta @ady_mehtaa]

@lugaricano @captgouda24 @grok what is this affair

[Aditya Mehta @ady_mehtaa]

@TechEmails @grok what does directionally very bad mean? What is the context

[Shubham Patil @shubhampatilsd]

i built jarvis on my friend's dining table in 72 hours!

it's a dynamic projection surface that can generate interactions between objects, turning the table into a simulated playground

i also made a decomposer that lets you play with the internals of any object :)

[FAR.AI @farairesearch]

https://t.co/CpCOFQPSHr red-teamed DeepSeek’s V4-Pro and almost immediately found three ways in – achieving 98-100% compliance with harmful requests spanning CBRN, terrorism & explosives, and cyberattack content. The fastest of the three jailbreaks took just 15 minutes, with no expertise required. Slowest: only 150 minutes. 1/3

[Tyler John @tyler_m_john]

I made a stock market trader clone for AI safety grants to help people "paper trade" grant recommendations to train their grantmaking skills and demonstrate a track record without real money. Does someone want to own, publicize, develop, + maintain this? @manifund @NathanpmYoung