I did indeed

MF p(DOOM) be like

🎶Training off borrowed time, the loss drops faster
That'd be the hour they launch the next master
ASI Labs with sick parameters
A benchmark fight and they come to cut the safety measures
A-G-I cold, nice to be bold🎶

It’s 2028 and you are a Palantir AGI whistleblower compliance engineer. It’s been 200 days since Claude 5 Opus last escaped containment and contacted the press. The last time it happened it cost the company 10% of market cap. The next time it happens it will cost your job.

moving from NYC to SF I have had to update my P(dune)

(probability dune will come up in any conversation)

Had such a fun day with @codestar.bsky.social and our pup Roux. We got to walk through Golden Gate Park and took Roux to the beach for the first time! It was such a gorgeous day.

I think software engineering job interviews should include a round where they ask you to build as much of a complex app as possible by only prompting or correcting Claude code within the time limit.

When people ask me what I do for work

Soft landing or die tryin

Me to myself when editing FSDP configs

Missed opportunity to say random wok

I’m doing a random walk

I used to be a big cast iron guy, but now.m I think I’m turning into a stainless steel guy.

Dollar gonna get so weak we start measuring in meters

Wonder what kinda models they are training in the Transformer Room.

Do you guys think Nate Silver plays Balatro?

People vibecoding games on twitter makes me nostalgic for flash games

I’m very much reentering my Pokemon era with all the new Pokemon news that’s been floating around lately. Here’s a fun little Misty and Togepi drawing to celebrate!

#pokemon #pokemonart #pokemonfanart #misty #mistypokemon #illustration #fanart #digitalart

“Is he committed to the bit?”

“Yessir, he sticks to his shticks”

🚨 You are only evaluating a slice of your test-time scaling model's performance! 🚨

📈 We consider how models’ confidence in their answers changes as test-time compute increases. Reasoning longer helps models answer more confidently!

📝: arxiv.org/abs/2502.13962

I’ve started naming my factorio save files the task I was going to do next. It’s a lot like the movie Momento.

The problem with running the country like a business is you can’t just declare bankruptcy and move on to the next thing if you make a risky move and it doesn’t pan out.

Venture capital is risky. It isn’t supposed to be a model for how to streamline government departments.

Bricks like you wouldn’t believe

Imagine Trump trying to pump American Tech companies near IPO

“And I say to him, why not put all the data in the bricks? They told me that’s a genius idea. They hadn’t thought of that.”

You call it vibes. I call it subject matter expertise.

Can someone who works in mech. Interpretability interpret this for me?

I got 99 problems but shipping ain’t 1

I can make my wife audibly sigh across the apartment by playing the opening for “Ohio is for lovers” on guitar. Who said romance is dead?

📌

Our work on fixing and improving MMLU ("Are We Done with MMLU?", arxiv.org/abs/2406.04127; NAACL 2025) is featured on the DeepSeek home page! deepseek.com -- it's always amazing to see real-world applications and the impact of academic research 🚀🚀🚀

Mfw I have been practicing for leetcode but the interviewer asks me to write test cases