🇳🇵🇺🇸 co-founder https://xitter.gq/t.co/Vz65cBoc7N | ex @google AI | training data and evals for frontier labs | https://xitter.gq/t.co/2EFb7GLRu4
San Francisco
Joined May 2025
- Tweets2.1K
- Following856
- Followers641
- Likes17K
70% of Americans reach the top 20% in household income at least once in their working lives. That is an INSANE level of upward mobility.
Throwback to when I built this VR experiment in college. Could have made it into something HUGE for robotics simulation or something! Should have continued the project 🤣
It was for a class with some psych students, running an A/B behavioral experiment in VR. Built on Unity for Oculus Quest 2 (SO to @PalmerLuckey, great tech).
I spent a couple of non-stop nights building it since i was the sole CS guy but it was super fun. Big upgrade from Game Maker 7 when I was a kid.
Honestly, I wish AI was around when I was building it. I enjoy game dev, but there's a lot of grunt work, especially with testing.
P.S. This was edited and uploaded using my personal agent in 30 secs. That's part of the grunt work AI is saving me from nowadays.
Interesting read - there are a few big caveats though.
1/ It isn't entirely free from trade-offs. In one of the setups, a model dropped roughly 3 percentage points on Terminal-Bench.
2/ The cost savings are estimates, not actual production bills.
3/ A shorter context doesn't mean you can delete the history. A summary might forget that a customer changed 'delivery' to 'pickup only'. If the original context is gone, you risk failing the task.
"You Don't Need Long Context"
NVIDIA finds that LLM agents can complete complex long-horizon tasks without maintaining their entire conversation history, which reduces inference costs.
So instead of accumulating tokens, they use frequent context compaction to summarize progress and retrieve information when needed, even resetting the context after every agent action.
Across six long-horizon benchmarks, this preserves accuracy while achieving up to 8.8x lower estimated inference costs.
alphaxiv.org/abs/2611.05306
peak
I am not crazy! I know they sold us the same data. I knew it was the same. Down to the same typo in row 4,821. As if I could ever make such a mistake. Never. Never! I just – I just couldn't prove it. They covered their tracks, they got that idiot at the data brokerage to lie for them. You think this is something? You think this is bad? This? This chicanery? They've done worse. That exclusivity clause! Are you telling me that two vendors just happen to have exclusive rights to the same data? No! They orchestrated it! They sold it to us twice! And I bought it! And I shouldn't have. I made them a preferred vendor! What was I thinking? They'll never change. They'll never change! Ever since they were founded, always the same! Couldn't keep their hands off other people's data! But not our vendors! Couldn't be our preferred data vendors! Selling it back to us! And THEY get to be featured on our blog? What a sick joke! I should've stopped them when I had the chance! ...And you, you have to stop them!
*admin inflation - where AI is best suited to automate gruntwork and speed up care
Two thoughts: 1. Bounties for AI-generated or AI-assisted tasks detection need better tooling (Talk to me if you're building there)
2. Voice dictation tooling better suited for tasks authoring/reviewing needs to be built.
we're fucked. absolutely fucked. the ai trade is over. we all top blasted and holding bags now. i can't believe it. how could we have been so wrong. this market makes no sense. my spleen hurts. i haven't seen green in such a long time. i miss money. i miss smiling. i miss sleeping not checking my phone every 1 min. i miss waking up happy. i should've just stayed in cash like a good boy. im like icarus flew too close to the sun. my ego got the better of me. there is no alpha. only luck. lady fortuna has left us to fend for ourselves. faith does not work. praying does not work. i am tired. hedgies are waging war and we are caught in the middle. how do you survive this. when will it end.
Working with domain experts on training data, we keep running into two problems that generating more tasks doesn't solve.
1/ Which tasks should you start with? Generating variations is getting easier, but if the seeds cover a narrow set of situations, you can end up with 1000s of tasks that still lack coverage.
2/ Where do you need another expert's judgment? More tasks doesn't automatically mean more independent judgment. Variations can add useful coverage, but they don't substitute for another expert seeing the problem from diff lenses.
We want to use generation to get more out of good expert data, without falling into the trapping of thinking bigger dataset > broader dataset.
SF is just a few AI researchers in trench coats huh?
Some news: I’m joining OpenAI to lead product management for Core Products!
I have a lot of respect for OpenAI for showing the world that scaling works, for bringing AI to over a billion people, and for believing great products are core to the mission.
I’m excited to work with @thsottiaux, an incredible PM team, and everyone across OpenAI. Together, we’ll build agents that have your back and your team’s, and keep you in control, so you get more time for the people and work you love.
Lots to learn, and lots to build. Please keep your feedback coming!
.@axiommathai stock keeps rising 📈
Everything is a compute/data problem. Compute still needs a lot of work, as does evals/training data/rl envs. Medicine is definitely the north star for most, understandably. Love that our team is working hard to make further inroads to data problems - better models, benchmarks, and hill-climbing potentials
i feel so lucky to be a founder in the us. in nepal I'd never be able to do a startup at all.
Pratik Karki retweeted
Canva is laying off 22% of its workforce which is 1271 employees, and this is what their CEO said.
We need more founders like thissss.
