@darrowoah
Account based inUnited States
About this account
- Account based in
- United States
- Connected via
- United States App Store
Account-level information from X, not a live location or the device used for a specific post.
21 yo building @stanford
Palo Alto, California
Joined June 2016
- Tweets1.4K
- Following702
- Followers550
- Likes2.8K
curious why people opt for artificial analysis over lmarena and vice versa for gauging new model abilities
one of the most frustrating things with AI right now is asking a mid-tier model to analyze something, then asking a frontier model the exact same question and watching it completely dismantle the weaker model's answer
People assume human-generated data is the gold standard for datasets, but sometimes human data doesn't work well
A good example of this is in the ProCUA paper, where they found that training on 20k human attempts at computer-use tasks decreased a model's performance
Instead they used a VLM to generate feasible CUA data which improved OSWorld performance from 26% to 45.0%, while the human data dropped it to 8–10%
arxiv.org/pdf/2606.17321
Prices are ridiculously high for local AI hardware on RAM and GPUs right now but I don't see them going down anytime soon
astra ultra in codex is a great research assistant but I’m less sold on its research judgment
I had it do literature review + propose research ideas and every time I challenged an assumption, it came back apologizing and explaining what it had missed
right now I’m not seeing the judgment needed for open-ended exploration let alone recursive self-improvement.
context: I’m a student researcher not a professor or a researcher at a frontier lab
Darrow retweeted
It’s unbelievable how much talent has come out of Stanford’s first-year entrepreneurship programs (ASES/BASES).
One of my favorite teams just raised tens of millions, and I now know at least four or five people from the program who have raised a cumulative $60–70M, all within two or three years of coming to college!
Stanford’s startup ecosystem develops talent from day one in a way I’ve never seen anywhere else
Absolutely mindblown.
Darrow retweeted
it doesn’t matter whether or not you actually believe superintelligence will eventually cause human extinction. in either case, we should all care about AI safety and alignment going well, and we should do everything we can to make this happen. more below!
seems like many assume AI will develop the same way internet companies did: boom, crash, and then resurgence
I haven’t seen a good argument for why the crash is the necessary next step
highlander.sh is back up and serving Minimax FastH3 at $0.005/s for a limited time!
We've just updated our pricing for realtime video generation to $0.005 for the cheapest realtime video anywhere! Try it here highlander.sh
We just launched the cheapest way to generate realtime 768p video! Check us out on Product Hunt
producthunt.com/products/hig…
