trevor (taylor’s version)
@tmychow
Followers
3K
Following
110
Media
191
Statuses
2K
openai launched gpt-4.5 to great excitement. yet under 2 months after launch, they announced it would be deprecated. is pre-training over now?. @jackgwhitaker and i show:.1. pre-training scaling laws haven't bent.2. why the marginal dollar left it for RL.3. why it'll come back
9
12
155
RT @taylorswift13: And, baby, that’s show business for you. New album The Life of a Showgirl. Out October 3 ❤️🔥. .
0
151K
0
i explicitly pre-registered this before using it, but even so, i am shocked at how much i like gpt-5 when using it for coding. it doesn't hallucinate, it follows my instructions, and it does sensible non-reward-hacky things.
day 0 take is gpt-5 will punch above its weight in "doing stuff", compared to what you might think from the usual benchmarks. increasingly, what matters are instruction following / tool calling / long context / hallucination, which all score great. it's also priced very well.
0
0
11
another gold medal.
1/n I’m thrilled to share that our @OpenAI reasoning system scored high enough to achieve gold 🥇🥇 in one of the world’s top programming competitions - the 2025 International Olympiad in Informatics (IOI) - placing first among AI participants! 👨💻👨💻
0
0
7
RT @CalaverasAI: To our knowledge, we have sold more code tokens than any other data vendor. Follow this space. .
0
1
0
day 0 take is gpt-5 will punch above its weight in "doing stuff", compared to what you might think from the usual benchmarks. increasingly, what matters are instruction following / tool calling / long context / hallucination, which all score great. it's also priced very well.
0
1
11
i've known deniz and sherry since we were all at stanford, and they are among the grittiest and fastest moving founders i know - you should work for them!.
We are hiring!. I don't tend to talk about the success of @tamarindbio publicly, but we are experiencing incredible demand. Tens of thousands of scientists are regularly using the platform, and we're onboarding GPUs and people as fast as we can. We just posted multiple roles
0
0
9
RT @zhengdongwang: in which i get to play my version of 'overrated or underrated' with tyler cowen. listen to tyler's full talk below https….
0
11
0
relative to many other notions of superhuman swe, the idea of a "bash-only superhuman coder" feels like the one we have the most line of sight to. glad to see it being benchmarked!. cc @sambrashears.
What happens if you compare LMs on SWE-bench without the fancy scaffolds?.Our new leaderboard “SWE-bench (bash only)” shows you which LMs are the best at getting the job done with just bash. More on why this is important 👇
2
0
9
there are million dollar bills lying on the pavement.
I sold a 100B token code dataset to a bunch of AI companies recently. I'm looking to scale data things into a startup and as part of this I'm looking for people who deeply understand the browser platform to work on RL environments. DM me.
0
0
6
RT from 🔒."heard at icml: everyone *says* their top priority is data, but all their money is still going to architecture".
3
0
28
designarena [dot] ai is a good start, but really what i want is for someone to spend a month talking to really good designers and building an incredibly rich rubric based on that.
1
0
8
i wish we had frontend evals that were far more taste-loaded; even as automated coding is improving, most models continue to struggle to make me elegant and aesthetically pleasing frontends.
2
0
15
incidentally i think it is a very funny fact about the economics professions that a piece we first put out in jan 2023 is still a "recent working paper".
0
0
3
this week's cover story on the economist: what if AI made the world’s economic growth explode?
I had a lot of fun researching and writing this piece: what if AI made the world’s economic growth explode?. The macroeconomics of an extraordinary thought experiment:.
2
0
27
