Twitxy Saved

Ted Sanders

@sandersted

Research at OpenAI. Be kind to others, and yourself.
San Francisco, CA tedsanders.com Joined September 2009
1.6K Following 11K Followers 1.1K Posts

Latest posts 5

Ted Sanders @sandersted
the GPT math results exemplify life's long arc of increasing cooperation and specialization.

credit is due to no single mind; it diffuses across many thousands who published math papers, wrote code, manufactured chips, built data centers, etc.

the age of heroes is ending.
1 replies 1 reposts 60 likes 4.4K views View on X
Ted Sanders @sandersted
One cool thing about ARC-AGI-3 is that it measures cost until success or timeout, instead of cost of first attempt.

This is arguably more aligned with real life, and shows how costly models at high efforts can counterintuitively be cheaper, by needing less iteration.
ARC Prize Verified account @arcprize Sep 30
GPT-6.1 Sol from @OpenAI on ARC-AGI (Verified):

- ARC-AGI-3: 52.7%, $7.6K (standard harness), 96.4%, $4.4K (provider adapter harness)
- ARC-AGI-2: 94.2%, $0.25/task
- ARC-AGI-1: 98.5%, $0.06/task

Its 96.4% on v3 was comparable to GPT-6 Astra's 99.9% but at a 77% lower cost.
View quoted post on X
1 replies 6 reposts 79 likes 7.5K views View on X
Ted Sanders @sandersted
Suppose you get a magic lamp that can solve any math problem.

If math is useful work, using this lamp is good.

If math is status/fun/jobs, using this lamp is bad.

If math is a path to understanding, using this lamp may be both good or bad.

It's tricky because math is all 3.
5 replies 2 reposts 54 likes 4.1K views View on X
Ted Sanders @sandersted
OpenAI is an incredibly ambitious company, which is both positive and negative. We all agree we should focus more, but no one agrees on what. A few years ago, I thought it was way too early to invest in computer use, because AIs were so bad at it.

Astra proves how wrong I was.
2 replies 5 reposts 168 likes 8.5K views View on X
Ted Sanders @sandersted
Is Astra AGI?

1. no
2. yes, in the literal sense that it's artificial, general, and intelligent
3. no, but it's still not doing work like humans do
4. yes, but it's literally superhuman on tons of benchmarks
5. no, but it's subhuman on things that are hard to benchmark

all true