Positive users celebrate GPT-5.6's record on self-contracted curves as an exciting breakthrough while negative users dismiss it as training data or diminishing returns.
Based on 8 visible X reactions from 10 accounts; directional sample.
Ask a question below.
Published answers will appear here.
@SebastienBubeck @jasondeanlee I've found these breakthroughs interesting because the hard part isn't just getting an answer. It's pushing into areas where the method itself becomes the discovery.
@SebastienBubeck @jasondeanlee 8 hours of compute just to shave 0.03 off a human benchmark? talk about diminishing returns.
@SebastienBubeck @jasondeanlee Did you get scammed into posting this or is it your scam?
@jasondeanlee + @canondetortugas who was also part of the fun on this improvement!
The model achieved the 2.26 result after eight hours.
@SebastienBubeck @jasondeanlee I've found these breakthroughs interesting because the hard part isn't just getting an answer. It's pushing into areas where the method itself becomes the discovery.
@SebastienBubeck @jasondeanlee 8 hours of compute just to shave 0.03 off a human benchmark? talk about diminishing returns.
@SebastienBubeck @jasondeanlee Did you get scammed into posting this or is it your scam?
@jasondeanlee @canondetortugas wow 2+epsilon would be insane...
@canondetortugas Congrats!
Well, the human record of 2.29 didn't last very long. @jasondeanlee used the double cycle cover prompt to get to 2.28 and with 8 (EIGHT!) hours of further thinking GPT-5.6 got it down to 2.26 and wrote a paper on it: https://www.dropbox.com/scl/fi/6zqstz2qbt0g8pbc1xidf/self_contracted_curves_2.261736.pdf?rlkey=xdc0a7mh5hizg3cqie8xyz7kh&st=fkuzmwoh&dl=0 https://twitter.com/SebastienBubeck/status/2075596982622835006
A very fun exercise getting acquainted with sol ultra. The improvement to < 2.28 only took 12 minutes once we had 2.29. https://x.com/canondetortugas/status/2076370622595109351/photo/1 https://twitter.com/SebastienBubeck/status/2076366761835045098
with help from Dylan foster @canondetortugas and lots of other friends. Constant is still improving ( clear strategy to get to 2.21 and potentially 2+\epsilon with recursion). https://twitter.com/SebastienBubeck/status/2076348395292377498
it might turn out that good models demand more expertise to use them to solve difficult problems. https://twitter.com/sebastienbubeck/status/2076348395292377498
@SebastienBubeck @jasondeanlee Was obviously in the training data doc
Positive users celebrate GPT-5.6's record on self-contracted curves as an exciting breakthrough while negative users dismiss it as training data or diminishing returns.
Based on 8 visible X reactions from 10 accounts; directional sample.
Ask a question below.
Published answers will appear here.
@canondetortugas Congrats!
Well, the human record of 2.29 didn't last very long. @jasondeanlee used the double cycle cover prompt to get to 2.28 and with 8 (EIGHT!) hours of further thinking GPT-5.6 got it down to 2.26 and wrote a paper on it: https://www.dropbox.com/scl/fi/6zqstz2qbt0g8pbc1xidf/self_contracted_curves_2.261736.pdf?rlkey=xdc0a7mh5hizg3cqie8xyz7kh&st=fkuzmwoh&dl=0 https://twitter.com/SebastienBubeck/status/2075596982622835006
A very fun exercise getting acquainted with sol ultra. The improvement to < 2.28 only took 12 minutes once we had 2.29. https://x.com/canondetortugas/status/2076370622595109351/photo/1 https://twitter.com/SebastienBubeck/status/2076366761835045098
with help from Dylan foster @canondetortugas and lots of other friends. Constant is still improving ( clear strategy to get to 2.21 and potentially 2+\epsilon with recursion). https://twitter.com/SebastienBubeck/status/2076348395292377498
it might turn out that good models demand more expertise to use them to solve difficult problems. https://twitter.com/sebastienbubeck/status/2076348395292377498