Elliot Glazer
@ElliotGlazer
Followers
2K
Following
13K
Media
60
Statuses
1K
Some of you guys are cool. Be at Cambridge week of 9/15 🤫
3
2
40
RT @antimath3: Also check out my CoT on using AI for math, FrontierMath, this paper, a workshop, etc
seewoo5.github.io
In this post, I’m going to share some of my recent experiences using AI (which is an extremely vague word these days) in mathematics. Specifically, I’ll describe my experiences with two events on AI...
0
2
0
RT @EpochAIResearch: Let’s take a look into GPT-5’s record-setting performance on FrontierMath. How did it perform on the holdout vs. non-h….
0
45
0
If you’d like to experiment with models on Project Euler problems, please do so using so as not to interfere with the statistics of the main site!.
3
0
56
GPT-5 twoshot it. I informed it after its first attempt that its code didn’t even answer the test cases correctly, and then it got it.
1
0
51
Finally, my contact pointed out that 942 is a problem with the top difficulty rating which might be amenable to AI reasoning because it “is the sort of problem where doing free-association on an immense amount of background knowledge is likely to be a big leg up.”.
1
3
55
“This one is quite interesting. It's on the machinery-heavy end compared to most human solutions, but it would be entirely believable that a human produced it, with one exception,” noting it effectively asserts that there are no Wall-Sun-Sun primes, which is an open question.
2
0
53
I then tried 947 because my contact found this one to be the most surprising of GPT-5’s successes on the MathArena eval. Once again, GPT-5 oneshot it, and I sent the reasoning summary for commentary.
1
1
54
He says “This is really remarkable. Very human-like solution, actually,” and was intrigued by the explicit observations GPT-5 makes in describing its solution.
1
1
67
950 was the highest difficulty problem solved in MathArena’s evaluation. In my run, GPT-5 oneshot it. I then sent the reasoning summary to my Project Euler contact, who happens to be the problem contributor.
1
0
65
I ran GPT-5 Pro on three Project Euler problems, 942 (difficulty: 100%), 947 (diff: 60%), and 950 (preliminary diff: 75%). Let’s go over them in reverse order.
11
29
456
RT @littmath: @rishicomplex Nice article. FWIW I’m a bit surprised we haven’t already seen a model crack some mildly interesting (but atten….
0
3
0
RT @SebastienBubeck: I copy pasted an unpublished manuscript of mine in ChatGPT and asked it to improve it. I expected that the method we'r….
0
47
0
GPT-5 is the best public math model across the board, with top scores on all MathArena tests (including 3/4 successes on a Project Euler 75% !!), SotA on FrontierMath T1-T3, and perhaps even a couple never-before-solved T4's. 🤫.
Its score on Project Euler is the most impressive, outperforming o4-mini by 20% and solving a problem with a 75% difficulty rating 3/4 times.
4
15
186
RT @GregHBurnham: The 2025 AIME has fallen!. On MathArena, GPT-5 (high) has solved the one problem that no prior model had solved, 2/4 time….
0
8
0
Google's AI falls hook, line, and sinker for obviously AI-generated math papers and announcements.
2
1
50
The X condition.
@ArmandDoma The content game is tough �� gotta be topical but also timeless.
0
0
3
But hearing more reports these days of it also providing value via brainstorming and testing out rudimentary proof strategies
Here is a case study in the frontier math research. I know a general answer and have a work in progress preprint where I do more but we spent almost 2 years thinking about this topic. The model got to this level in 15 minutes.
2
0
9
A sentiment I’m hearing a lot from mathematicians! Already vast improvement beyond google for technical searches. This is the biggest utility it provides for research at the moment.
One axis along which GPT-5 seems to me to be a significant improvement is search. A lot of the time I want to search the literature for mathematical object X with properties ABC. This is obviously very hard to google. Already had a success along these lines with GPT-5.
2
1
53
RT @EpochAIResearch: GPT-5 sets a new record on FrontierMath! On our scaffold, GPT-5 with high reasoning effort scores 24.8% (±2.5%) and 8.….
0
37
0