https://x.com/alexwei_/status/1946477742855532918?s=46

(http://www.autoadmit.com/thread.php?thread_id=5752305&forum_id=2:#49114310)

### Key Observations on the Progress:

1. **From Near-Zero to Competitive Performance**

- Early versions of GPT-4 struggled to score even 1-2 problems on the AIME (American Invitational Mathematics Exam), which is significantly easier than the USAMO.

- Now, AI is not just solving AIME-level problems reliably but also tackling Olympiad-level questions, which require deep reasoning, creativity, and proof-writing.

2. **Why This is Surprising**

- Many experts (including myself) expected this level to take at least 2-3 more years. The fact that it happened this year suggests:

- Better training techniques (e.g., improved reasoning loops, synthetic data, and self-improvement methods).

- Stronger mathematical priors (models may now "understand" abstract patterns rather than just memorizing).

- Possibly, more sophisticated search/verification methods (though Gemini's exact approach isn't public).

3. **Is Contest Math "Close to Solved"?**

- **For AIME:** Probably yes. AI is likely near-superhuman at this level already.

- **For USAMO/IMO:** Not yet, but progress is accelerating. A 50% USAMO score is already better than many human contestants, but the hardest problems (e.g., IMO Q6-style combinatorics/number theory) still require breakthroughs in symbolic reasoning and long-term planning.

- **For Proof Writing:** AI still makes stylistic errors, but formal verification (e.g., Lean) can help close the gap.

4. **What’s Next?**

- **IMO Gold (60+ score) within 1-2 years?** If progress continues at this rate, it's plausible.

- **General Mathematical Research:** The real test will be whether these models can contribute to unsolved problems (e.g., combinatorics conjectures, lightweight formal math).

### Why This Matters Beyond Olympiads:

- **Education:** AI could become the ultimate tutor for advanced math.

- **Research:** Automated reasoning might assist in mathematical discovery.

- **AGI Benchmarks:** Math is a strong proxy for structured reasoning—this progress hints at broader capabilities.

It’s an exciting time! Wouldn’t be surprised if an AI wins an IMO gold medal by 2026.

(http://www.autoadmit.com/thread.php?thread_id=5752305&forum_id=2:#49115797)

"we are releasing GPT-5 soon but want to set accurate expectations: this is an experimental model that incorporates new research techniques we will use in future models. we think you will love GPT-5, but we don't plan to release a model with IMO gold level of capability for many months."

https://x.com/sama/status/1946569252296929727

(http://www.autoadmit.com/thread.php?thread_id=5752305&forum_id=2:#49115705)

I’m not entirely talking about guardrails here. Like the poaster above mentioned it’s also using way more compute/tokens/parameters than whatever distillation MoE we will receive as civilians.

(http://www.autoadmit.com/thread.php?thread_id=5752305&forum_id=2:#49116798)

are you dumb btw? you can't grasp the implications of this? AI being equivalent with the best programmers/mathmaticians in the world? Why would anyone use a normal lawyer when AI is the equivalent of having Dershowitz personally represent you with unlimited billing hours to your case for flat fee? it's all over

(http://www.autoadmit.com/thread.php?thread_id=5752305&forum_id=2:#49115777)

fuck that, total chaos is the only way

(http://www.autoadmit.com/thread.php?thread_id=5752305&forum_id=2:#49116800)

"We can confirm that Google DeepMind has reached the much-desired milestone, earning 35 out of a possible 42 points — a gold medal score. Their solutions were astonishing in many respects. IMO graders found them to be clear, precise and most of them easy to follow."

https://deepmind.google/discover/blog/advanced-version-of-gemini-with-deep-think-officially-achieves-gold-medal-standard-at-the-international-mathematical-olympiad/

(http://www.autoadmit.com/thread.php?thread_id=5752305&forum_id=2:#49119024)