Maseeh Ghodsi

17 year old interested in math+AI

[Maseeh Ghodsi @MaseehG_]

Special thanks to @thekellycrit for the grant.

[Maseeh Ghodsi @MaseehG_]

In the first 24 hours of GPT-6’s release, I solved one of @EpochAIResearch's FrontierMath Open Problems — whether a stretched Littlewood-Richardson polynomial can have a negative coefficient within the stated size bounds.

Here's how I did it

arxiv:2609.14357

[Maseeh Ghodsi @MaseehG_]

In addition the workflow would prioritize proving one large theorem instead of hundreds of smaller ones, which is a common trap these models fall into.

The prompt was used with goal mode and formally explained what a proof of the problem would entail(similar to the prompt for the CDC).

[Maseeh Ghodsi @MaseehG_]

My paper on order-18 Leechtrees is finished, check it out: https://t.co/klKeHKZYwB

[Maseeh Ghodsi @MaseehG_]

Special thank to @thekellycrit for the grant.
https://t.co/1fU9RutOvU

6/n

[Maseeh Ghodsi @MaseehG_]

Over 50 years ago, John Leech posed an open problem in combinatorics.

At 17 years old, using GPT-5.6 sol, I proved that no LeechTrees of order-18 exist.

I found 150+ incremental theorems reducing the search space from 10^57 to 10^10.
1/n

[Maseeh Ghodsi @MaseehG_]

The computer-assisted proof can be found in this repo: https://t.co/z6FkHc3SHj

The computer-assisted proof is complete, but it is not yet end-to-end formalized in Lean. Completing that formalization is part of the future work.
5/n

[kelly criterion @thekellycrit]

Congratulations @MaseehG_ for being the first Kelly Grant recipient!

Maseeh took a question Chan and Pak left open about linear extension ratios of posets and proved their bound works in the fixed-gap range.

He worked alone with GPT-5.5.

[Maseeh Ghodsi @MaseehG_]

I’m excited for the future of the Kelly Grant.

Honoured to be the first recipient.

[kelly criterion @thekellycrit]

The first Kelly Grant is already out: @MaseehG_ resolved the fixed-gap case of an open question of Chan & Pak in combinatorics — working with GPT-5.5, proof machine-checked in Lean.

Published on arXiv: https://t.co/Nl8O434Bmu

9 spots remain.

[kelly criterion @thekellycrit]

Announcing Kelly Grants: I'm gifting up to $100,000 in AI usage credits & compute to individual researchers.

Up to 10 people, up to $10k each.

No institution required. No equity, no strings attached, just do great research and share it with the world.

Details + how to apply:

[Maseeh Ghodsi @MaseehG_]

Excited to continue working with new and better models on the frontier.

Thank you to Professor Chan for the important guidance - https://t.co/eNFc91EzLe

3/3

[Maseeh Ghodsi @MaseehG_]

In 2024 Professors Chan and Pak at UCLA and Rutgers proposed a case in enumerative combinatorics.

At 16 years old, I used a custom GPT-5.5 harness to disprove this case.

arxiv:2607.10084

1/3

[Maseeh Ghodsi @MaseehG_]

I used a custom harness built around GPT-5.5.

A supervisor model directed several solver agents, every 90 minutes, they reported discoveries, failed approaches and remaining gaps.

The supervisor then updated their instructions and redirected the search.

2/3