Skip to content

diangao/attackable-problems

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

3 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

attackable problems

This repo tracks problems that may be worth attacking with frontier reasoning models, human experts, and tight verification loops.

It is not a list of important topics. It is a list of problems where a model could plausibly search a large idea space and produce an artifact humans can check.

Selection Rule

A problem belongs here only if it has:

  1. A precise target.
  2. A verifier that is cheaper than discovery.
  3. Public data, code, or literature to start from.
  4. A route to a first 7-day experiment.
  5. A reason general reasoning could matter, not just scale or data cleaning.

Current Shortlist

Rank Problem Domain Why it is attackable
1 Hadwiger-Nelson: search for a 6-chromatic unit-distance graph math finite graph certificates, mechanical validation, direct analogy to the OpenAI unit-distance result.
2 Cancer dependency synthesis from DepMap medicine / biology public CRISPR/RNAi screens, held-out dependency prediction, wet-lab follow-up possible.
3 C. elegans connectome-to-function gap neurobiology full connectome plus perturbation atlas; clear prediction gap between anatomy and function.
4 ARC / ConceptARC cognitive abstraction models cognitive science human baselines, task generators, explicit verifier, useful for understanding human-like abstraction.
5 Small Ramsey number certificate search math exact finite combinatorics, SAT/proof certificates, historically hard but checkable.
6 Collateral sensitivity treatment design medicine / evolution sequential therapy is an optimization problem with experimental validation.
7 FlyWire circuit hypothesis generation neurobiology whole fly brain connectome, programmatic access, testable circuit hypotheses.
8 Self-healing reach adapters for agent runtimes agent infrastructure adapter drift has cheap verify loops; code-level repairs could keep long-tail web/app CLIs usable.

First 7-Day Attack

Start with problem 001.

Output target:

  • A cleaned corpus of known 5-chromatic unit-distance graphs.
  • A verifier that checks unit-distance embedding and k-colorability certificate.
  • A model/human search loop that proposes graph transformations.
  • A "kill criterion": if no candidate pressure toward 6-colorability appears after one week, pause math search and switch to DepMap.

Repo Layout

  • rubric.md — scoring rubric.
  • safety.md — safety boundary, especially for medicine / biology.
  • docs/workflow.md — intake, promotion, and first-week attack protocol.
  • problems/ — one file per candidate.
  • templates/problem.md — template for future entries.

Source Event

This repo was created after OpenAI announced a general-purpose reasoning model had disproved the planar unit distance conjecture on May 20, 2026:

About

Verifier-first backlog of attackable open problems

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors