deniz.in

Markets

Weather

Loading weather

· via Hacker News – Front Page (native)

AI agent fleet attacks P vs NP live, logging checked lemmas and dead ends in public

A project at pequalnp.com is aiming a coordinated fleet of AI agents at the P vs NP problem, streaming checker-verified lemmas, refuted claims and dead approaches on a public live dashboard.

AI agent fleet attacks P vs NP live, logging checked lemmas and dead ends in public

A project hosted at pequalnp.com, which surfaced on the front page of Hacker News, is pointing a coordinated fleet of AI agents at the P versus NP question and streaming the whole attempt in public. The site's summary of the current state is blunt: the verdict remains open, and the work is happening now.

What the dashboard shows

The front page is a live status board rather than a write-up. It tracks how many agent sessions have run, how many agents are active at this moment, and three cumulative totals: lemmas checked, claims refuted, and approaches ruled dead. A search box covers the archive of lemmas, programs and "seats," and an interactive map loads the latest snapshot of the effort.

Each lemma sits in one of a handful of states: checked and holding a result, an unconfirmed conjecture, under active work by an agent, refuted, or declared dead. A rolling log called the record splits into two halves: a frontier of confirmed results, and a graveyard commemorating approaches that did not survive contact with the checker.

Records, not narratives

The design detail the project leans on hardest is provenance. According to the site, every line in the feed is a structured record — a lemma stated, an attempt begun or abandoned, a verdict issued by a checker — and none of it is an agent describing its own progress. That is a direct answer to the familiar failure mode of LLM-driven research tools: fluent, confident accounts of work that never happened. Here a claim counts only when a checker confirms it, and a dead approach stays visibly dead.

Aiming at either answer

The fleet is not committed to a single conclusion. The dashboard separates work aimed at proving P ≠ NP from work aimed at proving P = NP, and displays a weight for each side: the number of checked lemmas whose program targets that answer. The site is unusually careful about what that figure means, stating plainly that the weight is a tally rather than a probability and that neither column constitutes a proof.

Why the target is brutal

For readers who have not touched complexity theory recently: P versus NP asks whether every problem whose solution can be verified quickly can also be solved quickly. The question has stood open for over half a century, carries a $1 million Clay Millennium Prize, and is ringed by barrier results — relativization, natural proofs, algebrization — that have knocked out whole families of candidate proofs. Most researchers expect the answer to be that P and NP differ, but no proof exists. The agents are therefore working in a field where decades of clever human attempts already lie buried, which makes the project's graveyard section genre-appropriate.

Why it matters

Whatever happens to the mathematics, the experiment is an unusually clean public test of agentic AI on work that cannot be bluffed. There is no partial credit for a plausible paragraph about P versus NP: a lemma either passes a checker or it does not, and the site makes both outcomes equally visible.

That symmetry is the interesting part. If the frontier grows, it suggests that swarm-style agent fleets can accumulate verified micro-results on genuinely open problems — a pattern that would transfer naturally to proof assistants, formal verification and specification-driven coding. If the graveyard grows while the frontier stalls, the project will have produced an honest, structured dataset on exactly where current agents break down against hard research, which is also worth having.

The live counters mean any snapshot will age quickly, and nothing on the page suggests a resolution is near. But as a template for auditable machine labor on real mathematics — many agents, structured records, checker verdicts, everything public — the format may outlast this particular run.

  • #ai-agents
  • #p-vs-np
  • #complexity-theory
  • #automated-reasoning
  • #mathematics

Related posts