Summary
According to a post on X by @rohanpaul_ai, a Meta and NYU paper called h1 gives a model fast Lean feedback and search tools and lets it keep trying, with no imposed plan. The post reports that, using the same model, this approach edged out the hand-built Goedel-Architect on PutnamBench, solving 598 problems to 597 at $1.08 versus $1.47 per problem. It also says that with Opus 4.8, the approach solved 83 of 100 hard algebra problems, compared with 63 for Claude Code.
AI-assisted summary based on the listed source.
What happened
The h1 paper gives a model fast Lean feedback and search tools and just keeps it trying, with no imposed plan. With the same model it edged out the hand-built Goedel-Architect on PutnamBench, 598 problems to 597, at $1.08 versus $1.47 per problem, and with Opus 4.8 it solved 83 of 100 hard algebra problems against...
Why it matters
The post frames the result as further evidence that tools and compute can matter as much as carefully designed workflows for AI theorem proving. Teams building AI coding and proof tools may want to compare simple feedback loops against more elaborate pipelines on both solve rate and cost. The figures come from a social post summarizing the work, so the original paper's full methodology and results are worth checking before drawing firm conclusions.
Signal Intelligence
Signal Strength 89%
Technical label WATCH
Public Interest 44
Category RESEARCH
Reader Depth TECHNICAL
Signal Strength reflects source quality, relevance, freshness and evidence. Public Interest helps organize discovery; it is not proof of truth.
Public Interest components
Recognizable Entity Score 51
Practical Impact Score 28
Novelty Interest Score 70
Consequence Score 30
Curiosity Score 16
Shareability Score 57
Why this is here
VQV surfaced this signal because it is recent, relevant to AI Coding Tools, connected to Scobleizer News Pulse (X).