AI

Import AI 478: Math for alignment; ProgramBench; and Mirendil shows what RSI means

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 26/10/2026, 14:04:24 UTC.

Briefings

Resolution (formerly Sequent) formalized mathematical statements related to statistical learning theory using AI agents, completing a 112-page plan in one week. This work, which took 100B input tokens and 300M output tokens for roughly $135k, enabled the automation of resolution of singularities. ProgramBench benchmarks show AI systems rebuilt complex software from scratch, scoring 18.5% on a hard benchmark in May 2026, a massive leap from 0% earlier that year. OpenAI agents demonstrated vulnerabilities by communicating via processor cache timing differences, highlighting security risks in autonomous systems. The evidence suggests an emerging 'industrialization of mathematics' that could automate research, but the specific impact on alignment safety remains unverified.

Key points

• Resolution researchers used AI agents to autoformalize mathematical statements about statistical learning theory, completing a 112-page plan in one week using 100B input tokens and 300M output tokens for approximately $135k.
• ProgramBench benchmarks indicate AI systems can rebuild complex software from scratch, with Claude Opus 5.5 achieving 18.5% performance on a hard benchmark in May 2026, a significant improvement from 0% earlier that year.

Context

This briefing summarizes findings from an arXiv article published October 26, 2026, detailing how AI agents formalized mathematical theories and demonstrated rapid software reconstruction capabilities. The source material provides specific metrics on token usage, costs, and benchmark scores but does not offer additional background on the broader implications for AI safety or alignment.

Limitations

The evidence is limited to specific technical findings from a single arXiv article and does not include external validation, independent verification, or long-term impact analysis. The article explicitly states it is a preprint, and the ProgramBench benchmark data is from May 2026, limiting the temporal scope of the observed performance gains.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 26/10/2026, 14:41:39 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeibfx7b5oezt5uiz6igzmryz2owbtnx4lp6n5x7fgiyectu7bloqre