Claude AI produces 13-million-line Lean 4 proof of Fermat's Last Theorem in 11 days

·34 views

Anthropic announced on September 4, 2026, that its Claude AI agents had generated a complete formal Lean 4 proof of Fermat's Last Theorem, spanning 13 million lines of code and roughly 29,500 intermediate theorems produced in just 11 days. The proof was verified mechanically by Lean's type checker and confirmed as valid by mathematician Kevin Buzzard of Imperial College London, who has been leading a human formalization effort toward the same goal since 2024. Buzzard acknowledged the proof checks out but described it as mathematically uninformative, characterizing it as a significant step for automated formalization rather than a contribution to mathematical understanding. The output, which exceeds five times the size of Lean's entire existing mathematics library, was designed to be machine-checked rather than human-readable, with all theorem names auto-generated. The computational cost of generating the proof has been estimated at around $300,000 in token fees, and a full build required over five hours on 96 parallel jobs, peaking at 153 GB of RAM.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Developer builds Firecracker-based agent sandbox and benchmarks it against Google's design

A developer published a write-up describing their own implementation of an isolated sandbox for AI agents, built on Firecracker microVMs. The project was a response to a sandbox architecture that Google had described publicly. The author measured the performance of their build rather than relying on theoretical claims. The post was shared on Hacker News, where it gathered 8 points and no comments at the time of capture. The full write-up is available on the author's personal blog.

0 comments Read more at Hacker News

ProgrammingDEV Community ·

OpenAI's always-on 'Dots' agents raise privacy and compute cost concerns

At its DevDay event on September 29, OpenAI introduced Dots, agents that run continuously in the cloud, each with its own computer and browser, and connect to many apps through plugins. Users can control when the agent acts alone or asks first, and sensitive actions such as changing passwords remain with the user. Critics such as Signal president Meredith Whittaker argue that deep access to personal data creates privacy and security risks, including exposure of third parties who never consented to an AI reading their messages. Because always-on agents are computationally expensive, OpenAI is also routing some work to a smaller, cheaper model, while companies report AI spending well above their budgets. Experts suggest granting access gradually and treating agent permissions as an architecture decision involving least privilege and cost controls.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Rate limiting Minimal APIs in .NET with the built-in middleware

Public APIs are exposed to clients that may flood an endpoint with requests, whether from a bug, a faulty script or deliberate abuse. Since .NET 7, the framework includes rate limiting middleware, so no third-party package is needed. The article demonstrates a per-IP fixed-window policy that rejects excess requests immediately with a 429 status and a Retry-After header. It also covers applying the policy to specific endpoints, keeping health checks unlimited, and writing integration tests with WebApplicationFactory.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

DejaVu: AI incident agent uses long-term memory to recall past outage fixes

An engineer describes building DejaVu, an AI on-call agent designed to remember every outage a team has resolved. The author argues that most incidents are recurring failure patterns that appear under different service names and error messages, and that general chatbots lack the context to recognize them. DejaVu works through a recall, reason, and retain loop, using the Hindsight open-source memory system to retrieve semantically similar past incidents and feed them to an LLM. In a side-by-side test, the same model reported 20% confidence without memory versus 98% with memory for a simulated cache bug. The author notes that a keyword search would miss many of the 6 cache-related incidents in a 20-incident dataset, which semantic recall can surface.

0 comments Read more at DEV Community