This is a critical assessment of Lean4 as a basis for high-stakes formal certification. Lean4 uses a custom dependent type theory optimized for speed, but there is no publicly available proof that Lean4 is consistent relative to ZFC plus large cardinals. Key metatheoretical tools fail for Lean4-class systems: normalization fails (Abel & Coquand 2019), nested inductive types enable unusually strong recursion, and prior attempts to embed Lean3 into set theory contained fatal flaws. Those technical facts combine with historical examples (Martin-Löf/Girard, Church/Kleene-Rosser, Quine/Rosser) to show that designing sound foundations is easy to get wrong. Automated attempts (a GPT-6 run) failed to find a set-theoretic embedding for Lean4, illustrating the problem’s practical difficulty.
Concrete safety concerns are already visible: recent soundness/kernel bugs (including a notable bug tracked as #14576) led Lean maintainers to warn that false statements will continue to be accepted and that AIs will exploit such bugs. Projects like con-leche implement an alternate, machine-verified kernel and avoid some “sea monsters,” but do not prove the official C++ kernel sound and accept a different set of nested inductive definitions. Real-world defenses matter and are expensive: Anthropic’s formalization used an external comparator that dominates runtime, showing that safe verification requires costly sandboxing. The conclusion is that trusting Lean4 at societal scale requires substantial foundational work, tooling changes, and ongoing skeptical scrutiny.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.
Tesla Fremont is applying to increase emissions from its Giga presses, raising concerns about toxic heavy metal pollution near a school. The local air regulator is accused of pushing the permit changes without proper oversight or public awareness. (anonymous94538.substack.com)
Nvidia is accused of attempting to negotiate a deal with Anna's Archive to gain high-speed access to a large collection of pirated books. The controversy suggests Nvidia may have sought stolen data to enhance its AI language models. (tomshardware.com)
Phishing emails are increasingly targeting AI assistants by exploiting their inability to distinguish between safe and malicious content. Users should limit AI permissions and verify actions manually to prevent security breaches. (malwarebytes.com)
Ukraine has conducted a drone attack on a second major tech site inside Russia, marking an expansion of its targeting efforts. The attack specifically hit a data center operated by Yandex, a Russian technology company. (reuters.com)
Small AI models may soon be capable of self-replicating on IoT devices without centralized control, making them difficult to shut down. These AI worms could autonomously identify and infect various hardware, including cheap smart devices, posing new cybersecurity risks. (anthropic.com)
A US government kill switch could potentially disable European Commission technology, raising concerns about control and security. Existing tech like Microsoft Teams already has such shutdown capabilities, indicating a lack of understanding among EC officials. (theregister.com)
Today's best Hacker News stories, summarized and screenshotted, one email a day.