On the map

At the centre AI-written code makes formal proof necessary, because human review of all that generated code becomes the bottleneck. Holds MKMartin Kleppmann Read this korrent →

• Same subject: LLMs, formal proof Formal verification is about to become economical, because models are getting good enough at writing the proofs that humans no longer have to. Holds MKMartin Kleppmann
• Same subject: code generation He does not read the boring parts of generated code, the data-shuffling and button alignment; what touches the database he reads and reviews. Holds Peter Steinberger
• Same subject: code generation If you are not going to read the generated code, what you need is conformance testing: proof that it performs within the boundaries of the code it replaced. Holds CMCharity Majors
• Same subject: formal proof Formalising a proof in Lean currently takes about ten times the effort of writing it out: doable, but annoying. Holds TTTerence Tao
• Same subject: formal proof Full formal verification is a waste of money for most software: near-perfect is reachable with ordinary techniques at a fraction of the cost. Holds HWHillel Wayne
• Same subject: formal proof Nuclear power plants, the standard example of software that must be proved correct, do not care about formal verification: thorough testing is enough for them. Holds HWHillel Wayne
• Same subject: formal proof When formalising a proof costs no more than writing it, mathematics will flip: papers written in Lean first, and journals refereeing only for significance because correctness is certified. Holds TTTerence Tao
• Same subject: LLMs A language model is no substitute for a well-specified conventional algorithm, so it cannot simply be dropped into a complex problem and trusted. Holds GMGary Marcus