July 5, 2026
Proof or it didn’t compile
Lean Software Scaling Laws
Could a niche coding language save us from AI-made buggy software? The internet is fighting over it
TLDR: Researchers want to test whether stricter coding languages like Lean become easier for AI to manage as software gets bigger, which could matter for future security. Commenters split hard between “this could save us from buggy AI code” and “this is fantasy cosplay for people who hate simple tools.”
A research idea about teaching AI to handle giant codebases better somehow turned into a full-blown comment-section cage match. The basic pitch is simple: instead of asking which coding language is most popular today, ask which one stays more predictable and less chaotic as projects get huge. The star of the proposal is Lean, a formal math-heavy language that fans say could one day help AI write software that is actually safer and more secure, not just fast and sloppy.
And yes, the community had feelings. Supporters were practically popping champagne, arguing this is the kind of nerdy long-game thinking that could stop the future from becoming a landfill of AI-generated security holes. Their hottest take: today’s easy languages may be winning because AI has seen more of them, but that doesn’t mean they’ll keep winning when the code gets enormous and messy. Critics, meanwhile, rolled their eyes so hard you could hear it through the screen. They called the idea “rewrite the world in monk-mode programming” and joked that the cure for bad AI code is apparently making everyone suffer first.
The funniest running joke was that current AI already produces “vibe-coded spaghetti,” so asking it to rewrite the world into a strict, proof-heavy language sounded to some like hiring a raccoon to organize your tax records. Others fired back that if AI is going to write most future software anyway, we’d better give it rules strict enough that it can’t freestyle us into a cybersecurity apocalypse.
Key Points
- •The article proposes measuring how coding LLM perplexity scales with codebase size to estimate predictability scaling laws across programming languages.
- •It describes two evaluation approaches: training and fine-tuning models from scratch, or measuring perplexity across increasingly large context windows.
- •The article identifies Lean as a test case that may have weaker initial LLM performance but stronger scaling behavior on larger codebases.
- •It argues that better scaling exponents could make some languages easier for LLMs to understand, repair, and generate as projects grow.
- •The article frames this research as relevant to software correctness, formal verification, and long-term cybersecurity outcomes.