I think this only helps with security vulnerabilities (bugs that are not vulnerabilities are not that impactful, and I don’t see this generalizing beyond ordinary software security pre-takeoff). Plausibly after LLMs can rewrite whole codebases in a different language without making them worse, as the next step after that they could also be taught to rewrite them in a much richer dependently typed language in a way that obsessively tracks all security properties that come to mind.
LLMs still can’t rewrite codebases (well enough to matter). So this is probably not yet a 2026 thing, but 2027-2028 is plausible, and it probably mostly helps with the still-mostly-speculative future problem of more effective LLM-enabled hacking.
I think this only helps with security vulnerabilities (bugs that are not vulnerabilities are not that impactful, and I don’t see this generalizing beyond ordinary software security pre-takeoff). Plausibly after LLMs can rewrite whole codebases in a different language without making them worse, as the next step after that they could also be taught to rewrite them in a much richer dependently typed language in a way that obsessively tracks all security properties that come to mind.
LLMs still can’t rewrite codebases (well enough to matter). So this is probably not yet a 2026 thing, but 2027-2028 is plausible, and it probably mostly helps with the still-mostly-speculative future problem of more effective LLM-enabled hacking.