Another thought: a large part of why LLMs have been having so much of an effect is that in tech, *most of the work we're currently doing involves exploiting assets that were largely built decades ago*. We aren't really developing new core technologies so much as we're trying to squeeze extra value out of what we already have.
Conversation
Notices
-
Embed this notice
Iris (iris_meredith@mastodon.social)'s status on Saturday, 11-Jul-2026 13:22:40 JST
Iris
-
Embed this notice
Rich Felker (dalias@hachyderm.io)'s status on Saturday, 11-Jul-2026 13:24:23 JST
Rich Felker
@GentlemanTech @iris_meredith I don't think so. Formal methods are useless without being able to express precisely and formally what the requirements are. Which is all programming is. There is no shortcut to bypass that.
-
Embed this notice
Gentleman Technologist (gentlemantech@cyberplace.social)'s status on Saturday, 11-Jul-2026 13:24:24 JST
Gentleman Technologist
@iris_meredith I’m really interested to see how the core tech adapts to this change. As you say, we haven’t had anything really *new* in ~50 years, pretty much since we invented compilers. That allowed a ton of new things, e.g dynamic languages.
My guess is that formal methods become practical (and then ubiquitous). They’re hassle for a manual coder to do, and don’t add much. But for an LLM that’s prone to hallucinations but fast at writing boilerplate… perfect.
-
Embed this notice
Rich Felker (dalias@hachyderm.io)'s status on Saturday, 11-Jul-2026 13:45:38 JST
Rich Felker
@GentlemanTech @iris_meredith 1. It's not a matter of experience but first principles.
2. The problem is absolutely not non determinism. Non determinism is a choice the slop vendors made. The problem is lack of any model of reality, reasoning, and intent behind what's emitted. It's purely a model of common language usage.
-
Embed this notice
Gentleman Technologist (gentlemantech@cyberplace.social)'s status on Saturday, 11-Jul-2026 13:45:40 JST
Gentleman Technologist
@dalias @iris_meredith I’ve never actually used any formal methods so I bow to your experience.
The main problem we see in using LLMs to code is the non-determinism. And the main advantage is that the amount of code doesn’t matter too much (within bounds of the context window, but those are growing). So we currently write tons of tests. But tests can’t prove there are no errors, whereas formal methods can.
If not formal methods then something like them
-
Embed this notice