Elektrine lite

← Feed

@kornel@mastodon.social

Post #942389

2026-03-18 12:18 UTC

In today's Coding Black Mirror episode: what if AI code generators get sped up to emit tens of thousands of tokens per second? For now we're pretending LLMs are just sloppy devs and try to review the output, which is already exhausting. When it gets to the point of being able to rewrite the whole project on every keystroke, source code won't be the source any more, but a discardable compiler output. Switching languages will be like applying Photoshop filters?

Replies (6)

  • @kornel@mastodon.social 2026-03-18 13:23

    Visual Basic's `On Error Resume Next` can be galaxy-brained to `On Error LLM("rewrite the program to stop doing that")`

    Open ##2769025

  • @scottgal@hachyderm.io 2026-03-18 12:24

    @kornel@mastodon.social already happening Taalas has an ASIC with 4b class models already. 16k toks/s

    Open ##2769026

  • @iainvdw@mastodon.social 2026-03-18 12:36

    @kornel@mastodon.social I’m afraid predictability and idempotency will be things of the past if that ever happens. Switching from one language to another, back and forth for several times, is most likely going to act like a destructive Photoshop filter or image compression, losing more and more details in the process. How can we rely on that for critical business logic or sensitive output?

    Open ##2769027

  • @kornel@mastodon.social Speed is not much of a bottleneck at the moment. Most frontier models (e.g. GPT-5.4, Claude) produce around 60 tokens/s. That's not exactly blazing but… Rust repo currently is ~840 MBs. It would take a model about little under 2 months to generate that much output. That's already several orders of magnitude less than it took us to get here. The bottleneck is the quality of the output and ability to specify the constraints. It's not that easy to describe Rust in sufficient details so that AI could actually write a decent Rust.

    Open ##2769037

  • @jarek@sunny.garden 2026-03-18 14:05

    @kornel@mastodon.social my half-joking take has been that creating good specifications will get more and more important, and as a result we might be blessed with UML once again

    Open ##2769039

  • @jsbarretto@social.coop 2026-03-18 14:37

    @kornel@mastodon.social Code is not the artifact and has never been the artifact though: the real artifact is the specification the code implements and the practices and rituals that allow it to be comprehended by human beings. Sometimes that specification exists in a form that can be written down, and sometimes it's an informal set of ideas in somebody's head, but the code is functionally impossible to integrate into wider systems without it. LLMs might remove the need for boilerplate - code that is superfluous with respect to the specification - but that's probably all they'll do.

    Open ##2769040