Conversation

Notices

1. Embed this notice
   Thomas 🔭🕹️ (thomasfuchs@hachyderm.io)'s status on Friday, 24-Jul-2026 01:27:30 JST Thomas 🔭🕹️

   I wrote about how LLMs don't know how to make software, and even got to quote Weizenbaum. Twice!

   https://mir.aculo.us/do-llms-know-how-to-make-software-no/

   • Embed this notice
     Thomas 🔭🕹️ (thomasfuchs@hachyderm.io)'s status on Friday, 24-Jul-2026 01:38:26 JST Thomas 🔭🕹️

     in reply to

     • Michael Kohl 🇦🇹🇹🇭

     @citizen428 the guy was so far ahead it's not even funny :)

   • Embed this notice
     Michael Kohl 🇦🇹🇹🇭 (citizen428@chaos.social)'s status on Friday, 24-Jul-2026 01:38:28 JST Michael Kohl 🇦🇹🇹🇭

     in reply to

     @thomasfuchs It's insane that the second Weizenbaum quote is 50 years old! If you leave off the citation, one could easily imagine this being written as a reaction to LLMs.

   • Embed this notice
     Thomas 🔭🕹️ (thomasfuchs@hachyderm.io)'s status on Friday, 24-Jul-2026 02:03:42 JST Thomas 🔭🕹️

     in reply to

     • cthos 🐱

     @cthos hence I’m mentioning randomness as input

   • Embed this notice
     cthos 🐱 (cthos@mastodon.cthos.dev)'s status on Friday, 24-Jul-2026 02:03:43 JST cthos 🐱

     in reply to

     @thomasfuchs LLMs are stochastic (not deterministic) algorithms, they don't always produce the same next token for the same input (unless you set the temperature to 0).

     Edit: but the stochastic bit is at the very end where it uses the temperature setting as a weight to determine which of the most likely tokens to output based on randomness.

   • Embed this notice
     Andy Balaam (andybalaam@mastodon.social)'s status on Friday, 24-Jul-2026 02:32:01 JST Andy Balaam

     in reply to

     @thomasfuchs this has lots of overlap with what I wrote yesterday https://artificialworlds.net/blog/2026/07/22/llms-produce-plausible-noise/

   • Embed this notice
     Thomas 🔭🕹️ (thomasfuchs@hachyderm.io)'s status on Friday, 24-Jul-2026 02:32:01 JST Thomas 🔭🕹️

     in reply to

     • Andy Balaam

     @andybalaam oh no you got one of those "LLMs work just like brains" replyguys as commenter lol

   • Embed this notice
     cthos 🐱 (cthos@mastodon.cthos.dev)'s status on Friday, 24-Jul-2026 23:27:11 JST cthos 🐱

     in reply to

     • pixxl

     @pixxl @thomasfuchs But it does, by weighting the probability of the token it will select at any given step towards 100% (at 0 T): https://techscribr.github.io/posts/llm-temperature/

     But I don't think the average person who knows what an LLM is would agree that "the LLM algorithm is just the bit that spits out vector weights" when every implementation out there includes the sampler.

   • Embed this notice
     Thomas 🔭🕹️ (thomasfuchs@hachyderm.io)'s status on Friday, 24-Jul-2026 23:27:11 JST Thomas 🔭🕹️

     in reply to

     • cthos 🐱
     • pixxl

     @cthos @pixxl My point is that randomness does not exist in the algorithms themselves, it's always an input. If you set it to 0 there's no randomness and the algorithm is strictly deterministic.

     There's also implementation details like possible race conditions or minuscule differences in measuring other inputs such as time-limiting parallel processing which could influence the outcome; those are essentially also random inputs.

   • Embed this notice
     pixxl (pixxl@mastodon.social)'s status on Friday, 24-Jul-2026 23:27:13 JST pixxl

     in reply to

     • cthos 🐱

     @cthos @thomasfuchs LLMs are deterministic, as they are just machine learning models. The way they're used (by the control system) could be non-deterministic. In the end it depends on the control system since they could work in any way, but using the same prompt and the same seed will always produce the same output with most systems. Arguably the control system (the thing that picks the next token from the distribution) is not part of the model itself.

   • Embed this notice
     pixxl (pixxl@mastodon.social)'s status on Friday, 24-Jul-2026 23:27:13 JST pixxl

     in reply to

     • cthos 🐱

     @cthos @thomasfuchs Temperature itself has nothing to do with determinism either. It all depends on the RNG used by the control system which picks tokens. If it's a regular seeded RNG then using the same prompt and same seed will produce identical responses.