John Hardy1d ago

Re: SQL Lite doesn't provide unit tests

This is now looking extremely forward-looking. To defend itself from vibe-lifting, this open source project chose not to provide unit tests.

This acts to discourage cloning projects trying to move in and create vibe coded copies. It is extremely hard to clone SQL Lite and this was a deliberate policy on their part.

Expect to see that happen more and more on high profile open source projects.

#vibelifting #vibelaundering

Show threadTraveling OSM Salesgirl Problem ✊🏳️‍🌈🏳️‍⚧️🇺🇦23h ago

@jhlagado I have at least one friend (who is anti-AI, incidentally) who doesn’t consider something to be freedom-respecting (free/libre/“open source”) software…

  • if it takes a nontrivial amount of effort to build
  • if it doesn’t have documentation
  • if it doesn’t have tests

(Among other things.)

I understand the sentiment - legally granting the four freedoms doesn’t mean much if the author has made every effort to subvert their exercise in practice.

There’s no way around it - genLLMs in their current form must be outlawed.

  • Any public-facing genLLM must be freedom-respecting software (i.e. the source must be released under a freedom-respecting license);
  • it must declare its training data;
  • the training data must be used with the permission of its authors;
  • the weights must be released as freedom-respecting data;
  • and the weights must be considered derivative works of the training data for the purpose of licensing.

And this won’t happen as long as people in tech are being indifferent, passive, defeatist/inevitabilist, or (worst of all) idiotically embracing genLLMs.

Show threadJonas Schäfer

@contrapunctus @jhlagado I think just requiring the models to release their training data etc. is not enough.

Training of (serious) models requires so many resources that it's not viable for anyone to do.

LLMs, to be freedom-respecting, must, in my opinion, be trainable and usable on commodity hardware in reasonable time.

Mar 31 at 1:46pmWeb
Show threadJohn Hardy21h ago

@jssfr @contrapunctus

I would argue that even releasing all of the weights is not sufficient. There would be an argument that if you could reproduce absolutely all the initial conditions and all the seeds for all the random number generators as well as the neural networks weights, you might be able to reproduce the exact result and deterministically derive from the initial prompt. However, that's not practical and not the way people use LLMs.

Show threadJonas Schäfer21h ago
@jhlagado @contrapunctus Exactly.
Show threadJohn Hardy21h ago

@jssfr @contrapunctus

I made a quick explainer video discussing this issue here (6 minutes)

https://m.youtube.com/watch?v=2pf7zfCuV5w

AI, Code, and the Monkey Selfie

YouTube
Show threadJonas Schäfer21h ago
@jhlagado @contrapunctus "This video is AI-generated and therefore public domain." the irony.
Show threadJohn Hardy21h ago

@jssfr @contrapunctus

The US Copyright Office has described an LLM as a black box which is non-deterministically prompted by a human but the result has no human author.

Without a human author, no copyright is possible and therefore no licencing scheme freedom maximising or not has any relevance. The generated code is automatically considered public domain and this is not the basis of open source.