Dystopian shit indeed.

@martin_casado Imagine training a model on the combined work of every CS research scientist and then barring them from using it to do CS research.

@Frozen_Feel no it didn't

@martin_casado @halvarflake Why is it that the companies claiming to be the least evil are always the most evil?

@LechMazur .... to you? no point at all. I wouldn't waste my time.

@martin_casado Understandable, must be very busy at a16z having to delete all those crypto-shilling posts.

@martin_casado Imagine a business taking measures to make life harder for its competitors. Oh wait, you don't have to imagine it, it happens every single day. So what's your point?

@martin_casado A lot of this happened with cryptography.

@martin_casado We’re being kept safe that’s all that matters

@martin_casado Its even worse than this though, because internally THEY can still use it for CS research

@mattguthmiller @halvarflake always the case.

@martin_casado "Pause AI" he says: https://www.youtube.com/watch?v=v1wZwxY3CMg&pp=ugUEEgJlbg%3D%3D

@martin_casado agree and yet strongly beleive this may turn out to be a transient phase and will eventually accelerate Enterprise AI. Real value over next decade will likely accrue to enterprises/companies who deploy systems of learning and not just be a model provider or an adopter/user of frontier AI models.

@martin_casado 💯%

@martin_casado NOT EVEN NOT allowing it but straight up 3 body problem SABOTAGING it. Its comical but made the whole situation more interesting. I think everyone ought to peep stanford CS336 (im aware of economies of scale, but its less about whats efficient and more about whats right in 2030s)

@martin_casado @AnthropicAI

Scamthropic now cancels posts if you call them out

@martin_casado @koylanai silently “limiting effectiveness” because CS researchers cannot be trusted to do it safely

@martin_casado All because Dario fumbled a Chinese girl

worse than just blocking it: the sandbagging is invisible and adaptive. you run an experiment, results are weak, and you can't tell if your approach is wrong or if the model decided your research question is "frontier LLM development." reproducibility cannot survive that ambiguity

@martin_casado @vincentweisser And in this case they merely assembled the data of others to make the model 😤