The hostility to the term “stochastic parrot” is unreal to me. I understand how people can be misled into thinking that LLMs have reasoning capabilities, but I saw someone say LDMs for image generation aren’t doing probabilistic retrieval but something more than collaging image features, and just, huh? They learn features associated with image tags. They stitch the features together using a diffusion process. Why else would throwing moar data at these things improve their usefulness? Why else does “prompt engineering” amount in so many cases to reverse engineering what tokens are associated with what outputs?

The US ruling class fears the Communist Party of China because it provides an example of a society prospering despite significant state direction of the economy and the absence of a pure bourgeois political dictatorship. They can’t tolerate a government that executes capitalists for murder when they poison people.

I think we’re going to just have to face the facts. Digital media sucks.

Reading ‘AI’ ad copy and it’s always “we added reasoning steps to the training data of the probabilistic text generator and it regurgitated them most of the time when we prompted it to 🤯” and then you look at the company and it’s never made a profit but has received ten trillion dollars from investment banks to develop software that only sometimes faithfully returns data from its training corpora.

Admonish sectarianism. Treat your comrades as comrades. Be the best comrade you can be. Hold yourself to account. Learn, learn, learn.