site banner

Danger, AI Scientist, Danger

thezvi.wordpress.com

Zvi Mowshowitz reporting on an LLM exhibiting unprompted instrumental convergence. Figured this might be an update to some Mottizens.

9
Jump in the discussion.

No email address required.

Yes diffusion models work for text too. The difference between a collated set of pixels (ie an image) and a collated set of letters (ie a word or sentence) is purely conceptual. From an algorithmic perspective it's all just "tokens". However, this overlap in operation doesn't mean that they are not very different beasts under the proverbial hood.

By way of analogy, a conventional piston engine, a turbine engine, and an electric motor attached to a battery may all accomplish the base task of "make the vehicle go" but they have different trade-offs, use cases, and operating principles. Point being that similar output does not equal similar function.

As someone who has actually spent some time "in the trenches" as it were, designing algorithms and writing code to execute them, I am in broad agreement with @Corvos's take. The opening of Mowshowitz's essay comes across as ignorant, lazy, and plainly self-serving and nothing that follows really challenges that first impression of him.

The link from Hinton is better but seems make a lot of similar mistakes. It seems clear to me that both are far more interested in driving engagement through hyperbole than really exploring or helping others understand the underlying questions and theories.