doph

"LLMs can't know anything because they're not connected to the real world"

also: "we know LLMs aren't useful because we read that in a study"

kind of a beautiful symmetry

I'm having a similar issue with embedded Rust. The HAL for ESP chips changes weekly (which is exciting!) but it makes any LLM suggestions bad if they aren't grounded by referencing the current crate in your project (which is something they are frustratingly reluctant to do).

my pet theory is that Rust's trait system is the cause of many of the LLM problems. there are many and nuanced ways for structs, even e.g. a tuple, to accumulate methods. it's easy to imagine this making it hard for an LLM to memorize how to interact with even basic types.

Seems like the Chinese Room has given lots of people access to the notion of a sufficiently advanced Markov chain. I think generalization is the real tripping hazard. When people use the "next token predictor" slur, they attack generalization and imply memorization, even if they don't know it.

someone who needed a crossover cable this urgently is clearly deserving of our sympathy if not respect

In referencing Netflix's dumbing down of their scripts, the author links not to the original article at n+1 magazine, not to a short synopsis of that article at PCmag, but to a blog post that itself is an even shorter synopsis of the synopsis. To do that in the context of this essay is baffling.

totally agree, and have seen the same [self]awareness in other coverage - that the coverage itself is what they seek. but why can't any of you take your own advice here? honestly, what will it take for media to resist spreading the killers' messages to future killers?