The problem is one of complexity, the whole being greater than the sum of its parts. Electrical engineering creates a thinking that is opposite to what you need for understanding biology or human culture.

This is relevant because LLMs have basically swallowed all written human culture. Engineering is about creating parts that interact in a clean, well-defined way, like a MOSFET transistor or a REST API. The goal is that that when you combine the parts, and they all operate within specs, you can still predict what the whole system will do.

So saying LLMs are just a step away from the Pereceptron (or an equivalent filter) is mathematically true, but almost useless. It's like saying that a modern 3D game is basically just transistors, logic, CPUs and memory, nevermind the pattern in that memory or how it came to be. You really can't ignore all the cultural inputs and feedback loops if you want any kind of useful understanding what this thing is.

When you say "cross-entropy loss" people without stats background go to Wikipedia, take a glance, and adjust their mental model to "inscrutable magic".

Thinking of the main difference as the trade-off in how much CPU, memory and storage is allowed is not really wrong.

The part that is wrong is to think of gzip as a method that might reach similar complexity or generalization. And more importantly, to ignore the advanced way how training data gets curated or generated for (instructed, chain-of-thought) LLMs. But even then. The mental model that the LLM's goal is text compression is not wrong. The question to ask next is what kind of text it is expecting to compress.

Or like a machine learning lab claiming SOTA on a benchmark, beating a well-known method that they re-implemented, possibly with bugs, on their private dataset, for millions of compute. But you don't get the source to check, and they don't release any intermediate results or ablation experiments. Aka, from the outside you can't distinguish it from corporate marketing.

Comments URL: https://news.ycombinator.com/item?id=49737326

Points: 50

# Comments: 9

Compare that to being asked for advice, to explain directly something you know or have done, with the expectation of the person asking to learn something. Like asking why some decision was made in the past, based on what information and context, instead of confidently saying that it was done wrong. Both will eventually get you the same information.

Or a login form that gets hidden after login, but clears the username and password only when you click "login back in". (Bonus points if the backend also enforces a 5min session timeout "for security".)

AI can mirror this style of writing, but it can't mirror the clarity of thought behind it.

The author is: https://sigmoid.social/@pynicolas@mastodon.social

Comments URL: https://news.ycombinator.com/item?id=49134351

Points: 2

# Comments: 2