Holding the LLM Stack in Your Head | Hacker News Reader

Holding the LLM Stack in Your Head

thegustafson.com

9 pointsyayitswei5 comments

I appreciate your openness in stating upfront about the providence of the information here and applaud it.

However I have a pressing question, why should anyone trust this. I understand the process of self talk with claude can be useful for your own understanding. But this is a published result and is only really good when you've distilled it. Ideally you did real testing on it for validation.

The feeling I get from reading these articles, is "Well maybe this is true, but I don't know if I can trust it, so should I go and read a textbook?" If the goal is to distill learning here into something digestible for the reader. Then you as the author should provide proof that the statements are correct for me as the reader. I can't really tell from what's written here if you did that or not. If its just a bunch of generated articles which were useful for you in making, but for me as a reader I don't know if I get any value from them.

A common retort to this idea, is "Why should we trust anything a person says, people can be wrong too." However we typically would look for "proof of work" which AI really sidesteps here

In the first paragraph, "The goal isn't rigor, it's intuition that survives contact with real systems".

"I wrote this series as a learning exercise, in collaboration with AI (Claude Opus 4.8). It's me thinking out loud with a model to understand the stack end to end, not an authoritative reference. Take it with a grain of salt, and double-check anything you plan to rely on."

The author admits as much. It's basically him prompting, "Hey Claude write a long series about LLM tech." Waste of time.

Also I just realized, the information might be subtly wrong or misleading because Anthropic makes Claude dumber when trying to reverse engineer LLMs, which is hilarious. And the author is basically saying hey, I don't know if any of this is correct but, like, enjoy reading it! Hah.

Looks like good value, interactive toys add a lot of intuition. I would love to see this expanded to cover other modalities, not just text. And then expanded to non LLM, specialized models, like tts, tabfm, etc.

bug: /toybox#06 clears page

Why would I want to read this, given the author doesn't trust it and hasn't verified the content of this HUGE series of generated articles?

Of what possible use is it? If someone believes Claude, surely they'd just have a conversation with claude themselves. Is this intended just to function as SEO spam, or is there some other use?

Generating massive amounts of AI generated slop like this breaks all the assumptions of the internet, including on sites like HN.

Flagged as slop.

Reply on news.ycombinator.com