Hacker News Reader

Comment by bootwoot | Hacker News Reader

True. But also -- how do humans do it? There are docs and there's other similar driver code. I wouldn't be surprised if Claude could build new driver code sight-unseen, given the appropriate resources

View on HN

Humans do it with access to the register-level data sheets, which are only available under NDA, and usually with access to a logic analyzer for debugging.

Usually, the problem with developing a driver isn't "writing the code," it's "finding documentation for what the code should do."

... and then figuring out where the hardware company cheapened out and created a whole unfixable mess (extra fun when you first ship your first 10k batch and things start failing after the vendor made a "simple revision"). Then finding a workaround.

> But also -- how do humans do it?

Probably a mix of critical thinking, thinking from first principles, etc. You know, all things that LLM's are not capable of.

Except it often is the case that when you break down what humans are doing, there are actual concrete tasks. If you can convert the tacit knowledge to decision trees and background references, you likely can get the AI to perform most non-creative tasks.

If you have to hold the LLM's hand to accomplish a task, using human intelligence to do so, you can't consider the task performed by AI.

I half agree. But two points: 1) if you can formalize your instructions ... then future instances can be fully automated. 2) You are still probably having the AI perform many sub-tasks. AI-skeptics regularly fall into this god-of-the-gaps trap. You aren't wrong that human-augmented AI isn't 100% AI ... but it still is AI-augmentation, and again, that sets the stage for point 1 - to enable later future full automation on long enough timecycles.

> if you can formalize your instructions

Isn't that...code?

No. Think of all engineering disciplines that aren't software. Those all depend on human-language formal instructions.

Formal instructions paired by tables are almost as rigid as code. Btw normal engineering disciplines have a lot of strict math and formulas. Neither electrical nor mechanical engineering runs on purely instructions.

The non-software engineering disciplines I'm thinking of rely on blueprints, schematics, diagrams, HDLs, and tables much more than human language formal instructions. More so than software engineering.

Disagree, they rely on both equally, not much more on one of them. Consider the process of actually building a large structure with only a set of such diagrams. The diagrams primarily cover nouns (what, where, using what), whereas the human language formal instructions cover the verbs (how, why, when). You can't build anything with only one of the two.

And sure, the human language formal instructions often appear inside tables or diagrams, that doesn't make them anything less so.

This is based on having worked with companies that do projects in the 10 figure range.

Scientific method. There are many small discoveries humans make that involve forming a hypothesis, trying something out, observing the results, and coming to a conclusion that leads to more experimentation until you get to what you actually want. LLMs can’t really do that very well as the novel observations would not be in their training data.

> But also -- how do humans do it?

Intelligence.

Reply on news.ycombinator.com