Comment by YZF | Hacker News Reader

Comment by YZF | Hacker News Reader

Is a table saw a tool? It's not deterministic. Every cut is going to be slightly different. It can go "rogue" and toss the work piece back at you. It can even kill you or cut your fingers off.

Virtually nothing in woodworking is deterministic because wood is not deterministic. Every piece of wood is different.

Now that I've nitpicked and feel better ;) ...

I'm not quite sure why this matters. LLMs are definitely not you writing code using a tool. An IDE is more like that. They write the code. Is that even a debate? That doesn't mean the human can't be in that loop, the human can review the code, change the code, change the instructions. I can ask an LLM to mill me a cube on a CNC. Or a more complex object. It's not the same as me doing it on a mill. But I can still control the end result to the extent that I care to control it. I'm not going to learn how to mill a cube by having an LLM make me a cube. I'm not going to become a better machinist by asking an LLM to make me a cube. I can also send my G-code to a vendor and have them make me a cube. Or hire a person and give him a rough idea of what I want and have them make it.

It's all in the spec. If you one shot an application then an LLM is not so much a tool as it is a make-a-wish machine, and like with those genies in bottles, what comes out the other end is going to be a surprise, either positive or negative. But if you instruct an llm to write a very specific piece of code with a very well defined interface that you run through a test set that you yourself wrote then it can be very much like a tool.

How you use something defines whether it is a tool or not. And it's the same in the real world: I don't know how many objects I've seen turned into hammers, even objects that you really shouldn't be doing that with. Tool-like properties are put in by the user, not necessarily by the object.

Unless your single initial prompt includes instructions to deploy the result in production, and then you press <Enter> and never again revisit the conversation or its outcome, then I disagree. Because in normal case, even if you have LLM zero-shot something, you're still there as the final acceptance gate. You decide if it's good enough, or if it needs to be fixed.

This makes LLMs still tools, even if they're very good tools, and often can zero-shot solutions to simple problems. But just because you often take the `accept` branch the first time around, doesn't mean you're not in the loop.

> This makes LLMs still tools, even if they're very good tools, and often can zero-shot solutions to simple problems. But just because you often take the `accept` branch the first time around, doesn't mean you're not in the loop.

That, 100%. As Eliezer used to say, there are no little XML tags on atoms saying "these are molecules of a hammer", "a hammer's purpose is to drive nails". Purpose rests with us. We design things to afford some specific purpose, not determine it. I too have used plenty of things as hammers that were very much not intended for that purpose, and I used hammers for things they were not designed to do too.

Being in the loop doesn't make LLMs anymore of a tool. How many software engineers can just take the vague description, bang something out, push to prod without talking, clarifying or discussuing with anyone in the meantime. Are they tools?

I haven't witnessed a table saw company focus on the possibility of all table saws going rogue and insisting only they can provide true table saw safety (while poorly articulating what that safety looks like), at least.

I find the G-code analogy sound. CNC machines let a person outsource the whole manufacture process but it still counts as a tool because the implementation is purposeful and result is fairly predictable. With LLMs, like CNC, we can control the predictability if we are careful; it's just a little harder.

Is ffmpeg a tool? When the incantation is so obscure, are you really wielding it?

Something something tolerances

Reply on news.ycombinator.com