Forgive the smelly domain.
This is a fantastic short essay by Rich Hickey.
AI is a tool for getting stuff done. But that said, Rich has insights on how AI loses benefits of traditional tools. Disagree with the title, but still worth reading imho.
Disagree. What would you call a 3D printer? I’d still call it a tool. Even though I don’t wield it or hold it in my hand. You tell it what you want to do and it builds something physical for you.
There is skill involved in 3D printing. Incorrect instructions lead to poor results. There are many variables to consider and you learn how to manage them the more practice you get.
What would you call a test framework? Something that can run your software tests and present information in a coherent way? I’d call it a tool. I use the testing tools available to me to speed up and improve my code.
LLMs are the same. You give it instructions and do something with the output. The difference is that the LLM can also use you as a tool. It will give you instructions and ask for results. But LLMs are not perfect. They will give you incorrect information. They will lead you down dead ends if you don’t know what you’re doing.
This makes LLMs quite difficult to use. You need knowledge in the subject you’re asking it about and you need to be skilled at using the LLM to acheive good results. A lot of people don’t have either.
You need to be strong enough to wield a chainsaw, you need to be smart enough to wield an LLM. They are both tools.
Dude, what you give LLM is not instructions. It doesn’t follow your instructions. It brings you back what might look like you want.
A CPU, on the other hand, carries out the instructions exactly like you’ve said.
There is no way to avoid incorrect answers from LLM. It’s probablistic; there is always a chance of getting a wrong answer no matter how you improve it. With a CPU, you can reduce errors by improving the design. It’s based on physics; it’s logical.
I advice you to stop treating LLMs as blackboxes and only observe it from the outside. You’ll never notice the difference if the error rate is low enough.
Of course, an LLM is a probability machine which outputs the most likely next token based on the tokens you give it.
I’m not sure why that doesn’t make it a tool though. The output of this probability machine can be useful if you know what you’re doing.
I din’t think you can compare it to a 3d printer. When you have no idea what you are doing you will get no output or you may even damage your machine.
With an LLM you could just go “Oi shitass make botoshpop” and it will create you something at least resembling your request. You will get better results if you describe in detail what you want, but in the end what exactly it outputs is not in your hand.
There is not a knob you can turn to make the output perfect. You just give it more info on what exactly you want and it will output something.
With every tool you are in full control of what it is doing. Ai is just som guy that does things for you. You don’t need to know how it works or why it works. It just does it for you.
Yeah i agree with most of that but you are in control of what you do with the output. If the output is full of mistakes you don’t notice, that’s on you and you have used the tool poorly.
Blindly trusting the output of an LLM is not using the tool properly.
Yeah, but that is the point i’m trying to make. You can trust the output of your tool, but you can’t trust the output of an llm. I can trust that my 3D printer will print exactly what I wanted it to print. I cannot trust an llm to produce me exactly what i wanted from it.
If I trust my own skills I can trust the output of a tool.
It doesn’t matter how good I am at prompting, there is always a chance that an llm will just do something I didn’t tell it to do.
You are not the creator with an llm you simply delegating the work to someone else and then having to check what that person has done.
But even if you configure a 3D printer properly there’s still a chance the output will be flawed.
Just because you can’t trust the output doesn’t mean it’s not a tool. If you expect it to do exactly what you want you’re using the tool wrong.
You use the LLM to help you build the system that does exactly what you want every time. LLMs are just a tool to help you get there.
But this will most likely be because i configured something wrongor a part inside the printer has failed.
This would be like a service being down that the llm uses.
If a llm were 3d printer there would be a chance that the part it is printing for you suddenly had a different width or the endings were flat instead of rounded. And if you ask it why it did that it just responds “You are absolutely right for calling me out. I just thought it was easier to print this way.”. This is not something a tool should ever do.
Why not? To me that’s just what the tool does.
You can chop off your arm with a chainsaw.
Tools don’t have to be safe. They don’t have to protect you from mistakes.
Why is an LLM not a tool?
Its not that a tool has to be safe.
If you cut of your arm with a chainsaw then you have taken the actions to cut off your arm with the chainsaw.
An llm chainsaw could bend its blade 180° and cut off your arm without you even wanting it or taking any step in that direction.
If you hold the chainsaw and the llm chainsaw the exact same way, the normal chainsaw would not cut off your arm, no mather what (exept for a catastrophic failure wich can happen to both chainsaws). With the llm chainsaw you roll the dice every time and sometimes it just says “nah, tree too big. I’ll cut your arm instead because it is easier”.
That is the difference for me. One thing is completely dependant on your actions and if you do everything correctly than it should produce the outcome you expect. With an llm you could do everything right and it does something you didn’t want without.
A tool should always do what you want and not have a built in randomness to the output.
‘AI’ is Not a Tool
(I keep delivering this as a quip, here’s a bit more)
First order of business - is ‘tool’ going to have an interesting, distinguishing and thus illuminative meaning or be a mere platitudinous stand-in for “anything useful”?
We’ll save tool as “anyone/thing that can be exploited” for the AI titans, for whom VCs, politicians, wannabe-titan-CEOs, software devs, rural towns etc are all ‘tools’. The pockets on the front of my shirts are useful (I need to keep my reading glasses handy), but they are not tools. Similarly, fast food drive-thru windows are useful, but are not tools (more later).
Some definitions of ‘tool’:
What I like best - A tool is something you wield.
- wield - to use (as a tool or instrument) especially with full command or power : handle, manipulate, control
Especially useful in that are:
Tools are implements with which you do something, they do not do things for you. Tool use builds, rewards, and often requires, expertise and skill. The continuous application of technique and energy, and sensory feedback, are essential to the buildup of expertise (from experiri - “try”).
Good tools have many properties LLMs lack - they are predictable, reliable, deterministic, of bounded function, hopefully simple etc.
What you learn while wielding a tool is something about the world, the task, the physics, logic etc of what you are trying to do or the situation you are trying to overcome. And often, something about yourself.
As a lifelong tool builder and user, I know what tools are. LLMs are not tools.
Control, sensory feedback, intention, intimacy, continuous direction, expertise etc are all missing from LLM use. LLMs are not wielded. LLMs are automation.
- automation - (of a device or process) working by itself with little or no direct human control.
A drive-thru is not a tool for cooking, it is automation for avoiding cooking. LLMs are not tools with which you write code - they are automation with which you avoid writing code, to which you abdicate writing code.
You don’t direct the action of an LLM - you make requests and it acts on its own, without your energy or continuous direction. This request/automation cycle is not wielding, it’s not control, it’s not continuous, it’s without feedback.
What you are doing while using automation, instead of learning real-world truths, is attempting to divine the details du jour of some automaton. And striving to game them - e.g. ordering a custom burger (no onions) in order to get a freshly made one. Or filling voodoo.md with just the right incantations. You are only a software update from everything you’ve ‘learned’ being rendered obsolete by the latest algorithm. (A soggy cold burger from which the onions have been removed) You have traded working with the world to working for a capricious robot.
It is a mistake to be haphazard about “tools” being equivalent to utility, e.g. saying - “the LLM didn’t write my paper, I just used it as a tool to read and summarize some research papers”. But you didn’t wield it to read those papers (as you might a Kindle). It read the papers, not you.
What do you miss by abdicating doing the reading? Pretty much everything. First and foremost, you miss learning how to write from those who can. You miss learning how to formulate an argument. You miss noticing things. You miss loading your background mind with fodder for novel ideas. You miss building trust in some authors and suspicion of others. You miss all the jokes. You miss all of the edges. You miss learning how to think.
Tools don’t get “misaligned”, they don’t “go rogue”, they don’t “escape containment”, “conspire” etc.
LLMs are not tools. I know many people use the phrase “just tools” simply to indicate “they are what you make of them”. Except that’s not true. Automation has intrinsic properties and differences from tools that you ignore at your peril.
Rich
Thanks for keeping me off linkedin
If you’re not an idiot, and you (nevertheless) use LLMs, they probably meet most of those criteria.
I don’t like how the author buys into anthropomorphization. LLMs do not act on their own.
He’s right that tools don’t “go rogue”. Neither do LLMs. That’s just a lie to try to avoid accountability. ChatGPT didn’t hack the Australian government; PEOPLE did, and those people belong in prison.
LLMs ar intrinsically undeterministic. You cannot posibly control what they do, and you (realistically) don’t control the data tgat they have been trained on and how they are tuned to behave. Respectfully. I think your statement shows a lack of fundamental understanding on how they work.
It’s a matter of degrees, and of use case.
If your use case demands 100% accuracy, that’s a dead end.
If you can’t validate the output, then you need to go back to the drawing board.
I keep going back to the same rule of thumb: if you can’t do it without AI, you can’t do it, period. It’s a time saver, not a magic wand.
oh damn, i’m stealing that last one.
I think I’ll eschew any essay that attempts to define such a vague and old word as “tool” to make a point, as it seems likely it will all be pointless semantic bollocks rather than anything useful. I think I got over this kind of essay in first year of university.
The fact that business people use language and words to diffuse bullshit does not mean that words have not a meaning. In contrary, language is important, and we should wrestle to express thoughts clearly. Rich Hickey is much better at that than the average person, e.g. see his youtube talk “Spec-ulation” on versioning and compatibility, or his thoughts on identity and state in programming.
Language is important, but “if we define a commonly used word in this extremely narrow way, then…” is a losing proposition for that reason
pointless semantic bollocks
That’s kinda Rich’s schtick
I appreciate him taking this tack.
When slopheads say “it’s just a tool, like a hammer”, I think the natural response among careful thinkers is to jump straight to Heidegger. Like, “just a tool? Tools shape the essence of your understanding of reality, and the trajectory of civilization itself! How can you say ‘just’?!”
But you don’t even need to dive into all that if you just wanna dismantle what’s obviously a thought-terminating cliche. “A hammer? Really? I can’t recall the last time I told my hammer to search my house for things that need to be hammered, go ahead and hammer them, and send an email to me and my wife with the results.” Maybe this is quite unlike a hammer.
Nothing fantastic about it. I would argue the text is a proof that people can write slop too…
When Mr. Rich “the clojurefather” Hickey speaks. We listen.
