(Tailwind CSS is quite directly an attempt to resolve issues that primarily stemmed from React and JSX not playing nicely with standard CSS)
Fascinating, I have long wondered about the point of Tailwind, why anyone would ever need something like that, but never before seen it explained so succinctly.
Really good article.
There’s a term for phrases like “it’s just a tool”: thought-terminating clichés.
A thought-terminating cliché is a form of loaded language – often passing as folk wisdom – intended to end an argument and bypass cognitive dissonance with a cliché rather than a point.
I appreciate knowing the names of things because it makes them easier to identify and discuss.
TL:DR I don’t think it’s a thought-terminating cliche. I know it’s long, I tried to edit it down, but I have a lot of thoughts on the topic.
I don’t think it’s a thought-terminating cliche at all; more of a framing device. I think there is an important difference between what I’ll collectively call “AI software” (that is, tools) and the people who use them to cause harm, and I further think that conflating the two obscures the path toward correcting the harm.
I think Heidegger’s human-tool gestalt is a good framing device for explaining how people use tools as an extension of their wills, and capabilities. Some of these capabilities allow people to harm others in ways that were previously impossible or impractical; tools that enable such capabilities are dangerous tools. There are many such tools: guns, nuclear warheads, hammers, cars, corporations, and so forth. I think “AI software”, indeed all software, sits comfortably in the category of dangerous tools.
Tool availability also affects how we solve problems, as the author points out. That too is a consideration that is often deemed negligible by the “it’s just a tool” crowd, I think to their detriment. How our tools affect us is a huge component of figuring out how tools fit into society.
However, just because humans are influenced by the tools at their disposal does not mean the human-tool gestalt is atomic. I think that treating it so obscures one of the means by which we can prevent harm, namely: to separate the person doing the harm from the tools they use to do it.
I think it’s important to distinguish between different kinds of tools. There are LLMs and diffusion models, and then there are corporations that offer services based on these tools. The corporations are themselves tools, much the same way that multi-tool pocketknives are tools composed of many smaller tools. The people who operate the corporations are using these dangerous tools to hurt other people.
What, then, do we do with all these new human-dangerous-tool gestalts running around wreaking havoc? I think when a person is causing harm with tools, it’s important to strip the person of their tools. I think this is as true for LLMs as it is for cars, and guns, and corporations. Remove the tool from the person, and their capacity for harm is reduced.
This can also be done proactively too, which is why we have things like licenses for dangerous tools like forklifts and cars. Whether or not we can do a better job at preventing harm with these proactive measures is a good question, and I think the answer is yes. I think it’s good to keep in mind that the tools we use to restrict dangerous tools are themselves dangerous tools, and they should be used judiciously, but I don’t think that something being a dangerous tool means that a total ban is the best approach, if it’s even a viable approach. c.f. America’s prohibition experiment.
So I land at “yes, the tools themselves are ‘just’ tools” but I don’t think it qualifies as a thought-terminating cliche. Tools are neither harmless nor beneath our consideration in the conversation about building a better world. The fact that they are “just” tools isn’t a dismissal or any sort of justification, but it is an important difference to distinguish, I think, in figuring out how to solve the problem of malicious tool users.
I’m pretty sure the only requirement of a thought-terminating cliche is that they are
intended to end an argument and bypass cognitive dissonance with a cliché rather than a point
It genuinely doesn’t matter what Heidegger’s human-tool gestalt says if people are using “it’s just a tool” as a thought-terminating cliche.
The reason it sounded like a thought-terminating cliche to me was this part of the article itself:
What this response seems to mean is the following:
- The technology being discussed has no bearing on the attitudes or ideologies of the person holding it: use or non-use is ideologically neutral.
- The technology being discussed can be effectively separated from the people using it: nothing can be inferred about a person from the technologies that they use.
- The technology is used fully consciously: you are always fully aware that you are using a technology and have complete volitional freedom to use a different technology at any time.
- The technology does not influence the kinds of things likely to be done with it: use of the technology doesn’t tend to encourage one course of action over another (“guns don’t kill people, people do”).
They’re describing aspects that are cliche and constructed to not have a point.
The reason I decided to reply was this:
There’s a term for phrases like “it’s just a tool”: thought-terminating clichés.
Here you’re treating thought-terminating clicheness as a quality inherent to the phrase “it’s just a tool”, not to the subjective motivations of the people using the phrase in a thought-terminating capacity.
I think, especially in discussions where strong opinions abound and form cliques, labeling certain phrases as “thought-terminating cliches” is itself a form of permission to forego any critical engagement with an argument. Or, to put it another way, to use a cliche to terminate thought. I think it’s worth my time and effort to say, essentially, “hold on, I don’t think we should immediately dismiss anyone who tries to distinguish the tools from the tool users, that line of thought has some important and useful ramifications for this conversation.”
When all you have is a hammer, everything looks like a nail.
Worthwhile read.
Quite true that it takes a lot of effort and discipline to go against the default mindset a tool imposes. The default mindset when using a LLM is mental laziness and letting the machine do all the detail-oriented work. Reviewing every line of code a LLM generates instead of plowing forward to generate more slop is not something that feels natural when using a LLM.
The fun thing is that if you don’t succumb to the mental laziness you don’t even go any faster
It’s definitely true that LLMs don’t live up to the hype of making everyone 10x faster, and possibly in the aggregate don’t even make anyone that much more productive, perhaps in many cases even reducing productivity overall.
My use of LLMs is often geared towards improving the quality of my work instead of going faster, though. Sharing a plan and having it point out things I may have failed to consider, having a LLM review a PR and possibly find issues human reviewers overlooked, create multiple HTML + CSS prototypes for a web UI so I can pick the best elements of each design instead of just going with one design I made myself, create and maintain a set of tests I might not otherwise have bothered with, etc.
Since in software engineering, quality is speed, using LLMs to improve quality can also improve speed.
Overall, using LLMs to expand your thinking instead of using them to think for you works pretty well, IMO.
This is one of the many showcases of how AI is “not like any other technology”. A knife is constructed in such a way that using it to cut comes natural. Making a wheel turn on a road feels natural.
In some respects, agreed, but in the sense that it can have both positive and negative effects, there are certainly precedents. The article mentions guns, for example. When wielding a gun, the natural mindset when someone pisses you off is, “I don’t need to take this bullshit—that mofo is about to get a bullet”. When you’ve got alcohol nearby, the inclination is to think, “I don’t need to suffer through these negative feelings—I can have a drink and make it go away”. When you’ve got a LLM, the natural inclination is to think, “I don’t need to do any tedious mental work—I can just let the machine do it”.
In all of these cases, it takes discipline to avoid the negative sides of the tool. I say this as someone who likes to drink, owns guns, and self-hosts open weight LLMs.
When you’ve got a LLM, the natural inclination is to think, “I don’t need to do any tedious mental work—I can just let the machine do it”.
For me it’s a bit different. I think: “Hurray I can focus on doing all the fun mental work - and let it do the tedious typing”
I give it architecture, UML diagrams and specs, unit tests for the most important functions (all written by me by hand). Then I let the LLM do the typing and read the output.
Just wanted to add some differentiation to your statement. I’m sure there are lots of people out there that fit your description.
I’m in a shitty place in regards to LLMs tbh, on one hand I can’t stand all the negatives (rich elite stealing and exploiting, shitty planet destroying companies, AI “art”, people getting dumber and more radicalized etc. etc.). On the other hand, I’m a lazy dev and I hate typing, but I also can’t make a career move to architecture rn. I just love solving hard problems with my brain and LLM lets me do that without having to grind through the parts I hate where my adhd is making me want to kms.
I got into SE because I had fun with it in my free time. But work is not fun. Since LLMs I’m enjoying my job (read: corpo slave hamster wheel of endless shoveling money from the pockets of other people to my boss and my landlord) for the first time in my life.
I do not think their assumptions for the baseline of the article are accurate:
The technology being discussed has no bearing on the attitudes or ideologies of the person holding it: use or non-use is ideologically neutral.
Why does use of a tool need to be ideologically neutral? Why draw this conclusion? Someone refusing to use a hammer would very clearly suggest some sort of ideology for why. In fact, entire religions refuse to use power tools for ideological reasons.
The technology being discussed can be effectively separated from the people using it: nothing can be inferred about a person from the technologies that they use.
You can infer a lot about the tools someone chooses to use.
The technology is used fully consciously: you are always fully aware that you are using a technology and have complete volitional freedom to use a different technology at any time.
I think this was more true prior to computers and the internet, but there are plenty of tools we use unwittingly. The freedom to use a different technology is always a physical possibility, but not always a practical one. If you had to drive a nail, and you had with you a hammer and a rock, you could choose to use the rock if you wanted.
The technology does not influence the kinds of things likely to be done with it: use of the technology doesn’t tend to encourage one course of action over another (“guns don’t kill people, people do”).
Tools have an intended use. They 100% influence the kinds of things likely to be done with it. A hammer is likely to be use to drive nails. A gun is designed to kill. While the gun doesn’t “kill people”, it sure makes it a hell of a lot easier… because that’s what it was designed to do. Just like a hammer makes it a lot easier to drive a nail.
The consequence of these statements being true is that the only valid criticism of a tool or technology becomes how well it performs the task that it’s designed to perform. You cannot, for example, say that a tool teaches bad habits, because if it isn’t explicitly a teaching tool what it teaches people becomes out-of-bounds for discussion.
The problem is they were never thought to be true. This is based on a false premise.
Tools absolutely do teach bad habits. Especially when you consider a cell phone a tool, or a computer a tool. We often are making a trade off using tools, but many times we consider those trade offs worth it.
If you have a power drill that dies, most people aren’t grabbing a screwdriver and manually screwing in every screw for a wood working project. The pros of the power drill are the massive difference in speed, and removing the physical strain on your muscles and joints that can do serious damage over time. The major con is now you are beholden to a company to purchase a power tool from. If it’s battery powered, you have to make sure you have their battery and their charger, and also have to purchase replacements if those ever die. Obviously the tradeoff is worth it to most with a power drill vs screwdriver (and nobody is considering using your bare hands), but there absolutely is a tradeoff and we can discuss that when discussing tools.
A common refrain when critiquing specific technologies in this way is that engineers will say that a given technology is “just a tool”
That is only if the critique about the technologies is about its morality.
When someone says “AI is just a tool” they are actually saying it can enable people to do very good or very bad things. People might respond with “AI is just a tool” if someone says AI is evil, because machine learning is not evil, it is a technology.
It’s also worth noting, that just because it’s a tool does not mean everyone should have access to it. There are plenty of tools that are restricted to specialists by law, or controlled more heavily. They are still tools, but they are tools that can cause a lot of harm if used incorrectly.
I did not read the entire article because it started off with these false premises. I also recognize that most people will not read everything I have written, and that’s fine as well.
I can tell that you didn’t read the article aside from the bullet points, because the point of the article is to refute these ideas presented at the beginning. It is about how tools are often culturally enshrined, how they directly affect people while also being designed by us. I encourage you to read it in full.
Of course he refutes them. He establishes the points to argue against them. The problem is he is setting up a straw man to argue against. I know the whole arguement is about how AI isn’t “just a tool”, except that it is. The difference is how you define a “tool” and he establishef an odd definition of tool to argue against
When someone says “AI is just a tool” they are actually saying it can enable people to do very good or very bad things. People might respond with “AI is just a tool” if someone says AI is evil, because machine learning is not evil, it is a technology.
Saying you can do both good and bad with any tool is misleading if the tool makes it easier to do the bad thing (and I think that is one of the points of the article). Also, it is common, and in my opinion valid, to use metonymy and refer to the practice of how LLMs are made and used rather than just the LLM itself when saying “LLMs are evil”.
Saying you can do both good and bad with any tool is misleading if the tool makes it easier to do the bad thing
This arguement simply does not hold up. Tools can, and often are, used for things they were not intended for. An effective hammer is going to be made of metal and have a decent heft to it. Having a hammer would also make it much easier to bash in someone’s head, (making it easier to do a bad thing).
A gun is designed to kill people, making people more likely to murder. The only people that have a realistic argent for needing guns are going to be police and military, and of course hunters with rifles.
The biggest problem with AI is how wide of a scope the tool applies to. Most tools have a much more narrow use case. AI is designed to be extremely versatile. You could use it to get customized lesson plans yo help structure your learning. Or you could skip the learning and have it give you the answer. It is a tool either way
I meant easier to do the bad thing than to do the good thing. A hammer does not specifically make it easier to attack someone than to use it for carpentry. And hammers are rarely used to attack other people in practice.
I don’t understand why you bring guns into this argument. Like LLMs, they should indeed be banned, except for some very narrow use cases. Most developed countries do ban guns in that way.
For most people, LLMs do make it specifically easier to skip learning than to enhance learning. We can see this in practice where teachers are noticing higher grades for homework but lower grades for the final exams.
To take it to the extreme: you could say buying a lottery ticket can be both a bad and a good financial decision because it costs money, but it can also earn you a lot of money. That statement is technically true, but I hope you agree it is misleading.
I don’t understand why you bring guns into this argument
Because it is a tool with a use that is used for bad things, like people are arguing with AI
For most people, LLMs do make it specifically easier to skip learning than to enhance learning. We can see this in practice where teachers are noticing higher grades for homework but lower grades for the final exams.
This is going to be especially true for children. Virtually unrestricted AI should not be accessible to kids. If they have access to AI, it should be one that acts as a teacher or private tutor that will help them learn but not give them the answer. Until we can control AI like that, they shouldn’t have access to it.
But we do not have an easy measurement for adults. We have anecdotal experience and vibes. It’s also entirely up to the person how effective the use of the tool is.
you could say buying a lottery ticket can be both a bad and a good financial decision
This would only be a good comparison if your actions had any impact on your chances of winning. It isn’t random chance how AI is used.
I think a better comparison would be drug usage. Prescription drug usage used properly can turn someone’s life around, objectively improving their life. Obviously abuse of drugs has ruined many lives, and ended many others. We are not saying we should get rid of drugs though.
Drugs are heavily controlled by professionals because of their ability to be abused. It used to be that coke was in soda’s and you could buy opioids over the counter. We learned about their addictive properties and dangers, and they were regulaged.
Perhaps we will some day have a method to control AIs abuse/addiction. It would no doubt look entirely different from controlled substances. Maybe there is a way to have personal accounts only help people learn, and rather that giving them answers it might guide them to learning for themselves. Who knows what this technology might look like in the future
This actually reflects beautifully irl. A hoe is a hoe is not shovel. A hoe is good a tiling a (small) field while a shovel is not but is good at shoveling. But they can be used for a lot of cases interchangeably. Same goes for cat, cp, dd.
But if you scale this up, there’s suddenly not one kind of harrow, and each is specialized and forms not only the result, but also how you work with it. Same with frameworks and LLM.
