Sometimes, people who had shown a glimmering of understanding, manage to demonstrate that they too do not get it. Today’s example is from Sharknado III’s President himself, Mark Cuban.
If you have zero education, but learn how to ask AI models the right questions , in many jobs you will be able to outperform someone with an advanced degree, but who is unwilling to use Large Language Models. Just takes a smartphone, curiosity to experiment and a mindset to learn.
— Mark Cuban (@mcuban.bsky.social) February 17, 2025 at 1:59 PM
Frederick Brooks, in 1996, wrote a wonderful and thoughtful article1 on the framing of computing, focusing especially on “computer science” and “artificial intelligence”. And by ‘framing’ here, I mean, what conceptual framework helps us to understand how these fields work.
Many of the challenges in the broad field of computing are analyzed in this article: is it science or engineering? Is it about discovery or creation? Brooks’s answer is, computing is broadly about making tools, we are toolmakers, nothing more, and nothing else. “An honourable calling”, in his terms.
But there’s also some Tolkein in there. Brooks builds on Tolkein, following Dorothy Sayers, and describes what computing folks do as “subcreation” – we create, but only within the framework within which we ourselves are made.
A large language model is one of these tools, and as such, it exists entirely within the framework of human language and human culture. It is a subcreation, not a source of universal wisdom. The question is, what does this kind of tool help us to do?
Brooks moves on to discuss how contemporary AI (this was 1996, remember) had evolved in a positive way, from the original and sketchy goal of building giant brains into something else: building tools that were assistants and advisors. Instead of attempting to ‘be’ an entire entity (as the ‘general AI’ folks would have it), they help people.
Brooks’s thesis was:
IA (intelligence amplifying) > AI (artificial intelligence)
If what we are building, and what LLMs are, is about amplifying human intelligence, how does this reflect on Cuban’s point?
Well, let’s re-frame large language models as an amplifier2.
A person with zero knowledge of a task, with an LLM, might perform marginally better than a person of zero knowledge without an LLM. But they will always be substantially off worse than a person with good knowledge of the task, probably with or without an LLM.
A good coder will always produce better code than a non-coder with an LLM.
What an LLM might bring is productivity. A non-coder with an LLM might well be more productive than a non-coder without, but they are inevitably going to be inferior to a good coder with an LLM. The same goes for a writer, an accountant, a graphic designer, and even, possibly, a CEO.
In other words, a large language model will not help a stupid person to become smart. It’s a tool, an amplifier – it simply enables them to use what capacities they do have, more productively. It simply makes that stupid person capable of generating exponentially more stupidity.
That’s the point about an amplifier: if there isn’t any signal, all you get is noise – white noise34.
Seeing ourselves as toolsmiths, Brooks argues, also changes the way we need to work.
“If the computer scientist is a toolsmith, and if our delight is to fashion power tools and amplifiers for minds, we must partner with those who will use our tools, those whose intelligences we hope to amplify” (Brooks, 1996, p64).
Brooks’s argument also matches my personal experience: we should never foist our tools onto users, we should collaborate with them. In fact, that is how we learn to be better. Working with other people, on their problems, helps us: we get to work on challenging, interesting, and real problems, not toy ones; it keeps us honest and away from hype; and, not least – it is more fun.
Brooks, F. (1996). The Computer Scientist as Toolsmith II. Communications of the ACM, March 1996, 39(3), pp. 61-68. This article was based on his acceptance lecture for the first ACM Allen Newell Award, which was given in 1994. ↩
Now you might disagree. Consider: image generation. Is it not actually creating images? Well, no. it’s using a prompt, and those who embrace generative AI embrace these prompts as being critical to a good outcome. An image generation AI is amplifying a prompt into an image. ↩
A lovely and ironic pun, entirely due to Andrew Robinson. And meaningful, too: as language model outputs tend to be biased to the stereotypical. ↩
An even older computing person might simply describe this as: garbage in, garbage out. And they’d not be wrong. ↩