I’m going to kick off this issue by going back to my favorite piece, which I’ve cited on here before — a 1985 interview with Joseph Weizenbaum (the father of the chatbot and a pioneer of commercial computing systems). I love this interview precisely because — in stark contrast to today’s crop of programmers — Weizenbaum still remembers that making the computer do a thing is merely a step in creating value, and it is neither the first step nor the most important one.
The computer has almost since its beginning been basically a solution looking for a problem. People come to MIT and say, “You have there a very wonderful instrument which solves a lot of problems. Surely there must be problems in my establishment — in this case the educational establishment — for which your wonderful instrument is a solution. Please tell me for what problems your wonderful instrument is a solution.”
The questioning should start the other way — it should start with the question of what education is supposed to accomplish in the first place. Then one might ask, in terms of what it's supposed to do, what are the priorities? What are the most urgent problems? And once one has identified the urgent problems, then one can perhaps say, “Here is a problem for which the computer seems to be well-suited.”
The productization of thoughtfulness
It is a great irony that 41 years later, it is precisely chatbots that have turned this sober perspective on its head. It has never been easier to fool yourself that extruding a “functional” prototype is the same as building a valuable product, and that everything else between you and wealth is just a sales problem.
But it would be a mistake to call this an AI-era phenomenon. The industry spent decades cultivating fertile ground in which today’s prototype mania could take root. Corporations carefully peeled the things they wanted (methods for making software widgets) away from the humanist context in which those methods were originally born, and stamped their own brand on each one. Instead of design, research, and thinking, there is only product design, product research, product thinking. And now, there is not even that; one only has to take a look at their LinkedIn feed’s many self-appointed “AI product managers” and “AI-native designers” who spend all their time pretending they invented “build to learn” yesterday.
There are few bigger cliches in the industry than complaining about the crass commercialization of craft. But the priorities of the computer people are hardly being overridden by the iron law of the profit motive! There has never been a more money-losing technology than the large language model, and yet it is aimed mainly at creating business problems rather than solving them:
LLMs mostly seem to first and foremost approach problem solving as a process of obliging the prompt writer. In other words, the top objective is to give an answer, no matter how wrong it is, quickly and efficiently. The frictionlessness is kind of the point, and runs counter to so much design and development work. It makes a great colleague for an impatient product owner who needs copy asap, but then whose job is it to check assumptions and identify moments when research is needed, etc?
AI boosters acknowledge this problem only in passing: it is early days, today’s more powerful models are larger than yesterday’s, so if we keep making the models larger, eventually they will get so powerful that all the problems will become solved, somehow. Possibly by creating AGI (or is it ASI now?). And perhaps some companies think this already exists, because we are now seeing models have unprecedented access to internal systems. Meta’s much-hyped Muse, for example, can see its own file system, and (because it’s an “everything app” and guardrails are more of a suggestion) will happily dump that file system for any user who asks. Multiple people have now combed through the code and there are some very funny moments there, so check it out.
Nevertheless, the mystique of Artificial Very Special Boy Intelligence has accomplished one thing: to allow OpenAI and Anthropic (and now Google) to get away with admitting that they are literally doing cybercrime every day and leverage that into an argument for reduced regulation rather than simply go to jail:
Instead of OpenAI being prosecuted for creating malware that hacked another company, press releases, news outlets, media personalities, and lawmakers refer to “rogue models” as if they acted on their own. Instead of researchers being questioned about their companies’ habit of plagiarizing academics’ work or using customer data to train models without consent, the public’s imagination is redirected to fears about what the future might hold upon the arrival of fictional superintelligent machines.
To be normal about LLMs, we must be able to be normal about computers
No one I know is excited about the incoming GPT-6, possibly because they have things like “a memory” that allow them to compare the hype around previous model releases to what was actually delivered, and how quickly Sam Altman turned around from “this one is the one” to “this one is crap, actually, and the next one will be the one.” Instead, the real eggheads are stoked for something called Jev, which has achieved the ground-breaking intellectual leap of using normal technology we already had to do things cheaply and reliably like we always could.
New larger models are only driving up the total cost of inference, in exchange for diminishing returns. But the frontier of “what if LLMs were just normal technology and could be combined with other normal technology” remains pretty much unexplored, because developers are blinded by the hype of AI as a special thing rather than just another SaaS product. But SaaS itself has been in a slump — indeed, has been relying on AI as the last hypergrowth opportunity — because it has reached exactly the same frontier: everyone involved has forgotten that computers are just a normal medium:
"Tech" is always couched in terms of new capabilities rather than the affordances of a medium. I long for the day that computers are no longer "tech" but just another expressive medium like film, then it will be possible to talk about those affordances.
Waking up from the dream that “computer” is a special class of thing may just be the breakthrough that we need to unwind four decades of decay and embrace the perspective that was so obvious to Weizenbaum: the actual product is not digital, and value does not originate in bits. In light of that, this guidance for business people looking to commission some software is extremely clear-headed: it buttresses the obligatory “you can use AI to prototype” in between “you must first define the business logic that you need” and “you can’t just hand over the AI’s prototype to a dev and ask them to ‘build this.’”
Normal technology is an empowering force
Some readers may remember an earlier era’s (valiant but ultimately failed) attempt to wake computer people up from the Web’s doom spiral: the motherfucking website (which was down for a decade, but is now back). Its banner has been picked up by the modern motherfucking website, a is a nice reminder that tech has not actually stood still all those years, and that the modern stuff is actually possible to incorporate into your code without bloating it with a million calories of Javaslop.
via Andy Bell
But the project I’m personally keeping an eye on — one that is truer to the vision at the top of this article than anything else I’ve seen — is Bianca Wylie’s Less Computer, a series of strategies aimed at decentering computers from our decision-making. This piece is my favorite (so far):
This strategy is sometimes understood as starting conversations about computers with the question "what problem are we trying to solve?" But sometimes the issues we live with are better understood as ongoing issues rather than problems. For this reason, it's a flag to start any conversation about computers - it's better to start them with the context in which computer use is being proposed.
Reading through the piece brought back a memory, or perhaps a mosaic of many identical memories: a design workshop with a sentence that starts with “how might we…” and inevitably ends with “…using the power of generative AI?”
Do the difficult thing, and cut that last part off. Baking a solution into your vision statement will never serve you, in the end.
— Pavel at the Product Picnic