The 1990s Web vs. today's AI - by Arnold Kling
In My Tribe
SubscribeSign in
The 1990s Web vs. today's AI<br>I was more engaged with the former
Arnold Kling<br>Aug 03, 2026
33
25
Share
I am old. I was in my forties when the Web led the Internet breakthrough of the 1990s. Because I got very excited and involved with the Web, my model for thinking about AI comes from that era.<br>Here is where I come down today: if on a scale for disruption of 1 to 10 where the Industrial Revolution is a 10 and the 1990s Web is an 8, AI is a 6.<br>The difference is that the Web is a participatory technology that benefits from Metcalfe’s Law and Reed’s Law. AI is a centralized technology that advances along narrow metrics.<br>Improvements in AI are like improvements in personal computing in the 1980s. As Intel advanced from the 8088 chip to the 286 chip to the 386 chip, and so on, the metrics for PCs indicated greater capability. But to most users, faster processors delivered diminishing returns.<br>Even if a better PC made you more productive and made me more productive, it didn’t make us more productive. It did not provide the synergy that arrived with network computing.<br>Robert Metcalfe was a network architect. He worked on the problem of getting bits to travel from node to node without colliding with one another. It is intuitively obvious that this problem becomes more complex the more nodes there are on the network. But Metcalfe also was convinced that with more nodes and greater complexity the network became more valuable.<br>David Reed went further. He argued that the more groups and subgroups that could be created on a network, the more valuable that network would become. The combinatorics of this favor a large network even more than Metcalfe’s Law.<br>You can look up the exact formulas proposed by Metcalfe and Reed on Wikipedia. Wikipedia, flawed as it is on politically charged topics, is an example of the participatory nature of the Web. The Web became a home for both legacy businesses and innovators to contribute ideas and capabilities.<br>Relative to the Web, AI is sterile. Those of us who do not work in the major labs are users, not contributors. AI progress consists of what takes place in the labs. To use another analogy from my 1990s era, imagine how little progress we would have made if, rather than the Internet, we would have been limited to America Online, Prodigy, and CompuServe.<br>Until now, human progress has come from our collective brain. Individually, nobody knows how to make a pencil. But when we can cooperate, whether in small groups, large organizations, markets, or computer networks, we can expand knowledge in all sorts of directions.<br>AI as it operates today truncates this learning process. Claude and I can have a chat in which I propose a theory of the Great Depression that is outside of the consensus that Claude usually works with. And during the conversation Claude may concede the merits of my approach. But when it leaves the conversation, Claude will have learned nothing from me. When the next person asks it about the Great Depression, it will revert to its previous stance.<br>So the AI labs are trying to do away with the what Joseph Henrich called The Secret of our Success. Implicitly, they are assuming that we no longer need to share ideas with one another and to contest ideas with one another. They just have to throw more data and computer resources at problems.<br>Perhaps this new way of trying to make progress will work, and collective intelligence will not be needed going forward. But as of now I would bet against that. If AI progress is not stalling already, then I fear that it will soon.<br>Share
Subscribe
33
25
Share
Previous
Discussion about this post<br>CommentsRestacks
In My Tribe reply rules
Andy in TX
8h
Liked by Arnold Kling
I think you are confusing general LLM models (which dd learn a tiny bit, not zero, as your interactions cause really, really small changes in the underlying probability tables) and the models aimed at specific tasks and that have local post training at least. Take a large law firm that has combined its files of contracts into a local model. That seems to offer learning by "us", at least the "us" inside the firm.
Reply
Share
Kurt
7h
Suppositions on what it is, isn't, can't, or might be...are premature.
Reply
Share
23 more comments...
TopLatestDiscussions
No posts
Ready for more?
Subscribe
© 2026 Arnold Kling · Privacy ∙ Terms ∙ Collection notice<br>Start your SubstackGet the app<br>Substack is the home for great culture
This site requires JavaScript to run correctly. Please turn on JavaScript or unblock scripts