Controversial take: that COBOL vibecoding thingy is actually good news for IBM.
I expect a hard reversion into professional knowledge, bc that's what the machines can't automate.
The last 40 years were essentially a long run towards replacing professional knowledge with a cheaper simulation thereof, with mixed results on the micro level and disastrous results on the macro level.
R to @oliverbeige: Norbert Wiener, The Human Use of Human Beings, 1950. Two years after Cybernetics, things were already going down the drain.
https://archive.org/details/humanuseorhumanb0000norb/page/188/mode/2up
Norbert Wiener, merchant of doom.
Parallel Paths to Cognition: The Carnegie-UCSD Connection in Artificial Intelligence is out now on Seemingly Incompatible.
https://carnegietech.substack.com/p/parallel-paths-to-cognition-the-carnegie #econtwitter #machinelearning
R to @oliverbeige: And that's really all you need to know.
It's because LLMs use gradient descent to traverse jagged search spaces.
Almost every criticism of AIs supposed inadequacy is a reflection of the critic’s own inability to prompt AI effectively.
If you hammer away at a piano for a week and get upset it cant produce Chopin. Thats on you.
Shattering the knowledge frontier.
The whole "soon we're all gonna be unemployed and unemployable" doomsaying also played out at the dawn of the personal computer age, but luckily nature provided infinite work for everyone in the form of compliance documentation and exploding administrative overhead.
Herb Simon and Al Newell's "General Problem Solver", according to Grok.
New blog post:
Autodiff through function types: Categorical semantics the ultimate backpropagator, by @_julesh_
https://cybercat.institute/2026/02/20/categorical-semantics-ultimate-backpropagator/
Gemini is getting tired of reasoning...
The Library of Babel: Combinatorial Optimization at Carnegie is now out on Seemingly Incompatible.
https://carnegietech.substack.com/p/the-library-of-babel-combinatorial #econtwitter
"The words or the language do not seem to play any role in my mechanism of thought. The psychical entities which seem to serve as elements in thought are certain signs and more or less clear images which can be voluntarily reproduced and combined." — Einstein to Hadamard, 1945.
The reason why the Ringbahn works so well as a synecdoche for the attention economy is that Berlin's tariff area A is all attention economy, no information economy. Which is also the reason why Berlin comes across like a circus sideshow.
There are two alternatives:
- Handwritten exams with zero technology aid in a tightly controlled environment.
- You can use whatever you want, but the task is tailored so that it cannot be solved by an LLM alone, and a human alone would need months or years to solve it.
"A Science of the Artificial: Herbert Simon on Design" is now out on Seemingly Incompatible.
https://open.substack.com/pub/carnegietech/p/a-science-of-the-artificial-herbert
For example, you could ask if and why prediction markets outperformed poll aggregators over the last three US presidential elections. That's something you can answer in a few hours using an LLM as research assistant.
https://x.com/i/status/2000620128669700563
“In the autumn of 1912, Ernst Zermelo stood before a small gathering of mathematicians in Cambridge and announced a result that would haunt the twentieth century. Chess, he declared, was solved, at least in principle. One of the two players, White or Black, must possess a winning strategy, or else both can force a draw. The game is finite; the tree of possibilities, though vast, terminates. Therefore, by the iron logic of mathematical induction, the outcome is determined before the first pawn moves.
The audience received this news with the peculiar mixture of satisfaction and unease that accompanies theorems of pure existence. Zermelo had proved that an answer existed without providing any hint of what that answer might be. His proof was what mathematicians call non-constructive: it demonstrated the existence of a winning strategy the way one might prove that a prime number greater than a googolplex exists: true, certainly, but utterly unhelpful to anyone hoping to find it.
This was Zermelo’s curse, and it would echo through the decades that followed. The chess tree contains more positions than atoms in the observable universe. Knowing that a perfect strategy exists tells us nothing about how to play well. The game is finite, but it might as well be infinite for any creature bound by time and matter. The gap between existence and construction, between knowing that an answer is there and actually finding it, would become the central drama of a field that did not yet have a name.”
Paths, Trees, Flowers, Conflicts captures the golden age of combinatorial optimization (and graph theory), the winding path towards recognizing that not all puzzles are easy to solve just because they are finite. Part 1, from Zermelo to Harary, is out now. #econtwitter #MLtools
https://carnegietech.substack.com/p/paths-trees-flowers-conflicts-the
"Herbert Simon’s causality papers from the early 1950s have achieved recognition far beyond their original econometric context, becoming foundational for multiple fields that barely existed when he wrote them. The intervention-based conception of causality that Simon formalized has become the dominant framework in program evaluation, causal inference, and policy analysis.
Computer scientists building causal discovery algorithms cite Simon’s 1952 and 1953 papers as pioneering the graphical approach to representing causal structure. Philosophers analyzing counterfactual reasoning trace their frameworks to Simon’s emphasis on intervention and invariance. Researchers in machine learning developing methods for inferring causation from data draw on Simon’s insights about identifiability.
Yet the synthesis Simon never completed, between formal causal structure and bounded causal learning, remains unfinished even as modern approaches have developed tools Simon lacked."
Herbert Simon and Causality, now out as part of Seemingly Incompatible: Bounding Rationality at Carnegie Tech's Graduate School of Industrial Administration, 1949-74. #econtwitter #MLtools
https://carnegietech.substack.com/p/herbert-simons-causality
We have two shorthands for our GenAI projects, "80/20" and "120/40", for "80% of human quality at 20% of human effort" vs "120% quality at 40% effort". The former is known as "AI slop", and public sentiment is coalescing around the idea that that's all GenAI is capable of. 1/
In which we introduce the least known game in game theory, the game that explains (almost) all of civilization: the protector-provider game, aka swords against ploughshares.
An Economic Pattern Language: On Governance, by @oliverbeige. 100% handwritten.
https://econpatterns.substack.com/p/on-governance
Frank Rosenblatt's 1958 Perceptron paper cites Hayek's Sensory Order. Peak cyberpunk.
The overarching template of tectonic shifts in the Western world is that the coalition of the attention economy and the information economy against the production economy (the firewall) bc the information economy starts realizing that the attention economy was lying all the time.
The reason why GenAI can't render text has a lot to do with why gradient descent can't solve combinatorial optimization problems. https://oliverbeige.medium.com/generative-ai-and-combinatorial-optimization-part-2-why-text-based-image-generators-cant-render-36229b4ec0fa
Def.: Kruger-Dunning Effect, the popular attempt by academic riffraff to distinguish themselves from non-academic riffraff by citing the imaginary Dunning-Kruger Effect. Named after Kruger-Dunning (1999), a highly cited but rarely read piece of pseudoscientific piffle.
If I were to offer a very naive expression of the functioning of memory I would call it a lossy stochastic compression of past experiences that maximizes our ability to successfully navigate future experiences. Apparently that's also what immune memory does.