This morning Anthropic shipped a model called Fable 5. Strip the branding and it’s a file. Some weights, billions of numbers. On the launch page Stripe says the file compresses months of engineering into days.
The file refuses to help you build a competing file, and it won’t tell you when it’s refusing. To use the file you accept that your traffic is kept for 30 days, even if your contract said zero retention. There is a second version of the file with the restraints removed, and it goes only to people on a list. Days before shipping it, the company that made it asked its rivals to agree to a brake pedal.
These are strange terms for a file. They make sense once you work out what the numbers stand in for.
In school I read that the judiciary is a pillar of the economy, and I didn’t see why. A judgment is one line on paper, and behind the line are decades of cases and centuries of precedent, refined into an instinct for where it should fall.
The Hutter Prize approaches knowledge through compression. It rewards losslessly compressing a fixed sample of Wikipedia, counting the programs as well as the archive. A model that assigns higher probability to the text lets a coder represent it in fewer bits. Prediction and compression are linked mathematically. [0] Hutter connects this to intelligence by combining universal prediction with decision theory in AIXI. The prize turns that connection into a measurable research problem. [1]
I read Zen and the Art of Motorcycle Maintenance. Pirsig compares analysis to a knife, using a motorcycle to explain what he means. A parts catalogue divides the machine into components you can order and replace. When you are tracing a fault, you may need to group those components differently, according to how they work together. Both descriptions are useful, but for different purposes. We become familiar with one way of dividing things and begin to treat it as obvious. Pirsig asks us to notice that division again. Before we can reason about the parts, we have already made a choice about what counts as a part.
We call it judgment in law, taste in design, rationality, analytical thought. The words overlap. In each case, some of the work is deciding which distinctions matter and what follows from them. Pirsig gives that work an image, a knife.
In forming a hypothesis, we choose which parts of a problem to examine and how they might be related. Anthropic reports that in blinded comparisons, its scientists preferred Mythos 5’s molecular biology hypotheses to those from Opus-class models about 80% of the time. The company sells access to these models by the token.
People already pass on what they know through teaching. A model’s responses can also be used as training examples for another model. This process, called distillation, can reproduce some of the model’s useful behavior [2], and Anthropic has added safeguards to limit it.
David Deutsch argues that humans are universal explainers, capable in principle of understanding any explanation an artificial general intelligence could understand. A machine might have more memory or work faster, while we can use computers to extend our own abilities. On that view, continued improvement would be possible for both. We still have to find out how quickly these particular models can improve and what that takes.
Now reread the terms. A knife that refuses to make competing knives. A knife that goes dull in your hand without saying so. A knife that keeps a record of every cut for 30 days. A sharper knife reserved for a list. Rivals asked to agree on how sharp any knife may get. The question underneath is who gets to use these models, and under what terms. Much of that is being settled through contracts, safety policy and prices. Those are the proxies.
[0] Hutter’s prize FAQ explains the coding argument and why program size is included. Delétang et al., Language Modeling Is Compression (2023), applies the prediction–compression equivalence to language models. ↩︎
[1] Marcus Hutter, A Theory of Universal Artificial Intelligence based on Algorithmic Complexity (2000). AIXI is an idealized, uncomputable agent. The step from prediction to action includes a reward objective and sequential decision theory. ↩︎
[2] Buciluǎ, Caruana, and Niculescu-Mizil, Model Compression (2006), trains smaller models using predictions from an ensemble. Hsieh et al., Distilling Step-by-Step! (2023), uses language-model labels and rationales to train smaller task-specific models. ↩︎