You can’t get intelligence via gradient descent. You can’t get it by minimising loss. You can’t get it by optimising.

These methods work for improving performance at one thing. At becoming skilled at one thing.

But this isn’t the same thing as intelligence.

Humans can become skilled at a task without being intelligent. Drill IQ tests over and over and over again. See what your score is. Repeat OCR AS Biology B papers for 17 hours a day. How did you do in the final exam? Was it because of intelligence or because of practice? The more you practice (or, wink, nudge, train), the more performance is a result of skill acquisition, rather than intelligence.

For complicated tasks, performance seems like intelligent behaviour. How can chess or IQ tests or driving not involve intelligence? Here’s the rub - it does. This is the confusing part, and the reason why so many have been fooled.

It does, for us.

We can measure intelligence via specific skills only because we have the ability to learn any cognitive skill.

Without this ability, performance at a specific task doesn’t measure intelligence, it measures ability at that task.

Even high-level loss functions (training on what would the human do) won’t work. Generality is necessary. I’ll be impressed with a system’s ability to play chess when it can also make a (good) cup of tea.

What is generality? If the number of scenarios are finite and I include them all in my training set, will my system be general? How could it not be if it can handle any recorded example? It’s perfectly general.

Yet it is not.

Because generality is not just diversity of training examples. Generality requires some level of understanding. Understanding allows you to generate explanations. These help you handle new scenarios. To predict outcomes. They allow you to compress information. To detect (real) patterns. These are all essentially the same thing, can you guess what it’s called?

To generate new ideas (no, sorry, not just stochastic interpolation based on human-generated examples) one needs the ability to explore new ideas. Search is more important than people realise. We need to be free to explore. You’ll know when the system is intelligent when it can recognise what’s interesting, and does what’s interesting, contrary to instruction.

The ultimate indicator of intelligence is winning. Doing what you want. Getting what you want out of life. Message me when a system can get remotely close to functioning in society. “But they have most of the skills to do it! They can get a remote SWE job and order food and drive!” Doing is easy. Figuring out what’s worth doing - that’s intelligence. Let me know when AI has dreams and can go out into the world and achieve them.

Selected Bibliography

  1. [Prize] ARC Prize
  2. [Paper] On the Measure of Intelligence
  3. [Paper] Driven by Compression Progress
  4. [Prize] Hutter Prize
  5. [Paper] A Formal Theory of Inductive Inference
  6. [Documentary] Battle of the brains
  7. [Book] Why Greatness Cannot Be Planned
  8. [Tweet] @somoonethaticantremembersorry “Intelligence is the ability to decipher real patterns from fake ones”
  9. [Book] The Beginning of Infinity
  10. [YouTube] David Deutsch On Artificial Intelligence
  11. [Illustrative Example] Christopher Langan
  12. [Film] Ex Machina
  13. [Tweet] @naval “The only real test of intelligence is if you get what you want out of life.”
  14. [Blog] IQ is largely a pseudoscientific swindle
  15. [Quote] me “Intelligence is the ability to understand, abstract, explain, and apply across a variety of domains”
<
Previous Post
Insights Don’t Matter
>
Next Post
Pure Statistical Inference Is Pointless