Crossposted from the AI Alignment Forum. May contain more technical jargon than usual.

Thanks to Linchuan Zhang, Jack Ryan and Carl Shulman for helpful comments and suggestions. Remaining mistakes are my own.

Epistemic status: “There is something fascinating about [anthropics]. One gets such wholesale returns of conjecture out of such a trifling investment of fact.” (Mark Twain, Life on the Mississippi)

Attempts to predict the development of artificial general intelligence (AGI) sometimes use biological evolution to upper bound the amount of computation needed to produce human level intelligence, e.g. in Ajeya’s use of biological anchors. Such attempts have mostly ignored observation selection effects. Shulman and Bostrom’s 2012 paper How Hard is Artificial Intelligence? analyzes how evolutionary arguments interact with various ways of reasoning about observation selection effects, drawing evidence from timings of evolutionary milestones and instances of convergent evolution. This post is a summary of key arguments advanced by the paper; see the original paper for citations.

More concretely, suppose evolutionary processes produce human-level intelligence on 1/10 or 1/10^1000 planets that develop life. Call the former case “easy” and the latter case “hard.” The paper attempts to determine whether the facts about evolution on Earth can distinguish between evolving intelligence being easy or hard.

Recall two common forms of anthropic reasoning:[1]

  • Self-Sampling Assumption (SSA): Observers should reason as if they were a random sample from all actually existent observers in their reference class. Observers should reason as if they have an equal probability of being in any world with observers, regardless of the number of observers in that world. Worlds where a higher fraction of observers are “like you” are more probable.
  • Self-Indication Assumption (SIA): Observers should reason as if they were a random sample from all possible observers. Observers should reason as if they have a probability of being in a world proportional with the number of observers it contains. Worlds where a higher number of observers are “like you” are more probable.

For more discussion, see Katja's Anthropic Principles or Bostrom’s Anthropic Bias.

Key takeaways:

  • Universes where evolution is easy have vastly more intelligent observers than universes in which intelligence is hard. Since SIA a priori favors universes with many observers over universes with few, SIA assigns almost dogmatically high prior credence to evolution being easy.
    • Caveat: Intelligent life could run cheap computer programs that simulate observers, so any universe with a few instances of intelligent life could have an immense number of observers. SIA cannot strongly distinguish between universes where nearly all resources are used for observers, which is possible both with intelligence in every solar system or only a few times per galaxy. SIA therefore advises that evolution is relatively easy, but can still be some orders of magnitude more difficult than once per solar system.
  • Given SSA, our observation of humans having evolved cannot alone distinguish between evolution being easy or hard. However, under both SSA and SIA, whether or not intermediaries to intelligence evolved multiple times can provide evidence about evolution’s overall difficulty. If evolution is easy, we would expect predecessors to intelligence to have evolved more than once. Evolutionary developments that have occurred multiple times cannot be subject to massive anthropic distortion.
    • The Last Common Ancestor (LCA) between humans and octopuses, estimated to have lived 560 million years ago, had an extremely primitive nervous system. However, octopuses have extensive central nervous systems and display sophisticated behaviors like memory, visual communication, and tool use.
      • Other examples of complex learning and memory include corvids (crows and ravens, LCA about 300 million years ago) and elephants (LCA about 100 million years ago).
    • Caveat: Non-human intelligence might not be as scalable as human intelligence, despite appearances of generality. If this is the case, these examples might not represent an instance of convergent evolution since they and humans might be substantially different.
    • Caveat: The LCA of octopuses and humans already contained a nervous system, which only evolved once in Earth’s history and thus might be arbitrarily difficult to evolve.
    • Caveat: The LCA of various other intelligent organisms and humans might have some undetected property that predisposed both organisms towards intelligence. Similarly, the LCA of all animals with eyes contained opsins proteins, which might have been extremely difficult to evolve.
  • If a task requires many steps of varying difficulty to accomplish and you condition on the task being done in about the time expected for accomplishing a single step, each of the steps is expected to take about the same amount of time. In effect, conditioning on the task being done in a short time prohibits any of the steps taking very long and the truncated distributions for tasks of varying difficulty are similar.
    • Example: Suppose you need to pick two locks, and the first takes uniform [0, 20] seconds and the second takes uniform [0, 1000] seconds, if you condition both taking <20 seconds, then you know the second lock took less than 20 seconds, so you cannot distinguish between the second lock taking uniform [0, 1000] seconds or uniform [0, 20] seconds.
    • Crucially, since the conditional distributions are roughly uniform, the time gap between when the last of these steps was completed and the end of the interval is equal in expectation to the time any of the hard steps took. Given that the Earth remains habitable for about a billion years and is about 5 billion years old, the expected number of “hard steps” in evolution is around 5. We can rule out hypotheses that postulate large numbers of evolutionarily hard steps because they predict intelligent life evolving much later than a billion years before the Earth stops being habitable. The LCA between humans and chimps was 6 million years ago, making it improbable that scaling up brains contained any hard steps.
      • Caveat: This argument cannot distinguish between evolutionary steps being hard (~1 billion years) or extremely hard (~100 billion years), since anthropic conditioning implies these would have taken the same amount of time historically.

  1. Armstrong roughly argues that SSA corresponds to making decisions under average utilitarianism while SIA corresponds to making decisions under total utilitarianism. ↩︎

New to LessWrong?

New Comment
2 comments, sorted by Click to highlight new comments since: Today at 1:13 PM

Consider the following charts:

Chart 1

Chart 2

Chart 1 shows the encephalization quotient (EQ) of various lineages over time, while Chart 2 shows the maximum EQ of all known fossils from any given time. (Source 1, Source 2. Admittedly this research is pretty old, so if anyone knows of more recent data, that'd be good to know.)

Both of these charts show a surprising fact: that the intelligence of life on Earth stagnated (or even decreased) throughout the entire Mesozoic Era, and did not start increasing until immediately after the K/T event. From this it appears that life had gotten stuck in a local equilibrium that did not favor intelligence; i.e. the existence of dinosaurs (or other Mesozoic species) made it impossible for any more intelligent creatures to emerge. Thus the K/T event was a Great Filter: we needed a shock severe enough to dislodge this equilibrium, but not so severe as to wipe out all the lineages from which intelligence could evolve.

If this is true, then the existence of ravens and elephants today is not much evidence that evolving intelligence is easy, because they exist for the same reason that humans do.

None of this considers octopuses. It would be interesting to see if their brain size history follows similar curves as for the vertebrates illustrated above (but since they're made up of soft tissue we may never know). If so, then that would confirm the view that evolving intelligence is difficult. On the other hand, it's hard to imagine that the marine ecosystem would've been affected by the K/T event in the same way that the terrestrial was. Or maybe octopuses are themselves what is suppressing the evolution of greater intelligence among marine invertebrates.

I'm going to have some criticism here, but don't take it too hard :) Most of this is directed at our state of understanding in 2012.

I think a way to do better is not to mention SSA or SIA at all, and just talk about conditioning on information. Don't even have to say "anthropic conditioning" or anything special - we're just conditioning on the fact that sampling from some distribution (e.g. "worlds with intelligent life who figure out evolution") gave us exactly our planet. (My own arguments for this on LW date from c. 2015, but this was a common position in cosmology before that.)

This gives you information that is more "anthropic" than SSA, but more specific than SIA. We can now ask probabilistic questions entirely in the language of conditional probabilities, which tells you more about what empirical questions are important. E.g. "What us the probability that octopus-level intelligence evolves on an earth-like planet in the milky way, conditional on some starting distribution over models of evolution, and further conditional on a sampling process from planets with human-level intelligence returning Earth?" The task is simply to update the models of evolution by reweighting according to how well they predict that sampling from our reference class give us.

Also, assuming all distributions are uniform gives one an unrealistic picture of timing there at the end. Think about what happens if the distributions are Poisson!

Footnote: Armstrong argues something more niche than that, because he's not talking about a "normal" CDT agent doing averaging/totalling, he's talking about an ADT agent doing averaging/totalling, and these are very different baseline agents!