How could someone expect AI in (median) 30 years and expect the same 30 years later?

• If you assume their expectations are normally distributed, this seems absurd; half the probability mass of their expectations is gone and the rest is grouped much closer to the present, so they'd have a much sooner estimate if they were being consistent.
• However, if their expectations are geometrically distributed, they'd have just over a 2% expectation of success in a given year. If you didn't have a mechanistic explanation of how AGI might come about - and most people don't and haven't - you could just think that we might discover what we need for it next year, or just as likely the year after, but it would take an unlikely breakthrough for that to happen.
• Equivalently, if you think we're waiting to find the right paradigm, we could have a 3% chance to find that each year (median = 22 years) and then 8 years to follow through on it afterwards. You'd have to increase your probability moderately over time, as the right paradigm might have been found and you just can't tell it's arrived yet, but your median time-to-AGI would stay consistent over time. (I've noticed many people here think something like this, though with a higher chance than 3%.)
• Having an estimate like this and stating your median value is misleading; "pause our efforts in a few years because we were wrong and we need to recalibrate" is sensible advice when you pass the bulk of a normal distribution, but it doesn't reflect any change in this model. If you say "AI 2027" and there is no AGI in 2027, people will not react like it was 50/50.
• That being said, I'm inclined to put a lot more than 3% chance annually on a breakthrough in the current conditions, so I don't think people should be predicting 25-30 years anymore. (I think a market crash is fairly likely and also don't think it would change the calculation much.)

A guess as to how lifeforms evolve defensive toxins (i.e. become poisonous):

• Being poisonous doesn't benefit an individual animal much; if you taste bad, a predator might give up in the middle of eating you, but you're still hurt and likely to be attacked by another predator.
• However, genes are selected for only in terms of the animals that do reproduce; a poisonous animal can make its kin, who share the relevant genes, more likely to reproduce in a few ways:
  • Many insects, like ants, have only a few individuals responsible for reproducing; the metabolic cost of producing poison makes worker ants less efficient, but reduces how many get eaten, and it has little effect on the queen (who probably won't be eaten regardless).
  • Healthy animals are in general less likely to be attacked and far more likely to successfully reproduce than unhealthy ones; the poison pushes a few more animals into the 'unhealthy' category, but makes all the animals less attractive to predators. If the poison isn't very costly, it should be able to spread by chance up to the point where this is effective in small populations. (Healthy animals can generally afford a modest additional cost to avoid predators; stotting is a behavioral equivalent.)
  • Much of the metabolic cost of a poison is in compensating for its negative effects in your own body; if you're consuming a poisonous animal regularly, it's cheap to reuse its poison. (Poison dart frogs do this.)
  • More costly poisons would presumably evolve over time from weaker poisons, in the same way that other defensive features are exaggerated over time.