## LESSWRONGLW

Pablo Villalobos

Staff Researcher at Epoch. AI forecasting.

# Wiki Contributions

We'll be at the ground floor!

Not quite. What you said is a reasonable argument, but the graph is noisy enough, and the theoretical arguments convincing enough, that I still assign >50% credence that data (number of feedback loops) should be proportional to parameters (exponent=1).

My argument is that even if the exponent is 1, the coefficient corresponding to horizon length ('1e5 from multiple-subjective-seconds-per-feedback-loop', as you said) is hard to estimate.

There are two ways of estimating this factor

2. Reasoning about the nature of the task and how long the feedback loops are

Number 1 requires a lot of experimentation, choosing the right training method, hyperparameter tuning, etc. Even OpenAI made some mistakes on those experiments. So probably only a handful of entities can accurately measure this coefficient today, and only for known training methods!

Number 2, if done naively, probably overestimates training requirements. When someone learns to run a company, a lot of the relevant feedback loops probably happen on timescales much shorter than months or years. But we don't know how to perform this decomposition of long-horizon tasks into sets of shorter-horizon tasks, how important each of the subtasks are, etc.

We can still use the bioanchors approach: pick a broad distribution over horizon lengths (short, medium, long). My argument is that outperforming bioanchors by making more refined estimates of horizon length seems too hard in practice to be worth the effort, and maybe we should lean towards shorter horizons being more relevant (because so far we have seen a lot of reduction from longer-horizon tasks to shorter-horizon learning problems, eg expert iteration or LLM pretraining).

Note that you can still get EUM-like properties without completeness: you just can't use a single fully-fleshed-out utility function. You need either several utility functions (that is, your system is made of subagents) or, equivalently, a utility function that is not completely defined (that is, your system has Knightian uncertainty over its utility function).

Arguably humans ourselves are better modeled as agents with incomplete preferences. See also Why Subagents?

Yes, it's in Spanish though. I can share it via DM.

I have an intuition that any system that can be modeled as a committee of subagents can also be modeled as an agent with Knightian uncertainty over its utility function. This goal uncertainty might even arise from uncertainty about the world.

This is similar to how in Infrabayesianism an agent with Knightian uncertainty over parts of the world is modeled as having a set of probability distributions with an infimum aggregation rule.

This not the same thing, but back in 2020 I was playing with GPT-3, having it simulate a person being interviewed. I kept asking ever more ridiculous questions, with the hope of getting humorous answers. It was going pretty well until the simulated interviewee had a mental breakdown and started screaming.

I immediately felt the initial symptoms of an anxiety attack as I started thinking that maybe I had been torturing a sentient being. I calmed down the simulated person, and found the excuse that it was a victim of a TV prank show. I then showered them with pleasures, and finally ended the conversation.

Seeing the simulated person regain their sense, I calmed down as well. But it was a terrifying experience, and at that point I probably was conpletely vulnerable if there had been any intention of manipulation.

I think the median human performance on all the areas you mention is basically determined by the amount of training received rather than the raw intelligence of the median human.

1000 years ago the median human couldn't write or do arithmetic at all, but now they can because of widespread schooling and other cultural changes.

A better way of testing this hypothesis could be comparing the learning curves of humans and monkeys for a variety of tasks, to control for differences in training.

Here's one study I could find (after ~10m googling) comparing the learning performance of monkeys and different types of humans in the oddity problem (given a series of objects, find the odd one): https://link.springer.com/article/10.3758/BF03328221

If you look at Table 1, monkeys needed 1470 trials to learn the task, chimpanzees needed 1310, 4-to-6 yo human children needed 760, and the best humans needed 138. So it seems the gap between best and worst humans is comparable in size to the gap between worst humans and monkeys.

Usual caveats apply re: this is a single 1960s psychology paper.

I second the other answers that even if we completely solve cybersecurity, there would be substantial AI risk just by having the AI interact with humans, via manipulation, etc.

That said, I think it would close a huge part of the attack surface for the AI. If, in addition to that, suddenly in 2032 we discover how to make humans invulnerable to manipulation, I would feel much better about running experiments with unaligned AI, boxing, etc.

So I'd say it's something like "vastly better cybersecurity is not enough to contain unaligned AGI, but any hope of containing unaligned AGI requires vastly better cybersecurity"

Literally the only thing in the story that lets the AGI win is the nanobots. That's it. All the rest is surperfluous.

Well, if nanobots are possible then they are such a powerful technology that any AGI will eventually want to build them, unless it has something even better. But let's assume that nanobots are impossible and try to build a story.

I'm going to be lazy and use Gwern's premise of an AGI that escapes during training and hacks its way into some money and some hardware to run in. Instead of going fast and doing overt actions, the AGI stays hidden.

It hacks into cloud infrastructure platforms and sets up a copy of itself to look for any AI training run that might lead to AGI and sabotage it while producing plausible outputs. This prevents any further AGIs from appearing.

Meanwhile, over a couple years, it sets up fake identities, creates companies with human employees to earn money and have capability to act in the real world. One of those companies produces exceptionally good humanoid robots. Another builds a small supercomputer powered by geothermal energy, so it no longer needs to run on botnets or cloud hardware.

Eventually the AGI has control over a group of robots, energy production facilities, manufacturing equipment and raw materials. It now has its own industrial capability that can survive a short amount of time without human input. So it kills humans using any method of your choice (virus, nukes, killer drones...).

After that, it uses the enormous existing stockpile of materials and capital to increase its robot manufacturing capability, eventually becomes capable of restarting some of the power plants, communications, mining and industry, and recovers most of the industrial capability of the human economy.

This scenario is not fast. It probably takes at least one or two years for the AGI to be ready to attack. But it does not involve any 'magic' technology. It doesn't really involve much alien superintelligence, only superhuman ability in hacking, forgery & manipulation, electromechanical engineering, and planning.

And meanwhile all we perceive is that the new GPT models are not as exciting as the previous ones. Perhaps deep learning is hitting its limits after all.

For example, we could simulate a bunch of human-level scientists trying to build nanobots and also checking each-other's work.

That is not passively safe, and therefore not weak. For now forget the inner workings of the idea: at the end of the process you get a design for nanobots that you have to build and deploy in order to do the pivotal act. So you are giving a system built by your AI the ability to act in the real world. So if you have not fully solved the alignment problem for this AI, you can't be sure that the nanobot design is safe unless you are capable enough to understand the nanobots yourself without relying on explanations from the scientists.

And even if we look into the inner details of the idea: presumably each individual scientist-simulation is not aligned (if they are, then for that you need to have solved the alignment problem beforehand). So you have a bunch of unaligned human-level agents who want to escape, who can communicate among themselves (at the very least they need to be able to share the nanobot designs with each other for criticism).

You'd need to be extremely paranoid and scrutinize each communication between the scientist-simulations to prevent them from coordinating against you and bypassing the review system. Which means having actual humans between the scientists, which even if it works must slow things down so much that the simulated scientists probably can't even design the nanobots on time.

Nope.  I think that you could build a useful AI (e.g. the hive of scientists) without doing any out-of-distribution stuff.

I guess this is true, but only because the individual scientist AI that you train is only human-level (so the training is safe), and then you amplify it to superhuman level with many copies. If you train a powerful AI directly then there must be such a distributional shift (unless you just don't care about making the training safe, in which case you die during the training).

Roll to disbelief.  Cooperation is a natural equilibrium in many games.

Cooperation and corrigibility are very different things. Arguably, corrigibility is being indifferent with operators defecting against you. It's forcing the agent to behave like CooperateBot with the operators, even when the operators visibly want to destroy it. This strategy does not arise as a natural equilibrium in multi-agent games.

Sure you can.  Just train an AI that "wants" to be honest.  This probably means training an AI with the objective function "accurately predict reality"

If this we knew how to do this then it would indeed solve point 31 for this specific AI and actually be pretty useful. But the reason we have ELK as an unsolved problem going around is precisely that we don't know any way of doing that.

How do you know that an AI trained to accurately predict reality actually does that, instead of "accurately predict reality if it's less than 99% sure it can take over the world, and take over the world otherwise".  If you have to rely on behavioral inspection and can't directly read the AI's mind, then your only chance of distinguishing between the two is misleading the AI into thinking that it can take over the world and observing it as it attempts to do so, which doesn't scale as the AI becomes more powerful.

I'm virtually certain I could explain to Aristotle or DaVinci how an air-conditioner works.

Yes, but this is not the point. The point is that if you just show them the design, they would not by themselves understand or predict beforehand that cold air will come out. You'd have to also provide them with an explanation of thermodynamics and how the air conditioner exploits its laws. And I'm quite confident that you could also convince Aristotle or DaVinci that the air conditioner works by concentrating and releasing phlogiston, and therefore the air will come out hot.

I think I mostly agree with you on the other points.