Previously Staff Researcher at Epoch, these days I’m mostly thinking about better frameworks for understanding superintelligence.
Pablo Villalobos
Karma: 605
I just want to flag that there is another family of non-independent decision rules called multiplier preferences or more generally variational preferences which result in a certain kind of adversarial robustness to Knightian uncertainty. Multiplier preferences in particular have a very natural interpretation in terms of information geometry and can be seen as roughly the dual of active inference. Where active inference says “my utility function can be seen as a prior”, or “I can deviate from my policy during planning by paying a KL cost”, multiplier preferences say “Nature can adversarially deviate from my model of it by paying a KL cost”.
This applies to basically any form of competition beyond the amount needed to avoid pricing power.
For example, if you are a job seeker, you can dedicate your time to learning actual skills and looking for employers with unmet needs, or you can rack up your credentials to be candidate #1 instead of #2 in an oversubscribed hiring process.
That doesn’t make these activities unproductive in most cases: if customers or employers pick you it’s because you are providing some marginal value, but the marginal value might be low compared to the amount you capture
I think this is partly cope. Middle management at large companies might have a significant political component but large companies still have much higher labor productivity than small ones, they still represent like a quarter of OECD economic output, and the average middle manager is probably still creating very significant amounts of marginal value despite the political infighting.
Yes, there might be very few middle managers at the top of Forbes’ list. But look at millionaires, the people at the top 10% of the wealth distribution in the US. Most middle managers will probably be there, along with other highly paid professionals. And if you found a startup and it fails, you won’t make the 10%, which is what happens in the overwhelming majority of cases.
So purely in terms of wealth and creating value for society, marginal improvements in middle management seem quite valuable. Sure, being a founder might have much higher EV, but also vastly higher variance. And risk aversion is behaviorally indistinguishable from just having a different utility function.
Speaking of which, you might want to be immortal and go to the moon, but most people don’t. You could argue that if they’d read the right books or had the right parent/teacher/friend or had more vision, they would also want that, but at that point you’re just saying that everyone should be playing the game you like, instead of the one they like.And I dispute the idea that there’s less politics, signaling or strategic/conflictive behavior at the “winner’s bracket”. Look at Sam Altman, look at the status competitions and fighting for credit in the highest halls of science. Look at states for God’s sake. Do you really think Vladimir Putin or Donald Trump or Xi Jinping are not in the winner’s bracket, that they will have less real power than the first person to figure out how to solve aging?
Physicists might have found the secret knowledge of how to create nuclear weapons that nobody else had, but after they had the bright idea the bottlenecks were capital, labor, natural resources, and the ability to manage their combination efficiently, and the physicists were not the ones who got to control the weapons in the end.
I think something like your thesis might be true in terms of actually having good counterfactual impact on reality vs merely capturing the resulting wealth, power and prestige, what you call leading the parade. But that doesn’t mean that you get to both have the impact and lead the parade by pursuing just the impact part!
Personal view as an employee: Epoch has always been a mix of EAs/safety-focused people and people with other views. I don’t think our core mission was ever explicitly about safety, for a bunch of reasons including that some of us were personally uncertain about AI risk, and that an explicit commitment to safety might have undermined the perceived neutrality/objectiveness of our work. The mission was raising the standard of evidence for thinking about AI and informing people to hopefully make better decisions.
My impression is that Matthew, Tamay and Ege were among the most skeptical about AI risk and had relatively long timelines more or less from the beginning. They have contributed enormously to Epoch and I think we’d have done much less valuable work without them. I’m quite happy that they have been working with us until now, they could have moved to do direct capabilities work or anything else at any point if they wanted and I don’t think they lacked opportunities to do so.
Finally, Jaime is definitely not the only one who still takes risks seriously (at the very least I also do), even if there have been shifts in relative concern about different types of risks (eg: ASI takeover vs gradual disempowerment).
We’re in the nearby bar, Casa Remigio, since the theater is occupied
I suspect the analogy does not really work that well. Much of human genetic variation is just bad mutations that take a while to be selected out. For example, maybe a gene variant slightly decreases the efficiency of your neurons and makes everything in your brain slightly slower
I stand corrected. Although the broader point about share prices noisily approximating a discounted expected cash flow which can be added or multiplied still holds
There is a sense in which the price approximates an intrinsic property of the shares that you can add up or multiply by the number of shares. Each share gives you a vote in the shareholder assembly and an equal portion of the dividends. If you had all the shares, you would own the company and in principle could pay yourself as much as the company can afford in dividends.
How much the company can afford to pay in dividends in the future is basically how much net operating profit after taxes (NOPAT) the company will have.
If you have a prediction of the future NOPAT of the company, it implies a present value for the whole company and its shares assuming all of it is cashed out as dividends. It is commonly assumed that in most cases the market price of shares oscillates around a rational expectation of future NOPAT, in which case it would be a reasonable approximation to something that you can semantically multiply by the number of shares to get the overall value of the company.
The arguments you make seem backwards to me.
All this to say, land prices represent aggregation effects / density / access / proximity of buildings. They are the cumulative result of being surrounded by positive externalities which necessarily result from other buildings not land. It is the case that as more and more buildings are built, the impact of a single building to its land value diminishes although the value of its land is still due to the aggregation of and proximity to the buildings that surround it.
Yes, this is the standard Georgist position, and it’s the reason why land owners mainly capture (positive and negative) externalities from land use around them, not in their own land.
Consider an empty lot on which you can build either a garbage dump or a theme park, each of equivalent economic value. Under SQ, the theme park is built as the excess land value is capture by the land owner. Under LVT, the garbage dump is built as the reduced land values reduces their tax burden. The SQ encourages positive externalities, LVT encourages negative externalities.
This seems wrong. The construction of a building mainly affects the value of the land around it, not the land on which it sits. Consider the following example in which instead of buildings, we have an RV and a truck, so there is no cost of building or demolishing stuff:
There’s a pristine neighborhood with two empty lots next to each other in the middle of it. Both sell for the same price. The owner of empty lot 1 rents it to a drug dealer, who places a rusty RV on the lot and sells drugs in it. The owner of empty lot 2 rents it to a well-known chef who places a stylish food truck on the lot and serves overpriced food to socialites in it.
Under SQ, who do you think would profit from selling the land now? The owner of lot 2 has to sell land next to a drug dealer that a prospective buyer can do nothing about. The owner of lot 1 has to sell land next to delicious high-status food, and if a buyer minds the drug dealer he can kick him out. Who is going to have an easier time selling? Who is going to get a higher price?
Now, suppose there is a LVT. If the tax is proportional to the selling price of the land under SQ (as it ideally should), which owner is going to pay more tax?
The case of the theme park and garbage dump is exactly the same, with the added complication of construction / demolition costs. An LVT should be proportional to the price of the land if there were no buildings on top of it (and without taking into account the tax itself), so building a garbage dump is not going to significantly reduce your tax payments.
In such a way, a land value tax has a regularisation effect on building density, necessitating a spread of concentration.
There are several separate effects here, if you are a landowner. Under LVT:
You are incentivized to reduce the density in surrounding land
You are incentivized to build as densely as possible within your own land to compensate the tax
Under SQ:
You are incentivized to increase the density in surrounding land
You are not incentivized to increase density in your own land
The question is, which of these effects is bigger? I would say that landowners have more influence over their own land than over surrounding land, so a priori I would expect more density to result from an LVT
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
Empirically fitting scaling laws for whatever task we care about
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).
See Knightian Decision Theory. Part I
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.
Galaxy-Brained
This design will inspire you to conjure grander and more baroque intellectual works