I agree that it’s hard, but there are all sorts of possible moves (like LessWrong folks choosing to work at this future regulatory agency, or putting massive amounts of lobbying funds into making sure the rules are strict)
If the alternative (solving alignment) seems impossible given 30 years and massive amounts of money, then even a really hard policy seems easy by comparison
Eliezer gives alignment a 0% chance of succeeding. I think policy, if tried seriously, has >50%. So it’s a giant opportunity that’s gotten way too little attention
I’m optimistic about policy for big companies in particular. They have a lot to lose from breaking the law, they’re easy to inspect (because there’s so few), and there’s lots of precedent (ITAR already covers some software). Right now, serious AI capabilities research just isn’t profitable outside of the big tech companies
Voluntary compliance is also a very real thing. Lots of AI researchers are wealthy and high-status, and they’d have a lot to lose from breaking the law. At the very least, a law would stop them from publishing their research. A field like this also lends itself to undercover enforcement
I think an agreement with China is impossible now, because prominent folks don’t even believe the threat exists. Two factors could change the art of the possible. First, if there were a widely known argument about the dangers of AI, on which most public intellectual agreed. Second, since the US has a technological lead, it could actually be to their advantage.
It sounds like Eliezer is confident that alignment will fail. If so, the way out is to make sure AGI isn’t built. I think that’s more realistic than it sounds
1. LessWrong is influential enough to achieve policy goals
Right now, the Yann LeCun view of AI is probably more mainstream, but that can change fast.
LessWrong is upstream of influential thinkers. For example:
- Zvi and Scott Alexander read LessWrong. Let’s call folks like them Filter #1
- Tyler Cowen reads Zvi and Scott Alexander. (Filter #2)
- Malcolm Gladwell, a mainstream influencer, reads Tyler Cowen every morning (Filter #3)
I could’ve made a similar chain with Ezra Klein or Holden Karnofsky. All these chains put together is a lot of influence
Right now, I think Eliezer’s argument (AI capabilities research will destroy the world) is blocked at Filter #1. None of the Filter #1 authors have endorsed it. Why should they? The argument relies on intuition. There’s no way for Filter #1 to evaluate it. I think that’s why Scott Alexander and Holden Karnofsky hedged, neither explicitly endorsing nor rejecting the doom theory.
Even if they believed Eliezer, Filter #1 authors need to communicate more than an intuition to Filter #2. Imagine the article: “Eliezer et al have a strong intuition that the sky is falling. We’re working on finding some evidence. In the meantime, you need to pass some policies real fast.”
In short, ideas from LessWrong can exert a strong influence on policymakers. This particular idea hasn’t because it isn’t legible and Filter #1 isn’t persuaded.
2. If implemented early, government policy can prevent AGI development
AGI development is expensive. If Google/Facebook/Huawei didn’t expect to make a lot of money from capabilities development, they’d stop investing in it. This means that the pace of AI is very responsive to government policy.
If the US, China, and EU want to prevent AGI development, I bet they’d get their way. This seems a job for a regulatory agency. Pick a (hopefully narrow) set of technologies and make it illegal to research them without approval.
This isn’t as awful as it sounds. The FAA basically worked, and accidents in the air are very rare. If Eliezer’s argument is true, the costs are tiny compared to the benefits. A burdensome bureaucracy vs destruction of the universe.
Imagine a hypothetical world, where mainstream opinion (like you’d find in the New York Times) says that AGI would destroy the world, and a powerful regulatory agency has the law on its side. I bet AGI is delayed by decades.
3. Don’t underestimate how effectively the US government can do this job
Don’t over-index on covid or climate change. AI safety is different. Covid and climate change both demand sacrifices from the entire population. This is hugely unpopular. AI safety, on the other hand, only demands sacrifices from a small number of companies
For now, I think the top priority is to clearly and persuasively demonstrate why alignment won’t be solved in the next 30 years. This is crazy hard, but it might be way easier than actually solving alignment
Is there a good write up of the case against rapid tests? I see Tom Frieden’s statement that rapid tests don’t correlate with infectivity, but I can’t imagine what that’s based on
In other words, there’s got to be a good reason why so many smart people oppose using rapid tests to make isolation decisions
Could you spell out your objection? It’s a big ask, having read a book just to find out what you mean!
Short summary: Biological anchors are a bad way to predict AGI. It’s a case of “argument from comparable resource consumption.” Analogy: human brains use 20 Watts. Therefore, when we have computers with 20 Watts, we’ll have AGI! The 2020 OpenPhil estimate of 2050 is based on a biological anchor, so we should ignore it.
Longer summary:
Lots of folks made bad AGI predictions by asking:
How much compute is needed for AGI?
When that compute will be available?
To find (1), they use a “biological anchor,” like the computing power of the human brain, or the total compute used to evolve human brains.
Hans Moravec, 1988: the human brain uses 10^13 ops/s, and computers with this power will be available in 2010.
Eliezer objects that:
“We’ll have computers as fast as human brains in 2010” doesn’t imply “we’ll have strong AI in 2010.”
The compute needed depends on how well we understand cognition and computer science. It might be done with a hypercomputer but very little knowledge, or a modest computer but lots of knowledge.
An AGI wouldn’t actually need 10^13 ops/s, because human brains are inefficient. One example, they do lots of operations in parallel, which could be replaced with fewer operations in series.
Eliezer, 1999: Eliezer mentions that he too made bad AGI predictions as a teenager
Ray Kurzweil, 2001: Same idea as Moravec, but 10^16 ops/s. Not worth repeating
Someone, 2006: it took ~10^43 ops for evolution to create human brains. It’ll be a very long time before a computer can reach 10^43 ops, so AGI is very far away
Eliezer objects that the use of a biological anchor is sufficient to make this estimate useless. It’s a case of a more general “argument from comparable resource consumption.”
Analogy: human brains use 20 Watts. Therefore, when we have computers with 20 Watts, we’ll have AGI!
OpenPhil, 2020: A much more sophisticated estimate, but still based on a biological anchor. They predict AGI in 2050.
How the new model works:
Demand side: Estimate how many neural-network parameters would emulate a brain. Use this to find the computational cost of training such a model. (I think this part mischaracterizes OpenPhil’s work, my comments at the bottom)
Supply side: Moore’s law, assuming
Willingness to spend on AGI training is a fixed percent of GDP
“Computation required to accomplish a fixed task decreases by half every 2-3 years due to better algorithms.”
Eliezer’s objections:
(Surprise!) It’s still founded on a biological anchor, which is sufficient to make it invalid
OpenPhil models theoretical AI progress as algorithms getting twice as efficient every 2-3 years. This is a bad model, because folks keep finding entirely new approaches. Specifically, it implies “we should be able to replicate any modern feat of deep learning performed in 2021, using techniques from before deep learning and around fifty times as much computing power.”
Some of OpenPhil’s parameters make it easy for the modelers to cheat, and make sure it comes up with an answer they like:
“I was wondering what sort of tunable underdetermined parameters enabled your model to nail the psychologically overdetermined final figure of ’30 years’ so exactly.”
Can’t we use this as an upper bound? Maybe AGI will come sooner, but surely it won’t take longer than this estimate.
Eliezer thinks this is the same non-sequitur as Moravec’s. If you train a model big enough to emulate a brain, that doesn’t mean AGI will pop out at the end.
Other commentary: Eliezer mentions several times that he’s feeling old, tired, and unhealthy. He feels frustrated that researchers today repeat decades-old bad arguments. It takes him a lot of energy to rebut these claims
My thoughts:
I found this persuasive, but I also think it mischaracterized the OpenPhil model
My understanding is that OpenPhil didn’t just estimate the number of neural network parameters required to train a human brain. They used six different biological anchors, including the “evolution anchor’, which I find very useful for an upper bound.
Holden Karnofsky, who seems to put much more stock in the Bio Anchors model than Eliezer, explains the model really well here. But I was frustrated to see that the write-up on Holden’s blog gives 50% by 2090 (first graph) using the evolution anchor, while the same graph in the old calcs gives only 11%. Was this model tuned after seeing the results?
My conclusion: Bio Anchors is a terrible way to model when AGI will actually arrive. But I don’t agree with Eliezer’s dismissal of using Bio Anchors to get an upper bound, because I think the evolution anchor achieves this.
What particular counterproductive actions by the public are we hoping to avoid?
Zvi just posted EY’s model
I should’ve been more clear…export controls don’t just apply to physical items. Depending on the specific controls, it can be illegal to publicly share technical data, including source code, drawings, and sometimes even technical concepts
This makes it really hard to publish papers, and it stops you from putting source code or instructions online
Why isn’t there a persuasive write-up of the “current alignment research efforts are doomed” theory?
EY wrote hundreds of thousands of words to show that alignment is a hard and important problem. And it worked! Lots of people listened and started researching this
But that discussion now claims these efforts are no good. And I can’t find good evidence, other than folks talking past each other
I agree with everything in your comment except the value of showing EY’s claim to be wrong:
Believing a problem is harder than it is can stop you from finding creative solutions
False believe in your impending doom leads to all sorts of bad decisions (like misallocating resources, or making innocent researchers’ lives worse)
Belief in your impending doom is terrible for your mental heath (tbh I sensed a bit of this in the EY discussion)
Insulting groups like OpenAI destroys a lot of value, especially if EY is actually wrong
If alignment were solved, then developing AGI would be the best event in human history. It’d be a shame to prevent that
In other words, if EY is right, we really need to know that. And know it in way that’s easy to persuade others. If EY is wrong, we need to know that too, and stop this gloom and doom
I agree. This wasn’t meant as an object level discussion of whether the “alignment is doomed” claim is true. What I’d hopes to convey is that, even if the research is on the wrong track, we can still massively increase the chances of a good outcome, using some of the options I described
That said, I don’t think Starship is a good analogy. We already knew that such a rocket can work in theory, so it was a matter of engineering, experimentation, and making a big organization work. What if a closer analogy to seeing alignment solved was seeing a proof of P=NP this year?
In fact, what I’d really like to see from this is Leverage and CFAR’s actual research, including negative results
What experiments did they try? Is there anything true and surprising that came out of this? What dead ends did they discover (plus the evidence that these are truly dead ends)?
It’d be especially interesting if someone annotated Geoff’s giant agenda flowchart with what they were thinking at the time and what, if anything, they actually tried
Also interested in the root causes of the harms that came to Zoe et al. Is this an inevitable consequence of Leverage’s beliefs? Or do the particular beliefs not really matter, and it’s really about the social dynamics in their group house?
I don’t agree with the characterization of this topic as self-obsessed community gossip. For context, I’m quite new and don’t have a dog in the fight. But I drew memorable conclusions from this that I couldn’t have gotten from more traditional posts
First, experimenting with our own psychology is tempting and really dangerous. Next time, I’d turn up the caution dial way higher than Leverage did
Second, a lot of us (probably including me) have an exploitable weakness brought on high scrupulously combined with openness to crazy-sounding ideas. Next time, I’d be more cautious (but not too cautious!) about proposals like joining Leverage
Third, if we ever need to maintain the public’s goodwill, I’ll try not to use words like “demonic seance”… even if I don’t mean it literally
In short, this is the sort of mistake worth learning about, including for those not personally affected, because it’s the kind of mistake we could plausibly make again. I think it’s useful to have here, and the right attitude for the investigation is “what do these events teach us about how rationalist groups can go wrong?” I also don’t think posting a summary would’ve been sufficient. It was necessary to hear Geoff and Anna’s exact words
So is this an accurate summary of your thinking?
You agree with FDT on some issues. The goal of decision theory is to determine what kind of agent you should be. The kind of agent you are (your “source code”) affects other agents’ decisions
FDT requires you to construct counterfactual worlds. For example, if I’m faced with Newcomb’s problem, I have to imagine a counterfactual world in which I’m a two-boxer
We don’t know how to construct counterfactual worlds. Imagining a consistent world in which I’m a two-boxer is just as hard as imagining a one where objects fall up, or 2+2 is 5
You get around this by constructing counterfactual models, instead of counterfactual worlds. Instead of trying to imagine a consistent world that caused me to become a two-boxer, I just make a simple mental model. For example, my model might indicate that my “source code” causes both the predictor’s decision and mine. From here, I can model myself as a two-boxer, even though that model doesn’t represent any possible world
I think 2, 3, and 8 are true but pretty easy to overcome. Just get someone knowledgeable to help you
4 (low demand for these essays) seems like a calibration question. Most writers probably would lose their audience if they wrote about it as often as Holden. But more than zero is probably ok. Scott Alexander seems to be following that rule, when he said that we was summarizing the 2021 MIRI conversations at a steady drip so as not to alienate the part of his audience that doesn’t want to see that
I think 6 (look weird) used to be true, but it’s not any more. It’s hard to know for sure without talking to Kelsey Piper or Ezra Klein, but I suspect they didn’t lose any status for their Vox/NYT statements