This is only true if, for example, you think AI would cause GDP growth. My model assigns a lot of probability to ‘AI kills everyone before (human-relevant) GDP goes up that fast’, so questions #7 and #8 are conditional on me being wrong about that. If we can last any small multiples of a year with AI smart enough to double GDP in that timeframe, then things probably aren’t as bad as I thought.
To emphasize, the clash I’m perceiving is not the chance assigned to these problems being tractable, but to the relative probability of ‘AI Alignment researchers’ solving the problems, as compared to everyone else and every other explanation. In particular, people building AI systems intrinsically spend a degree of their effort, even if completely unconvinced about the merits of AI risk, trying to make systems aligned, just because that’s a fundamental part of building a useful AI.
I could talk about the specific technical work, or the impact that things like the AI FOOM Debate had on Superintelligence had on OpenPhil, or CFAR on FLI on Musk on OpenAI. Or I could go into detail about the research being done on topics like Iterated Amplification and Agent Foundations and so on and ways that this seems to me to be clear progress on subproblems.
I have a sort of Yudkowskian pessimism towards most of these things (policy won’t actually help; Iterated Amplification won’t actually work), but I’ll try to put that aside here for a bit. What I’m curious about is what makes these sort of ideas only discoverable in this specific network of people, under these specific institutions, and particularly more promising than other sorts of more classical alignment.
Isn’t Iterated Amplification in the class of things you’d expect people to try just to get their early systems to work, at least with ≥20% probability? Not, to be clear, exactly that system, but just fundamentally RL systems that take extra steps to preserve the intentionality of the optimization process.
To rephrase a bit, it seems to me that a worldview in which AI alignment is sufficiently tractable that Iterated Amplification is a huge step towards a solution, would also be a worldview in which AI alignment is sufficiently easy (though not necessarily easy) that there should be a much larger prior belief that it gets solved anyway.
There is a huge difference in the responses to Q1 (“Will AGI cause an existential catastrophe?”) and Q2 (“...without additional intervention from the existing AI Alignment research community”), to a point that seems almost unjustifiable to me. To pick the first matching example I found (and not to purposefully pick on anybody in particular), Daniel Kokotajlo thinks there’s a 93% chance of existential risk without the AI Alignment community’s involvement, but only 53% with. This implies that there’s a ~43% chance of the AI Alignment community solving the problem, conditional on it being real and unsolved otherwise, but only a ~7% chance of it not occurring for any other reason, including the possibility of it being solved by the researchers building the systems, or the concern being largely incorrect.
What makes people so confident in the AI Alignment research community solving this problem, far above that of any other alternative?
On the other hand, improvements on ImageNet (the datasets alexnet excelled on at the time) itself are logarithmic rather than exponential and at this point seem to have reached a cap at around human level ability or a bit less (maybe people got bored of it?)
The best models are more accurate than the ground-truth labels.
Are we done with ImageNet?https://arxiv.org/abs/2006.07159
Yes, and no. We ask whether recent progress on the ImageNet classification benchmark continues to represent meaningful generalization, or whether the community has started to overfit to the idiosyncrasies of its labeling procedure. We therefore develop a significantly more robust procedure for collecting human annotations of the ImageNet validation set. Using these new labels, we reassess the accuracy of recently proposed ImageNet classifiers, and find their gains to be substantially smaller than those reported on the original labels. Furthermore, we find the original ImageNet labels to no longer be the best predictors of this independently-collected set, indicating that their usefulness in evaluating vision models may be nearing an end. Nevertheless, we find our annotation procedure to have largely remedied the errors in the original labels, reinforcing ImageNet as a powerful benchmark for future research in visual recognition.
Figure 7. shows that model progress is much larger than the raw progression of ImageNet scores would indicate.
I think this is wrong, but I’m having trouble explaining my intuitions. There are a few parts;
You’re not doing Solomonoff right, since you’re meant to condition on all observations. This makes it harder for simple programs to interfere with the outcome.
More importantly but harder to explain, you’re making some weird assumptions of the simplicity of meta-programs that I would bet are wrong. There seems to be a computational difficulty here, in that you envision 2n small worlds trying to manipulate 2m other worlds, where m>n. That makes it really hard for the simplest program to be one where the meta-program that’s interpreting the pointer to our world is a rational agent, rather than some more powerful but less grounded search procedure. If ‘naturally’ evolved agents are interpreting the information pointing to the situation they might want to interfere with, this limits the complexity of that encoding. If they’re just simulating a lot of things to interfere with as many worlds as possible, they ‘run out of room’, because 2m≫2n.
Your examples almost self-refute, in the sense that if there’s an accurate simulation of you being manipulated at time t+1, it implies that simulation is not materially interfered with at time t, so even if the vast majority of Solomonoff inductions have attempted adversary, most of them will miss anyway. Hypothetically, superrational agents might still be able coordinate to manipulate some very small fraction of worlds, but it’d be hard and only relevant to those worlds.
Compute has costs. The most efficient use of compute is almost always to do enact your preferences directly, not manipulate other random worlds with low probability. By the time you can interfere with Solomonoff, you have better options.
To the extent that a program P is manipulating predictions so that another other program that is simulating P performs unusually… well, then that’s just how the metaverse is. If the simplest program containing your predictions is an attempt at manipulating you, then the simplest program containing you is probably being manipulated.
IRV is an extremely funky voting system, but almost anything is better than Plurality. I very much enjoyed Ka-Ping Yee’s voting simulation visualizations, and would recommend the short read for anyone interested.
I have actually made my own simulation visualization, though I’ve spent no effort annotating it and the graphic isn’t remotely intuitive. It models a single political axis (eg. ‘extreme left’ to ‘extreme right’) with N candidates and 2 voting populations. The north-east axis of the graph determines the centre of one voting population, and the south-east axis determines the centre of the other (thus the west-to-east axis is when the voting populations agree). The populations have variances and sizes determined by the sliders. The interesting thing this has taught me is that IRV/Hare voting is like an otherwise sane voting system but with additional practically-unpredictable chaos mixed in, which is infinitely better than the systemic biases inherent to plurality or Borda votes. In fact, if you see advantages in sortition, this might be a bonus.
The latter is the source for human perplexity being 12. I should note that it tested on the 1 Billion Words benchmark, where GPT-2 scored 42.2 (35.8 was for Penn Treebank), so the results are not exactly 1:1.
FLOPS don’t seem to me a great metric for this problem; they are often very sensitive to the precise setup of the comparison, in ways that often aren’t very relevant (the Donkey Kong comparison emphasized this), and the architecture of computers is fundamentally different to that of brains. What seems like a more apt and stable comparison is to compare the size and shape of the computational graph, roughly the tuple (width, depth, iterations). This seems like a much more stable metric, since scale-based metrics normally only change significantly when you’re handling the problem in a semantically different way. In the example, hardware implementations of Donkey Kong and various sorts of software emulation (software interpreter, software JIT, RTL simulation, FPGA) will have very different throughputs on different hardware, and the setup and runtime overheads for each might be very different, but the actual runtime computation graphs should look very comparable.
This also has the added benefit of separating out hypotheses that should naturally be distinct. For example, a human-sized brain at 1x speed and a hamster brain at 1000x speed are very different, yet have seemingly similar FLOPS. Their computation graphs are distinct. Technology comparisons like FPGAs vs AI accelerators become a lot clearer from the computation graph perspective; an FPGA might seem at a glance more powerful from a raw OP/s perspective, but first principles arguments will quickly show they should be strictly weaker than an AI accelerator. It’s also more illuminating given we have options to scale up at the cost of performance; from a pure FLOPS perspective, this is negative progress, but pragmatically, this should push timelines closer.
I disagree with that post and its first two links so thoroughly that any direct reply or commentary on it would be more negative than I’d like to be on this site. (I do appreciate your comment, though, don’t take this as discouragement for clarifying your position.) I don’t want to leave it at that, so instead let me give a quick thought experiment.
A neuron’s signal hop latency is about 5ms, and in that time light can travel about 1500km, a distance approximately equal to the radius of the moon. You could build a machine literally the size of the moon, floating in deep space, before the speed of light between the neurons became a problem relative to the chemical signals in biology, as long as no single neuron went more than half way through. Unlike today’s silicon chips, a system like this would be restricted by the same latency propagation limits that the brain is, but still, it’s the size of the moon. You could hook this moon-sized computer to a human-shaped shell on Earth, and as long as the computer was directly overhead, the human body could be as responsive and fully updatable as a real human.
While such a computer is obviously impractical on so many levels, I find it a good frame of reference to think about the characteristics of how computers scale upwards, much like Feynman’s There’s Plenty of Room at the Bottom was a good frame of reference for scaling down, considered back when transistors were still wired by hand. In particular, the speed of light is not a problem, and will never become one, except where it’s a resource we use inefficiently.
Scaling Language Model Size by 1000x relative to GPT3. 1000x is pretty feasible, but we’ll hit difficult hardware/communication bandwidth constraints beyond 1000x as I understand.
I think people are hugely underestimating how much room there is to scale.
The difficulty, as you mention, is bandwidth and communication, rather than cost per bit in isolation. An A100 manages 1.6TB/sec of bandwidth to its 40 GB of memory. We can handle sacrificing some of this speed, but something like SSDs aren’t fast enough; 350 TB of SSD memory would cost just $40k, but would only manage 1-2 TB/s over the whole array, and could not push it to a single GPU. More DRAM on the GPU does hit physical scaling issues, and scaling out to larger clusters of GPUs does start to hit difficulties after a point.
This problem is not due to physical law, but the technologies in question. DRAM is fast, but has hit a scaling limit, whereas NAND scales well, but is much slower. And the larger the cluster of machines, the more bandwidth you have to sacrifice for signal integrity and routing.
Thing is, these are fixable issues if you allow for technology to shift. For example,
Various sorts of persistent memories allow fast dense memories, like NRAM. There’s also 3D XPoint and other ReRAMs, various sorts of MRAMs, etc.
Multiple technologies allow for connecting hardware significantly more densely than we currently do, primarily things like chiplets and memory stacking. Intel’s Ponte Vecchio intends to tie 96 (or 192?) compute dies together, across 6 interconnected GPUs, each made of 2 (or 4?) groups of 8 compute dies.
Neural networks are amicable to ‘spatial computing’ (visualization), and using appropriate algorithms the end-to-end latency can largely be ignored as long as the block-to-block latency and throughput is sufficiently high. This means there’s no clear limit to this sort of scaling, since the individual latencies are invariant to scale.
The switches themselves between the computers are not at a limit yet, because of silicon photonics, which can even be integrated alongside compute dies. That example is in a switch, but they can also be integrated alongside GPUs.
You mention this, but to complete the list, sparse training makes scale-out vastly easier, at the cost of reducing the effectiveness of scaling. GShard showed effectiveness at >99.9% sparsities for mixture-of-experts models, and it seems natural to imagine that a more flexible scheme with only, say, 90% training sparsity and support for full-density inference would allow for 10x scaling without meaningful downsides.
It seems plausible to me that a Manhattan Project could scale to models with a quintillion parameters, aka. 10,000,000x scaling, within 15 years, using only lightweight training sparsity. That’s not to say it’s necessarily feasible, but that I can’t rule out technology allowing that level of scaling.
It might be possible to convince me on something like that, as it fixes the largest problem, and if Hanson is right that blackmail would significantly reduce issues like sexual harassment then it’s at least worth consideration. I’m still disinclined towards the idea for other reasons (incentivizes false allegations, is low oversight, difficult to keep proportionality, can incentivize information hiding, seems complex to legislate), but I’m not sure how strong those reasons are.
I agree this makes a large fractional change to some AI timelines, and has significant impacts on questions like ownership. But when considering very short timescales, while I can see OpenAI halting their work would change ownership, presumably to some worse steward, I don’t see the gap being large enough to materially affect alignment research. That is, it’s better OpenAI gets it in 2024 than someone else gets it in 2026.
This constant seems to be very small, which is why compute had to drop all the way to ~$1k before any researchers worldwide were fanatical enough to bother trying CNNs and create AlexNet.
It’s hard to be fanatical when you don’t have results. Nowadays AI is so successful it’s hard to imagine this being a significant impediment.
Excluding GShard (which as a sparse model is not at all comparable parameter-wise)
I wouldn’t dismiss GShard altogether. The parameter counts aren’t equal, but MoE(2048E, 60L) is still a beast, and it opens up room for more scaling than a standard model.
Robin Hanson argued that negative gossip is probably net positive for society.
Yes, this is what my post was addressing and the analogy was about. I consider it an interesting hypothesis, but not one that holds up to scrutiny.
Lying about someone in a damaging way is already covered by libel/slander laws.
I know, but this only further emphasizes how much better paying those who helped a conviction is. Blackmail is private, threat-based, and necessarily unpoliced, whereas the courts have oversight and are an at least somewhat impartial test for truth.
Gwern’s claim is that these other institutions won’t scale up as a consequence of believing the scaling hypothesis; that is, they won’t bet on it as a path to AGI, and thus won’t spend this money on abstract of philosophical grounds.
My point is that this only matters on short-term scales. None of these companies are blind to the obvious conclusion that bigger models are better. The difference between a hundred-trillion dollar payout and a hundred-million dollar payout is philosophical when you’re talking about justifying <$5m investments. NVIDIA trained an 8.3 B parameter model as practically an afterthought. I get the impression Microsoft’s 17 B parameter Turing-NLG was basically trained to test DeepSpeed. As markets open up to exploit the power of these larger models, the money spent on model scaling is going to continue to rise.
These companies aren’t competing with OpenAI. They’ve built these incredibly powerful systems incidentally, because it’s the obvious way to do better than everyone else. It’s a tool they use for market competitiveness, not as a fundamental insight into the nature of intelligence. OpenAI’s key differentiator is only that they view scale as integral and explanatory, rather than an incidental nuisance.
With this insight, OpenAI can make moonshots that the others can’t: build a huge model, scale it up, and throw money at it. Without this understanding, others will only get there piecewise, scaling up one paper at a time. The delta between the two is at best a handful of years.
If OpenAI changed direction tomorrow, how long would that slow the progress to larger models? I can’t see it lasting; the field of AI is already incessantly moving towards scale, and big models are better. Even in a counterfactual where OpenAI never started scaling models, is this really something that no other company can gradient descent on? Models were getting bigger without OpenAI, and the hardware to do it at scale is getting cheaper.
Legalizing blackmail gives people with otherwise no motivation to harm someone through the sharing of information the motive to do so. I’m going to take that as the dividing line between blackmail and other forms of trade or coercion. I believe this much is generally agreed on in this debate.
If you’re going to legalize forced negative-sum trades, I think you need a much stronger argument that assuming that, on net, the positive externalities will make it worthwhile. It’s a bit like legalizing violence from shopkeepers because most of the time they’re punching thieves. Maybe that’s true now, when shopkeepers punching people is illegal, but one, I think there’s a large onus on anyone suggesting this to justify that it’s the case, and two, is it really going to stay the case, once you’ve let the system run with this newfound form of legalized coercion?
Before I read these excerpts, I was pretty much in the ‘blackmail bad, duh’ category. After I read them, I was undecided; maybe it is in fact true that many harms from information sharing comes with sufficient positive externalities, and those that do not are sufficiently clearly delimited to be separately legislated. Having thought about it longer, I now see a lot of counterexamples. Consider some person, who:
had a traumatic childhood,
has a crush on another person, and is embarrassed about it,
has plans for a surprise party or gift for a close friend,
or the opposite; someone else is planning a surprise for them,
has an injury or disfiguration on a covered part of their body,
had a recent break-up, that they want to hold out on sharing with their friends for a while,
left an unkind partner, and doesn’t want that person to know they failed a recent exam,
posts anonymously for professional reasons, or to have a better work-life balance,
doesn’t like a coworker, but tries not to show it on the job.
I’m sure I could go on for quite a while. Legalizing blackmail means that people are de-facto incentivized to exploit information when it would harm people, because their payout stops being derived from the public interest, through mechanisms like public reception, appreciation from those directly helped by the reveal of information, or payment from a news agency, and becomes proportional almost purely to the damage you can do.
It’s true that in some cases these are things which should be generally disincentivized or made illegal, nonconsensual pornography being a prime example. In general I don’t think this approach scales, because the public interest is so context dependent. Sometimes it is in the public interest to share someone’s traumatic childhood, spoil a surprise or tell their coworker they are disliked. But the reward should be derived from the public interest, not the harm! If we want to monetarily incentivize people to share information they have on sexual abuse, pay them for sharing information that led to a conviction. And if you’re not wanting to do that because it causes the bad incentive to lie… surely blackmail gives more incentive to lie, and the accuser being paid requires the case never to have gone to trial, so is worse on all accounts.
Apple’s launch events get pretty big crowds, a lot of talk, and a lot of celebration.
Putting aside the general question, is OpenAI good for the world, I want to consider the smaller question, how do OpenAI’s demonstrations of scaled up versions of current models affect AI safety?
I think there’s a much easier answer to this. Any risks we face from scaling up models we already have with funding much less than tens of billions of dollars amounts to unexploded uranium sitting around, that we’re refining in microgram quantities. The absolute worst that can happen with connectionist architectures is that we solve all the hard problems without having done the trivial scaled-up variants, and therefore scaling up is trivial, and so that final step to superhuman AI also becomes trivial.
Even if scaling up ahead of time results in slightly faster progress towards AGI, it seems that it at least makes it easier to see what’s coming, as incremental improvements require research and thought, not just trivial quantities of dollars.
Going back to the general question, one good I see OpenAI producing is the normalization of the conversation around AI safety. It is important for authority figures to be talking about long-term outcomes, and in order to be an authority figure, you need a shiny demo. It’s not obvious how a company could be more authoritative than OpenAI while being less novel.
I think the results in that paper argue that it’s not really a big deal as long as you don’t make some basic errors like trying to fine-tune on tasks sequentially. MT-A outperforms Full in Table 1. GPT-3 is already a multi-task learner (as is BERT), so it would be very surprising if training on fewer tasks was too difficult for it.
If the issue is the size of having a fine-tuned model for each individual task you care about, why not just fine-tune on all your tasks simultaneously, on one model? GPT-3 has plenty of capacity.