Director of Epoch, an organization investigating the future of Artificial Intelligence.
Currently working on:
Macroeconomic models of AI takeoff
Trends in Artificial Intelligence
Forecasting cumulative records
Improving forecast aggregation
I am also one of the coordinators of Riesgos Catastróficos Globales, a Spanish-speaking network of experts working on Global Catastrophic Risks.
I also run Connectome Art, an online art gallery where I host art I made using AI.
Strong upvoting because I want to incentivize people to write and share more summaries.
Summaries are awesome and allow me to understand the high level of papers that I would not have read otherwise. This summary in particular is well written and well-formatted.
Thanks for writing it and sharing it!
As further reading I recently came across Research as a Stochastic Decision Process, which discusses another systematic approach to research.
Summary copy pasted from the article:
Many of our default intuitions about how to pursue uncertain ideas are counterproductive:
We often try easier tasks first, when instead we should try the most informative tasks first.
We often conflate a high-level approach with a low-level instantiation of the approach.
We are often too slow to try to disprove our own ideas.
Building frameworks that reify the research process as a concrete search problem can help unearth these incorrect intuitions and replace them with systematic reasoning.
I like this post a lot; it is succint and provides a great actionable for EAs to act on.
Stylistically I would prefer if the Organization section was broken down into a paragraphs per section to make it easier to read.
I like that you precommited to a transparent way of selecting the new causes you present to the readers and limited the scope to 15. I would personally have liked to see them broken up in sections depending on what method they were chosen by.
For other readers who are eager for more, here there are other two that satisfy the criteria but I suppose did not make it to the list:
Atomically Precise Manufacturing (cause area endorse by two major organizations—OPP and Eric Drexler from FHI)
Aligning Recommender Systems (cause profile with more than 50 upvotes in the EA forum)
I do agree with your assesment, and I would be medium excited about somebody informally researching what algorithms can be quantized to see if there is low hanging fruit in terms of simplifying assumptions that could be made in a world where advanced AI is quantum-powered.
However my current intuition is there is no much sense in digging in this unless we were sort of confident that 1) we will have access to QC before TAI and that 2) QC will be a core component of AI.
To give a bit more context to the article, Pablo and me originally wrote it because we disagreed on whether current research in AI Alignment would still be useful if quantum computing was a core component of advanced AI systems.
Had we concluded that quantum ofuscation threatened to invalidate some assumptions made by current research, we would have been more emphatic about the necessity of having quantum computing experts working on “safeguarding our research” on AI Alignment.
Thank you for writing this up—always good to see criticism of key ideas.
I want to contest point 4.
The fact that we can decompose “Good done / extra person or $” into three factors that can be roughly interpreted as Scale, Tractability and Neglectedness is not a problem, but a desirable property.
In ultimate instance, we want to evaluate marginal cost effectiveness ie “Good done / extra person or $”. However this is difficult, so we want to split it up in simpler terms.
The mathematical equation that decomposes the cost serves as a guarantee that by estimating all three factors we will not be leaving anything important behind.
I’m currently working on explicating some of these factors, but some examples would be drastic climate change, long-lived radionuclides, increase in persistent pathogens.
Can you explain the bit about long-lived radionuclides?
How would they be produced? How would they affect “technological carrying capacity”?
Note that we believe that quantum supremacy has already been achieved.
As in, the quantum computer Sycamore from Google is capable of solving a (toy) problem that we currently believe unfeasible in a classical computer.
Of course, there is a more interesting question of when will we be able to solve practical problems using quantum computing. Experts believe that the median for a practical attack on modern crypto is ~2035.
I regardless believe that outside (and arguably within) quantum cryptanalysis the applications will be fairly limited.
The paper in my post goes in more detail about this.
Thank you so much for your kind words and juicy feedback!
Google has already deployed post-quantum schemes as a test
I did not know about this, and this actually updates me on how much overhead will be needed for post quantum crypto (the NIST expert I interviewed gave me an impression that it was large and essentially would need specialized hardware to meet performance expectations, but this seems to speak to the contrary (?))
here may be significant economic costs due to public key schemes deployed “at rest”
To make sure I understand your point, let me try to paraphase. You are pointing out that:
1) past communications that are recorded will be rendered insecure by quantum computing
2) there are some transition costs associated with post quantum crypto—which are related to for example the cost of rebuilding PGP certificate networks.
If so, I agree that this is a relevant consideration but does not change the bottom line.
Thank you again for reading my paper!
I think we broadly agree.
I believe that chemistry and material science are two applications where quantum computing might be a useful tool, since simulating very simple physical systems is something where a quantum computer excels at but arguably significantly slower to do in a classical computer.
On the other hand, people more versed on material science and chemistry I talked to seemed to believe that (1) classical approximations will be good enough to approach problems in these areas and (2) in silico design is not a huge bottleneck anyway.
So I am open to a quantum computing revolution in chemistry and material science, but moderately skeptical.
Summarizing my current beliefs about how important quantum computing will be for future applications:
Cryptoanalysis ⇒ very important for solving a handful of problems relevant for modern security, with no plausible alternative
Chemistry and material science ⇒ Plausibly useful, not revolutionary.
AI and optimization ⇒ unlikely to be useful, huge constraints to overcome
Biology and medicine ⇒ not useful, systems too complex to model
It is not intended to be a calibrated estimated, though we were hoping that it could help others make calibrated estimations.
The ways that a calibrated estimate would differ include:
1. The result is a confidence interval, not a credence interval (most places in the paper where it says probability is should say confidence, I apologize for the oversight), so your choice of prior can make a big difference to the associated credence interval.
2. The model is assuming that no discontinuous progress will happen, but we do not know whether this will hold. (Grace, 2020) estimates a yearly rate of discontinous breakthroughs on any given technology of 0.1%, so I’d naively expect a 1-(1-0.1%)^20 = 2% chance that there is such a discontinuous breakthrough for quantum computing in the next 20 years.
3. The model makes optimistic assumptions of progress—namely that a) the rate of exponential progress will hold for both the physical qubit count and the gate error rate, b) there is no correlation between the metrics in a system (which we show it is probably an optimistic assumption, since it is easier to optimize only one of the metrics than both) and c) we ignore the issue of qubit connectivity due to lack of data and modelling difficulty.
If I was pressed to put a credence bound on it, I’d assign about 95% chance that EITHER the model is basically correct OR that the timelines are slower than expected (most likely if the exponential trend of progress on gate error rate does not hold in the next 20 years), for an upper bound on the probability that we will have RSA 2048 quantum attacks by 2040 of <5% + 95% 5% ~= 10%.
Either case, I think that the model should make us puzzle over the expert timelines, and inquire whether they are taking into account any extra information or being too optimistic.
EDIT: I made an artihmetic mistake, now corrected (thanks to Eric Martin for pointing it out)
The citation is a link: (Grace, 2020)
Just in case: https://aiimpacts.org/discontinuous-progress-in-history-an-update/
Without entering into too many sensitive details, when I have looked at the output of similar programs I have noticed that I was excited about the career path of 1 out of every 3 participants.
But a) I dont know how much of it was counterfactual, b) when I made the estimation I had an incentive to produce an optimistic answer and c) it relies on my subjective judgement, which you may not trust.
Also worth noting that I think that the raw conversion rate is not the right metric to focus on—the outliers usually account for most the impact of these programs.
Brainstorming some concrete examples of what everyday longtermism might look like:
> Alice is reviewing a CV for an applicant. The applicant does not meet the formal requirements for the job, but Alice wants to hire them anyway. Alice visualices the hundreds of people making a similar decision to hers. She would be ok with hiring this specific applicant, because she trusts her instincts a lot. But she would not trust 100 people in a similar position to make the right choice; ignorning the recruiting guidelines might disadadvantage minorities in a illegible way. She decides that the best would be if everyone chose to just follow the procedure, so she chooses to forego her intuition to favor better decision making overall.
> Beatrice is offered a job as a Machine Learning engineer to help the police with some automated camera monitoring. Before deciding whether to accept, she seeks out open criticism of that line of work, and tries to imagine what are some likely consequences of developing that kind of technology, both positive and negative. After doing a balance she realizes that while it will most likely be positive there is a plausible chance that it will enable really bad situations, and rejects the job offer.
> Carol reads an interesting article. She wants to share it in social media. She could spend some effort paraphrasing the key ideas in the article, or just share the link. She has internalized that spending one minute summarizing key ideas might well be worth a lot of time saved to her friends who otherwise could use her summary to decide whether to read the whole article. Out of habit she summarizes the article as best as she can, making it clear who she genuinely thinks would benefit from reading the article.
Why would research on ‘minor’ GCRs like the ones mentioned by Arepo be harder than eg AI alignment?
My impression is that there is plenty of good research on eg effects of CO2 on health, the Flynn effect and Kessler syndrome, and I would say its much higher quality than extant X risk research.
Is the argument that they are less neglected?
:surprised:
How can I put footnotes on my posts?!?!
I feel like invoking worldview diversification here is discussing things at the wrong level.
Is like saying “oh its ok that you believe in intersectionality, because from a worldview diversification perspective we want to work on many causes anyway”, and failing to address the fundamental disagreement that within their worldview a intersectionalist does not find cause prioritization useful.
Like, I feel the crux of intersectionality is about different problems being interwoven in complex, hard-to-understand ways. So as OP pointed out, if you believe this you’ll need to address all problems at once by radically restructuring society.
Meanwhile, the crux of worldview diversificationists is that we are not certain of our own values and how they will change, so it is better to hedge your bets by compromising between many views.
I find your steelman convincing (would love more intersectionalists to confirm though!).
Re: downsides of intercause prioritization. Beyond making people feel bad about their work, systematic prioritization can systematically misallocate resources, while a more informal, holistic and intersectional approach is less likely to make this kind of mistake.
Arguably, while EAs are very well aware of the importance of hit-based giving, they are overly focused on a few cause areas. Meanwhile my (naive) impression is that intersectionalists are succesfully tackling a much wider array of problem areas and interventions, from community help to international aid and political lobbying.
I do not think it is a stretch to think that prioritization frameworks are partly to blame for cause convergence in the EA community.
Thank you for writing this!
While this is not the high note of the paper, I read with quite some interest your notes about heavy tailed distributions.
I think that the concept of heavy tailed distributions underpins a lot of considerations in EA, yet as you remark many people (including me) are still quite confused about how to formalize the concept effectively, and how often it applies in real life.
Glad to see more thinking going into this!
What is EAA? Effective Animal Advocacy?
Totally second the motion of empowering junior researchers to ask for research ideas here and somewhere else.
Also, I’d encourage you to write down your own research agenda in the form of a blogpost listing some open questions in this forum!
It will be useful for other researchers and you will get interesting feedback on your ideas :)