Brief note about this post: I am a graduate student working near the area of quantum computing hardware. Recently, I have been trying to figure out what to do with my career, and came across this 80,000 Hours post that mentioned AI hardware. I figured I might be able to work in this area, so I’ve spent a little time (~100 hours) looking into this topic. This post is a summary of my initial takeaways from exploring this, as well as an open invitation to comment/critique/collaborate on my personal career plans.
Many thanks to Changyan Wang and for feedback on parts of this post and helpful edits from Malte Hendrickx and Eric Herboso. All remaining mistakes are my own.
I first came across the idea of working on AI hardware from 80,000 Hours (80k), from their post “Some promising career ideas beyond 80,000 Hours' priority paths”, where they offer a few reasons to go into AI hardware:
“Some ways hardware experts may be able to help positively shape the development of AI include:
- More accurately forecasting progress in the capabilities of AI systems, for which hardware is a key and relatively quantifiable input.
- Advising policymakers on hardware issues, such as export, import, and manufacturing policies for specialized chips. (Read a relevant issue brief from CSET.)
- Helping AI projects in making credible commitments by allowing them to verifiably demonstrate the computational resources they’re using.
- Helping advise and fulfill the hardware needs for safety-oriented AI labs.”
In sections 1–4, I will try to give my understanding of each of these ideas a little more deeply and speculate on what sort of career path may lead in that direction (one section corresponding to each of the four points above). In section 5, I will then try to summarize the types of careers that could work on these problems. In section 6, I will discuss some small tests one could perform to check to try out these different careers. I will then finish up in section 7 with my current thinking about my own career plans in light of this.
Note also the below advice from the 80k post (emphasis my own):
“If you do take this path, we encourage you to think carefully through the implications of your plans, ideally in collaboration with strategy and policy experts also focused on creating safe and beneficial AI.”
I have not done this careful thinking, and would love to collaborate with strategy and policy experts.
Discussion of topic
A classic example of a forecast in the space of computer hardware is Moore’s Law, which predicted that the number of transistors on a computer chip would double every two years. One reason EA might be interested in hardware trends like this is for the purpose of forecasting AI timelines. I think the most comprehensive forecasting in this space is being done by Ajeya Cotra at the Open Philanthropy Project. Her report is the culmination of a number of detailed forecasts, including how the price of computer power will change over time. A forecast of the cost of computer power, in turn, requires a forecast of the cost and abilities of AI hardware. As described in section 4, there is increasing investment in innovative technologies for AI hardware, so the most detailed forecasts in AI hardware might require more than an extrapolation of Moore’s Law. (Also, for discussion of the forecasting being done at OpenAI, see, for instance, Danny Hernandez’s podcast with 80k and the links in the show notes)
At first glance, it seemed to me that the existence of Ajeya’s report demonstrates that the EA community already has enough people with sufficient knowledge and access to expert opinion that, on the margin, adding one expert in hardware to the EA community wouldn’t improve these forecasts much. I think an argument against this initial reaction is that subject matter experts can probably have a better understanding of blind spots and an intuition about unknown unknowns. Indeed, in his 80k podcast, Danny Hernandez says “the kind of person who I’d be most interested in trying to make good forecasts about Moore’s law and other trends, is somebody who has been building chips for a while or has worked in building chips for a while. I think there aren’t that many of those people.”
Some examples of career paths in computer hardware that would work toward forecasting:
- Working on broad forecasts (like Ajeya’s) as a superforecaster. It seems that at least Open Philanthropy Project and OpenAI employ people working on this, and I think some of the policy-focused organizations (discussed in the next section) are interested in this type of work. I think there are several paths that lead here, though some prerequisites may be (1) having enough experience in the fields of hardware and AI to avoid blind spots and know who the subject matter experts are, and (2) experience as a forecaster.
- Being a subject matter expert on narrow topics in AI hardware trends. Ideally, this would be someone on the very cutting edge of AI hardware, which I think would include professors and more senior staff at companies like NVIDIA.
Discussion of topic
This topic was also touched on in the podcast with Danny Hernandez, where he spoke about how experts in hardware could influence the safe development of of AI, stating:
“Trying to work with governments or other sorts of bodies that might be able to regulate AI hardware or perhaps create the kinds of incentives that would make an advance at the right times and the right places… it’d be reasonable to try starting now with that kind of thing in mind. But that’s pretty speculative. I know less about that than the forecasting type thing.”
An example of the interplay between AI hardware and policy is the brief from the Center for Security and Emerging Technology (CSET) referenced in the 80k post from section 0. This brief builds the case why AI hardware has unique instrumental value in the AI policy space and how to use it. Unlike software, which is decentralized and hard to regulate, the equipment to make the most advanced computer chips is much more centralized. Therefore, carefully crafted policy can regulate the distribution of AI hardware, providing a leverage point to regulate the development of AI more generally. The brief utilizes a relatively deep understanding of the state-of-the-art in AI hardware, identifying exactly which companies would need to be involved, and making recommendations on what class of equipment to target.
A series of Future Perfect newsletters (including Nov 13; Nov 20; and especially Dec 04, 2020) outlines a case that there is some long-hanging fruit in enacting effective policy in Washington, DC. So, I am cautiously optimistic that people interested in hardware policy can do a lot of good in this space (for further discussion of this, see section 5.)
- 80k has a lot of advice on AI policy on their AI Policy priority path writeup, where they mention careers paths including working at top AI labs, joining a think tank, working for the US government, working in academia, or working in party politics.
- My understanding is that one way to slice the space of careers in policy is between government roles and non-government roles. Government roles are closer to where the decisions are being made, but are also more well-suited to certain backgrounds than others.
- Danny gives one picture of a career path in this area. He discussed how at some places, like OpenAI, you can build a relationship with the company by reaching out to a current employee to build a relationship with them as your informal mentor, and then eventually converting that relationship into a job offer. Since this may not be feasible in some policy roles, one path he described was: “So you just apply to the informal places first and you walk up the chain. Sometimes there’s a way to get some minimum credential. I think like a public policy masters or something is kind of one way where people get a credential quite quickly that makes them seem reasonable. So it’s like you could be somebody that has one of those and has a background in hardware and then all of a sudden you’re like one of the most credentialed people there is. It could happen pretty quickly”
3. Hardware Security and Increased Coordination
Discussion of topic
These ideas were also discussed in the 80k podcast with Danny Hernandez. I think the reason these two topics are lumped together is that, if you want to improve coordination, one likely necessary condition is being able to trust each other, and security guarantees are one way to build trust. There is a paper Toward Trustworthy AI Development: Mechanisms for Supporting Verifiable Claims that fleshes out 10 mechanisms to implement toward this end (very brief summary from OpenAI here). The three mechanisms they report under the heading of hardware (and their proposed actions) are:
- Secure hardware for machine learning
- Proposed action: Industry and academia should work together to develop hardware security features for AI accelerators or otherwise establish best practices for the use of secure hardware (including secure enclaves on commodity hardware) in machine learning contexts.
- High-precision compute measurement
- Proposed action: One or more AI labs should estimate the computing power involved in a single project in great detail and report on lessons learned regarding the potential for wider adoption of such methods.
- Computing power support for academia
- Proposed action: Government funding bodies should substantially increase funding for computing power resources for researchers in academia, in order to improve the ability of those researchers to verify claims made by industry.
The other type of increased coordination brought up in the podcast with Danny is trying to get big companies to sign up for the Windfall Clause. The motivation behind the Windfall Clause is to “address several potential problems with AI-driven economic growth. The distribution of profits could compensate those rendered faultlessly unemployed due to advances in technology, mitigate potential increases in inequality, and smooth the economic transition for the most vulnerable.” The proposed solution to this brought up in the FHI document is “an ex ante commitment by AI firms to donate a significant amount of any eventual extremely large profits.” One example of this type of commitment is the OpenAI LP. It seems there also could be a winner-takes-all competition for the company making the chips that create transformative AI, so AI hardware companies are likely an effective place to target this type of policy.
- For career paths in hardware security, I think there is some mainstream research being done in this area, but I’m not sure if it is the type of research that would address the mechanisms in the above paper. I would love to learn more about the state of this field from someone with more experience. Regarding careers that would give you influence of Windfall Clause type coordination, in the 80k podcast Danny suggests the way to go about having this kind of influence is to be a founder/early employee at a startup, or to at least have a close relationship with an executive.
4. Advising and Fulfilling Hardware Needs
Discussion of topic
I’m not sure if 80k has expanded on this point anywhere else, but I think one reasonable interpretation is that this would be something like working in industry and being a contact point for the EA community and organizations like OpenAI. If anyone did want to contact an expert, people in the EA community generally know your name and can direct others toward you.
I think someone in this role could also be proactive about keeping EA organizations up to date about the state-of-the-art. Further, as we gain a better understanding of how EAs can most effectively influence the development of AI, it seems reasonable that there will be an increased utility to have EAs working directly on AI hardware.
There is a large list of established companies and startups in the AI hardware space on James Wang’s twitter account. Note that some of the companies on this list are working on technologies that are unlike what’s worked on in mainstream computer architecture. Some of the new types of hardware that I’ve heard about are:
- Photonics: It’s hard to get an idea exactly the state of the technology in industry in this field since there are a lot of secretive start-ups, but there’s the sales pitch in this video. I have compiled a list of the companies I know if working in photonics chips for AI here.
- Quantum computing: There is a significant academic research effort in this area as well as a growing list of companies working toward commercial devices. Note there are some reasons one may be skeptical of the contribution of quantum computing to AI Alignment in general, see this post from Jaime Sevilla and links within. (I don’t know enough to have a strong opinion personally.)
- Other technologies beyond the current industry standards that are cataloged in the IEEE IRDS (though I don’t know which of the many technologies listed are as relevant to AI hardware as the two listed above)
Note, as discussed in the podcast with Danny, there could also be risks associated with working directly on AI hardware, for instance it could just accelerate AI timelines without making anything safer.
- I think this would involve working at one of the industries like those listed above and maintaining involvement in the EA community. I think the most clear path to working on AI hardware would be gaining experience in computer architecture, but many of the different technologies could be approached from different directions.
5. Some Example Career Paths
Given the problems in AI Hardware listed above, here are some career paths I think one could take to work on them. When possible, I’ll try to highlight a person that has actually been in this role, I would love to hear more examples of possible role models.
- University professor doing research at the cutting edge of AI hardware. I think some possible research topics could be: anything in section 3, computer architecture focusing on AI hardware, or research in any of the alternative technologies listed in section 4. I would love to learn about what other areas are important and who the leaders are in all these areas.
- Academia working on AI Policy and Strategy. 80k has a lot of resources on this career path here. Also see the 80k podcast with Allan Dafoe.
- Government research in places like IARPA. Jason Matheny was the head of IARPA and explains some of the roles in this space in this talk (note, some of these roles can be a <5 year tour of duty as part of a completely different career, and still have a really high impact).
- Think tanks like CSET: My impression is that this is mostly policy focused and has less of a technical focus compared to IARPA. See, e.g., the 80k podcast with Helen Toner. My impression is that many roles in think tanks are not designed to be career-long roles, but jumping off points to careers in other roles in government.
- Office of Science and Technology Policy (OSTP): I think of two different approaches here:
- As described in the 80k podcast with Tom Kalil, one way to get into a highly influential organization like this is to work as an adviser to politicians.
- Another example of this type of career is that of the physicist Jake Taylor. My understanding is that he took a sort of “tour of duty” type of role at the OSTP, where he was a big factor in the White House’s increased interest in quantum computing. This resulted in the billion dollar National Quantum Initiative (NQI). While the direct analogy of this for AI was passed in the FY21 NDAA, I think this still highlights the amount of impact one can have pushing an idea forward.
- Industry: See section 4 for a list of possible companies to work at.
- Startups may be especially interesting because of the coordination aspects discussed in section 3. Anyone interested in going the startup route may want to pick a grad/undergrad program in a department that has especially good resources for entrepreneurship.
- Forecasting at an organization like Open Philanthropy Project or OpenAI to influence funding and policy (similar to Ajeya Cotra and Danny Hernandez, as described in section 1).
6. Some Small Tests in this Area
Some ways to make small tests in the technical side of things:
- REUs for undergrads
- Online courses (for instance, someone in the silicon photonics industry highly recommended this course on edX course for an introduction to their field)
Some ways to make small tests in the non-technical side of things include:
- AAAS Science and Technology Policy Fellowship (STPF)
- Governance of AI Fellowship
- FHI Research Scholars Programme or their Summer Research Fellowship
7. My career plans
Given this information, here’s how I have been thinking about what I should try with my career plans. Critical comments are especially welcome on this, also open to DMs.
First, I plan to do more exploration before I graduate (planned in spring 2022) by
- Gaining experience in photonics from the edX course mentioned above in spring 2021
- Expanding my experience with real AI hardware doing an internship in summer 2021
- Gaining experience in tech policy by joining a reading group
- Applying for the AAAS STPF during June–November 2021, (for the positions starting in September 2022) and, if accepted, do that after graduation.
These experiences will probably update my thoughts on my career significantly. Specifically, I think my experience in the STPF (including not being accepted) would update me significantly about my comparative advantage for policy. However, following from the 80k career guide, with my current experience here are my plans A/B/Z:
- Plan A: Try for a “Jake Taylor” type career, staying involved with technical research but take “tour of duty” roles in government. I think one possible path would be to gain experience in industry after grad school in either photonics or quantum computing. After, say, five years, apply to be a program manager at DARPA or IARPA.
- Plan B: If straddling tech and policy is untenable, stick to the government/policy side, and try for a “Jason Matheny” type career (who is the former director of IARPA and the current founding director of CSET).
- Plan Z: Apply to industry/national lab jobs in quantum computing, and re-evaluate how I will have my impact guided by section 1, 3, and/or 4.
I would also be interested if anyone has opinions about whether academic roles might be more impactful than roles in industry or government. As I see it, the main reason to go into academia is an argument of comparative advantage, but it seems to me that it may give no more opportunities to do good than a role in industry or government.