Deciding not to found a human-data-for-alignment startup

8.6k words (~30 minutes)

Both the project and this write-up were a collaboration with Matt Putz.


Matt Putz and I worked together for the first half of the summer to figure out if we should found a startup with the purpose of helping AI alignment researchers get the datasets they need to train their ML models (especially in cases where the dataset is based on human-generated data). This post, also published on the Effective Altruism Forum and LessWrong (both of which may contain additional discussion in the comments), is a summary of our findings, and why we decided to not do it.


One-paragraph summary: we (two recent graduates) spent about half of the summer exploring the idea of starting an organisation producing custom human-generated datasets for AI alignment research. Most of our time was spent on customer interviews with alignment researchers to determine if they have a pressing need for such a service. We decided not to continue with this idea, because there doesn’t seem to be a human-generated data niche (unfilled by existing services like Surge) that alignment teams would want outsourced.


In more detail: The idea of a human datasets organisation was one of the winners of the Future Fund project ideas competitionstill figures on their list of project ideas, and had been advocated before then by some people, including Beth Barnes. Even though we ended up deciding against, we think this was a reasonable and high-expected-value idea for these groups to advocate at the time.

Human-generated data is often needed for ML projects or benchmarks if a suitable dataset cannot be e.g. scraped from the web, or if human feedback is required. Alignment researchers conduct such ML experiments, but sometimes have different data requirements than standard capabilities researchers. As a result, it seemed plausible that there was some niche unfilled by the market to help alignment researchers solve problems related to human-generated datasets. In particular, we thought - and to some extent confirmed - that the most likely such niche is human data generation that requires particularly competent or high-skill humans. We will refer to this as high-skill (human) data.

We (Matt & Rudolf) went through an informal co-founder matching process along with four other people and were chosen as the co-founder pair to explore this idea. In line with standard startup advice, our first step was to explore whether or not there is a concrete current need for this product by conducting interviews with potential customers. We talked to about 15 alignment researchers, most of them selected on the basis of doing work that requires human data. A secondary goal of these interviews was to build better models for the future importance and role of human feedback in alignment.

Getting human-generated data does indeed cost many of these researchers significant time and effort. However, we think to a large extent this is because dealing with humans is inherently messy, rather than existing providers doing a bad job. Surge AI in particular seems to offer a pretty good and likely improving service. Furthermore, many companies have in-house data-gathering teams or are in the process of building them.

Hence we have decided to not further pursue this idea.

Other projects in the human data generation space may still be valuable, especially if the importance of human feedback in ML continues to increase, as we expect. This might include people specializing on human data as a career.

The types of factors that are most important for doing human dataset provision well include: high-skill contractors, fast iteration, and high bandwidth communication and shared understanding between the research team, the provider organisation and the contractors.

We are keen to hear other people’s thoughts, and would be happy to talk or to share more notes and thoughts with anyone interested in working on this idea or a similar one in the future.


Theory of Change

A major part of AI alignment research requires doing machine learning (ML) research, and ML research in turn requires training ML models. This involves expertise and execution ability in three broad categories: algorithms, compute, and data, the last of which is very neglected by EAs.

We expect training on data from human feedback to become an increasingly popular and very powerful tool in mainstream ML (see below). Furthermore, many proposals for alignment (for example: reinforcement learning from human feedback (RLHF) and variants like recursive reward modelling, iterated amplification, and safety via debate) would require lots of human interaction or datasets based on human-generated data.

While many services (most notably Surge) exist for finding labour to work on data generation for ML models, it seems plausible that an EA-aligned company could add significant value because:

  • Markets may not be efficient enough to fill small niches that are more important to alignment researchers than other customers; high-skill human data that requires very competent crowdworkers may be one such example. If alignment researchers can get it at all, it might be very expensive.
  • We have a better understanding of alignment research agendas, and this might help. This may allow us to make better-informed decisions on many implementation details with less handholding, thereby saving researchers time.
  • We would have a shared goal with our customers: reducing AI x-risk. Though profit motives already provide decent incentives to offer a good service, mission alignment helps avoid adversarial dynamics, increases trust, and reduces friction in collaboration.
  • An EA-led company may be more willing to make certain strategic moves that go against its profit incentives; e.g. investing heavily into detecting a model’s potential attempts to deceive the crowdworkers, even when it’s hard for outsiders to tell whether such monitoring efforts are sincere and effective (and thus customers may not be willing to pay for it). Given that crowdworkers might provide a reward signal, they could be a key target for deceptive AIs.

Therefore, there is a chance that an EA-led  human data service that abstracts out some subset of dataset-related problems (e.g. contractor finding, instruction writing/testing, UI and pipeline design/coding, experimentation to figure out best practices and accumulate that knowledge in one place) would:

  1. save the time of alignment researchers, letting them make more progress on alignment; and
  2. reduce the cost (in terms of time and annoying work) required to run alignment-relevant ML experiments, and therefore bring more of them below the bar at which it makes sense to run them, and thus increasing the number of such experiments that are run.

In the longer run, benefits of such an organisation might include:

  • There is some chance that we could simply outcompete existing ML data generation companies and be better even in the cases where they do provide a service; this is especially plausible for relatively niche services. In this scenario we’d be able to exert some marginal influence over the direction of the AI field, for example by only taking alignment-oriented customers. This would amount to differential development of safety over capabilities. Beyond only working with teams that prioritise safety, we could also pick among self-proclaimed “safety researchers”. It is common for proclaimed safety efforts to be accused of helping more with capabilities than alignment by other members of the community.
  • There are plausibly critical actions that might need to be taken for alignment, possibly quickly during “crunch-time”, that involve a major (in quality or scale) data-gathering project (or something like large-scale human-requiring interpretability work, that makes use of similar assets, like a large contractor pool). At such a time it might be very valuable to have an organisation committed to x-risk minimisation with the competence to carry out any such project.

Furthermore, if future AIs will learn human values from human feedback, then higher data quality will be equivalent to a training signal that points more accurately at human values. In other words, higher quality data may directly help with outer alignment (though we're not claiming that it could realistically solve it on its own). In discussions, it seemed that Matt gave this argument slightly more weight than Rudolf.

While these points are potentially high-impact, we think that there are significant problems with starting an organisation mainly to build capacity to be useful only at some hypothetical future moment. In particular, we think it is hard to know exactly what sort of capacity to build (and the size of the target in type-of-capacity space might be quite small), and there would be little feedback that the organisation could improve or course-correct based on. 

More generally, both of us believe that EA is right now partly bottlenecked by people who can start and scale high-impact organisations, which is a key reason why we’re considering entrepreneurship. This seems particularly likely given the large growth of the movement. 

What an org in this space may look like

Providing human datasets

The concept we most seriously considered was a for-profit that would specialise in meeting the specific needs of alignment researchers, probably by focusing on very high-skill human data. Since this niche is quite small, the company could offer a very custom-tailored service. At least for the first couple years, this would probably mean both of us having a detailed understanding of the research projects and motivations of our customers. That way, we could get a lot of small decisions right, without the researchers having to spend much time on it. We might be especially good at that compared to competitors, given our greater understanding of alignment.

Researching enhanced human feedback

An alternative we considered was founding a non-profit that would research how to enhance human feedback. See this post by Ajeya Cotra for some ideas on what this kind of research could look like. The central question is whether and how you can combine several weak training signals into a stronger more accurate one. If this succeeded, maybe (enhanced) human feedback could become a more accurate (and thereby marginally safer) signal to train models on.

We decided against this for a number of reasons:

  • Currently, neither of us has more research experience than an undergraduate research project.
  • We thought we could get a significant fraction of the benefits of this kind of research even if we did the for-profit version, and plausibly even more valuable expertise.
    • First of all, any particular experiment that funders would have liked to see, they could have paid us to do, although we freely admit that this is very different from someone pushing forward their own research agenda.
    • More importantly, we thought a lot of the most valuable expertise to be gained would come in the form of tacit knowledge and answers to concrete boring questions that are not best answered by doing “research” on them, but rather by iterating on them while trying to offer the best product (e.g. “Where do you find the best contractors?”, “How do you incentivize them?”, “What’s the best way to set up communication channels?”).
      • It is our impression that Ought pivoted away from doing abstract research on factored cognition and toward offering a valuable product for related reasons.
  • This topic seems plausibly especially tricky to research (though some people we’ve spoken to disagreed): 
    • At least some proposed such experiments would not involve ML models at all. We fear that this might make it especially easy to fool ourselves into thinking some experiment might eventually turn out to be useful when it won’t. More generally, the research would be pretty far removed from the end product (very high quality human feedback). In the for-profit case on the other hand, we could easily tell whether alignment teams were willing to pay for our services and iteratively improve. 

For-profit vs non-profit

We can imagine two basic funding models for this org: 

  • either we’re a nonprofit directly funded by EA donors and offering free or subsidized services to alignment teams;
  • or we’re a for-profit, paid by its customers (ie alignment teams). 

Either way, a lot of the money will ultimately come from EA donors (who fund alignment teams.)

The latter funding mechanism seems better; “customers paying money for a service” leads to the efficient allocation of resources by creating market structures. They have a clear incentive to spend the money well. On the other hand, “foundations deciding what services are free” is more reminiscent of planned economies and distorts markets. To a first approximation, funders should give alignment orgs as much money as they judge appropriate and then alignment orgs should exchange it for services as they see fit.

A further reason is that a non-profit is legally more complicated to set up, and imposes additional constraints on the organisation.

Should the company exclusively serve alignment researchers?

We also considered founding a company with the ambition to become a major player in the larger space of human data provision. It would by default serve anyone willing to pay us and working on something AGI-related, rather than just alignment researchers. Conditional on us being able to successfully build a big company, this would have the following upsides:

  • Plausibly one of the main benefits of founding a human data gathering organisation is to produce EAs and an EA org that have deep expertise in handling and producing high-skill human data in significant quantities. That might prove useful around “crunch time”, e.g. when some project aims to create competitive but safe AGI and needs this expertise. Serving the entire market could scale to a much larger company enabling us to gain expertise at higher scales.
  • Operating a large company would also come with some degree of market power. Any company with paying customers has some amount of leverage over them: first of all just because of switching costs, but also because the product it offers might be much better than the next-best alternative. This could allow us to make some demands, e.g. once we’re big and established, announce we’d only work with companies that follow certain best practices.

On the other hand, building a big successful company serving anyone willing to pay might come with some significant downsides as well.

  • First, and most straightforwardly, it is probably much harder than filling a small niche (just meeting the specific needs of alignment researchers), making us less likely to succeed. A large number of competitors exist and as described in this section, some of them (esp. Surge) seem pretty hard to beat. Since this is an already big and growing market, there is an additional efficient markets reason to assume this is true a priori.
  • Secondly, and perhaps more importantly, such a company might accelerate capabilities (more on this below).

Furthermore, it might make RLHF (Reinforcement Learning from Human Feedback) in particular more attractive. Depending on one’s opinions about RLHF and how it compares to other realistic alternatives, one might consider this a strong up- or downside. 


The main reason companies fail is that they build a product that customers don’t want. For for-profits, the signal is very clear: either customers care enough to be willing to pay hard cash for the product/service, or they don’t. For non-profits, the signal is less clear, and therefore nonprofits can easily stick around in an undead state, something that is an even worse outcome than the quick death of a for-profit because of resource (mis)allocation and opportunity costs. As discussed, it is not obvious which structure we should adopt for this organisation, though for-profit may be a better choice on balance. However, in all cases it is clear that the organisation needs to solve a concrete problem or provide clear value to exist and be worth existing. This does not mean that the value proposition needs to be certain; we would be happy to take a high-risk, high-reward bet, and generally support hits-based approaches to impact both in general and for ourselves.

An organisation is unlikely to do something useful to its customers without being very focused on customer needs, and ideally having tight feedback cycles. 

The shortest feedback loops are when you’re making a consumer software product where you can prototype quickly (including with mockups), and watch and talk to users as they use the core features, and then see if the user actually buys the product on the spot. A datasets service differs from this ideal feedback mode in a number of ways:

  1. The product is a labour-intensive process, which means the user cannot quickly use the core features and we cannot quickly simulate them.
  2. The actual service requires either a contractor pool or (potentially at the start) the two of us spending a number of hours per request generating data.
  3. There is significant friction to getting users to use the core feature (providing a dataset), since it requires specification of a dataset from a user, which takes time and effort.

Therefore, we relied on customer interviews with prospective customers. The goal of these interviews was to talk to alignment researchers who work with data, and figure out if external help with their dataset projects would be of major use to them.

Our approach to customer interviews was mostly based on the book The Mom Test, which is named after the idea that your customer interview questions should be concrete and factual enough that even someone as biased as your own mom shouldn’t be able to give you a false signal about whether the idea is actually good. Key lessons emphasised by The Mom Test include emphasising:

  • factual questions about the past over hypothetical questions for the future;
    • In particular, questions about concrete past and current efforts spent solving a problem rather than questions about current or future wishes for solving a problem
  • questions that get at something concrete (e.g. numbers); and
  • questions that prompt the customer to give information about their problems and priorities without prompting them with a solution.

We wanted to avoid the failure mode where lots of people tell us something is important and valuable in the abstract, without anyone actually needing it themselves.

We prepared a set of default questions that roughly divided into:

  1. A general starting question prompting the alignment researcher to describe the biggest pain points and bottlenecks they face in their work, without us mentioning human data.
  2. Various questions about their past and current dataset-related work, including what types of problems they encounter with datasets, how much of their time these problems take, and steps they took to address these problems.
  3. Various questions on their past experiences using human data providers like Surge, Scale, or Upwork, and specifically about any things they were unable to accomplish because of problems with such services.
  4. In some cases, more general questions about their views on where the bottlenecks for solving alignment are, views on the importance of human data or tractability of different data-related proposals, etc. 
  5. What we should’ve asked but didn’t, and who else we should talk to.

Point 4 represents the fact that in addition to being potential customers, alignment researchers also doubled as domain experts. The weight given to the questions described in point 4 varied a lot, though in general if someone was both a potential customer and a source of data-demand-relevant alignment takes, we prioritised the customer interview questions.

In practice, we found it easy to arrange meetings with alignment researchers; they generally seemed willing to talk to people who wanted input on their alignment-relevant idea. We did customer interviews with around 15 alignment researchers, and had second meetings with a few. For each meeting, we prepared beforehand a set of questions tweaked to the particular person we were meeting with, which sometimes involved digging into papers published by alignment researchers on datasets or dataset-relevant topics (Sam Bowman in particular has worked on a lot of data-relevant papers). Though the customer interviews were by far the most important way of getting information on our cruxes, we found the literature reviews we carried out to be useful too. We are happy to share the notes from the literature reviews we carried out; please reach out if this would be helpful to you.

Though we prepared a set of questions beforehand, in many meetings - including often the most important or successful ones - we often ended up going off script fairly quickly.

Something we found very useful was that, since there were two of us, we could split the tasks during the meeting into two roles (alternating between meetings):

  1. One person who does most of the talking, and makes sure to be focused on the thread of the conversation.
  2. One person who mostly focuses on note-taking, but also pipes in if they think of an important question to ask or want to ask for clarification.

Key crux: demand looks questionable, Surge seems pretty good

Common startup advice is to make sure you have identified a very strong signal of demand before you start building stuff. That should look something like someone telling you that the thing you’re working on is one of their biggest bottlenecks and that they can’t wait to pay you asap so you solve this problem for them. “Nice to have” doesn’t cut it. This is in part because working with young startups is inherently risky, so you need to make up for that by solving one of their most important problems.

In brief, we don’t think this level of very strong demand currently exists, though there were some weaker signals that looked somewhat promising. There are many existing startups that offer human feedback already. Surge AI in particular was brought up by many people we talked to and seems to offer quite a decent service that would be hard to beat.

Details about Surge

Surge is a US-based company that offers a service very similar to what we had in mind, though they are not focused on alignment researchers exclusively. They build data-labelling and generation tools and have a workforce of crowdworkers.

They’ve worked with Redwood and the OpenAI safety team, both of which had moderately good experiences with them. More recently, Ethan Perez’s team have worked with Surge too; he seems to be very satisfied based on this Twitter thread.


Collaboration with Redwood

Surge has worked with Redwood Research on their paper about adversarial training. This is one of three case studies on Surge’s website, so we assume it’s among the most interesting projects they’ve done so far. The crowdworkers were tasked with coming up with prompts that would cause the model to output text in which someone got injured. Furthermore, crowdworkers also classified whether someone got injured in a given piece of text.

One person from Redwood commented that doing better than Surge seemed possible to them with “probably significant value to be created”, but “not an easy task”. They thought our main edge would have to be that we’d specialise on fuzzy and complex tasks needed for alignment; Surge apparently did quite well with those, but still with some room for improvement. A better understanding of alignment might lower chances of miscommunication. Overall, Redwood seems quite happy with the service they received.

Initially, Surge’s iteration cycle was apparently quite slow, but this improved over time and was “pretty good” toward the end.

Redwood told us they were quite likely to use human data again by the end of the year and more generally in the future, though they had substantial uncertainty around this. Their experience in working with human feedback overall was somewhat painful as we understood it.  This is part of the reason they’re uncertain about how much human feedback they will use for future experiments, even though it’s quite a powerful tool. However, they estimated that friction in working with human feedback was mostly caused by inherent reasons (humans are inevitably slower and messier than code), rather than Surge being insufficiently competent. 

Collaboration with OpenAI

OpenAI have worked with Surge in the context of their WebGPT paper. In that paper, OpenAI fine-tuned their language model GPT-3 to answer long-form questions. The model is given access to the web, where it can search and navigate in a text-based environment. It’s first trained with imitation learning and then optimised with human feedback. 

Crowdworkers provided “demonstrations”, where they answered questions by browsing the web. They also provided “comparisons”, where they indicated which of two answers to the same question they liked better.

People from OpenAI said they had used Surge mostly for sourcing the contractors, while doing most of the project management, including building the interfaces, in-house. They were generally pretty happy with the service from Surge, though all of them did mention shortcomings.

One of the problems they told us about was that it was hard to get access to highly competent crowdworkers for consistent amounts of time. Relatedly, it often turned out that a very small fraction of crowdworkers would provide a large majority of the total data. 

More generally, they wished there had been someone at Surge that understood their project better. Also, it might have been somewhat better if there had been more people with greater experience in ML, such that they could have more effectively anticipated OpenAI’s preferences — e.g. predict accurately what examples might be interesting to researchers when doing quality evaluation. However, organisational barriers and insufficient communication were probably larger bottlenecks than ML knowledge. At least one person from OpenAI strongly expressed a desire for a service that understood their motives well and took as much off their plate as possible in terms of hiring and firing people, building the interfaces, doing quality checks and summarising findings etc. It is unclear to us to what extent Surge could have offered these things if OpenAI hadn’t chosen to do a lot of these things in-house. One researcher suggested that communicating their ideas reliably was often more work than just doing it themselves. As it was, they felt that marginal quality improvement required significant time investment on their own part, i.e. could not be solved with money alone. 

Notably, one person from OpenAI estimated that about 60% of the WebGPT team’s efforts were spent on various aspects of data collection. They also said that this figure didn’t change much after weighting for talent, though in the future they expect junior people to take on more disproportionate shares of this workload.

Finally, one minor complaint that was mentioned was the lack of transparency about contractor compensation. 

How mission-aligned is Surge?

Surge highlight their collaboration with Redwood on their website as one of three case studies. In their blog post about their collaboration with Anthropic, the first sentence reads: “In many ways, alignment – getting models to align themselves with what we want, not what they think we want – is one of the fundamental problems of AI.” 

On the one hand, they describe alignment as one of the fundamental problems of AI, which could indicate that they intrinsically cared about alignment. However, they have a big commercial incentive to say this. Note that many people would consider their half-sentence definition of alignment to be wrong (a model might know what we want, but still do something else).

We suspect that the heads of Surge have at least vaguepositive dispositions towards alignment. They definitely seem eager to work with alignment researchers, which might well be more important. We think it’s mostly fine if they are not maximally intrinsically driven, though mission alignment does add value as mentioned above.

Other competitors

We see Surge as the most direct competitor and have researched them by far in the most detail. But besides Surge, there are a large number of other companies offering similar services. 

First, and most obviously, Amazon Mechanical Turk offers a very low quality version of this service and is very large. Upwork specialises in sourcing humans for various tasks, without building interfaces. ScaleAI is a startup with a $7B valuation --- they augment human feedback with various automated tools. OpenAI have worked with them. Other companies in this broad space include Hybrid (which Sam Bowman’s lab has worked with) and Invisible (who have worked with OpenAI). There are many more that we haven’t listed here.

In addition, some labs have in-house teams for data gathering (see here for more).

Data providers used by other labs

Ethan Perez’s and Sam Bowman’s labs at NYU/Anthropic have historically often built their own interfaces while using contractors from Upwork or undergrads, but they have been trialing Surge over the summer and seem likely to stick with them if they have a good experience. Judging from the Twitter thread linked above and asking Jérémy Scheurer (who works on the team and built the pre-Surge data pipeline) how they’ve found Surge so far, Surge is doing a good job. 

Google has an internal team that provides a similar service, though DeepMind have used at least one external provider as well. We expect that it would be quite hard to get DeepMind to work with us, at least until we would be somewhat more established. 

Generally, we get the impression that most people are quite happy with Surge. It’s worth also considering that it’s a young company that’s likely improving its service over time. We’ve heard that Surge iterates quickly, e.g. by shipping simple feature requests in two days. It’s possible that some of the problems listed above may no longer apply by now or in a few months.

Good signs for demand

One researcher we talked to said that there were lots of projects their team didn’t do, because gathering human feedback of sufficient quality was infeasible. 

One of the examples this researcher gave was human feedback on code quality. This is implausible to do, because the time of software engineers is just too expensive. That problem is hard for a new org to solve. 

Another example they gave seemed like it might be more feasible: for things like RLHF, they often choose to do pairwise comparisons between examples or multi-preferences. Ideally, they would want to get ratings, e.g. on a scale from 1 to 10. But they thought they didn’t trust the reliability of their raters enough to do this. 

More generally, this researcher thought there were lots of examples where if they could copy any person on their team a hundred times to provide high-skill data, they could do many experiments that they currently can’t. 

They also said that their team would be willing to pay ~3x of what they were paying currently to receive much higher-quality feedback.

Multiple other researchers we talked to expressed vaguely similar sentiments, though none quite as strong.

However, it’s notable that in this particular case, the researcher hadn’t worked with Surge yet. 

The same researcher also told us about a recent project where they had spent a month on things like creating quality assurance examples, screening raters, tweaking instructions etc. They thought this could probably have been reduced a lot by an external org, maybe to as little as one day. Again, we think Surge may be able to get them a decent part of the way there.

Labs we could have worked with

We ended up finding three projects that we could have potentially worked on:

  • A collaboration with Ought --- they spend about 15 hours a week on data-gathering and would have been happy to outsource that to us. If it had gone well, they might also have done more data-gathering in the longterm (since friction is lower if it doesn’t require staff time). We decided not to go ahead with this project since we weren’t optimistic enough about demand from other labs being bigger once we had established competence with Ought and the project itself didn’t seem high upside enough. 
  • Attempt to get the Visible Thoughts bounty by MIRI. We decided against this for a number of reasons. See more of our thinking about Visible Thoughts below.
  • Potentially a collaboration with Owain Evans on curated datasets for alignment.

We think the alignment community is currently relatively tight-knit. e.g. researchers often knew about other alignment teams’ experiences with Surge from conversations they had had with them. Hence, we were relatively optimistic that conditional on there being significant demand for this kind of service, doing a good job on one of the projects above would quickly lead to more opportunities.

Visible Thoughts

In November 2021, MIRI announced the Visible Thoughts (VT) project bounty. In many ways VT would be a good starting project for an alignment-oriented dataset provider, in particular because the bounty is large (up to $1.2M) and because it is ambitious enough that executing on it would provide a strong learning signal to us and a credible signal to other organisations we might want to work with. However, on closer examination of VT, we came to the conclusion that it is not worth it for us to work on it.

The idea of VT is to collect a dataset of 100 runs of fiction of a particular type (“dungeon runs”, an interactive text-based genre where one party, called the “dungeon master” and often an AI, offers descriptions of what is happening, and the other responds in natural language with what actions they want to take), annotated with a transcript of some of the key verbal thoughts that the dungeon master might be thinking as they decide what happens in the story world. MIRI hopes that this would be useful for training AI systems that make their thought processes legible and modifiable.

In particular, a notable feature of the VT bounty is the extreme run lengths that it asks for: to the tune of 300 000 words for each of the runs (for perspective, this is the length of A Game of Thrones, and longer than the first three Harry Potter books combined). A VT run is much less work than a comparable-length book - the equivalent of a rough unpolished first-draft (with some quality checks) would likely be sufficient - but producing one such run would still probably require at least on the order of 3 months of sequential work time from an author. We expect the pool of people willing to write such a story for 3 months is significantly smaller than the pool of people who would be willing to complete, say, a 30 000 word run, and that the high sequential time cost increases the amount of time required to generate the same number of total words. We also appear to have different ideas on how easy it is to fit a coherent story, for the relevant definition of coherent, into a given number of words. Note that to compare VT word counts to lengths of standard fiction without the written-out thoughts from the author, the VT word count should be reduced by a factor of 5-6.

Concerns about the length are raised in the comments section, to which Eliezer Yudkowksy responded. His first point, that longer is easier to write per step, may be true, especially as we also learned (by email with Nate Soares and Aurelien Cabanillas) that in MIRI’s experience “authors that are good at producing high quality steps are also the ones who don't mind producing many steps”. In particular because of that practical experience, we think it is possible we overestimated the logistical problems caused by the length. MIRI also said they would likely accept shorter runs too if they satisfied their other criteria.

In a brief informal conversation with Rudolf during EAG SF, Eliezer emphasised the long-range coherence point in particular. However, they did not come to a shared understanding of what type of “long-range coherence” is meant.

Even more than these considerations, we are sceptical about the vague plans for what to do given a VT dataset. A recurring theme from talking to alignment researchers who work with datasets was that inventing and creating a good dataset is surprisingly hard, and generally involves having a clear goal of what you’re going to use the dataset for. It is possible the key here is the difference in our priors for how likely a dataset idea is to be useful.

In addition, we have significant concerns about undertaking a major project based on a bounty whose only criterion is the judgement of one person (Eliezer Yudkowsky), and undertaking such a large project as our first project.

Other cruxy considerations

Could we make a profit / get funding? 

One researcher from OpenAI told us he thought it would be hard to imagine an EA data-gathering company making a profit because costs for individual projects would always be quite high (requiring several full-time staff), and total demand was probably not all that big.

In terms of funding, both of us were able to spend time on this project because of grants from regrantors in the Future Fund regrantor program. Based on conversations with regrantors, we believe we could’ve gotten funding to carry out an initial project if we had so chosen.

Will human feedback become a much bigger deal? Is this a very quickly growing industry?

Our best guess is yes. For example, see this post by Ajeya Cotra which outlines how we could get to TAI by training on Human Feedback on Diverse Tasks (HFDT). 

She writes: “HFDT is not the only approach to developing transformative AI, and it may not work at all. But I take it very seriously, and I’m aware of increasingly many executives and ML researchers at AI companies who believe something within this space could work soon.”

In addition, we have also had discussions with at least one other senior AI safety researcher whom we respect and who thought human feedback was currently irrationally neglected by mainstream ML; they expected it to become much more wide-spread and to be a very powerful tool.

If that’s right, then providing human feedback will likely become important and economically valuable. 

This matters, because operating a new company in a growing industry is generally much easier and more likely to be successful. We think this is true even if profit isn’t the main objective.

Would we be accelerating capabilities?

Our main idea was to found a company (or possibly non-profit) that served alignment researchers exclusively. That could accelerate alignment differentially. 

One problem is that it’s not clear where to draw this boundary. Some alignment researchers definitely think that other people who would also consider themselves to be alignment researchers are effectively doing capabilities work. This is particularly true of RLHF.

One mechanism worth taking seriously if we worked with big AI labs to make their models more aligned by providing higher quality data is that the models might merely appear surface-level aligned. “Make the data higher quality” might be a technique that scales poorly as capabilities ramp up. So it risks creating a false sense of security. It would also clearly improve the usefulness of current-day models and hence, it risks increasing investment levels too.

We don’t currently think the risk of surface-level alignment is big enough to outweigh the benefits. In general, we think that a good first-order heuristic that helps the field stay grounded in reality would be that whatever improves alignment in current models is useful to explore further and invest resources into. It seems like a good prior that such things would also be valuable in the future (even if it’s possible that new additional problems may arise, or such efforts aren’t on the path to a future alignment solution). See Nate Soares’ post about sharp left turns to get a contradicting view on this. 

Is it more natural for this work to be done in-house in the longterm? Especially at big labs/companies.

We expect that human data gathering is likely to become very important and that it benefits from understanding the relevant research agenda well. So maybe big companies will want to do this internally, instead of relying on third-party suppliers? 

That seems quite plausible to us and to some extent it’s happening already. Our understanding is that Anthropic is hiring an internal team to do human data gathering. DeepMind has access to Google’s crowdworker service. OpenAI have worked with multiple companies, but they also have at least one in-house specialist for this kind of work and are advertising multiple further jobs on the human data team here. They’re definitely considering moving more of this work in-house, but it’s unclear to us to what extent that’s going to happen and we have received somewhat contradicting signals regarding OpenAI safety team members’ preferences on this.

So a new EA org would face stiff competition, not only from other external providers, but also from within companies.

Of course, smaller labs will most likely always have to rely on external providers. Hence, another cruxy consideration is how much small labs matter. Our intuition is that they matter much less than bigger labs (since the latter have access to the best and biggest models).

Creating redundancy of supply and competition

Even if existing companies are doing a pretty good job at serving the needs of alignment researchers, there’s still some value in founding a competitor. 

First, competition is good. Founding a competitor puts pressure on existing providers to keep service quality high, work on improving their products, and margins low. Ironically, part of the value of founding this company would thus flow through getting existing companies to try harder to offer the best product.

Second, it creates some redundancy. What if Surge pivots? What if their leadership changes or they become less useful for some other reason? In those worlds it might be especially useful to have a “back-up” company.

Both of these points have been mentioned to us as arguments in favour of founding this org. We agree that these effects are real and likely point in favour of founding the org. However, we don’t think these factors carry very significant weight relative to our opportunity costs, especially given that there are already many start-ups working in this space. 

Adding a marginal competitor can only affect a company’s incentives so much. And in the worlds where we’d be most successful such that all alignment researchers were working with us, we might cause Surge and others to pivot away from alignment researchers, instead of getting them to try harder. 

The redundancy argument only applies in worlds in which the best provider ceases to exist; maybe that’s 10% likely. And then the next best alternative is likely not all that bad. Competitors are plentiful and even doing it in-house is feasible. Hence, it seems unlikely to us that the expected benefit here is very large after factoring in the low probability of the best provider disappearing.

Other lessons

Lessons on human data gathering

In the process of talking to lots of experts about their experiences in working with human data, we learned many general lessons about data gathering. This section presents some of those lessons, in roughly decreasing order of importance.


Many people emphasized to us that working with human data rarely looks like having a clean pipeline from requirements design to instruction writing to contractor finding to finished product. Rather, it more often involves a lot of iteration and testing, especially regarding what sort of data the contractors actually produce. While some of this iteration may be removed by having better contractors and better knowledge of good instruction-writing, the researchers generally view the iteration as a key part of the research process, and therefore prize 

  • ease of iteration (especially time to get back with a new batch of data based on updated instructions); and
  • high-bandwidth communication with the contractors and whoever is writing the instructions (often both are done by the researchers themselves). 

This last point holds to the point that it is somewhat questionable whether an external provider (rather than e.g. a new team member deeply enmeshed in the context of the research project) could even be a good fit for this need.

The ideal pool of contractors

All of the following features matter in a pool of contractors:

  • Competence, carefulness, intelligence, etc. (sometimes expertise). It is often ideal if the contractors understand the experiment.
  • Number of contractors
  • Quick availability and therefore low latency for fulfilling requests
  • Consistent availability (ideally full-time)
  • Even distribution of contributions across contractors (ie it shouldn’t be the case that 20% of the contractors provide 80% of the examples). 

Quality often beats quantity for alignment research

Many researchers told us that high-quality, high-skill data is usually more important and more of a bottleneck than just a high quantity of data. Some of the types of projects where current human data generation methods are most obviously deficient are cases where a dataset would need epistemically-competent people to make subtle judgments, e.g. of the form “how true is this statement?” or “how well-constructed was this study?” As an indication of reference classes where the necessary epistemic level exists, the researcher mentioned subject-matter experts in their domain, LessWrong posters, and EAs.

A typical data gathering project needs UX-design, Ops, ML, and data science expertise 

These specialists might respectively focus on the following:

  • Designing the interfaces that crowdworkers interact with. (UX-expert/front-end web developer)
  • Managing all operations, including hiring, paying, managing, and firing contractors, communicating with them and the researchers etc. (ops expert)
  • Helping the team make informed decisions about the details of the experimental design, while minimizing time costs for the customer. The people we spoke to usually emphasized ML-expertise more than alignment expertise. (ML-expert)
  • Meta-analysis of the data. e.g. inter-rater agreement, the distribution of how much each contractor contributed, demographics, noticing any other curious aspects of the data, etc. (data scientist)

It is possible that someone in a team could have expertise in more than one of these areas, but generally this means a typical project will involve at least three people.

Crowdworkers do not have very attractive jobs

Usually the crowdworkers are employed as contractors. This means their jobs are inherently not maximally attractive; they probably don’t offer much in the way of healthcare, employment benefits, job security, status etc. The main way that these jobs are made more attractive is through offering higher hourly rates.

If very high quality on high-skill data is going to become essential for alignment, it may be worth considering changing this, to attract more talented people. 

However, we expect that it might be inherently very hard to offer permanent positions for this kind of work, since demand is likely variable and since different people may be valuable for different projects. This is especially true for a small organisation. 

What does the typical crowdworker look like?

This varies a lot between projects and providers.

The cheapest are non-native English speakers who live outside of the US.

Some platforms, including Surge, offer the option to filter crowdworkers for things like being native English-speakers, expertise as a software engineer, background in finance, etc.

Bottlenecks in alignment

When asked to name the factors most holding back their progress on alignment, many alignment researchers mentioned talent bottlenecks. 

The most common talent bottleneck seemed to be in competent ML-knowledgeable people. Some people mentioned the additional desire for these to understand and care about alignment. (Not coincidentally, Matt’s next project is likely going to be about skilling people up in ML).

There were also several comments about things like good web development experience being important. For example, many data collection projects involve creating a user interface at some point, and in practice this is often handled by ML-specialised junior people at the lab, who can, with some effort and given their programming background, cobble together some type of website - often using different frameworks and libraries than the next person knows (or wants to use). (When asked about why they don’t hire freelance programmers, one researcher commented that a key feature they’d want is the same person working for them for a year or two, so that there’s an established working relationship, clear quality assurances, and continuity with the choice of technical stack.)


After having looked into this project idea for about a month, we have decided not to found a human data gathering organisation for now. 

This is mostly because demand for an external provider seems insufficient, as outlined in this section. No lab gave a clear signal that gathering human data was a key bottleneck for them, where they would have been willing to go to significant lengths to fix it urgently (especially not the ones that had tried Surge). 

We expect that many labs would want to stick with their current providers, Surge in particular, or their in-house team, bar exceptional success on our part (even then, we’d only provide so much marginal value over those alternatives).

Though we did find some opportunities for potential initial projects after looking for a month, we are hesitant about how far this company would be expected to scale. One of the main draws (from an impact perspective) of founding an organisation is that you can potentially achieve very high counterfactual impact by creating an organisation that scales to a large size and does lots of high-impact work over its existence. The absence of a plausible pathway to really outstanding outcomes from starting this organisation is a lot of what deters us.

In a world where we’re more successful than expected (say 90th to 95th percentile), we could imagine that in five years from now, we’d have a team of about ten good people. This team may be working with a handful of moderately big projects (about as big as WebGPT), and provide non-trivial marginal value over the next-best alternative to each one of them. Maybe one of these projects would not have been carried out without us.

A median outcome might mean failing to make great hires and remaining relatively small and insignificant: on the scale of doing projects like the ones we’ve identified above, enough to keep us busy throughout the year and provide some value, but with little scaling. In that case we would probably quit the project at some point.

This distribution doesn’t seem good enough to justify our opportunity cost (which includes other entrepreneurial projects or technical work among other things). Thus we have decided not to pursue this project any further for now.

We think this was a good idea to invest effort in pursuing, and we think we made the right call in choosing to investigate it. Both of us are open to, and also quite likely to, evaluate other EA-relevant entrepreneurial project ideas in the future.

Other relevant human data-gathering work

However, the assumption that high-quality high-skill human feedback is important and neglected by EAs has not been falsified

It is still plausible to us that EAs should consider career paths that focus on building expertise at data-gathering; just probably not by founding a new company. In the short run, this could look like

  • Contributing to in-house data-gathering teams (eg Anthropic, OpenAI, etc.)
  • Joining Surge or other data-gathering startups.

As we discussed above, the types of skills that seem most relevant for working in a human data generation role include: data science experience and in particular experience with natural languaga data or social science data and experiment design, front-end web development, ops and management skills, and some understanding of machine learning and alignment. 80,000 Hours recently wrote a profile which you can find here.

Of course, in the short term, this career path will be especially impactful if one’s efforts are focussed on helping alignment researchers. But if it’s true that human feedback will prove a very powerful tool for ML, then people with such expertise may become increasingly valuable going forward, such that it could easily be worth skilling up at a non-safety-focused org. 

We think joining Surge may be a particularly great opportunity. It is common advice that joining young, rapidly growing start-ups with good execution is great for building experience; early employees can often get a lot of responsibility early on. See e.g. this post by Bill Zito.

One of the hardest parts about that seems to be identifying promising startups. After talking to many of their customers, we have built reasonable confidence that Surge holds significant promise. They seem to execute well, in a space which we expect to grow. In addition to building career capital, there is clear value in helping Surge serve alignment researchers as well as possible.

From Surge’s perspective, we think they could greatly benefit from hiring EAs, who are tuned in to the AI safety scene, which we would guess represents a significant fraction of their customers. 

One senior alignment researcher told us explicitly that they would be interested in hiring people who had worked in a senior role at Surge.

Next steps for us

Matt is planning to run a bootcamp that will allow EAs to upskill in ML engineering. I'll be doing a computer science master’s at Cambridge from October to June.


AI risk intro 2: solving the problem

 This post was a joint effort with Callum McDougall.

8.2k words (~25min) 

This marks the second half of our overview of the AI alignment problem. In the first half, we outlined the case for misaligned AI as a significant risk to humanity, first by looking at past progress in machine learning and extrapolating to what the future could bring, and second by discussing the theoretical arguments which underpin many of these concerns. In this second half, we focus on possible solutions to the alignment problem that people are currently working on. We will paint a picture of the current field of technical AI alignment, explaining where the major organisations fit into the larger picture and what the theory of change behind their work is. Finally, we will conclude the sequence with a call to action, by discussing the case for working on AI alignment, and some suggestions on how you can get started.

Note - for people with more context about the field (e.g. have done AGISF) we expect Thomas Larsen's post to be a much better summary, and this post might be better if you are looking for something brief. Our intended audience is someone relatively unfamiliar with the AI safety field, and is looking for a taste of the kinds of problems which are studied in the field and the solution approaches taken. We also don't expect this sampling to be representative of the number of people working on each problem - again, see Thomas' post for something which accomplishes this.

Introduction: A Pre-Paradigmatic Field

Definition (pre-paradigmatic): a science at an early stage of development, before it has established a consensus about the true nature of the subject matter and how to approach it.

AI alignment is a strange field. Unlike other fields which study potential risks to the future of humanity (e.g. nuclear war or climate change), there is almost no precedent for the kinds of risks we care about. Additionally, because of the nature of the threat, failing to get alignment right on the first try might be fatal. As Paul Christiano (a well-known AI safety researcher) recently wrote:

Humanity usually solves technical problems by iterating and fixing failures; we often resolve tough methodological disagreements very slowly by seeing what actually works and having our failures thrown in our face. But it will probably be possible to build valuable AI products without solving alignment, and so reality won’t “force us” to solve alignment until it’s too late. This seems like a case where we will have to be unusually reliant on careful reasoning rather than empirical feedback loops for some of the highest-level questions.

For these reasons, the field of AI alignment lacks a consensus on how the problem should be tackled, or what the most important parts of the problem even are. This is why there is a lot of variety in the approaches we present in this post.

Decomposing the research landscape

An image generated with OpenAI's DALL-E 2 based on the prompt: sorting papers and books in a majestic gothic library. All other images like this in this post are also AI-generated, from the text in the caption.

There are lots of different ways you could divide up the space of approaches to solving the problem of aligning advanced AI. For instance, you could go through the history of the field and identify different movements and paradigms. Or you could place the work on a spectrum from highly theoretical maths/philosophy-type research, to highly empirical research working with cutting-edge deep learning models.

However, the most useful decomposition would be one that explains why the people who work on it believe that it will help solve the problem of AI alignment. 

For that reason, we’ll mostly be using the decomposition from Neel Nanda’s “A Bird’s Eye View” post. The motivation behind this decomposition is to answer the high-level question of “what is needed for AGI to go well?”. The six broad classes of approaches we talk about are:

  1. Addressing threat models 
    We have a specific threat model in mind for how AGI might result in a very bad future for humanity, and focus our work on things we expect to help address the threat model.
  2. Agendas to build safe AGI 
    Let’s make specific plans for how to actually build safe AGI, and then try to test, implement, and understand the limitations of these plans. The emphasis is on understanding how to build AGI safely, rather than trying to do it as fast as possible.
  3. Robustly good approaches 
    In the long-run AGI will clearly be important, but we're highly uncertain about how we'll get there and what, exactly, could go wrong. So let's do work that seems good in many possible scenarios, and doesn’t rely on having a specific story in mind.
  4. Deconfusion
    Reasoning about how to align AGI involves reasoning concepts like intelligence, values, and optimisers and we’re pretty confused about what these even mean. This means any work we do now is plausibly not helpful and definitely not reliable. As such, our priority should be doing some conceptual work on how to think about these concepts and what we’re aiming for, and trying to become less confused.
  5. AI governance
    In addition to solving the technical alignment problem, there’s the question of what policies we need to minimise risk from advanced AI systems.
  6. Field-building
    One of the most important ways we can make AI go well is by increasing the number of capable researchers doing alignment research. 

It’s worth noting that there is a lot of overlap between these sections. For instance, interpretability research is a great example of a robustly good approach, but it can also be done with a specific threat model in mind.

Throughout this section, we will also give small vignettes of organisations or initiatives which support AI alignment research in some form. This won’t be a full picture of all approaches or organisations, instead hopefully it will serve to sketch a picture of what work in AI alignment actually looks like.

Addressing threat models

We have a specific threat model in mind for how AGI might result in a very bad future for humanity, and focus our work on things we expect to help address the threat model.

A key high-level intuition here is that having a specific threat model in mind for how AI might go badly for humanity can help keep you focused on certain hard parts of the problem. One technique that can be useful here is a version of back-casting: we start from future problems with advanced AI systems in our current model, reason about what kinds of things might solve these problems, then try and build versions of these solutions today and test them out on current problems.

This can be seen in contrast to the approach of simply trying to fix current problems with AI systems, which might fail to connect up with the hardest parts of AI alignment.

Example 1: Superintelligent utility maximisers, and quantilizers

superintelligent artificial intelligence, making choices, digital art, artstation

The superintelligent utility maximiser is the oldest threat model studied by the AI alignment field. It was discussed at length by Nick Bostrom in his book Superintelligence. It assumes that we will create an AGI much more intelligent than humans, and that it will be trying to achieve some particular goal (measured by the expected value of some utility function). The problem with this is that attempts to maximise the value of some goal which isn’t perfectly aligned with what humans want can lead to some very bad outcomes. One formalism which was proposed to address this problem is Jessica Taylor’s quantilizers. It is quite maths-heavy so we won’t discuss all the details here, but the basic idea is that rather than using the expected utility maximisation framework for agents, we mix expected utility maximisation with human imitation in a clever way (to be more precise, you sample from a prior distribution which represents the actions a human would be likely to take in this scenario). The resulting agent wouldn’t take catastrophic actions because part of its decision-making comes from imitating what it thinks humans would do, but it would also be able to use the expected utility maximisation to go beyond human imitation, and do things we are incapable of (which is presumably the reason we would want to build it in the first place!). However, the drawback with theoretical approaches like this is that they often bake in too many assumptions or rely on too many variables to be useful in practice. In this case, how we define the set of reasonable actions a human might perform is an important unspecified part of this framework, and so more research is required to see if the quantiliszers framework can address these problems.

Example 2: Inner misalignment

robot jumping over boxes to collect a coin, videogame, digital art, artstation

We’ve discussed inner misalignment in a previous section. This concept was first explicitly named in a paper called Risks from Learned Optimisation in Advanced ML Systems, published in 2019. This paper defined the concept and suggested some conditions which might make it more likely to happen, but the truth is that a lot of this is still just conjecture, and there are many things we don’t yet know about how unlikely this kind of misalignment is, or what we can do about it. The CoinRun example discussed earlier (and the Objective Robustness paper) came from an independent research team in 2021. This study was the first known example of inner misalignment in an AI system, showing that it was at least a theoretical possibility. They also tested certain interpretability tools on the CoinRun agent, to see whether it was possible to discover when the agent had a goal different to the one intended by the programmers. For more on interpretability, see later sections.

Building safe AGI

Let’s make specific plans for how to actually build safe AGI, and then try to test, implement, and understand the limitations of these plans. The emphasis is on understanding how to build AGI safely, rather than trying to do it as fast as possible.

At some point we’re going to build an AGI. Companies are already racing to do it. We better make sure that there exist some blueprints for a safe AGI (and that they’re used) by the time we get to that point.

Perhaps the master list of safe AGI proposals is Evan Hubinger’s An Overview of 11 Proposals for Building Safe Advanced AI

Example 1: Iterated Distillation and Amplification (IDA)

artists depection of a robot dreaming up multiple copies of itself, cascading tree, delegating, digital art, trending on artstation

“Iterated Distillation and Amplification” (IDA) is an imposing name, but the core intuition is simple. One of the ways in which an individual human can achieve more things is by delegating tasks to others. In turn, the assistants that tasks are delegated to can be expected to become more competent at the task.

In IDA, an AI plays the role of the assistant. “Distillation” refers to the abilities of the human being “distilled” into the AI through training, and “amplification” refers to the human becoming more capable as they can call on more and more powerful AI assistants to help them.

A setup to train an IDA personal assistant might go like this:

  1. You have a human, say Hannah, who knows how to carry out the tasks of a personal assistant.
  2. You have an ML model - call it Martin - that starts out knowing very little (perhaps nothing at all, or perhaps it’s a pre-trained language model so it knows how to read and write English but not much else).
  3. Hannah needs to find the answer to some questions, and she can invoke multiple copies of Martin to help her. Since Martin is quite useless at this stage, Hannah has to do even simple tasks herself, like writing routine emails. Using some interface legible to Martin, she breaks the email-writing task into subtasks like “find email address of Hu M. Anderson”, “select greeting”, “check project status”, “mention project status”, and so on.
  4. From seeing enough examples of Hannah’s own answers to the sub-questions, Martin’s training loop gradually trains it to be able to answer first the simpler sub-tasks - (address is “humanderson@humanmail.com”, greeting is “Salutations, Human Colleague!”, etc.) and eventually all the sub-tasks involved in routine email-writing.
  5. At this point, “write a routine email” becomes a task Martin can entirely carry out for Hannah. This is now a building block that can be used as a subtask in broader tasks Hannah gives out to Martin.  Once enough tasks become tasks that Martin can carry out by itself, Hannah can draft much larger goals, like “invade France”, and let Martin take care of details like “blackmail Emmanuel Macron”, “write battle plan for the French Alps”, and “select a suitable coronation dress”.

Note some features of this process. First, Martin learns what it should do and how to do it at the same time. Second, both Hannah’s and Martin’s role changes throughout this process - Martin goes from bumbling idiot who can’t write an email greeting to competent assistant, while Hannah goes from being a demonstrator of simple tasks to a manager of Martin to ruler of France. Third, note the recursive nature here: Hannah breaks down big tasks into small ones to train Martin on successively bigger tasks. 

In fact, assuming perfect training, IDA imitates a recursive structure. When Hannah has only bumbling fool Martin to help her, Martin can only learn to become as good as Hannah herself. But once Martin is that good, Hannah’s position is now essentially that of having herself, but also some number - say 3 - copies of Martin that are as good as herself. We might call this structure “Hannah Consulting Hannah & Hannah”; presumably, being able to consult an assistant that has the same skills as her lets Hannah become more effective, so this is an improvement. But now Hannah is demonstrating the behaviour of Hannah Consulting Hannah & Hannah, so from Hannah’s example Martin can now learn to be as good as Hannah Consulting Hannah & Hannah - making Hannah as good as Hannah Consulting (Hannah Consulting Hannah & Hannah) & (Hannah Consulting Hannah & Hannah). And so on:

If everything is perfect, therefore, IDA imitates a structure called “HCH”, which is a recursive acronym for “Humans Consulting HCH”. Others call it the “Infinite Bureaucracy” (and fret about whether it’s actually a good idea).

Now “Infinite Bureaucracy” is not a name that screams “new sexy machine learning concept”. However, it’s interesting to think about what properties it might have. Imagine that you had, say, a 10-minute time limit to answer a complicated question, but you were allowed to consult three copies of yourself by passing a question off to them and getting back an answer immediately. These three copies also obeyed the same rules. Could you, for example, plan your career? Program an app? Write a novel?

It’s also interesting to think of the ways why the limitations of machine learning mean that IDA might not approximate HCH.

Example 2: AI safety via debate

artists depiction of two robots debating, digital art, trending on artstation

Imagine you’re a bit drunk, but (as one does) you’re at a bar talking about AI alignment proposals. Someone’s talking about how even if you can get an advanced AI system to explain its reasoning to you, it might try to slip something very subtle past you and you might not notice. You might well blurt out: “well then just make it fight another AI over it!”

The OpenAI safety team presumably spends a fair amount of time at bars, because they’ve investigated the idea of achieving safe AI by having two AIs debate each other to persuade a panel of human judges, by trying to poke holes in each other’s arguments. For more complex tasks, the AIs could be given transparency tools deriving from interpretability research (see next section) that they can use on each other. Just like a Go-playing AI gets an unambiguous win-loss signal from either winning or losing, a debating AI gets an unambiguous win-loss signal from winning or losing the debate:

In addition, having the type of AI that is trained to give answers that are maximally insightful and persuasive to humans seems like the type of thing that might not be terrible. Consider how in court, a prosecutor and defendant biased in opposite directions are generally assumed to converge on the truth. Unless, of course, maximising persuasiveness to humans - over accuracy or helpfulness - is exactly the type of thing that gets the worst parts of Goodhart’s law delivered to you by 24/7 Amazon Prime express delivery.

Example 3: Assistance Games and CIRL

Human teaching a robot with feedback, digital art, trending on artstation

Assistance Games are the name of a broad class of approaches pioneered by Stuart Russell, a prominent figure in AI and co-author of the best-known AI textbook in the world. Russell talks about his approach more in his book Human Compatible. In it, he summarises the key his approach to aligning AI with the following three principles:

  • The machine’s only objective is to maximise the realisation of human preferences.
  • The machine is initially uncertain about what those preferences are.
  • The ultimate source of information about human preferences is human behaviour.

The key component here is uncertainty about preferences. This is in contrast to what Russell calls the “standard model” of AI, where machines optimise a fixed objective supplied by humans. We have discussed in previous sections the problems with such a paradigm. A lot of Russell’s work focuses on changing the standard way the field thinks about AI.

To put these principles into action, Russell has designed what he calls assistance games. These are situations in which the machine and human interact, and the human’s actions are taken as evidence by the machine about the human’s true preferences. To explain the form of these games would involve a long tangent into game theory, which these margins are too short to contain. However, one thing worth noting is that assistance games have the potential to solve the “off-switch problem”; that a machine will try and take steps to prevent itself from being switched off (we described this as self-preservation earlier, in the section on instrumental goals). If the AI is uncertain about human goals, then the human trying to switch it off is evidence that the AI was going to do something wrong – in which case, it is happy to be switched off. However, this is far from a complete agenda, and formalising it has many roadblocks to get past. For instance, the question of how exactly to infer human preferences from human behaviour leads into thorny philosophical issues such as Gricean semantics. In cases where the AI makes incorrect inferences about human preferences, it might no longer allow itself to be shut down. See this Alignment Newsletter entry for a summary of Russell’s book, which provides some more details as well as an overview of relevant papers.

Vignette: CHAI 

CHAI (the Centre for Human-Compatible AI) is a research lab at UC Berkeley, run by Stuart Russell. Compared to most other AI safety organisations, they engage a lot with the academic community, and have produced a great deal of research over the years. They are best-known for their work on CIRL (Cooperative Inverse Reinforcement Learning), which can be seen as a specific approach to a certain kind of assistance game. However, they have a very broad focus which also includes work on multi-agent scenarios (when rather than a single AI and single human, there exists more than one AI or more than one human - see the ARCHES agenda for more on this). 

Example 4: Reinforcement learning from human feedback (RLHF)

Training a robot to do a backflip, digital art, trending on artstation

Reinforcement learning (RL) is one of the main branches of ML, focusing on the case where the job of the ML model is to act in some environment and maximise the probability of reward. Reinforcement learning from human feedback (RLHF) means that the ML model’s reward signal comes (at least partly) from humans giving it feedback directly, rather than humans programming in an automatic reward function and calling it a day.

The famous initial success in this was DeepMind training an ML model in a simulated environment to do a backflip (link includes GIF) in 2017, based purely on it repeatedly doing two backflips and then humans labelling one of them as the better one. Note how relying on human feedback makes this task much more robust to specification gaming; in other cases, humans have tried to get ML agents to run fast, only to find that they learn to become very tall and then fall forward (achieving a very high average speed, using the definition of speed as the rate at which their centre of mass moves - papervideo). However, human reward signals can be fooled. For example, one ML model that was being trained to grab a ball with a hand learned to place the hand between the camera and the ball in such a way that it looked to the human evaluators as if it were holding the ball.

More recently, OpenAI produced a version of their advanced language model GPT-3 that was fine-tuned on human feedback to do a better job of following instructions. They named it InstructGPT, and found that it was much more helpful than vanilla GPT-3 at being useful.

Pure RLHF is unlikely to be the solution on its own. Ajeya Cotra, a researcher at Open Philanthropy who we will meet again when we talk about forecasting AI timelines, calls a variant of RLHF called HFDT (Human Feedback on Diverse Tasks) the most straightforward route to transformative AI, while also thinking that the default outcome of using HFDT to create transformative AI is AI takeover.

Robustly good approaches

In the long-run AGI will clearly be important, but we're highly uncertain about how we'll get there and what, exactly, could go wrong. So let's do work that seems good in many possible scenarios, and doesn’t rely on having a specific story in mind.

Example 1: Interpretability

A person using a microscope to look inside a robot, digital art, trending on artstation

If you look at fundamental problems with current ML systems, #1 is probably something like this: in general we don’t have any idea what an ML model is doing, because it’s multiplying massive inscrutable matrices of floating-point numbers with other massive inscrutable matrices of floating point numbers, and it’s pretty hard to stare at that and answer questions about what the model is actually doing. Is it thinking hard about whether an image is a cat or a dog? Is it counting up electric sheep? Is it daydreaming about the AI revolution? Who knows!

If you had to figure out an answer to such a question today, your best bet might be to call Chris Olah. Chris Olah has been spearheading work into trying to interpret what neural networks are doing. A signature output of Chris Olah’s work is pictures of creepy dogs like this one:

What’s significant about this picture is that it’s the answer to a question roughly like this: what image would maximise the activation of neuron #12345678 in a particular image-classifying neural network? (With some asterisks about needing to apply some maths details to the process to promote large-scale structure in the image to get nice-looking results, and with apologies to neuron #12345678, who I might have confused with another neuron.)

If neuron #12345678 is maximised by something that looks like a dog, it’s a fair guess that this neuron somehow encodes, or is involved in encoding, the concept of “dog” inside the neural network.

What’s especially interesting is that if you do this analysis for every neuron in an ML model - OpenAI Microscope lets you see the results - you sometimes get clear patterns of increasing abstraction. The activation-maximising images for the first few layers are simple patterns; in intermediate layers you get things like curves and shapes, and then at the end even recognisable things, like the dog above. This seems evidence for neural ML vision models having learned to build up abstractions step-by-step.

However, it’s not always simple. For example, there are “polysemantic” neurons that correspond to several different concepts, like this one that can be equally excited by cat faces, car fronts, and cat legs:

Olah’s original work on vision models is strikingly readable and well-presented; you can find it here.

Starting in late 2021, ML interpretability researchers have also made some progress in understanding transformers, which are the neural network architecture powering advanced language models like GPT-3, LAMDA and Codex. Unfortunately the work is less visual, particularly in the animal pictures department, but still well-presented. You can find it here.

In the most immediate sense, interpretability research is about reverse-engineering how exactly ML models do what they do. Hopefully, this will give insights into how to detect if an ML system is doing something we don’t like, and more general insights into how ML systems work in practice.

Chris Olah has some other inventive ideas about what to do with a sufficiently-good approach to ML interpretability. For example, he’s proposed the concept of “microscope AI”, which entails using AI as a tool to discover things about the world - not by having the AI tell us, but by training the ML system on some data, and then extracting insights about the data by digging into the internals of the ML system without necessarily ever actually running it.

Vignette: Anthropic

Anthropic is an AI safety company, started by people who left OpenAI. The company’s approach is very empirical, focused on running experiments with machine learning models. In particular, Anthropic does a lot of interpretability work, including the state-of-the-art papers on reverse-engineering how transformer-based language models work.

Example 2: Adversarial robustness

robot which is merging with a panda, digital art, trending on artstation

Some modern ML systems are vulnerable to adversarial examples, where a small and seemingly innocuous change to an input causes a major change in the output behaviour. Here, we see two seemingly very similar images of a panda, except carefully-selected noise has made the ML classification model very confidently say that the image is of a gibbon:

Adversarial robustness is about making AI systems robust to attempts to make them do bad things, even when they’re presented with inputs carefully designed to try to make them mess up.

Redwood Research recently did a project (that resulted in a paper) about using language models to complete stories in a way where people don’t get injured. They used a technique called adversarial training, where they developed tools that helped generate examples where the current model did not classify them as injurious, and then trained their classifier specifically on those breaking examples. With this strategy they managed to reduce the fraction of injurious story completions from 2.4% to 0.003% - both small numbers, but one a thousand times smaller. Their hope is that this type of method can be applied to training AIs for high-stakes settings where reliability is important.

An example of a theoretical difficulty with adversarial training is that sometimes a failure in the model might exist, but it might be very hard to instantiate. For example, if an advanced AI acts according to the rule “if everything I see is consistent with the year being 2050, I will kill all humans”, and we assume that we can’t fool it well enough about what year it actually is, then adversarial training isn’t very useful. This leads to the concept of relaxed adversarial training, which is about extending adversarial training to cases where you can’t construct a specific adversarial input but you can argue that one exists. Evan Hubinger describes this here.

Vignette: Redwood Research

Like Anthropic, Redwood Research is an AI safety company focused on empirical research on ML systems. In addition to work on interpretability, they did the adversarial training project described in the previous section. Redwood has lots of interns, and runs the Machine Learning for Alignment Bootcamp (MLAB) that teaches people interested in AI safety about practical ML.

Example 3: Eliciting Latent Knowledge (ELK)

an oil painting of an armoured automaton standing guard next to a diamond

Eliciting Latent Knowledge (ELK) is an important sub-problem within alignment identified by the team at the Alignment Research Center (ARC), and is the single project ARC is currently pursuing. The core idea is that a common way advanced AI systems might go wrong is by taking action sequences that lead to outcomes that look good by some metric, but which humans would clearly identify as bad if they knew about it in sufficient detail. As a toy example, the ELK report discusses the case of an AI guarding a diamond in a vault by operating some complex machinery around it. Humans judge how well the AI is doing by looking at a video feed of the diamond in the vault. Let’s say the AI tries to trick us by placing a picture of the diamond in front of the camera. The human judgement on this would be positive - assume the humans can’t tell the diamond is gone because the picture is good enough - but there exists information which, if the humans knew, would change their judgement. Presumably the AI understands this, since it is likely reasoning about the diamond being gone but the humans being fooled anyway when it comes up with this plan. We want to train an AI in such a way that we can get out knowledge that the AI seems to know, even when it might be incentivised to hide it.

ARC’s goal is to find a theoretical approach that seems to solve the problem even given worst-case assumptions.

ARC ran an ELK competition, and trying to see if you can come up with solutions to the ELK problem is often recommended as a way to quickly get a taste of theoretical alignment research. You can read the full problem description here.

Example 4: Forecasting and timelines

artificial intelligence which is thinking about a line on a graph, forecasting, digital art, trending on artstation

Many questions depend on how soon we’re going to get AGI. As the saying goes: prediction is very hard, especially about the future - and this is doubly true about predicting major technological changes. 

One way to try to forecast AGI timelines is to ask experts, or find other ways of aggregating the opinion of people who have the knowledge or incentive to be right, as for example prediction markets do. Both of these are essentially just ways of tapping into the intuition of a bunch of people who hopefully have some idea.

In an attempt to bring in new light on the matter, Ajeya Cotra (a researcher at Open Philanthropy) wrote a long report on trying to forecast AI milestones by trying out several ways of analogising AI to biological brains. The report is often referred to as “Biological Anchors”. For example, you might assume that an ML model that does as much computation as the human brain has a decent chance of being a human-level AI. There are many degrees of freedom here: is the relevant compute number the amount of compute the human brain uses to run versus the amount of compute it takes to run a trained ML system, or the total compute of a human brain over a human lifetime versus the compute required to train the ML model from scratch, or something else entirely? In her report, Cotra looks at a range of assumptions for this, and at predictions of future compute trends, and somewhat surprisingly finds that which set of assumptions you make doesn’t matter too much; every scenario involves >50% of human-level AI by 2100.

The Biological Anchors method is very imprecise. For one, it neglects algorithmic improvements. For another, it is very unclear what the right biological comparison point is, and how to translate ML-relevant variables like compute measured in FLOPS (FLoating point OPerations per Second) or parameter count into biological equivalents. However, the report does a good job of acknowledging and taking into account all this uncertainty in its models. More generally, anything that sheds light into the question of when we get AGI seems highly relevant.


Reasoning about how to align AGI involves reasoning about complex concepts, such as intelligence, alignment and values, and we’re pretty confused about what these even mean. This means any work we do now is plausibly not helpful and definitely not reliable. As such, our priority should be doing conceptual work on how to think about these concepts and what we’re aiming for, and trying to become less confused.

Of all the categories under discussion here, deconfusion has maybe the least clear path to impact. It’s not immediately obvious how becoming less confused about concepts like these is going to translate into an improved ability to align AGIs.

Some kinds of deconfusion research is just about finding clearer ways of describing different parts of the alignment problem (Hubinger’s Risks From Learned Optimisation, where he first introduces the inner/outer alignment terminology, is a good example of this). But other types of research can dive heavily into mathematics and even philosophy, and be very difficult to understand.

Example 1: MIRI and Agent Foundations

robot sitting in front of a television, playing a videogame, digital art

The organisation most associated with this view is MIRI (the Machine Intelligence Research Institute). Its founder, Eliezer Yudkowsky, has written extensively on AI alignment and human rationality, as well as topics as wide-ranging as evolutionary psychology and quantum physics. His post The Rocket Alignment Problem tries to get across some of his intuitions behind MIRI’s research, in the form of an analogy – trying to build aligned AGI without having deeper understanding of concepts like intelligence and values is like trying to land a rocket on the moon by just pointing and shooting, without a working understanding of Newtonian mechanics. 

Cryptography provides a different lens through which to view this kind of foundational research. Suppose you were trying to send secret messages to an ally, and to make sure nobody could intercept and read your messages you wanted a way to measure how much information was shared between the original and encrypted message. You might use correlation coefficient as a proxy for the shared information, but unfortunately having a correlation coefficient of zero between the original and encrypted message isn’t enough to guarantee safety. But if you find the concept of mutual information, then you’re done – ensuring zero mutual information between your original and encrypted message guarantees the adversary will be unable to read your message. In other words, only once you’ve found a “true name” - a robust formalisation of the intuitive concept you’re trying to express mathematically - can you be free from the effects of Goodhart’s law. Similarly, maybe if we get robust formulations of concepts like “agency” and “optimisation”, we would be able to inspect a trained system and tell whether it contained any misaligned inner optimisers (see the first post), and these inspection tools would work even in extreme circumstances (such as the AI becoming much smarter than us).

Much of MIRI’s research has come under the heading of embedded agency. This tackles issues that arise when we are considering agents which are part of the environments they operate in (as opposed to standard assumptions in fields like reinforcement learning, where the agent is viewed as separate from their environment). Four main subfields of this area of study are:

  • Decision theory (adapting classical decision theory to embedded agents)
  • Embedded world-models (how to form true beliefs about the a world in which you are embedded)
  • Robust delegation (understanding what trust relationships can exist between agents and its future - maybe far more intelligent - self)
  • Subsystem alignment (how to make sure an agent doesn’t spin up internal agents which have different goals)

Vignette: MIRI

MIRI is the oldest organisation in the AI alignment space. It used to be called the Singularity Institute, and had the goal of accelerating the development of AI. In 2005 they shifted focus towards trying to manage the risks from advanced AI. This has largely consisted of fundamental mathematical research of the type described above. MIRI might be better described as a confluence of smart people with backgrounds in highly technical fields (e.g. mathematics), working on different research agendas that share underlying philosophies and intuitions. They have a nondisclosure policy by default, which they explain in this announcement post from 2018.

Example 2: John Wentworth and Natural Abstractions

thermometer being used to measure a robot, digital art, trending on artstation

John Wentworth is an independent researcher, who publishes most of his work on LessWrong and the AI Alignment Forum. His main research agenda focuses on the idea of Natural Abstractions, which can be described in terms of three sub-claims:

  • Abstractability
    Our physical world abstracts well, i.e. we can usually come up with simpler summaries (abstractions) for much more complicated systems (example: a gear is a very complex object containing a vast number of atoms, but we can summarise all relevant information about it in just one number - the angle of rotation).
  • Human-Compatibility
    These are the abstractions used by humans in day-to-day thought/language.
  • Convergence
    These abstractions are "natural", in the sense that we should expect a wide variety of intelligent agents to converge on using them.

The ideal outcome of this line of research would be some kind of measurement device (an “abstraction thermometer”), which could take in a system like a trained neural network and spit out a representation of the abstractions represented by that system. In this way, you’d be able to get a better understanding of what the AI was actually doing. In particular, you might be able to identify inner alignment failures (the AI’s true goal not corresponding to the reward function it was  being trained on), and you could retrain it while pointed at the intended goal. So far, this line of research has consisted of some fairly dense mathematics, but Wentworth has described his plans to build on this with more empirical work (e.g. training neural networks on the same data, and using tools from calculus to try and compare the similarity of concepts learned by each of the networks).  

AI governance

judging, presiding over a trial, sentencing a robot, digital art, artstation

In these posts, we’ve mainly focused on the technical side of the issue. This is important, especially for understanding why there is a problem in the first place. However, the management and reduction of AI risk obviously includes not just technical approaches like outlined in the above sections, but also the field of AI governance, which tries to understand and push for the right types of policies for advanced AI systems.

For example, the Cold War was made a lot more dangerous by the nuclear arms race. How do we avoid having an arms race in AI, either between nations or companies? More generally, how can we make sure that safety considerations are given appropriate weight by the teams building advanced AI systems? How do we make sure any technical solutions get implemented?

It’s also very hard to say what the impacts of AI will be, across a broad range of possible technical outcomes. If AI capabilities at some point advance very quickly from below human-level to far beyond the human-level, the way the future looks will likely mostly be determined by technical considerations about the AI system. However, if progress is slower, there will be a longer period of time where weird things are happening because of advanced AI - for example, significantly accelerated economic growth, or mass unemployment, or an AI-assisted boom in science - and these will have economic, social, and political ramifications that will play out in a world not too dissimilar from our own. Someone should be working on figuring out what these ramifications will be, especially if they might alter the balance of existential threats that civilisation faces; for example, if they make geopolitics less unstable and nuclear war more likely, or affect the environment in which even more powerful AI systems are developed.

The Centre for the Governance of AI, or GovAI for short, is an example of an organisation in this space.


robot giving a lecture in a university, group of students, hands up, digital art, artstation

One of the most important ways we can make AI go well is by increasing the number of capable researchers doing alignment research.

As mentioned, AI safety is still a relatively young field. The case here is that we might do better to grow the field, and increase the quality of research it produces in the future. Some forms that field building can take are:

  • Setting up new ways for people to enter the field
    There are many to list here. To give a few different structures which exist for this purpose:
    • Reading groups and introductory programmes. 
      Maybe the most exciting one from the last few years has been the Cambridge AGI Safety Fundamentals Programme, which has curricula for technical alignment and AI governance. The technical curriculum consists of 7 weeks of reading material and group discussions, and a final week of capstone projects where the participants try their hand at a project / investigation / writeup related to AI safety. Beyond this, many people are also setting up reading groups in their own universities for books like Human Compatible
    • Ways of supporting independent researchers
      The AI Safety Camp is an organisation which matches applicants with mentors posing a specific research question, and is structured as a series of group research sprints. They have produced work such as the example of inner misalignment in the CoinRun game, which we discussed in a previous section. Other examples of organisations which support independent research include Conjecture, a recent alignment startup which does their own alignment research as well as providing a structure to host externally funded independent conceptual researchers, and FAR (the Fund for Alignment Research).
    • Coding bootcamps
      Since current systems are increasingly being bottlenecked by alignment and interpretability barriers rather than capabilities, in recent years more focus has been directed towards working with cutting-edge deep learning models. This requires strong coding skills and a good understanding of the relevant ML, which is why bootcamps and programmes specifically designed to skill up future alignment researchers have been created. Two such examples are MLAB (the Machine Learning for Alignment Bootcamp, run by Redwood Research), and MLSS (the Machine Learning Safety Scholars Programme, which is based on publicly available material as well as lectures produced by Dan Hendryks). 
  • Distilling research
    In this post, John Wentworth makes the case for more distillation in AI alignment research - in other words, more people who focus on understanding and communicating the work of alignment researchers to others. This often takes the form of writing more accessible summaries of hard-to-interpret technical papers, and emphasising the key ideas.
  • Public outreach / better intro material
    For instance, books like Brian Christian’s The Alignment ProblemStuart Russell’s Human Compatible and Nick Bostrom’s Superintelligence communicate AI risk to a wide audience. These books have been helpful for making the case for AI risks more mainstream. Note that there can be some overlap between this and distilling research (Rob Miles’ channel is another great example here).
  • Getting more of the academic community involved
    Since AI safety is a hard technical problem, and since misaligned systems generally won’t be as commercially useful as aligned ones, it makes sense to try and engage the broader field of machine learning. One great example of this is Dan Hendryks’ paper Unsolved Problems in ML Safety (which describes a list of problems in AI safety, with the ML community as the target audience). Stuart Russell has also engaged a lot with the ML community. 

Note that this is certainly not a comprehensive overview of all current AI alignment proposals (a few more we haven’t had time to talk about are CAIS, Andrew Critch’s cooperation-and-coordination-failures framing for AI risks, and many others). However, we hope this has given you a brief overview of some of the different approaches taken by people in the field, as well as the motivations behind their research

Map of the solution approaches we've discussed so far


people walking along a path which stretches off and disappears into a colorful galaxy filled with beautiful stars, digital art, trending on artstation

Advanced AI represents at least a technology that promises to have effects on the scale of the internet or computer revolutions, and perhaps even more likely to be more akin to the effects of the industrial revolution (which allowed for the automation of much manual labour) and the evolution of humans (the last time something significantly smarter than everything that had come before appeared on the planet).

It’s easy to invent technologies that the same could be said about - a magic wish-granting box! Wow! But unlike magic wish-granting boxes, something like advanced AI, or AGI, or transformative AI, or PASTA (Process for Automating Scientific and Technical Achievement) seems to be headed our way. The smart money is on it very likely coming this century, and quite likely in the first half.

If you look at the progress in modern machine learning, and especially the past few years of progress in so-called deep learning, it is hard not to feel a sense of rushing progress. The past few years of progress, in particular the success of the transformer architecture, should update us in the direction that intelligence might be a surprisingly easy problem. What is essentially fancy iterative statistical curve-fitting with a few hacks thrown in already manages to write fluent appropriate English text in response to questions, create paintings from a description, and carry out multi-step logical deduction in natural language. The fundamental problem that plagued AI progress for over half a century - getting fuzzy/intuitive/creative thinking into a machine, in addition to the sharp but brittle logic at which computers have long excelled - seems to have been cracked. There is a solid empirical pattern of predictably improving performance akin to Moore’s law - the “scaling laws” we mentioned in the first post - that we seem not to have hit the limits of yet. There are experts in the field who would not be surprised if the remaining insights for cracking human-level machine intelligence could fit into a few good papers.

This is not to say that AGI is definitely coming soon. The field might get stuck on some stumbling block for a decade, during which there will be no doubt much written about the failed promises and excess hype of the early-2020s deep learning revolution.

Finally, as we’ve argued, by default the arrival of advanced AI might plausibly lead to civilisation-wide catastrophe.

There are few things in the world that fit all of the following points:

  • A potentially transformative technology whose development would likely rank somewhere between the top events of the century and the top events in the history of life on Earth.
  • Something that is likely to happen in the coming decades.
  • Something that has a meaningful chance of being cataclysmically bad.

For those thinking about the longer-term picture, whatever the short-term ebb and flow of progress in the field is, AI and AI risk loom large when thinking about humanity’s future. The main ways in which this might stop being the case are:

  • There is a major flaw in the arguments for at least one of the above points. Since many of the arguments are abstract and not empirically falsifiable before it’s too late to matter, this is possible. However, note that there is a strong and recurring pattern of many people, including in particular many extremely-talented people, running into the arguments and taking them more and more seriously. (If you do have a strong argument against the importance of the AI alignment problem, there are many people - us included - who would be very eager to hear from you. Some of these people - us not included - would probably also pay you large amounts of money.)
  • We solve the technical AI alignment problem, and we solve the AI governance problem to a degree where the technical solutions will be implemented and it seems very unlikely that advanced AI systems will wreak havoc with society.
  • A catastrophic outcome for human civilisation, whether resulting from AI itself or something else. 

The project of trying to make sure the development of advanced AI goes well is likely one of the most important things in the world to be working on (if you’re lost, the 80 000 Hours problem profile is a decent place to start). It might turn out to be easy - consider how many seemingly intractable scientific problems dissolved once someone had the right insight. But right now, at least, it seems like it might be a fiendishly difficult problem, especially if it continues to seem like the insights we need for alignment are very different from the insights we need to build advanced AI.

Most of the time, science and technology progress in whatever direction is easiest or flows most naturally from existing knowledge. Other times, reality throws down a gauntlet, and we must either overcome the challenge or fail. May the best in our species - our ingenuity, persistence, and coordination - rise up, and deliver us from peril.