This document aims to be two things: a summary of the things that we learned from the Good Technology Project (GTP), and a post-mortem of the project itself.
I’m going to simply state my beliefs in this post, but I should clarify beforehand that I am not very certain about these, they are my current best guesses.
What was GTP?
GTP started in late 2015 when Richard Batty and I met up for coffee in Oxford. We ended up talking about entrepreneurship: both Richard and I were working in software, and we believed that entrepreneurship could provide a route to leverage our skills. But work on the concrete problem of how to actually do that was frustratingly sparse.
Once we started thinking about the area, we realised there was more to do. The goal of GTP became broadly to do whatever we could to “fix” technology entrepreneurship. We observed that technology entrepreneurship is very powerful, and yet the industry is not strongly incentivised to solve important problems. While they may prioritize well when it comes to profit, they don’t prioritize well when it comes to impact, even when this is an explicit goal.
We gradually ramped up our work on the project: in the middle of 2016 when we started making real progress, Richard quit his job to work part time on the GTP, supporting himself with freelancing. I also dropped down to four days a week at work, with one day on GTP. We continued in this way for about a year, which gave us quite a lot of flexibility, since Richard could adjust his hours easily, and I frequently worked Saturdays in addition to my one weekday.
Since then we tried a number of different approaches (which I’ll discuss below), before eventually stopping the project due to a combination of failure and lack of remaining steam.
A guiding model
Before we discuss what GTP did, I’m going to present an abstract model of how we believe the entrepreneurship process works. The point of this is to situate what we learned in terms of where we think it fits into the bigger picture.
The entrepreneurship landscape
We can think of the space of potential startups as a height map over a plane.1 The plane corresponds to roughly “where” the startup is, what problem it is solving, how it is solving it, etc. The “height” corresponds to how good the startup is by some measure (for now we’ll assume there is just one kind of “success”, and that our entrepreneurs actually care about impact as well as profit and so will try to maximize it as best they can).
Then what we want to do is to locate the highest peaks in this landscape. These correspond to the best opportunities to get whatever it is that we want - money and impact, in this case.
We have some broad beliefs about the landscape:
- There is a huge amount of variation globally: the highest peaks are much higher than the median.
- There is a lesser, but still substantial, amount of local variation: even within a region there is a big difference between the best and the median.
- Peaks are relatively sparse: most of the points are low.
These are based on the usual observations about the distributions of startup earnings/impact. While a significant portion of the variance there is going to be due to luck and other exogenous factors, I think that the most significant portion is due to the nature of the startup in question. So these aren’t entirely uncontroversial assumptions, but I’m not going to defend them further here simply because I don’t think it’s the most interesting part of this post.
Meet the locals
Many people are trying to navigate the landscape and find peaks. Commonly these people have an “area” of expertise in which they’re much better at surveying the landscape, and have a better ability to climb nearby hills. For example, someone who already has experience in retail business will be better able to assess opportunities that target retail businesses, and better able to tackle the problems along the way to making them a success.
These “locals”” correspond to individual entrepreneurs or potential entrepreneurs, and their situation in the landscape corresponds to their specialised knowledge, skills or beliefs (their “edge”, to borrow a term from EF).
A helpful distinction that EF makes is between people with a “domain” edge and people with a “technology” edge. This isn’t a hard-and-fast distinction, but the idea is that people with a domain edge have some advantage relating to the problem or the business idea at hand, whereas people with a technology edge have an advantage in some aspect of the implementation.
Since we think of startups that solve the same problem as being “close together” in our landscape, we can think of people with a domain edge as inhabiting a fairly compact region, whereas people with a technology edge occupy a cross-cutting strip of the landscape. If you know about distributed systems, then you might have an advantage in specific parts of both financial technology and medicine. (This isn’t tremendously important, but I find it helpful to visualise the complex ways in which people’s skills and knowledge might intersect.)
Far above the landscape are the “planners”. The planners still want to find the highest peaks, but they have very poor visibility of the landscape. They may be able to tell that one area is generally hillier than another, but little more than this. They can drop expeditions down onto the landscape, but this is costly and is generally unlikely to land on a peak.
When we think about deliberately starting high-impact ventures, we are like the planners. We are trying to deliberately zoom in on the peaks across the whole landscape, even though we’re not already situated in that area.
We can learn about areas in more detail, but this requires laborious research, and even then our understanding is likely to be inferior to that of the locals.
Implications of the landscape model
This model is generally pessimistic about the ability of the planners to get to the peaks that they want to find. Because they have poor visibility into the landscape, it is hard for them to actually explore effectively without going through the effort of actually “becoming” a local. Most successful startups instead come from locals who know the region, and happen to be in the vicinity of a substantial peak.
This leaves us in the rather unsatisfying state that Julia Galef outlines in “Can we intentionally improve the world?”. Julia contrasts two approaches: the top-down approach favoured by “Planners” (governments, foundations, effective altruists), and the bottom-up approach favoured by “Hayekians” (entrepreneurs of various stripes).
While the Planner approach has had some success, the prevailing wisdom in modern startup theory is that the Hayekian approach is the only viable one. Success (as Paul Graham argues) comes from using situated knowledge to find and solve problems that actually occur on the ground, rather than by trying to impose a top-down plan.
I broadly agree with these conclusions, but I think there’s some hope for “hybrid” top-down/bottom-up solutions, which I’ll outline later.
How general is this model?
The landscape model seems like it is a good fit in situations where:
- There is a large space of opportunities
- The opportunities are very varied by some metric
- There are many people with strong “local” knowledge
- Local optimization is reasonable at finding nearby peaks
- Global optimization is hard, and gaining enough knowledge to do local optimization is also hard
This applies not only to for-profit entrepreneurship, but to other exploratory domains, including charity entrepreneurship and perhaps even research. Exactly how hilly the landscape is or how strong the “local” advantage is will vary, but I think it may be a useful tool in other domains too.
So we’ve decided we want to improve the entrepreneurship process. What should we try?
Advise entrepreneurs directly
The first strategy we tried was to try and influence entrepreneurs directly. However, since we wanted to change what projects they worked on, we needed to get to them before they had committed to an idea.
To that end we teamed up with EF, a startup accelerator that takes individuals pre-idea, and helps them form teams and generate ideas.2 We pitched to the Summer 2017 EF cohort before they started, and we got an excellent response: over 10% of the cohort said they were interested in talking about how to have more of an impact with their startup.
However, the advice process itself was an almost total failure. There were three main problems:
Firstly, the number one thing that entrepreneurs wanted from us was a better idea to work on. We were simply unable to provide such ideas (or at least appropriate ones, see the next point).
This points towards one of the features of the landscape model: it is hard for planners to know an entrepreneur’s local area even as well as the entrepreneur, let alone better.
Secondly, it was more or less impossible to persuade people to move towards domains they didn’t know about. EF pushes very hard to get people to embrace their “edge”, and this means that it simply isn’t realistic either to get someone who knows about medicine to look at a finance idea, nor to get someone with flexible technical skills to work on an idea without a domain cofounder.
We think this is good advice! We looked at 50 potentially high-impact tech companies, and almost all of them had a founder who had a strong background in the area. Complete outsiders were rare.
Our observations of the cohort subsequently bore this out: everyone we knew ended up in a team where the problem was determined by the founder with the domain expertise.3
This points to another feature of the landscape model: entrepreneurs are locally situated by their existing background knowledge, and this is part of what lets them do what they do. Attempts to “move” them are likely to both meet with resistance and be ultimately counterproductive.
Finally, the number of people who were willing to actually contribute some work towards working out how to have a higher impact was almost zero (we had only one person from EF7 who substantially engaged). That’s not to say that they didn’t care about having an impact, but merely that it wasn’t a goal that could override the other pressures on them from the EF programme.
There were many other problems with our content and the way we presented our pitch, but these were the problems that looked like structural ones.
During this time we also developed a fair amount of object-level material about what we think are the important factors in a high-impact startup. We presented this in some workshops to EF participants, and at EAG Oxford 2016. This is perhaps of interest, but it’s not as relevant to the high-level picture so I’ve relegated it to a follow-up post.
So what do we say to entrepreneurs?
From this point on we had a somewhat ambivalent relationship to individual entrepreneurs. They tended to either be already working in broadly the right area, in which case we had nothing further to say to them (except some general advice about how to keep on track), or they were in an area that seemed unlikely to be high-impact, in which case we also had little to say to them.
Surprisingly, the most common advice we gave after this point was “try working or studying in an important area for a few years, then reconsider entrepreneurship”. This is particularly important for people with a technology edge, since you want to either develop some domain edge in an important area, or find a cofounder there. This advice wasn’t generally terribly well-received - people who’ve got to the point of thinking of themselves as potential entrepreneurs usually want to do something soon.
Glimmers of hope: serial entrepreneurs
By far our most successful individual advice process was with a repeat entrepreneur, who had already sold his first company. This made a huge difference: he had far more resources to deploy on the process, he had time, and most of all he cared enough to be willing to go in for another round because he wanted the change rather than the money.
He had a pre-existing inclination to work on climate change, but we were able to help him narrow down the area to negative emissions technologies. I don’t think he’s actually started anything yet, but perhaps it will turn out that we had an impact there.
Possibly targeting serial entrepreneurs as a market could work. I’d be excited to see Founders’ Pledge, or someone similar, looking into this.
Find better problems
One reason we couldn’t advise entrepreneurs well was because we didn’t have good problems to offer them. So perhaps we could get better at that?
At first blush, “what problem should I work on?” is a classic EA question - isn’t this just cause prioritization? However, if you look at the cause prioritization work that we’ve actually done, it’s all far too high-level to be action-guiding if you’re actually looking to do new things. Knowing that animal welfare is high-priority is useful, but then you want to know which are the most important parts of the problem, and then which are the most important parts of that problem, until you find something you can actually tackle.
I think there are a couple of reasons why this hasn’t been apparent. One is simply that EAs haven’t focussed on doing this kind of prioritization before. We have something to say to donors, early-career employees, and perhaps even later-career employees; but the “explorer” is a comparatively new personality for us.
Another reason is that EAs have often operated in an evaluating capacity. If you are being trying to find the best donation opportunity, you are in the position of evaluating an discrete set of existing opportunities (the thing has to be established enough that you can donate to it!). If you are being presented with existing ideas in this way and need to evaluate them, you can use a coarse cause prioritization as an initial filter and then move on to more laborious methods. However, if you’re looking for new ideas, then initially filtering down to a priority cause helps, but still leaves far too much work for an individual to realistically do.
So what we need is much more granular cause prioritization, ideally right down to the size of a problem that can be worked on by an individual or team.4
The most ambitious version of this might look like an atlas of the world’s problems, broken down by subproblems, and prioritized as best we can. This could be useful both for people working on solving problems, and also for establishing common vocabulary about just what they are. For example, Amodel et al’s “Concrete Problems in AI Safety” paper did wonders for just clarifying what the problems are, and was generally regarded as a very useful contribution for that reason.
Nice as this vision is, it number of problems:
- It would be a hilariously enormous amount of work to create and maintain.
- It’s not clear that it’s possible to nicely break down problems in this way.
- It’s unclear how the prioritization would work, or who would do it.
I pitched this idea to a few people and got some interest, but I never took it very far. I still think it could be valuable, and a reasonable MVP would be a thin, “vertical”, slice through a domain. I think health would be a good area for this, in that much of the needed material is already present. Indeed, I think this is one of the reasons that Charity Entrepreneurship has made so much progress with the top-down approach in health.
The existence of this problem and the difficulty of solving it also supports the landscape model, since it illustrates the problems that the planners have in understanding the landscape in enough detail to make decisions.
If we can’t do much with individual entrepreneurs, perhaps we can achieve more by targeting the institutions that play a role in the entrepreneurship process.
There are a priori reasons to think this might be a good strategy. Institutions already show signs of being more goal-driven and prioritizing than entrepreneurs: while entrepreneurs may be fixed on a particular idea or area, VCs operate more like experimenters, looking for promising areas and picking startups to fund as “trials”. So if there is already strategic thought going on at this level, we might be able to influence it.
One way in which startup accelerators and incubators in particular can affect the process is by changing the mix of people who become entrepreneurs. So the following idea suggests itself: pick an important area, and then bias recruitment towards people who already have the skills and inclination to work in that area (for example, you might recruit epidemiologists, development economists, and doctors, as well as technologists). If we can’t win a hand, perhaps we can stack the deck.
A minimal institutional solution would be to find an institution like EF and persuade them to let us bias their intake in this way. In practice, this is a difficult goal with a for-profit institution. Even if they might want to adjust their process to have more of an impact, anything that risks damaging the bottom line is dangerous. It’s possible that some existing institutions might be willing to try this – YC, for example, has been experimenting with some less profit-driven initiatives – but we didn’t manage to find any that we could work with.
We might expect to have more success with institutions with a more explicitly altruistic mission. We didn’t manage to talk to as many people in this area, but our experiences suggested that they were either altruistic but not interested in taking an effective altruist approach; or they were still too young and so were still overly worried about profitability. This approach might work if we could find the right institution, but we weren’t able to find one in the time we allowed ourselves.
A third alternative would be to start a completely new institution explicitly designed with this goal in mind (we called this “the Lab”, inspired by Bell Labs). This would also offer the opportunity to design the programme and the funding process to help keep things on track in the later stages of startup formation. This isn’t a new idea: I suggested something like this in a previous post, and Spencer Greenberg’s “Spark Wave” programme is a similar project (and is making some exciting progress).
Furthermore, an institutional solution has the potential to scale well, by replicating itself, in a way that can be harder to do if you are targeting individuals. Matt Clifford has a lovely metaphor for EF: startup investment is currently like running after lightning strikes, whereas EF is trying to build an electricity generator. We should do this for high-impact startups too, and then make lots of them.
I think this is a really exciting idea, but it’s heavy on operations, sales, risk, and hustle, so we don’t think that we are the right people to tackle it.5 Ideally, we could start such a project by splitting off an existing institution, piggybacking off its existing operations capacity and connections.6
We think that entrepreneurs start companies based on the experience of the founding teams. So that means that a way to increase the number of people who work on important problems is to have more entrepreneurs who are deeply familiar with those problems. Then when these people turn to starting companies, they will naturally solve the problems that they have become familiar with.
From an individual point of view, that means if you’re considering starting a social impact start-up in the future a good way of doing that would be to gain experience in an important problem area.
Concretely, we could develop a more detailed advice programme for potential entrepreneurs and individuals later in their career. 80,000 Hours would be the natural home for something like this, although they aren’t currently prioritizing it. Currently, a lot of their advice is focussed on people early in their careers, and making sure that they go into the right areas and build the right career capital. Helping to guide the process after people get further into their careers is a lot more work since it is quite career-specific, but not as much work as actually scoping out the problem space would be, since you are still assuming that people will find the most important problems themselves.
The problems with this approach is that it has a very long lead time, and it relies on a long chain of fallible steps between influencing someone’s career and them actually starting an important company. However, it could also be tried fairly easily as a natural extension of the work that 80,000 Hours is already doing.
Investors tend to be an influence on the ideas that get started even before start-ups try to seek funding. This is because entrepreneurs tend to be thinking about whether an idea will be interesting to investors as their developing it. So by changing what investors are looking for, we can try and persuade the bottom-up explorers to optimize for something different.
We are not sure exactly how effectively investors influence entrepreneurs but there seems to be a sense among entrepreneurs that certain things are ‘hot’ among investors at a particular time. A big enough philanthropic investor could therefore influence what gets started, although probably still only within broad areas.
A variant of this approach is the venerable one of offering prizes for companies tackling particular problems. This amounts to offering easy funding to anyone who works on a particular problem. Prizes are increasingly popular now (e.g. X-Prize), but we don’t have a good sense for how effective they are, or how scalable they are as a funding institution.
The main problem with this plan is that it relies heavily on finding sensible, wealthy investors.
The picture I’ve painted represents my best guess at the truth, but it has some big open questions.
Can you maximise for both profit and impact?
The landscape model implicitly assumes you are targeting one metric, but a startup that has an impact needs to be profitable too. Is it possible to do this? Is it practically possible to do this, given the pressure from investors etc.? Can we design governance structures that make this easier?
How come Charity Entrepreneurship can do top-down problem search?
Firstly, I think Charity Entrepreneurship is awesome, and one of the best things to come out of the EA movement in years. I am a huge fan.
How is it that they were able to do this top-down search in a way that I argued was, if not impossible, at least extremely hard? I think the answer is that they worked very hard; the non-profit sector is less efficient than the for-profit sector; and health is an unusually well-evidenced area. There already are big compendiums of the health problems in the world (see the Global Burden of Disease), and good evidence on which ones work. There’s still a big research and deployment task, and I think that is certainly more difficult for outsiders, but I think they provide a great example of it being possible to do this kind of thing.
However, I don’t think this scales, and I think it will be a lot harder in areas outside of health.
The previous section has given a brief history of the intellectual and practical progress of GTP, from its inception to its eventual death when we realised that none of the ideas that we thought were plausible were feasible for us to tackle.
What follows is a brief post-mortem of the project itself. Feel free to skip this if you’re not interested.
What went well?
Refinement of problems and concepts
The journey I’ve described to you may sound relatively clear-cut, but that’s because I’ve described it to you in terms of our current understanding of what happened. It actually took us a long time to come up with what we have, even though much of it seems obvious in retrospect.
This makes me think that many of the things that we came up with are in the camp of “surprisingly useful obvious truths”, which take a deceptively large amount of sweat to discover. So I hope you will not disdain our offerings.
I also think we did a good job of refining the problems that we were working on. We changed tack fairly frequently, and were fairly responsive to new evidence as we got it (with a lot of soul-searching along the way). I think we were guilty both of stopping things too early and of stopping things too late, but it could have been worse.
One thing that we found surprising was just how easy it was to make some progress, and say some things that had not obviously been said before. Even after spending a couple of years reading and talking about these problems, I still think that a lot of what we did was fairly original (of course, we might just have missed things!). I think the moral is that there are still uncharted areas to investigate, and if you’re in early then even amateurs like ourselves can make progress.
Richard and I worked very well together. I found that team cohesion and work habits made a huge amount of difference - GTP has been by far my most extensive and successful project to date.
I think we also did the right thing keeping the team to just the two of us. I don’t think we’d have been much more productive with more people, and we were able to keep scheduling to a minimum and be fairly spontaneous about meeting people.
Unexpected useful outcomes
In the course of our work we inevitably ended up looking at a lot of actual companies, and many of those at least looked like they might be high impact. We ended up keeping a list (warning: out of date) of these. This ended up being one of the things that people were most interested in, usually because they were looking for employers.
Personal discovery and development
Both Richard and I learned a lot during the process. In particular, I learned that I just really don’t like the kind of desk research we were doing, and I didn’t come around to it over time. That was in some ways a relief, and I’m aiming to avoid that kind of work in future.
Richard, on the other hand, really likes it, and is now doing much more of that kind of thing for 80,000 Hours.
Both of us found that actually running a project, while rewarding, was ultimately more difficult and stressful than we really wanted. I have gained even more respect for people who pull it off.
We got a surprising amount of value out of just writing everything down and putting it in Google Docs. The ability to refer back (and refer other people to it) is just invaluable. Writing documents was very helpful for mental clarification even if the end product was “disposable”.7
What went badly?
Lack of commitment
We both worked on GTP in our free time for about a year; and then Richard started working 2-3 days a week on it, and I moved to one day a week for another year.
Since we worked much better together, the fact that I was committing less time meant that we got a lot less done than we could have. In general, I think a lot of things would have progressed faster, been better, or generally more if I’d committed more time.
The project might well still have failed, but perhaps more interestingly.8
Especially unclear goals
Even for a “startup-like” project, our “product” shifted wildly, because our goals were so lofty. We couldn’t focus in on individual entrepreneurs as our target market, because we concluded that they were the wrong market. This meant that we were constantly finding our feet in new areas, or doing three things at once.
We had many meetings where the other party seemed very keen and interested, but then nothing actually materialized. We inevitably then ended up wasting a lot of time preparing for and thinking about projects that never happened. I think the lesson here is familiar: don’t count your customers until they’ve actually bought your product.
While I don’t think this went terribly, I do think that much it was largely useless. Our research mostly consisted of shallow reviews of one kind or another: either for entrepreneurs who we were trying to advise, to help them narrow down a field; or to help us figure out what fields to focus on; or as part of a number of abortive collaborations with institutions. In very few of these cases did anyone actually use the research that we’d done.
One surprising thing we learnt was that by far the most useful people to talk to were relatively junior EAs in a field. Unlike their seniors who had more breadth of experience, but often hadn’t thought about things through a prioritizing lens until we talked to them, even junior EAs have often thought about this a lot, and can be a great starting point.
Similarly, we got much better material from people working at foundations, who are used to looking at whole fields and assessing the importance and overall direction of research.
A positive effect of having done a lot of shallow research is that can give you enough “interactive knowledge” to talk to specialists and read papers with some understanding. This is very helpful if you are frequently dipping into an area.
Apart from recording things in Google Docs, our organization was pretty bad. We constantly attempted to record tasks in Asana, but the fluid nature of the work and my loathing of Asana made this an uphill struggle.
I think GTP as a whole has been a failure: it certainly failed to achieve its stated goals, although these were fairly ambitious. However, I think we’ve learned some useful things along the way.
My list of high-level takeaways would be:
- Entrepreneurship has a lot of potential, and there are many opportunities which could do with further exploration
- Influencing the entrepreneurship process is tricky, and may require applying leverage indirectly (via institutions, funders, etc.)
- More granular cause prioritization would be very useful for “explorers”
- The top-down/bottom-up conflict is real, but there is scope for hybrid solutions
I’m still very interested in this area, and I’d be very happy to talk to anyone about it in more detail. In particular, if you’re interested in working on any of the ideas in this document, please do get in touch.
I’d like to take a moment to thank some of the people who helped particularly with this project. In no particular order:
- Ben Clifford
- Ben Todd
- Eric Gastfriend
- Goodwin Gibbins
- Kit Harris
- Mario Pinto
- Matt Clifford
- Matt Gibb
- Max Dalton
- Naomi Morton
- Owen Cotton-Barratt
- Peter Hartree
- Rob Collins
- Sam Hilton
- Spencer Greenberg
Your generosity with your time and brainpower has been much appreciated.
We’ve become broadly convinced that EF is improving the process by getting involved earlier, and our startup theory has been significantly influenced by them. ↩
Another similar story about how Sparrho was founded: “I had a great postdoctoral researcher in my lab called Steve and he would spend about five or ten minutes every morning reading pre-prints from journals that he thought was relevant to the group. He knew what every single person was working on and he could recommend serendipitous papers that you can never find using a linear keyword search. Those kind of serendipitous recommendations lead to innovations. When I realised and I told Nilu that I had this problem and I had this great solution called Steve, Nilu’s background is machine learning so he asked why don’t we digitise Steve? Instead of just one person reading a couple of journals each day we now are using technology to do the same for tens of millions pieces of content.” ↩
People do start institutions like this from scratch (e.g. Matt and Alice with EF), but I fear survivorship bias. ↩
If that sounds like getting someone else to do all the hard work - yes, that’s the idea! More reasonably, we want to vary only the new component, rather than having to also leap all the other hurdles which are incidental to the core variation being tried. ↩
Don’t ask how many half-written documents there are in the “strategy” folder! ↩
The unfair version of this is to say “would the project have failed if Elon Musk had been running it?” If the answer is no, you weren’t working hard (or smart!) enough. ↩