Why I want to start a machine learning company
A reply to Ryan Carey, an effective-altruist researcher (later at the Future of Humanity Institute and the Center for Human-Compatible AI), who had asked Alyssa what she was actually trying to accomplish by starting a machine-learning company. The answer ranges from concrete (acquiring capital for projects beyond AI safety, such as anti-aging) to deeply meta: the EA and rationalist community has barely begun to answer questions like “under what conditions is a research team most effective?”, and the data needed to answer them well doesn’t exist in a form anyone can query. The proposed company is a stepping-stone — starting with customer relationship management software, where the data-extraction problem is comparatively easy, with an eye eventually toward medicine, science, and other harder domains.
My biggest targets in starting an ML company are: acquiring a greater ability to start other, more capital-intensive projects (I’ve discussed this on my blog at Budgeting in Effective Altruism ); going after key problems other than AI risk, like aging; answering a bunch of meta-questions that I think we’ll have to tackle to stop existential risk effectively.
The first two seem like fairly conventional wisdom (for our crowd), but the third isn’t that widely discussed yet, so I’d like to explain it a bit more. Superintelligent AI, whenever it’s eventually created, will be this very large, very complicated machine that’s a product of decades of both pure and applied research. So to build it effectively, we ought to have the answers to a bunch of questions about how research itself works, one of the most basic of which is “under what conditions is a research team most effective?”.
But even for a stunningly obvious question like that, nobody really seems to know how to answer it. Eg., every month or so, there’s a thread on Hacker News where people argue that the main way to help research is to increase government spending. In support of this idea, they’ll list major breakthroughs that were developed in US government labs — the transistor, the laser, radar, jet engines, GPS and so on. But if you take all of these breakthroughs and plot them out on a timeline, no matter which list you use, you’ll find that almost all of them happened between roughly 1935 and roughly 1975, with much less both before and after (please do try this yourself, if you like). Needless to say, this suggests increasing government funding in 2015 won’t have that much effect… but almost nobody even realizes this, let alone tries to come up with a reasonable theory to explain it, let alone succeeds. To quote historian Thomas Sowell, “Why do some groups, subgroups, nations, or whole civilizations excel in some particular fields rather than others? All too often, the answer to that question must be: Nobody really knows. It is an unanswered question largely because it is an unasked question.”
Take a famously successful research group, like Bell Labs. What was Bell Labs like? Well, for the most part, nobody really knows — all the archives are probably in forgotten file drawers somewhere, collecting dust in some government office building. A few historians have written books about it (eg. Jon Gertner, The Idea Factory ), but then you face the twin problems that (a) the historian will almost certainly leave out crucial details and (b) it’s extraordinarily easy to come up with theories that are wrong, but that seem to work, so long as you only look at one data point. For example, the top Amazon reviewer of that book explains Bell Labs’ success by saying, “They did this by gathering bright minds under one roof and giving them the freedom and time to pursue their ideas.” Yet, this sentence applies equally well to Princeton’s Institute for Advanced Study. And to quote Richard Hamming, “The Institute for Advanced Study in Princeton, in my opinion, has ruined more good scientists than any institution has created, judged by what they did before they came and judged by what they did after. Not that they weren’t good afterwards, but they were superb before they got there and were only good afterwards.” And this doesn’t even begin to address the more difficult questions, like “what research directions will do net good rather than harm” — eg. the Manhattan Project was famously productive, but lead directly to a world on the perpetual brink of nuclear annihilation.
And even dialing down the question difficulty further — hey, research is inherently messy and unpredictable, let’s ask how to do projects in a domain that we understand much better. Civil engineering is about as predictable as it gets, since every single piece of steel has to be carefully planned in advance or you get serious problems, and we’ve been doing it over and over, all over the world, for hundreds of years. And yet, if you look at subway construction projects, the cheapest projects cost (PPP-adjusted, per km of tunnel) a solid two orders of magnitude less than the most expensive ones. And nobody knows why. And I’ve yet to meet anyone that could predict, without being told the answer in advance, which ones are the cheapest and the most expensive — the answer is that the cheapest projects are in Spain and the most expensive ones in New York City, but no one will ever, ever guess that (again, please do try it yourself if you have time).
And the rationalist/x-risk crowd doesn’t seem to be doing much better at this than the status quo. Eg., for years now, some very smart people I know (who I have great respect for) believed that a big component of scientific success was doing very well in childhood math contests. I spent a few hours looking up the historical data here, and it turns out that this is false — a large number of such people become great mathematicians, and a smaller number go into various branches of software development, but essentially zero of them go into science (in fact, even among International Physics Olympiad winners, way more seem to go into pure math than into physics). I don’t want to make this idea sound stupid — it was certainly worth testing — but it was in fact wrong, and it was wrong in a way that anyone could disprove with two hours of Googling, and yet people still believed it for years.
So, what can we do? We need to figure out how to answer questions like this. And the data we’d need to do it well is more-or-less nonexistent. And the only ways we have of collecting that data would be mind-blowingly expensive — you can imagine the amount of time it would take to acquire and read through the personal notes of each of the US’s 20,000 physicists for every month for the last fifty years. Google’s Knowledge Graph would be perfect for something like this, if it had the data we needed — you could more-or-less just write code like “ DEFINE research_productivity(x) = {some function of x}; FOR ALL x,y LIST correlation(research_productivity(x), y(x)) ”, and narrow things down that way — but it doesn’t. There are a few custom-written pipelines at Google to import data from Wikipedia and so on, but nothing that could import data into the graph database from things like arbitrary Excel spreadsheets or freeform text. So a few months ago, while our project was in the middle of being canceled, I went looking through the academic literature to see who had addressed that problem. And I found that, back in 1998, there was a very interesting paper on precisely this topic called DIPRE . It was written by a Stanford grad student in computer science, named “Brin, Sergey M.”; unfortunately he doesn’t seem to have followed up on it, he must have gotten distracted or something. (A few other people have in fact followed up; the paper called KnowLife has a nice review of some relevant literature.)
Now, even with great software, doing all that research is still going to cost a lot, and that means you need a funding source. From a financial perspective, the most lucrative market for automated graph-based information extraction is (as my current best guess) medical data and internal data for large corporations. But the problem with those is that it’s playing on expert mode; the sales cycle is very long and difficult, and the information you want to analyze is horrendously complex (eg., ICD-10 alone has 68,000 separate codes). My model is that you need a gradual ramp-up through easier (but still profit-generating) problems that leads to solving the harder problems. Eg., I’d guess that doing sales for something medical but still fairly capitalism-driven, like laser eye clinics, is more-or-less the same as doing sales for other small businesses; from there you can move to individual doctors; from there you can move to small practices; and so on and so on until you get Kaiser and Medicare. Likewise for business management and government data. (Doing this for science, academic economics, journalism, etc. is likely to be extremely high-positive-externality but not Google-level profitable, so I’m ignoring those for financial purposes.)
So the place I’ve been looking at, where it seems fairly straightforward to get started, is in CRM (customer relationship management). From the data we have, the average salesman currently spends about 4–8 hours a week doing manual data entry on their CRM, and despite that still forgets to enter relevant data ~50–75% of the time (depending on which survey you believe). Per HubSpot’s 2014 survey, people’s #1 complaint about their CRM software is manual data entry; their #2 complaint is not being able to integrate their CRM’s data with data from other apps; and their #3 complaint is invalid/incorrect data. (Surveys can easily be biased, but this matches my own past experience in trying to find a CRM for MIRI and MetaMed very well.)
By itself, automating away all that data entry would be neat, but not transformative. The really powerful thing is that, as with studying science, it makes everything else you might want to do with data analysis way, way easier. Eg., CRM-based sales pipeline forecasting is something Base and several others have already built, but it becomes much easier and more accurate with lots of automatically-entered, high-quality data. CRM-based automatic lead scoring is something Infer has already built, but with better data it gets way easier and more powerful. Personality analysis is what Crystal , a startup Moshe’s friend Barry mentioned, already does (though it isn’t really a CRM), but the headroom you could get from better data (Crystal doesn’t even read from your inbox at all) plus using real academic psychology is “all of it”. I don’t think anyone’s even tried using recurrent neural nets for CRM yet, but that becomes much easier too, plus customizing products for each customer, plus marketplace matching of sellers to buyers, plus web analytics, plus social media integration, plus advertising system integration, etc., etc., etc.
The customer support space is interesting, IMO, because like CRM it’s (a) a significant market in financial terms, (b) there’s a large number of smaller businesses doing it, which are easier to sell to than giant Fortune 500s, and (c) the types of information in a customer support request are fairly similar to the information in other requests, which makes doing the information extraction and machine learning way easier (there might be, say, 30 common complaints, but this is much easier to handle than 68,000 ICD codes).