I Don't Believe in AI Venture Funds

November 2019 · 6 minute read

I’m not really into “AI VCs” on a fundamental technical level.

To me, they either went for a buzzword-pitch when raising money for the fund (as startups are quite used to do) or are trying to do something that doesn’t work.

What is AI useful for?

Current AI technology is very powerful when deployed for one goal: massively scale jobs that humans can already do pretty well.

Humans perform some tasks very slowly as the number of times they have to do them increases exponentially. This is the case of labelling spam emails and driving cars. A machine can improve the time needed to accomplish such things by orders of magnitude, due to its significantly higher computational power.

At the same time, the marginal cost of adding more machines and thus of gaining speed by concurrency is a fraction of adding more humans to solve the problem. That’s why scaling can happen economically too: it’s just cheaper to add a self-driving car to a fleet than a human driver.

Sometimes, when scaling what people are already doing, the AI algorithms take an approach that differs from the humans’ one, and this can lead to interesting insights (like AlphaGo’s famous 37th move). That’s a byproduct though, not the core value proposition like in the case of other branches of the same science, like Statistics, Psychometrics, Econometrics (more focused on causality and analysis, rather than correlation and prediction).

AI can pull off this kind of magic by finding patterns in large troves of (past) data. In the end, it’s all about being able to keep “in mind” gigabytes of data – be it thousands of past driving hours, thousands of past Chess matches, or thousands of past trading activities – and go through it looking for patterns in a reasonable time.

What today’s AI is very bad at is transferring domain expertise: Creating brand new concepts by getting what has been learnt in a domain and applying it to another one is not there yet. And this happens not only on a domain level but also on a time one: if what you’re looking for isn’t in the past, it can’t be predicted. In the industry, that’s the ‘overfit problem’.

(Incidentally, such transferability issue is one of the main bottlenecks stopping us from getting AGI and the so-called singularity)

How does this relate to the VC market context?

The market in which VCs operate in has a few characteristics we should consider:

First, successful startups follow a quite dramatic power-law curve, and the truly successful ones are in the order of 10s. The data is crazy sparse, and for such reason extracting the signal out of so much noise is an onerous – sometimes impossible – problem. Among many others, Mike Maples and Ben Horowitz wrote extensively about it.

Second, successful startups disrupt new markets. If any, past data tells us where not to go, since it represents mature opportunities that won’t yield the returns VCs usually promise. The next unicorn will happen because someone will transfer the same business methodology / some technology to a new domain.

Third, the startup investment market is extremely illiquid, deals happen slowly (think of HFT for a comparison, where speed is relevant indeed). Improving speed from days to minutes is irrelevant and pricing market data is not there.

These characteristics all lead to a conclusion quite relevant with how AI works: the data is not only sparse in terms of sample size, but also in terms of depth.

There are not so many well structured and relevant data columns that can describe startups uniformly.

Sure, all the engagement and financial metrics are there. Also, the startup’s market reach can be (somehow) quantified by using social numbers as proxies (I guess). Founders background, personality and preparation can also be taken into account, as the overall market quality in which they’re operating do (with a grain of salt, though, as described before).

However, all of this must be ultimately taken into account within its context. A 25% weekly retention rate after 16 weeks can mean something for a company, and something completely else for another. A 50-years-old founder can be a crazy bet for a Gen-Z social app (maybe), but a great fit for a retail logistics B2B company (maybe).

To me, this lack of deep data means that it’s more effective to operate with manual filters rather than to scale such filtering with a classifier. I can’t really see a case for scaling a fund’s deal-flow or assigning an investment probability score by using this kind of technology.

To me, it looks more like finding a problem for a sexy technology few people understand rather than the other way around.

Wrapping up: Why I don’t believe in the AI VC model

Startups are represented by sparse and thin datasets, their deals happen incredibly slowly, the past is misleading and must be considered within its context. Ultimately, this is a job that even humans can’t really accomplish well (“you can’t pick the winners”), and that’s why it’s incredibly difficult to automate.

AIs, on the other hand, are good at making things fast thanks to big and deep data, while are pretty bad at generalizing their domain expertise and at performing tasks humans can’t do even at a micro-level. They will disrupt low-skill jobs pretty fast and pretty dramatically. With the current state of technology, they won’t do it with high-skill ones.

The two worlds technically diverge.

Don’t get me wrong: Data can be most definitely a powerful tool for the venture market. I’m a data guy myself – if I’ll ever be working as a VC, Python will be my best friend.

However, I see the intersection between data and venture money as a successful bet if and when used analytically: supporting a team of humans through synthesized information and filters. There’s no automatic prediction there. It’s more about filtering top-down large amount of companies to make the deal-flow larger and more efficient. To me, this is not AI, nor ML.

Maybe I’ve just a naming problem?

A final note

Intellectual disclosure: for some of the reasons outline before, I’m deeply convinced that a VC should focus more on the quality of the deals it sees than on the quantity, and because of efficiency it should do it inbound rather done outbound.

Inbound+quality is diametrically opposite to marketing automation, and that’s why I’m writing from a position of skepticism: I don’t see the point of automating any part of the VC job. However, if that’s your thing, maybe using AI can be a pertinent technical choice.

A successful way of combining the two worlds together can be having an investment model that involves a ticket size significantly smaller than the market average and a yearly deal number significantly higher (in the orders of several hundred per year). That could be even more profitable than a traditional VC because it drowns the domain-transfer problem with sheer volume.

(Incidentally, that’s exactly the YC investment thesis: very large amount of small-size deals, as much automated as possible while creating the deal-flow by using inbound marketing)

That’s just my way of seeing things, though: an opinion as any other can be. Very much open to being disproven!

Follow up

Thank you for reading. If you have any questions, feel free to shoot me an email at blog@giansegato.com or tweet them @giansegato.

I've a non-lame 0-bullshit newsletter on startups, tech-enabled scalability and data-driven impact. Sign up to stay up to date and to keep this conversation open