A year of political surprises that wrong-footed almost every prediction market, and a year in which a Go program defeated a world champion in ways that surprised even the researchers who built it. The lesson is the same in both cases: the world's events have outpaced the consensus that was supposed to anticipate them.
On Brexit, and What Capital Cannot Be Wrong About
The June referendum result was, by the morning of the vote, considered impossible by every prediction market we monitor. By that evening, it was history. We have a London office that, depending on the eventual exit terms, may need to be substantially rethought. We had not modeled this scenario.
The capital-allocation lesson is uncomfortable but useful: the consensus probability assigned to a tail event is the probability the consensus assigns. It is not the probability of the event. We have known this in theory since 1997. We are knowing it more sharply now. The post-referendum reaction in our portfolio was instructive — several of our European companies saw multi-quarter operational disruptions as their cross-border partnerships were re-evaluated, while others were largely unaffected. The variance in outcomes was higher than our framework had suggested it would be, which means our framework had been quietly assuming away the kinds of jurisdictional shocks that the year delivered.
The London office itself remains in place. We have not yet made structural changes; we are waiting to see what the eventual exit terms imply. The team's mandate is unchanged. The longer-term question is whether London remains the right base for our European operations through the implementation of the new regulatory regime; we are reviewing this in 2017. For now, the office continues.
On the Election, and What We Have Stopped Predicting
The November US election delivered an outcome that, like Brexit, was inconsistent with the consensus probability distributions of the prediction markets, the polls, and the major financial institutions whose positioning depended on it. We had no edge in predicting it; we will not pretend otherwise.
What we have updated is not our political model — we do not have one — but our framework for evaluating regulatory risk. Categories of companies whose business models depend on the durability of trade frameworks, immigration norms, or international supply chains, now require a wider variance band in our underwriting. We are widening it. The change is operationally significant for our Asian portfolio, which has, until now, assumed cross-Pacific frictions that may not survive the next four years; for our European portfolio, which has assumed a stable EU-UK relationship that will not exist in 2019; and for our deep-tech portfolio, which has assumed the durability of immigration patterns that are now under direct political pressure.
None of these adjustments are dispositive. They simply require us to underwrite at lower confidence than before, which means lower prices on average and more conservative pacing. We are accepting the implications.
On AlphaGo, Read Slowly
The March match in Seoul, in which AlphaGo defeated Lee Sedol four games to one, is the most important event of 2016 to anyone with a long position in artificial intelligence. The result was decades ahead of where the AI research community had publicly claimed Go-playing was possible. Either the community was sandbagging — which we doubt — or the recent improvements in deep reinforcement learning have produced a discontinuity in capability whose implications run further than the game itself.
We are now actively meeting AI founders in a way we were not a year ago. We expect to deploy meaningful capital into this category over the next five years. The 2015 letter described our deep-tech mandate; the 2016 letter is, in retrospect, the year in which AI specifically became the largest single allocation within that mandate. We have made four commitments in 2016 that we would not have made before March — companies whose business cases depended on capabilities that the AlphaGo result demonstrated were closer to commercial reality than the consensus had assumed.
The most useful conversation we have had this year was with a portfolio CEO whose company had been ambivalent about whether to develop machine-learning capabilities internally. After watching the Seoul match together, the conversation shifted from "should we" to "how fast can we." Twelve months later, the company's machine-learning team is its largest engineering function. We expect this pattern to repeat across our portfolio over the next three years.
On the Sectors That Are Now Real
Three sectors that, three years ago, we would have characterized as speculative — autonomous vehicles, applied machine learning, and certain categories of synthetic biology — are now producing companies with real revenue. The transition from speculation to reality, in each case, happened faster than the consensus predicted. The pattern is the same in each: the underlying technical capability crossed a threshold, the cost curve flattened, and the founder cohort capable of commercializing the technology became visible.
We have positions in two of the three sectors. The third we are developing diligence in; we expect to make our first commitment in 2017 or 2018. The companies that will dominate these sectors in 2025 are, in most cases, being founded now. Our 2015-2017 vintages will be heavily weighted toward this cohort.
A Closing Note
The most useful posture, in a year when consensus has been wrong about the largest events, is humility about everything else the consensus is currently confident in. We are practicing it. The 2017 letter will, we expect, document where the practice succeeded and where it failed.
The Partners
Winzheng Family Investment Fund · December 2016