SBF, LoL, and risk management: a sidenote
A week ago, as Sam Bankman-Fried’s empire started to crumble into dust, the FT’s Alphaville section ran an exposé that shocked the world:
While many of the revelations over the last week have stunned amateur and expert observers alike, this one did not; as both a former pro LoL coach and a crypto expert trademark copyright, naturally, the first thing I did upon learning was op.gg him. He sucked (and was forthcoming about being so), haha funny, moved on.
However, the revelations made me look at the account again; and, in doing so, I found something that was extremely funny if you have sufficient context of all of:
a) LoL in general, and especially high-elo play.
b) The mechanics of Alameda/FTX and how it came crashing down.
c) SBF’s ‘effective altruism’ philosophy and the logic stemming from it.
It’s this:
Unfortunately, as I quickly found out when I tried to explain this to people,
1) I am a huge nerd.
2) There’s, like, two people this applies to.
However, there do seem to be more than two people in the world who would be interested in this, so in quite possibly the least important rejoinder to the whole SBF saga, I am going to explain what is so funny about that screenshot, and we’ll learn a little about everything along the way.
One of the many, many things that came out (or, rather, was brought to light) in the whole FTX implosion was a thread and later discussion that took place between SBF and engineer-cum-investment manager Matt Hollerbach on utility theory and something called Kelly bet sizing.
Matt put out a very good thread after the fact dissecting his arguments here, but the short version is this:
- SBF put out the thread (in December 2020) basically as an explanation of one of the core tenets of his personal philosophy in relation to effective altruism: summarised, if there’s a 51% chance of winning, pursue that over and over because it’s ++EV. You can agree or disagree with that, but it’s mostly logically consistent.
- The red flag here was how SBF treated Kelly bet sizing. In short, if we take 1) as true, you still wouldn’t put 100% of your wealth on that 51% chance because it wouldn’t represent a maximisation of your long-term earnings. If you’ve played poker, traded, played an esports game at a high level, or done any one of a number of other pursuits, you should already get the concept here; it’s just bet sizing.
Essentially, in his own words, SBF’s internal compass here put his tolerance for risk on a given bet at about 5 times that which Kelly bet sizing would say so — an immense appetite for risk, and an immense appetite for coinflips.
This risk appetite is reflected in how everything came crashing down in the end for SBF, FTX, Alameda, et. al. I’m going to avoid covering that in detail, because almost everything that’s worth saying — and can be said — in public has already been summarised in this post on website milky eggs dot com for some reason.
The one thing I will add/emphasise is that a large part of what made the FTX balance sheet start falling apart was that it had made a succession of venture capital bets on almost every project that asked for it in the Solana ecosystem in 2020–22 — bets that were precisely the sort of ‘5% chance of working for a 25x return’ that SBF gushes over there. In a just world, it would have been the fraud re: comingling of deposits, blatant misuse of customer funds, etc. that stood alone in bringing it down. In this world, all of this largely went unanswered for until all of those projects started to fail and hence even the book value of said Solana assets didn’t hold up anymore. Just something to bear in mind and get silently/vocally angry about.
In any case, the point is that the risk appetite here wasn’t a sideshow to the whole disaster, it was central to it. Perhaps obvious, but just a thought to make sure to hold onto.
So, onto LoL. If you’re still reading, you’re probably not one of the people in the audience who knows LoL at all (because you’ll have got what you came for anyway), so I’ll try to explain this as simply as possible.
LoL is a 5v5 game where your goal is ultimately to destroy the enemy’s base. In the early days of LoL (and even moreso in other MOBAs like DotA), there was a ton of variations in terms of strategies you could pursue towards that end.
However, the early game in particular has become much more regimented over time. There are three routes to the enemy’s base in the game, known as lanes, and players are arranged as follows:
1 player each in the top lane.
1 player each in the mid lane.
2 players each in the bottom lane.
1 player in the ‘jungle’.
In the first 10–15 minutes of the game, the four ‘lane’ players will spend 90% of their time in said lane, faced up against the corresponding four players on the other side (with the ‘jungler’ as a wildcard).
This is an extremely deliberate design choice on the part of the game developers, because by enforcing this structure, the developers can ensure a more consistent playing experience for players from game to game.
Each lane, and its assigned role, develops a certain feel, rewards and punishes certain virtues and vices, and hence players often develop a certain identity around said role, and in particular, the way that they personally play it.
So, why’s the SBF screenshot funny? Well, SBF primarily played support — one of the two roles in the bottom lane alongside the AD carry. This was not, initially, by choice. One of the more overlooked pieces of the lore in the post-FTX reckoning has been SBF’s college blog from 2011–12.
It’s about half utilitarian musings and half baseball sabermetrics, which is about the most 2011 blog thing possible, and his third-to-last post at the end of 2012. He did, however, made two posts on it after that. One of these was a post on game theory with explicit references to both LoL and MTG, which deserves an entire dissection in itself, but he signs off with this:
AD carries were, historically, the rockstars of LoL; they were protected by the presence of the support in the early game, had the fewest responsibilities in the mid game, and dictated who would win in fights in the late game. The role demanded the most ‘mechanical’ ability of any role, and either the smallest or most specific subset of game knowledge, depending on how polite you feel like being.
That he wanted to be the rockstar is, despite appearances, not that appealing or interesting; just setting the context. He then mentions that he should maybe move to bruiser, which essentially means the top lane role.
However, what we can see from that graph is that where he ended up actually playing most often is the support role. I would imagine this is probably a function of the role being the least contested and hence the quickest to queue into, since he still played every role except mid (the most popular one) quite a bit, but let’s explain support.
The inverse of the ADC in many ways, support is generally the least demanding role mechanically, and in general is a place where you can do a lot with little compared to other roles. However, a truly great support player needs to have knowledge beyond knowledge — a read and instinct for the game that ends up surpassing the conscious. (To state my own bias here: I was never a support player in my Diamond peak personal play, but my positional coaching work on TSM was with the support player at the time, so I’m particularly well-accustomed with how it works at the top level.)
However, as mentioned, there are different ways to approach the support position. There are three things to note here. The first is the fact that his ‘gold’ and ‘CS’ scores are very, very high. This is generally not something you particularly want to see from a support player; most high elo supports will average around 20–30 CS a game.
Why? CS is short for ‘creep score’; essentially, little guys come into your ‘lane’, you kill them, you get gold for it. Creeps come in at a finite rate. If you, the support, take creeps, then that’s usually a creep that your AD carry didn’t get to take; and, because the support player gets advantages from not taking creeps, and because the ADC is typically the player on the team who gets the most benefit from getting a higher share of the total available gold, that’s -EV for the both of you (unless you are so, so much better than your ADC that you’ll use the extra resources more effectively, which, hint: he wasn’t).
The second is his KDAs. KDAs don’t perfectly correspond to effectiveness, but a low KDA, especially on support, is a sign that something’s weird. Let’s say you fought the enemy bottom lane, 2v2, over and over, and each time, all 4 of you died. You would each accrue:
0–2 kills or assists (both enemy laners)
1 death (yourself)
giving you any of these scores:
2/1/0 (2.0 KDA)
1/1/1 (2.0 KDA)
0/1/2 (2.0 KDA)
A 2–3 KDA is hence generally normal in the long term. 1.52, 1.54, 1.04, 1.20…not so normal. There are roles in which you’d expect to have a lower KDA (particularly top lane with certain playstyles), but supports are more likely to have inflated KDAs than low ones.
The third point is the champions he choses — Zyra and Brand. This is the fun part. See, the support player goes with the AD carry because the AD carry is weak early on. This, in turn, means that the way that the early game plays out in the bottom lane is largely dictated by the actions and approach of the support player, and the tools they have through their champion.
For a champion to get played as support, it is hence useful if they offer a lot of flexibility in how the lane can be played — aggressively or passively, to attack the enemy ADC or protect your own ADC, and so on. In particular, most supports need to be able to offer some explicit protection to their ADC; make them quicker, heal them, get in the way of attacks, whatever.
Then there’s what’s known as AP supports. These are champions, played as support, that basically just do damage (AP, or ability power, being one of the two types of damage in the game). They may have some slight utility outside of that, but mostly, they just do damage.
Now, when this comes off, it can be great; if they get an advantage, they can ‘snowball’ (i.e. win -> advantage -> win) the lane to the point where the whole game can be decided by it. Usually, it’s not great. If they even go even against a conventional bottom lane, that’s usually a disadvantage, because AP supports are generally balanced around playing in a different lane and getting gold for themselves — so, in the mid and late-game, you now have MUCH less impact than the enemy support does.
Everyone, historically-speaking, hates AP supports, to the point that if a Zyra or Brand shows up in pro play, something’s gone badly wrong. (Zyra was widely played in season 7 of LoL in 2017 but has not seen substantial high-level play since; Brand has only ever seen fleeting play). The calculus on this is not tricky; if you’re substantially down in an even scenario on picking a champion, you usually shouldn’t be playing that champion, for the same reason that most chess players don’t open with 1.a3 even though it’d definitely take someone out of book.
Further to this, AP support players — at least for the period I was active in LoL — always drew ire among high-ranking players even when they won. Why? Given the all-or-nothing nature of the champion pick, and how it locked things in strategically, it would mean that the AP support player would always try to effectively win or lose the game in the early game, typically in 2v2 or at most 3v3 battles around the bottom lane in the first 10 or so minutes. If you were in the other two lanes, and often even in the jungle — tough luck, you now have very little influence on the result of the game. (for reference, LoL is very much a ‘win-more’ game where an early lead in one area of the map can mean the entire game is decided)
This sort of player was an entire archetype onto itself. It wasn’t always just AP support players, to be clear, but because of the nature of the lane and of the role, they were most closely associated with it. I can remember so many conversations from team houses throughout my career where this pro or that pro would complain about the latest such player climbing the local ranked ladder.
That was, of course, a part of it too — if executed well enough, while -EV on a long enough timescale and against good enough players, you would see plenty of players rise to the top of the ladder in their regions on the basis of this type of play. A few would stick around, either because they changed their approach when they ‘made it’, or it wasn’t a totally accurate label in the first place. Most, however, would end up sinking back.
Now, do you want to guess what the term for that particular archetype of player was?
Go on.
‘Coinflip player’.
Funny how things go, huh?
