Learning from Overdetermined Failure

Learning from Overdetermined Failure

In an interview with Tim Ferriss, Peter Thiel makes the argument that it's hard to learn from failure because it's overdetermined: a company can fail from doing one important thing wrong, like selling a product with negative gross margins or making a blood test that doesn't actually work. Or they can fail because they did fifty things wrong. If you "learn from failure," you do learn something, but it's hard to tell if you've learned enough. (Even for the lessons above, there are limits to how much you can learn: before it charged for transactions, Paypal was structurally unprofitable because there wasn't enough float; and in some kinds of software, failing to perform a stated function can mean fulfilling an unstated one—in expense management software, for example, great UX saves employees time but costs companies money.[1)

You can also look at the failure question by inverting it. Here's a nice little story from Liar's Poker:

[Junk bond pioneer Michael] Milken often spoke to students at business schools. On these occasions he liked, for dramatic effect, to demonstrate how hard it actually is to put a large company into bankruptcy. The forces interested in keeping a large company afloat, he argued, are far greater than those that wish to see it perish. He'd present the students with the following hypothetical situation. First, he'd say, let's locate our major factory in an earthquake zone. Then let's infuriate our unions by paying the executives large sums of money while cutting wages. Third, let's select a company on the brink of bankruptcy to supply us with an essential irreplaceable component in our production line. And fourth, just in case our government is tempted to bail us out when we get into trouble, let's bribe a few indiscreet foreign officials. That, Milken would conclude, is precisely what Lockheed had done in the late 1970s. Milken had purchased Lockheed bonds when the company looked to be heading for liquidation and had made a small fortune when it was saved in spite of itself...

What can you learn from that? Nothing matters and business is fake? Maybe it's specific to the company, or to 1970s defense contracting in general—perhaps "indiscreet foreign officials" are not the only corrupt parties out there. Maybe big companies really are that durable, and the successful ones we see are the cumulative result of a) massive economies of scale, coupled with b) slightly-less-massive levels of internal dysfunction.

All of these answers can be interesting, but none of them are useful outside of narrow domains like staying sane while working at a company that seems to be falling apart but that also keeps on growing.

In addition to the overdetermination problem, there are two other issues that make it hard to learn from failure: intersubjectivity is hard, and narrative bias is brutal.

Intersubjectivity just means that people have very different ideas of what vague terms like "working as hard as you can," "always treating customers fairly," or "building the best possible product" looks like in practice. One person's "best" is not the same as another's, and one person's idea of what level of goodness rounds up to "best" can also vary. This makes positive platitudes basically useless; some people who advocate a reasonable work-life balance have an unbalanced view of what that entails.

Narrative bias also throws things off. Consider a company that raises money in a hot funding market and spends almost all of it on salespeople to sell their product to other companies whose budgets all come from that same hot funding market. Imagine this company blew 3.4% of its biggest funding round on a party. And then suppose it used aggressive marketing tactics—when a competitor was holding an event at Cannes, they booked all the taxis in Nice for the night and had sales reps stationed in each one to pitch their product to people going to the competitor's event. This could be a cautionary tale about business aggression and excess, but it's actually the story of the early days of Salesforce. The company was aggressive, both in terms of how they pitched their product and who they chose to sell it to. That aggression sometimes came back to bite them; many of their customers went bankrupt in the aftermath of the dot-com bubble, and they switched from monthly to annual billing in part because they were worried they'd run out of cash. The reason it's not a cautionary tale is that it worked, but it worked because the company was able to recognize which of its choices were mistakes, and which were only temporarily useful.

If failure is overdetermined, and prone to narrative bias and tricky intersubjectivity issues, is there anything we can actually learn from it?

We can, by inverting these. For example, one way to look at that narrative bias is that it's a tool you can use to identify points when a company was one bad decision away from failure. Those can give useful lessons. Salesforce's choice to switch from monthly to annual billing, for example, embeds some important lessons:

  • Cash flow matters more than you think, because an easy funding environment lets businesses defer tough decisions about cash.
  • Customers may choose one product over another by a slim margin, but the cost of switching is higher than that, so raising prices or restructuring deals is not as intimidating as it looks.
  • Companies can be fond of the ways they're differentiated from their competitors—early Salesforce touted the benefits of the low-commitment monthly-billing model—but companies can only be differentiated if they're not dead.

Inverting intersubjectivity means admitting that someone's retrospective statements about their subjective experience are not likely to be a great guide to your own actions, and can be heavily discounted. Many people who are unusually effective in one domain underestimate how much their behavior seems compulsive to people who aren't (this applies equally well to salespeople and coders, though for very different behaviors—trying to befriend everyone you meet and running a simulation to create the optimal Wordle strategy are both things that will strike some people as perfectly normal and others as deeply weird).

Overdetermination is the hardest one to invert, both in theory and in practice. If bad companies often have many bad things wrong with them, then the inversion of this is to make no mistakes whatsoever, which is both pointlessly over-broad and too high a bad. But there's a narrower version: part of the investment process is often narrowing a company down to a handful of questions and getting very good answers to them. There are, for example, lots of arguments to be had about individual decisions Netflix makes, and the success or failure of specific shows, but the only broadly important questions are 1) can they continue to expand their subscriber count, 2) have they reached the limits of their pricing power, and 3) with both of these put together, are they the default high bidder for shows and movies people want to watch? Their newest top 20 holder has presumably been sitting on answers to those questions for a while, and reached the point where the quantifiable answers indicated that Netflix was trading at a favorable price.

At the company level, this means breaking down essentials into three broad categories:

  1. Things that have to be done, but that will never be a competitive advantage,
  2. A very short list of things to maximize rather than satisfice, and
  3. An idiosyncratic list of fairly arbitrary things to maximize because they symbolize point 2.

Some companies have a culture of obsessing over seemingly minor details: insisting that every task, however minor, have a single directly-responsible individual; demanding that people adhere to formatting guidelines for documents, presentations, and financial models; insisting on rapid turnaround for emails; turning every one-off decision into a policy and putting that policy in a wiki; etc. You can view this as control-freak behavior from founders, and there's probably some of that in the mix. But it's also a form of risk control: if failures tend to do multiple things wrong, then one response is to make a list of things that matter and insist on doing all of them exactly right. Picking relatively minor details to obsess over is a sort of enforced ritual acknowledgement of the important ones—it would be incredibly embarrassing to spend an hour formatting a presentation if that presentation had a material mistake in it, so over-investing in the form increases the cost of getting the substance wrong.

This kind of obsession turns out to be a form of failure-proofing a company. A company that fails gets almost everything wrong, so the opposite of that is to do everything right, but the implementation of that is to carefully define both "everything" and "right."

Diff Jobs

Some recent roles from companies in the Diff network:

  • A company tackling fraud is looking for a Snr. Content Manager to help with SEO, writing content, PR and website strategy. (Remote, European time)
  • A startup offering a new kind of product in the FinTech/InsurTech space is looking for a full-stack Node.js developer/API developer. Experience coding in an AWS serverless environment is a bonus. (remote, US time, Toronto).
  • An alternative data company is looking for data scientists, data analysts and data consultants. Senior and junior roles. (NY, remote)
  • A startup fractionalizing the ownership of real estate is looking for full stack engineers (Node.js/Typescript). (remote, US time)
  • A firm that helps investors use unique data sources to find and refine investment ideas is looking for analysts, data scientists and engineers who can collect data and analyze it to add insight to trades. (US, remote)

Elsewhere

Blackstone Buys the Dip

Blackstone Group reported healthy earnings yesterday, and suggested that they're looking at opportunities among beaten-down tech stocks ($, FT). A cash-burning growth company is a PE opportunity disguised as a VC opportunity: in growth mode, it's ramping up operating expenses to buy a durable stream of cash flows, while in levered-steady state mode it's reducing that growth investment and harvesting those high-margin cash flows instead.

As has been observed many times, especially recently: investors tend to give money to managers when recent results are good and pull it out when they're bad, so the return of the average dollar invested in a given fund is usually worse than the fund's returns over time. That's especially apparent in an industry with a fast capital deployment cycle. But PE's investing pace is slow compared to ETFs or hedge funds, and a PE fund's life is typically longer than a full economic cycle. It's possible that this is one driver of private equity returns: the easiest year to raise money is the year before prices drop, and if they continuously raise larger and larger funds, then their assets under management always peak when deals are most attractive. (The tradeoff being, of course, that they have to find a way to get acceptable returns throughout the cycle.)

Growing an Ecosystem

Calling industries "ecosystems" is a good habit, because it's a reminder that ecosystems work through flows rather than accumulations. The resources that get used are continuously getting reused somewhere else. In tech, one important part of this is that experience plus cash from one cycle turns into venture money and advice for the next cycle; at IPO, Apple's cap table had money from the early chip boom (Mike Markkula) and the earlier defense-tech explosion (Henry Singleton). This is happening in Africa, too, with founders recycling money from earlier liquidity events into angel investments in new companies. An acquisition that looks early from the perspective of a US company can be much more material in the context of a lower cost-of-living country where getting to scale is substantially cheaper, and the faster cycle from founding to potential exit means that a new venture ecosystem can spin up faster than used to be possible. So the current funding and acquisition environment is especially conducive to new startup clusters forming.

Shortages are Demand-Side, Not Supply-Side (Again)

The standard argument on the automotive chip shortage is that car companies halted production early and restarted it late, and they're still catching up. Almost two years since the start of responding to the pandemic, that explanation is getting a little old, and another one is starting to matter more: the chip shortage is the result of many more chips per car, not just swings in car production. 2021's total automotive chip shipments were well above 2020 but also slightly above the pre-Covid growth trend, so what we're seeing is more the result of the long lead time for adding new capacity than the relatively short lead time for adjusting existing capacity.

Pricing out Values

There are gaps in behavioral norms between countries, and sometimes there's pressure to close these, which usually takes the form of (ineffective) rhetoric or (hard-to-target) sanctions. A good Coasian might ask: why don't richer countries just pay poorer countries to set different rules? That has all sorts of difficult political and diplomatic ramifications if it's done directly, but it can also happen as an indirect result of the interactions between disclosure rules and supply chains: Nestlé is paying cocoa farmers directly as part of an effort to get them to stop using child labor ($, FT). European governments can't directly ask other countries not to use child labor, but they can require European companies to disclose whether or not any of it is used in their supply chain, putting companies in a position where the cheapest thing to do is to pay extra in order to have clean disclosures.


  1. Expensify and Concur can coexist in a world where this is true: Expensify can sell to companies that have an expense policy that aligns with what kinds of spending are either valuable to the business or are a cheap way to improve morale, and Concur can sell to companies that have an ostensibly generous expense policy and expect it to be abused.