Optimism for Realists: The Optimism Heuristic
I’d like to make optimism more widely available, because focusing your attention on wild success outcomes is the easiest way to 100–1000x your impact. Optimism is not some incremental 10–20% improvement that’d be boring.
Naive optimism, the unrealistic belief that wild success is inevitable or highly likely, seems very valuable to those that hold it. Naive optimism is the primary route to optimism overall, but since it relies on suspension of disbelief, it’s not available to everyone. It wasn’t available to me, and so I’ve had a longer road to optimism than most.
The optimism heuristic is a clear path for realists (people who strive for accurate beliefs) towards getting the key benefit of optimism. In a nutshell, it consists of doing the following to decide between projects:
- Do your best to estimate equivalently likely wild success outcomes for each project (say the 95th percentile outcome).
- Then work on the project with the best 95th percentile outcome (value/cost).
The optimism heuristic is provably [2] good at maximizing expected value in situations where upsides vary by orders of magnitude and the downside is bounded. And key domains like startups, philanthropy, and science have such a distribution where the majority of returns are in the wild success outcomes.
A long-standing topic of debate that I have been deeply interested in, is the benefits of optimistic beliefs vs realistic beliefs. And when I came across the optimism heuristic, I felt a rush of insights and lots of tingling. Sources of tingling:
- Optimism, which I’ll define as focusing on wild success outcomes, is not just helpful for rallying [3] talent, customers, and money around you. Wild success outcomes are the correct outcomes to be giving attention to when trying to maximize impact given unbounded upside and bounded downside. So in those domains, I think the optimism heuristic is almost always worth considering and should often be the main decision criteria.
- I had been using the optimism heuristic all the time to prioritize my own work, but I hadn’t realized I was using some form of optimism. I’m at OpenAI, because the wild success outcome of succeeding in our mission would, in my opinion, be the most important work that’s ever been done.
- I believe one of the most important, under-leveraged truths is that people can quickly improve their probabilistic judgment based forecasts this heuristic relies on. This belief of mine caused me to train hundreds of the staff at Twitch in this domain, which eventually led me to my current role at OpenAI. So where someone else might see this heuristic as something cute that sounds smart, I see something pragmatic and deployable in an organization.
- OKR’s, never felt quite right to me. I thought it was because of the grading bit, which I’ve heard is actually political and weird. But no, that wasn’t it. My beef was that OKR’s were pushing me to focus on ~30th [4] percentile outcomes, which sometimes was a good idea, but often was not.
The optimistic atmosphere of Silicon Valley, that encourages people to focus on work that would otherwise seem too ambitious, is one of the Valley’s most valuable resources.
If I could hit a button and simply remove the naive optimism trait from someone, I would not hit the button. However, I think they may be in a local optimum. Because:
- It’s important to learn from reality. If they don’t learn sufficiently from reality, they could get stuck on something that initially looked quite promising (10% chance 100x anything you’ve ever done in impact), but now has a vanishingly small chance of such an outcome (less than .1%)
- Competent realists often reject naive optimism. They are worth working with, so it’s valuable to be able to explain why a project is obviously good in language a realist will accept. Naive optimism, statements like “this next project will be a wild success, 100%,” unnecessarily risk massive amounts of credibility with realists. Failing once, after such a statement is rough. Two or three failures in a row given those expectations can be catastrophic (team exodus, the leader has to leave company, etc).
- The other major benefit of naive optimism is the charisma you get from exuding massive amounts of self-belief. I think that’s a good and common objection people make while defending naive optimism. However, self-belief can be decomposed from optimism and cultivated separately. I’ve seen realists with an unshakeable belief in their own competence. They exude that same charisma that naive optimism yields. I’m not claiming this is easy. It’s just part of my argument against naive optimism being globally optimum.
Ok. If you’re into this. If you’re picking up what I’m putting down. Here are some suggested next steps.
- Use the optimism heuristic to help you evaluate or advocate for a project you are excited about.
- Use the optimism heuristic to evaluate your current job/role vs other things you could do.
- Take calibration training, so that your 95% predictions are reliable.
Thanks to Darragh Buckley, Josh Albrecht, Tom Brown, Alex Matzner, and Nadia Egbal for feedback on this post.
Endnotes
- Exit amounts were estimated based on seed-db. Only companies seed-db classified as exits or dead were considered.
- The optimism heuristic has provably good performance in a classic computer science problem, the Multi-Arm bandit. The Multi-Arm bandit is a way of formalizing explore-exploit trade-offs: given a bunch of slot machines with unbounded, unknown payoff characteristics, which one do you pull. I don’t remember any of the other good performing heuristics from Algorithms to Live by, where I learned about this heuristic because they all felt way more complicated.
- Rallying people around you is worth A LOT in terms of hiring, fundraising, sales, etc. But if rallying people was the only benefit of optimism, that’d leave me feeling kind of gross about optimism.
- Since OKR’s are goals you can do 70 percent of the time, there’s only 30 percent of worlds they can’t be accomplished in, so they are 30th percentile outcomes. That said. OKR’s are great for coordinating people to incrementally improve something that is working (product market fit, a clearly successful line of research, etc). Google likes them, lots of companies like them, they are probably the current best choice for a company setting goals. I used to think the grading bit is why I felt weird about them, which I’ve heard is actually political and weird (I think the KR’s should all be observable predictions).