My goal is to maximize my expected value as a startup founder over the lifetime of my career — the next 30 years or so. First I became a world class engineering talent, then with my first startup I practiced execution in a known market so that I wouldn’t be learning both how to manage a new market and execution of a startup at the same time.

In this post I want to talk about the approaches I can take.

Constraining the Problem

Before I talk about the concrete paths I want to talk about how I’m bounding my decision tree. The topics of interest to me are ML broadly (especially small self-learning models), spatial computing (AR especially), health & longevity, full spectrum education, and robotics. Among these the timing is decent for many of these but the highest potential is in ML due to timing. The market is absurdly interested in GenAI but there are other powerful trends as well.

You can see 🚀The “Rocketship” Era — Going Big on Startup v2 for a bit of flavor.

Strategic Approaches

Ultimately, I want to work on foundational problems because the world needs to be re-framed. Most of our problems stem from poor foundational framing. For example, the flaw with education is it doesn’t go far enough. Who you are is more important than what you know — what you know is a second order effect. But our education system does not know what types of people it is trying to produce and so it fails to produce excellence of any kind. This type of framing problem exists at numerous institutional and technological levels.

How do I do that? How do I start making progress, building trust, and building capital I can deploy to change hearts and minds? How can I persuade industries to change? I feel the best way, as with most things, is to be the example and model change. Then collect accelerants as various kinds of capital. The best known vehicle for this is a startup and here are a few ways to go big.

1. Machine gun iterate on the idea pipeline. There is no shortage of opportunity. As I write this in 2024 GenAI and ML in general is getting insane funding. A rinky dink app like Characters.ai raised $150 million at a $1 billion valuation with no revenue and an unclear future. It feels like the internet era. There are billion dollar opportunities around every corner. The real question is what do I want to do for the next 10 years? Or can I build a buyable startup? The advantage of this approach is deliberately practicing finding founder market fit finding. I get to practice my intuition about what is attractive and what is not. I can do this with relatively shallow to start ideas that can scale. I need to make a secondary decision about the duration of this. The upside is more indications of capability with the potential for one of these seeds growing into a real opportunity.. The downside is time lost.
2. The “Lucky Palmer”. Disappear into my garage and emerge with breakthrough tech. I have a couple ideas that I think are huge in potential. The upside is having a true moat and being a first-order foundational player. The downside is half of invention is timing and these are extremely high risk especially with my limited resources. My interest here are in re-framing how ML is done and focusing on embodied small models that can bootstrap their learning one dimensions at a time. I won’t go into too much detail, you can read my white paper 🧠Universal Learning Machines.
3. Join a high profile startup building new products. For example, I could join the new products team at OpenAI. I would see how they make their cake, I would get paid to build frontier products, I would identify talent/make connections, I’d get a nice halo effect — investors love throwing money at people from high profile teams and cos, and I would get inspiration. The downside is the opportunity cost. Not only would there be time to make sure I can interview and land these roles but there is finding the right roles and then there is the cost of learning about all the special knowledge that doesn’t generalize and is unique to their organization. Financially this is the most conservative route.

What razors can we produce for ourselves though to make a decision? What has the highest expected value? In order to answer that question we need to frame the quantum of progress and the measure of value. I really just need trust. Trust from the right people to tackle the big problems I want to. All resources necessary will stem from that trust.

What is the surest way to build trust? Create a smash success such that people believe you can repeat it.

Expected Value

Let’s talk about expected value now. Assuming you can stay in the game long enough you should think in bets and that means using expected value.

The biggest risk with the “Lucky Palmer” is that I make zero incremental progress on building trust. In fact I probably hurt trust if I fail. I thought I had break out success potential but it failed to materialize. The odds of that happening are overwhelmingly high. Even though I feel certain I am on to a core framing problem in ML with 🧠Universal Learning Machines I need to navigate a few 0 → 1 technology breakthroughs. I think all the ingredients are out there. Certainly, this thesis, which is years old, has been validated by several indications in the marketplace. If I had the trust already there is no question I would do this. There are strong arguments for this having the highest expected value — it just depends on how you calculate the likelihood of success. I would call it a $100 billion over the next 10 years with a <0.1% chance of succeeding. So, expected value about $100,000 million.

The “Machine Gun” is interesting. I see on the order of a $100 million outcome as pretty high. I see it as at least a 20% chance of success. Just because I’ll iterate until I find fit and will only consider ideas with that potential. That gives me a $20 million expected value.

The third option increases my odds on the “Machine Gun” and may have a slight impact on the “Lucky Palmer” at the cost of time. I’m viewing this as a backup option.

If I think the expected value for the “Lucky Palmer” is higher than what’s the decision?

Expected value is great when you get to play many hands such that you can eventually win. But if you can’t last long enough it ceases to matter. Let’s say it takes 1-2 years to kickoff a “Lucky Palmer”, then in my career, I get about 10 shots. Let’s say my odds double every time I complete an iteration. I still have a high potential of ending up with no success. And if I fail in the early days using my own capital I may cause trust to evaporate rather than compound impacting downstream odds of success.

Options

Now, these don’t have to be exclusive they could be sequenced in time. The most conservative sequencing is (3), (1), then (2). I’ll call this Sequence 1.

Sequence 2: The most aggressive approach is (2) then (3), then cycle back to (2). It’s an enormous all-in kind of bet. I could chunk it down, tweak some knobs and dials — abort if I don’t see it bearing fruit. But this is the kind of endeavor that needs a full commitment. It’s going to be hard enough without existential questioning along the way.

Sequence 3: A logical next step would be to do (1). Get traction and try to exit early to cycle into (2). Perhaps shooting for a buyable startup and filtering for only ideas of that nature. Something that would get assimilated by a platform, Apple, Google, etc. This build trust, builds a skill base, and sets me up for (2). The thing about my current “Lucky Palmer” ideas that I have right now are that I likely have no more than 5 years before someone is going to realize the possibility see The Next 7 Moves in AI. It’s also going to be incredibly challenging to do without help to split the problem. I would want to raise about $5 million to have enough iterations and to move as fast as possible and move the odds of success upward from <0.1% to at least 1%.

With only one life to play it makes sense for my own sanity to do (3) or (1). I’ll be happier and less stressed. Less existential angst. If I had 100 lives to allocate I would do (2). If I could get the capital with no loss of trust I would do (2).

One of the challenges I’ve had in the last 3 months working towards Startup 2.0 is I got sucked down the rabbit hole seeing the potential of (2), after years of thinking about it I finally started to see the ingredients in papers that could actually produce it. But I also realized just how daunting it would be. I would basically be doing research that would lead to foundational models and algorithms. The $5 million would just test the thesis. Given my lack of official ML background it would be right for 3rd parties to be skeptical. It would take a leap of faith in me as a person that I am (1) not hallucinating and that (2) I could seduce the right talent.

Decisioning

I need more cycles anyway to sort the ingredients into a recipe first and foremost. The real decision is whether to go full time on thesis or not at this moment. The first step to pursue (2) would be to finish the white paper 🧠Universal Learning Machines and crystallize the plan. I understand the problem, and I believe all the puzzle pieces exist to solve it, but putting together truly 0→1 technology is non-trivial. Note: depending on when you are reading this it may not be viewable.

At the end of the day learning is bounded by sleep cycles. Discovery is bounded by learning. This is why I always like to have an A-thread where you practice what you know and a B-thread where you incubate your next cocoon-butterfly moment. There is a maximum to how much neurons can shift at night and the amount of learning time to cause that adaption is capped at about four hours of intense study and engagement — heavy diminishing returns after that. There will be minor learning to calibrate and understand sub-problems on your A-thread so it becomes effectively a 50/50 split of those four hours. I like to use Saturday mornings as my B-thread time to minimize context dropping inter-day.

So, let’s distill this into a weekly schedule.

• M-F: 0900 - 1900 ← A-Thread (Startup Machine Gun)
• S: 0900 - 1300 ← B-Thread (🧠Universal Learning Machines)
• S: Off / Float ← I keep this as A/B-Thread optional, otherwise I start to rebel, often I still do work though

Personally I view all of my life as integrated and I work 24/7. I sleep so I learn. I workout so I can work more. I have fun so I don’t rebel and can step away to gain objectivity. But this is ~50 hours of A-Thread butt-in-chair using a computer time. There may be another hour or two of directed thinking that happens as I go through the morning wind up and wind down.