In ancient Greece, there was a class of itinerant professionals who made their living selling their tutoring expertise to statesmen and nobility.¹ These professionals were called sophists, after the Greek sophía, meaning "wisdom", although the term sophist typically has the negative connotation of one who prizes the beauty or artistry with which an argument is articulated over the authenticity of its reasoning. A famous early Sophist philosopher, Protagoras, was notorious for his claim to "make the weaker argument the stronger"

Anyone who has forayed into the startup world, particularly coming from an engineering-oriented academic background, has likely been struck by the preponderance of Sophists it tends to attract. In the winding path towards any world-changing invention or company, there is undoubtably value in storytelling ability, and a penchant for Sophistry is a useful tool in the entreprenuer's toolbelt, but in a world where AI has replaced consensus for any concept of fundamentals, I personally have found it hard not to worry that Sophistry is doing more harm than good to this grand pursuit.

One of the more frequent rhetorical devices that has circulated throughout Silicon Valley in the last few years has been the idea of "The ChatGPT Moment for X", a hypothetical moment in which as pre-revenue startup releases a low-key research preview that suddenly turns the company into a trillion-dollar behemoth.

It's particularly noteworthy that this rhetorical device has supplanted the once-favored "Uber for X". These are two shorthands for conveying why throwing massive amounts of capital at something will result in outsized returns:

• Uber for X: Burn money to bootstrap a defensible two-sided marketplace from which one can ultimately extract ousized rents from network effects
• ChatGPT Moment for X: Burn money to build some incredible technology that will suddenly become super valuable at some unspecified date many years in the future

The basic reason for the shift in rhetoric is that a lot of smart capital allocators are able to scrutinize arguments of the former flavor, since network effects and retention are quantifiable metrics, but are basically unable to independently reason about arguments of the latter flavor - after all, if they were, they likely would be spending their time building those incredible technologies instead of allocating capital.

On top of this, venture capital is in a reactionary period, where a lot of people who were skeptical of the latter flavor of companies got badly burned when they missed out on early investments in OpenAI or Anthropic. This means that even if smart capital allocators are suspicious of these types of companies, they know that their own limited partners will drag them for missing out on the next Anthropic.

The problem with such Sophist arguments is that there comes a point in which these very expensive science projects do ultimately need to resolve in one direction or another. Whether or not the project is able to accumulate lots of capital is only correlated with the likelihood of that project's success by the depth of the capital allocators' understanding of the technical difficulties involved. This correlation is often extremely superficial and in some cases anticorrelated when the experts being consulted to diligence a given project have their own scientific biases and gaps.

This all came to a head this week with the release of the most recent Astra model from OpenAI.² I was working on my own startup during a period in which "The ChatGPT Moment for Robotics" was motivating capital allocators to pump huge sums of money into research labs that had the flavor of being the next OpenAI or Anthropic, oriented towards solving all physical labor. To the extent that these research labs have run their models through independently-verifiable benchmarks, Astra has largely shown a step-function improvement on anything that can be solved via open-loop control (which includes a lot of important manipulation benchmarks), while being a model that anyone with an OpenAI subscription can try out themselves. This was put succinctly by Lerrel Pinto:

Now, of course, there's a risk of prematurely updating one's priors just because some hot new model came out. But assuming that this is in fact a paradigm shift for robotics, it's not a very surprising one - it's essentially just the bitter lesson in action. Accepting that Astra or whatever future models any of the big AI labs put out will just be better than any models trained by physical intelligence neolabs simply means accepting that there is nothing particularly special about the nature of intelligence required to control a robot compared with the nature of intelligence required to write code or control a computer.

Since I am dispositionally the type of person who likes to try and make money off of what I believe to be a contrarian perspective on a as-of-yet-unfolding field of technology, a few months ago, I actually made a proper ex ante bet on this thesis with a junior employee at Physical Intelligence³ to the effect that Unitree would in ten years be a higher-revenue company than Physical Intelligence.⁴

But putting aside my own schadenfreude at seeing obnoxious Sophists get hit in the face by reality, this whole episode gets at the heart of something that's been bothering me for a while, namely, that as someone whose differentiated skillset is largely oriented towards training and deploying large AI models, basically any startup idea that I would work on is likely to ultimately be eaten up by one of the big labs. Or, put differently, the best decision for me to make to maximize my own expected value is to work at an AI lab.

In this way it feels like as an AI researcher, I am on the bad side of Roko's Basalisk, but instead of having the threat of infinite simulated torture hanging over me, it is instead an acute awareness of the ramifications of the Bitter Lesson and a belief that venture capital today is basically structured to pour mountains of cash onto theses which are largely predicated on motivated thinking and a fear of missing out, rather than any particular understanding of fundamentals. And while it may be possible for Sophist arguments to the contrary to still raise lots and lots of money - it may even be possible that some of these bets do materialize modest revenue - I simply cannot bring myself around to accept that any of these companies will ultimately be able to pay back their investors, short of resorting to contrived acquisitions or SPACs.

There are some bright spots in all of this, however. It is how I have personally been reconciling my belief that AI researchers like myself are destined to soon go the way of the mathematician, becoming unable to contribute to meaningfully advancing the frontier. These are as follows:

1. The input costs for creating value are likely to go to zero very quickly. Would-be entrepreneurs will no longer need to hire any other humans, nor raise outside capital, in order to build great new things.
2. The advantages of venture capital are likely to go to zero as well, and the cost to accepting outside capital are likely to outweigh the benefits it brings. More money doesn't buy you more speed in a world where everyone moves at the pace of the frontier.⁵
3. Silicon Valley is still, and will continue to be, the best place in the world to build a company. Applying a mechanistic understanding to company-building, and thereby correctly front-running the AI adoption curve, will likely continue to be massively a winning strategy.
4. There are still a huge number of problems to solve, and it's getting more and more tractable to solve them.
5. I have the huge priviledge, for now, to be able to work at the company that will win the biggest in all of this, and in a world where my job transforms from being an AI researcher into something more like a forward-deployed engineer, there will still be interesting and exciting things to do.