Mr. Market

The Rise of the Resident Eccentric in Tech

rothko_quote_500px

Red, John Logan 2009

Companies are hiring journalists now, which is great. But they also need to be hiring poets, novelists, sculptors, screenwriters, painters, philosophers, and other contra-logical weirdos.

I know I sound insane right now but there are already murmurings that companies are putting a premium on creative acumen, and this trend will likely only become more pronounced in the next six months or so.

IMO companies that invest in anomalous creative talent will succeed in ways too intangible to plagiarize and turn programmatic, and the ones who don't will either blend in (which is death) or always be late.

Why I think that

The most valuable questions in business right now have no correct answers:

These are questions that have no known answer. They are predictive, which should be a fine job for LLMs. But they also require near-constant pattern disruption to be truly effective, which is hard for LLMs.

That makes these some of the most high-leverage questions in tech today.

You'll notice that several companies with clear advantages in markets with highly-sophisticated buyers are already investing a lot of time and energy into answering these questions (ex: PostHog, Bun.)

These companies aren't winning BECAUSE they make unusual aesthetic choices. But it is helping them find the right customers faster and win over buyers who care about product philosophy, design, and ethos in addition to pure functionality.

Marketing/brand/etc. used to be also-ran pseudowork. What changed?

For most of computing history, questions of brand, taste, inspiration, and affect have been considered soft questions. You can't optimize for these answers against a loss function because there's no ground truth to compute loss against.

Because there's no 'right' answer, many in tech seem to believe finding the best possible answer to these questions isn't all that important. For the most part, you could just find a 'good enough' answer and go with it. Precision wasn't required, so things would work fine even if the answer you gave was just 'average' rather than 'perfect.'

Finding exceptional craftsmen to answer these kinds of questions (ex: marketers, brand people, sales guys, copywriters) was dismissed as less crucial to the success of the business than finding world-class technical hires. You NEEDED a top 1% engineer to succeed, but any marketer in the top 50% would probably do.

Now we have thinking machines.

LLMs have gotten incrementally better at executing on tasks where it is possible to find the correct answer through gradient descent. This includes data analysis, coding, deep research, and other 'hard sciences'. They're far from perfect at this, but they're definitely getting better.

At the same time, they're getting incrementally worse at knowledge work that requires creativity and often illogical, unnecessary novelty. This includes things like trust-building, brand marketing, writing, and pattern-interrupting aesthetics.

Finding a top 1% engineer is still crucial, obviously, and probably always will be. But now, you also have to worry about finding a resident eccentric.

Why gradient descent has an inverse relationship with creativity

We run into two problems when we try to use gradient descent to make machines better at creative disciplines.

The first problem is that gradient descent is mostly local. At each step, it reads the gradient, finds the nearest minimum, and moves accordingly.

In aesthetic space, visionary work isn't the nearest minimum. It's also not the MOST improbable position (that's just randomness).

It's a point in a different basin in an improbable position that, once revealed, is surprising but inevitable. Not the average, obvious choice, but also not a random choice, or even the total opposite choice.

Imperfect metaphor but top 1% creative work is like a phase transition in physics: it's when water turns to ice. It transforms your thinking. It makes new things possible. Best case, it's paradigm-exploding. But in the world of aesthetics, the phase transition point is always changing based on context.

The second issue is the loss function problem itself. When you can't specify loss analytically, you approximate it with human feedback. RLHF is the dominant approach: train a reward model on human rankings and then optimize against the reward model.

But this won't work either. The reward model still averages over the preferences of all human annotators, so you still end up with an average.

Once LLMs start doing something, it's already lame

Obviously there are times when you can ask an LLM to do creative work and it gives you something you haven't seen before.

But a month or even a week later, this fun trick will be infuriating to everyone who spends a meaningful amount of time on the internet. It's like the opening shot of a Wes Anderson movie. Cool the first time, but the tenth time, who cares and also shut up. My poetry professor used to say "never repeat your successes." This is why.

A model trained on what is currently working will produce output that reaches market already past peak, because the optimized approximation hits scale precisely when the original has saturated the field.

This makes the model a lagging indicator of aesthetic resonance.

Nobody even registers stuff they see on the internet anymore

Now add Shannon information, which is just the idea in information theory that the content of a message is based on how surprising it is. Basically, the more improbable a set of information is, the more 'information content' it has, and the more likely you'll actually receive a message.

'The sun came up' has almost no information content, whereas 'the sun never came up' has a lot.

Digital channels today, of course, are absolutely jammed up with probable information. We rarely see things we don't expect to see, because our algorithms are primed to show us things it already knows we really like or really hate (or fear or whatever).

As a result, most content we see online is pattern-matched and discarded before conscious attention ever metabolizes it.

The way to increase the chances your target customer will actually download what you're saying is to violate the expected pattern in some way.

Why poets and other artists are so adept with ambiguous decision-making

So how do you disrupt the pattern? You git gud at answering really ambiguous questions in a way that is technically right but also sufficiently anomalous.

Artists are good at this. The following is true of any artist, but we'll use poets for simplicity's sake and because it's the world I best understand.

The poet has two primary instruments:

Every poem is a series of decisions made in the absence of correct answers. Rather than trying to find the 'right' answer, the poet must weigh a series of stronger and weaker choices.

The internal tools they use to make these assessments are disparate, which is also good for doing imaginative work: they draw from a singular interiority, years of obsessive practice, and a lifetime of reading/other artistic study.

This is exactly the cognitive operation that produces remarkable brand work (as in, something people even remark on lol. Not necessarily good or bad.)

The loss function for a poem doesn't exist until the poem does, and even then it's not computable. It's the private phenomenological response of a reader who may or may not have encountered the right poem at the right moment in their life.

In art, ambiguity is nothing to be afraid of or 'solve' like it is in other disciplines. The ambiguity is a fixture of the material you work with to create something great. It's essential to meaning-making.

And so the serious practitioner has trained their pattern-recognition machinery in conditions of maximal difficulty: ambiguous signal, no ground truth, almost no market feedback, and a community of evaluators with high and heterogeneous standards. This training produces a sensitivity to what's working that is probably quite uncommon, and more likely to give you a 'unique' edge in the market.

The skills gap between an artist and a capable generalist doing creative work is at least as wide as the gap between a passionate craftsman engineer and a competent professional who learned to code but hasn't really kept up with the field.

Companies recognize the latter gap intuitively; they have built entire hiring processes around finding evidence of it. They almost never apply the same scrutiny to creative hires.

But doing so IMO would increase success across a bunch of different axes in bottom-up markets with highly sophisticated buyers.

The taste-gated market

This brings us to a caveat.

It's not always necessary to hire a resident eccentric. Some markets can do without an expert pattern disruptor just fine. Others, though, will benefit from them in a big way.

B2B purchasing decisions at massive, slow-moving old companies are committee work. Everybody on the procurement team optimizes against a specification, whether that be a list of must-have features, compliance needs, pricing, or a set of integrations.

This is where AI is great. It can instantly write RFP responses that hit every requirement.

But when you're dealing with highly sophisticated buyers, aesthetic acumen pays off big. Let's use developer markets as an example.

Why is POV and aesthetic philosophy important for them? (aka why is it ok to have fun w it.)

IMO four reasons:

These buyers also do not like AI in their online spaces. (I know what you're thinking: who the hell does? But one look at LinkedIn and it's clear many verticals do not care. Ex: vibecoders don't give a shit.)

AI is not ALWAYS detectable. Everyone has been fooled. But depending on where you're distributing your message, get caught once and you've done serious damage.

In general, buyers who care about aesthetics include: devs, designers, high-autonomy creatives, research scientists, and other discerning technical practitioners. Buyers who don't care: procurement at huge companies, finance people (for the most part), buyers in highly regulated industries.

That said, optimizing for highly sophisticated buyers often eventually makes you appealing to the rest. You get credibility by virtue of being the darling of the most discerning verticals in tech. (See: Stripe.)

The best part about this strategy is that it is unscalable

With the rise of 'Head of Storytelling' and 'Head of Narrative' roles etc., it's clear there's some motion toward identifying uniquely discerning creatives and letting them loose on highly sophisticated buyers.

But it's probably not going to be everybody, at least for a while. This is a very good thing for the companies who do it.

It's not always important to be a first mover. But with brand, where trust and identity builds over time, it's never too early.