The end of incrementalism: how AI will reward maximalist start-ups

In the 30 years since Tim Berners-Lee launched the world wide web, we’ve seen two innovation cycles.

The first was the consumer internet. During this rough and tumble period, it paid to stalk the fringes of the web: the internet was a blank canvas for inventing new consumer behaviors. Products went from offbeat curiosities one year to ubiquitous facts of life the next. Iconic successes seemed outright strange at first: Amazon (wait days to receive a product you’ve never seen), eBay (buy beanie babies from someone thousands of miles away), Google (trust an algorithm to answer your questions), LinkedIn (publicly post your resume), Facebook (share personal updates with people you haven’t seen in years), Airbnb (stay in a stranger’s home), Uber (get into a stranger’s car), Doordash (hire a stranger to pick up your food). Some of the also-rans were even weirder like Boo.com (3D avatar shopping assistants) and Flooz (digital currency for e-commerce).

As the well of consumer ideas ran dry, tech entered a second, more staid cycle of enterprise software creation. While this wave had its iconoclasts too, start-ups innovated more on how to solve known problems and less on what problems to solve. Snowflake followed the well-trod path of bigger, faster, cheaper databases – just phenomenally better than anyone else. In a stark contrast to the subversive culture of the consumer internet, Cloudera and Confluent emerged from the corridors of big tech (Yahoo and LinkedIn, respectively). Most other B2B SaaS unicorns made their fortunes generating business efficiencies. No wonder the median age of founders drifted upwards.

AI marks a possible third innovation cycle. Will it look more like the first or the second? Investor activity over the past year suggests the latter. But I think it will more closely resemble the early internet. AI should reward start-ups that are maximalist, consumer-like, and weird while penalizing those that are incremental, procedural, and professionalized.

In particular, I expect breakout AI companies to:

• Find new problems to solve instead of solving old problems more efficiently: With traditional enterprise software categories mostly saturated, incumbents will apply AI in all the obvious places. Disruptive companies will need to embrace truly net new thinking to find green pastures.

• Drive a renaissance in consumer software: Consumer software is poised for a comeback. Users are eager to interact with AI generated content. LLMs have unfurled a blank canvas for start-ups to (re)define the internet.

• Invent counterintuitive paradigms for human-machine interaction: The AI maximalists are most likely to win. Think autopilots versus copilots and software built for machines as well as humans.

Here are a few ways that AI’s evolutionary pressure may change company building.

Before: B2B SaaS is the most predictable way to build a venture-scale business.

Soon: AI is raising the barriers to entry for B2B SaaS start-ups.

For the past decade, identifying a sluggish enterprise incumbent, and grinding out a better software solution feature by feature was a fairly predictable path for company building. The trouble for AI start-ups looking to apply that formula is that the old incumbents have been largely replaced by spritely upstarts. Moreover, the new incumbents are well positioned to fight back. Unlike the shift from desktop to mobile or on-prem to cloud, companies don’t have to overhaul their tech stacks to adopt AI. Anyone with LLM API access can layer AI functionality into their product. So it’s no surprise that early enterprise AI successes (Github Copilot, Notion AI, Adobe Firefly) came from scaled players.

Where there is white space, expect vicious competition. Since network effects are rare in enterprise software, first mover advantage is weaker. This makes it more mimetic than consumer software: the instant one company experiences customer traction, a dozen competitors spring up – look at how crowded investor market maps for AI are already.

Founders will continue to build category-defining enterprise software applications, of course, but AI is raising the bar. New entrants will need to be particularly strange, risky, or rule-bending to deter competent incumbents and copycats.

Before: Consumer software start-ups are a fool’s errand since big tech monopolizes attention.

Soon: AI content is sparking a consumer software renaissance.

Consumers are curious to engage with AI shoggoths. Nearly all the foundation model companies that have achieved commercial scale are consumer or prosumer applications. ChatGPT, Midjourney, and Character.ai reached tens or hundreds of millions of MAUs in a couple months. The latent demand for edgier concepts like an AI therapist or AI lawyer is also considerable. Instagram, TikTok, Google, and YouTube are perhaps not quite the end-of-history attention monopolies we thought.

Underlying this success is AI’s exceptional creativity and (at least simulated) empathy. In certain cases, AI can create things that are more interesting than what our friends post on social media or will talk to us about topics that other humans won’t. These qualities are catalyzing new internet behaviors – most emblematically, the coalition demanding unfettered access to AI waifus.

AI thus seems more likely to be a disruptive innovation for consumer platforms than enterprise ones. AI companions might steal attention share from social media. Intelligent assistants could disintermediate aggregators like Expedia, Booking.com, OpenTable, and Yelp. Even the search market is up for grabs.

Before: Human-in-the-loop is the dominant UX paradigm.

Soon: AI autopilots are ubiquitous.

Contrary to popular wisdom, I doubt “copilot” interfaces are more than an intermediate step. AIs that do things for us are more useful than AIs that look over our shoulders.

Since GPT-4 is already better than most humans at plenty of tasks, humans will increasingly be a bottleneck for machine intelligence. A necessary one in some cases, but a counterproductive rate limit in others. We could try to solve mental health by giving human therapists better tools – or someone could build a therapy chatbot with mass appeal. Github Copilot can help human software engineers double the lines of code that they write, but to 100x programming output we need AI software engineers. At the limit, humans should be ideators and validators with everything in between done by autopilots.

The future of human-computer interfaces is one-to-many, not one-to-one. 

Before: Software is designed for humans.

Soon: Software is designed for humans and machines.

As AI autopilots proliferate, they will eventually become the majority of software users. Much as primitive web crawlers spurred the creation of the Robots.txt protocol, companies will need to design software optimized for machines as well as humans.

A few adaptations jump to mind:

• Designers will spec products to optimize for AI capabilities in addition to GUI ergonomics.

• We must build systems for managing, distinguishing, and permissioning AI and human identities.

• Many software interfaces went from single player to human-to-human multiplayer during the cloud era. They may evolve again into human-to-AI multiplayer interfaces in the AI era.

At the limit, software users might rarely know whether they’re interacting with humans or machines. What if new social networks solved the cold start problem with AI simulacra? And would you really care if the most entertaining Reddit posts were written by AI bots? In fact, wouldn’t the mystery make it more fun?

Conclusion

Much digital ink has been spilled pronouncing AI to be a sustaining, not disruptive, innovation. This understates its potential.

AI is spurring new products – intelligent assistants, human-esque companions, interactive codebases – that could dramatically change how humans interface with technology. These start-ups already earn billions of revenue between them. That doesn’t sound like a sustaining innovation to me.

Yes, there may be fewer multibillion dollar database companies started in the 2020s than in the 2010s. But there are plenty of opportunities for founders to build important companies with AI – perhaps bigger than we’ve seen in a long time. They’re just going to have to get weirder.

Thank you to Jacob Effron, Alex Mackenzie, Belen Mella, and Sam Ching for their feedback.