How to Build an AI Startup

The Insider's Guide to Building a World-Changing, Category-Dominating AI Startup

AI is transforming every industry on the planet. As a result, building an AI startup has never been a more exciting opportunity.

But it's also never been more competitive. For every breakout success, there are dozens of AI startups that crash and burn.

What separates the winners from the losers?

To find out, we dove deep into the early days of category-defining companies like Anthropic, OpenAI, Runway, and Midjourney. We uncovered the tactical playbooks they used to beat the odds.

In this no-BS guide, we'll give it to you straight. No recycled truisms or 30,000 foot views. Just battle-tested advice from founders and PMs who have been in your shoes and come out the other side.

---

Introducing Spencer

To help you rise above the noise, I have teamed up with Spencer Shulem. Spencer is the founder of an AI startup himself, BuildBetter.ai, and former PM + UX designer.

He’s grown the product to 21K product teams and has been building on GPTs since 2020 (way before most), so he’s the ideal expert for this type of advice.

---

Today’s Post

Words: 5,051 | Est. Reading Time: 22 mins

In this insider's tell-all, you'll learn:

1. The pitfalls that cause 90%+ of AI startups to crash and burn

2. 🔒 The 3 pillars every breakout AI startup is built on (miss one and you're doomed)

3. 🔒 How to find an innovation that will leave competitors in the dust

4. 🔒 Secrets of pricing & positioning

5. 🔒 How to hire an all-star team

---

1. Avoiding the Startup Graveyard

Ideas are cheap. Execution is everything. And boy do most teams botch it. Even heavily funded startups like Rabbit ($65M), Humane ($230M), or Inflection ($1.5B) are fizzling out. The failure rate is brutal.

After analyzing dozens of AI startup post-mortems, we identified the most common pitfalls:

Pitfall 1: Falling for Shiny Object Syndrome

When a shiny new model drops, it's tempting to drop everything and pivot. Resist that urge.

For example, when DALL-E 2 was released in 2022, dozens of "DALL-E for X" startups immediately popped up trying to apply the tech to everything from gaming to interior design.

The startups we studied succeeded by staying laser-focused on their target user and use case, only integrating new tech when it demonstrably moved the needle on key metrics.

For instance, Anthropic has been working on its constitutional AI technology for years, despite many flashy new approaches emerging. That focus allowed them to make rapid progress on their approach to safe and robust AI assistants.

Takeaway: Choose one hairball problem and one promising technical approach. Filter out the noise and nail the execution. Be extremely selective about chasing shiny new objects.

Pitfall 2: "It Works In The Lab"

An impressive research result doesn't equal a working product. The chasm between lab and production swallows startups whole.

Turning prototypes into products takes massive investments in infra, devops, UI, documentation and go-to-market. It's an all-hands, multi-year slog.

If you think your new technique will magically sell itself, get your head out of the clouds. Budget 10X more time and money for post-lab productization.

As Steve Jobs said:

There is just a tremendous amount of craftsmanship in between a great idea and a great product.

This was the Rabbit and Humane mistake. They got a good demo and good commercials. But the actual product didn’t work.

On the other other hand, Cohere spent two years building a robust serving platform for their language models, including capabilities for fine-tuning, prompt engineering, data hosting, and more. This foundational work enabled their self-serve API to reliably handle billions of requests from day one.

Takeaway: Plan to spend at least 50% of your time on post-research productization. Set up CI/CD, beta programs, chaos engineering, and exception tracking early. Performance, security and compliance can't be afterthoughts.

Pitfall 3: Irresponsible Deployment

In the rush to market, many AI startups fail to put adequate safeguards in place around data privacy, security, fairness, transparency and robustness. When things inevitably blow up, they feign ignorance.

Take the case of Clearview AI, a facial recognition startup that illicitly scraped billions of social media photos. When this was exposed, the company faced major backlash and legal consequences.

Or consider Zillow, which lost over $500M on its iBuying business because its home valuation model couldn't handle a shifting market. The company failed to stress test the AI system's limits before deploying it at scale.

Takeaway: Engage cross-functional stakeholders to understand potential risks early and often. Build technical and human oversight into your dev and deployment workflows. Communicate honestly about limitations. Better safe than sorry.

Pitfall 4: Prioritizing Flash over Function

Many AI startups get enamored with their own tech and lose sight of what users actually care about. They churn out flashy demos that generate reams of AI content but don't solve real problems.

Take the example of Quixey, a startup that aimed to build a deep learning-powered "search engine for apps." Its demos looked impressive, enabling users to find apps based on vague natural language queries.

But in practice, most people just found apps through the standard app store search and categories. All of Quixey's fancy ML wasn't actually helping users discover apps better. As a result, the company eventually shut down after burning through over $160M.

Successful startups like Runway laser-focus on their users' gnarliest problems from day one. They went deep on customer discovery with video creators to find the workflows that burn hundreds of hours and thousands of dollars. Then they cut the time and cost by 10X, not just 10%.

Takeaway: Strive to make users say "how did I ever live without this?" Fall in love with the problem, not your tech. Keep digging until you find an acute pain point, then reshape the workflow to eliminate the friction. Anything less than 10X improvement is not enough.

Pitfall 5: Raising Too Much, Too Fast

VC can seem necessary as an AI founder. In fact, Aakash was advising an AI founder last week who proclaimed:

All my competition is raising VC. I’m going to get crushed if I don’t!

But have you heard the story of Olive AI? The company was valued at $4B two years ago after raising $1B.

Unfortunately, the cash didn’t guarantee success. The “champagne and cocaine” mentality brought the company to its knees. Olive AI CEO Sean Lane proudly proclaimed, “we’ve pivoted 27 times.” But, ultimately, that just meant it raised too much too fast.

Today, Olive is dead.

On the other hand, successful startups like Cohere bootstrapped for two years before raising a $40M Series A. This allowed them to deeply validate their self-serve model and hit $1M ARR before taking on venture capital. With strong fundamentals in place, they could then scale with confidence.

Takeaway: Consider bootstrapping or raising a small seed round to get to product-market fit. Validate a pricing model that aligns with customer value. Don't scale until you have strong customer love, healthy unit economics, and a repeatable growth engine. Premature scaling kneecaps your runway.

Now that we've gotten the failure modes out of the way, here's the real inside baseball on how to win...