Sam Altman warned OpenAI will ‘steamroll’ AI startups. I run one. Here’s why I’m not worried

Awesome, right? Yes, but it can also be equal parts terrifying. 

Spend any amount of time online in tech communities and you are sure to see headlines declaring that the most recent AI breakthrough spells the end for most AI startups. In a space where breakthroughs happen by the week, this can be tiresome.

It is not just tech journalists, opinionists, and fire-stokers making these declarations. It also comes directly from the people responsible for launching these breakthroughs. A few months back on the 20VC podcast, OpenAI CEO Sam Altman said his company will steamroll any startup or product trying to build in and around their blast radius.

The general idea is that too many AI startups today are built atop of the “foundational models” produced by the likes of OpenAI and Anthropic. As these major players produce new models with boundary-pushing capabilities, the less-known companies that benefit from these capabilities will become obsolete. 

For example, if the newest OpenAI release lets you upload a PDF to ChatGPT to “chat with it,” what happens to all the companies that offer that as their core service? The prevailing wisdom today is that those hypothetical companies are going to be left for dead. 

I run an AI startup called Consensus. It’s basically like Google Scholar + ChatGPT. Our goal is to make it easy to consume and search for peer-reviewed scientific research. To date, we have avoided the shrapnel of the AI behemoths. But with OpenAI announcing their upcoming foray into search, some would argue our days are numbered.

I am here today to tell you that the future for AI startups is bright. The doomsday headlines we see about startups are just like attention-grabbing headlines in any other industry—mostly for show. 

Here are three reasons why AI startups are not doomed by every subsequent AI breakthrough: 

Most companies start as a ‘thin wrapper’

Being labeled “a thin GPT wrapper” is the biggest insult you can thrust upon an AI startup in 2024. A thin wrapper refers to a product with very little real technology built themselves that is propped up by being built on top of someone else’s technology.

“Thin wrapper” companies do exist and some will certainly get steamrolled by future iterations of OpenAI’s models. Just ask the team at Jasper AI. 

Jasper is an AI copywriting tool built using OpenAI models. In the pre-ChatGPT-hype world, their tool was lauded and they soared to a billion-dollar valuation. When the world became privy to ChatGPT, most users realized they could get the exact same functionality directly from the source—and Jasper’s revenue (and valuation) tumbled. 

However, being a company that is built with third-party technology at the core is not inherently a bad thing. No founder should worry about being a “thin” wrapper at the start of their product journey. In fact, being a thin wrapper in your early days is sometimes an outright necessity for new products in order to get off the ground. It is simply your job as a startup to turn your “thin” wrapper into a “thick” one over time through design, user interface, new features, services, branding, etc.

This phenomenon is not new. If we applied the same scrutiny we do today to AI startups to previous iconic companies, we would have called them thin wrappers over various third-party technologies at their inception, too:

Salesforce is a thin interface wrapper over an Oracle database.

Box is just a thin wrapper over AWS. 

Zoom is just a thin wrapper over Mac and PC cameras. 

Delta is a thin wrapper over Boeing planes.

And so on. 

A new capability directly from OpenAI is also a new capability for your hypothetical startup. It is simply your job as a startup to build enough schlep around that capability to make it compelling and useful enough for users to incrementally pay for. As that technology that you rely on improves, so does your product. 

Most things start as a thin wrapper. It is not a sin. The only sin is staying a thin wrapper. 

The difference between good and great is infinite

There is a proliferation of remarkable AI product demos online today. Despite this, there is a seemingly gigantic lag in the number of AI products that actually delight and solve problems when they are in the hands of their users. 

This is because artificial intelligence in its current form is a connoisseur of “good enough.” 

When something is “good enough,” and it is in a guard-railed demo setting, it can appear to be magic. Large language models (LLMs) have lowered the cost of marginal intelligence in products to near zero. Build a simple user interface, slap on a feature or two, and then add in a few API calls to OpenAI and you, quicker than ever before in human history, will have something that looks like an amazing product. 

Unfortunately—or fortunately from the founder perspective—building great software products is still incredibly hard. Before LLMs, the crux of the work required to build an amazing software product was an amalgamous stew of hundreds of factors like deep customer understanding, elegant design that requires taste—not technical acumen—and thousands and thousands of lines of code that handle every possible edge case that a user may encounter when using your tool.

None of that has changed. Just because it is now easy to build something that looks like a great software product does not mean it is now easy to actually build a great software product.

Let’s look at the difference between Google’s much-maligned “AI overviews” and the fast-soaring AI search startup Perplexity. 

By some definitions, Perplexity is not a “defensible” product. At the highest level, Perplexity is just LLMs interacting with search results. In a world with LLMs everywhere, couldn’t the greatest search engine of all time just throw an LLM summary over results and send Perplexity to its startup death? They can surely try, and try they did. So far, those efforts have been unsuccessful.

Software products don’t actualize at the highest level. They are a massive collection of details, and those details make the difference in how they solve the problems of their users. 

Perplexity has nailed the details: Its user interface has character but is ruthlessly simple. When you arrive on its search page, your cursor immediately gets put in the search box. Its response time is near instantaneous and is even equipped with a delightful loading screen. 

Google’s AI Overviews lack the same obsessive touch as Perplexity. In turn, they have not garnered the same user love. This is the difference between good and great—zoomed out, they may look the same, but zoomed in they are miles apart. 

I could write an entire book on the observation of the deceptively-minimal-but-actually-gigantic difference between pretty good and truly great across life domains. It exists everywhere and software is no different. AI today is “good enough” commoditized. It does not come close to commoditizing truly great, and that should be reassuring for all of us aspiring to build great products. 

Specialization matters

For as long as startups have existed, they have been advised to—and found success by—focusing their initial efforts on narrow problems. Niche problems rarely have markets large enough to maintain the attention of incumbents for them to completely solve. This creates the space for startups to come in, innovate, succeed, and then eventually expand. 

This is part of our hypothesis at Consensus. Google Scholar is the most-used academic search tool in the world, yet it is pretty universally disliked. This is because it is a side quest of Google—it has never given the problem of scientific search the tender love and support it deserves. A startup like ours can provide that. It is quite literally all we care about doing. Google cares about a million other things. 

The time-tested advice about niche problems is not suddenly defunct in the age of AI. It will continue to ring true when building products in the shadows of foundational model companies like OpenAI. If all that mattered was the raw technological horsepower of a product, then startups would never succeed against incumbents with deeper pockets and better technology. What actually matters is the details of your product—from the core feature set to the checkout process—that show a user you are there to solve their specialized problem. 

As famed AI investor and former GitHub CEO (and current Consensus investor) Nat Friedman posted to X recently: “People hire a janitor service to clean their office. They don’t hire a generic labor service, even though it’s basically the same thing.” – advice for AI startups.

If you just measured raw capabilities, a person off the street and an employee of a cleaning service company are effectively identical. The only difference is some packaging, some cheap materials (cleaning supplies), a sprinkle of expertise, and a trust that this person has solved your exact problem before. That difference will drive 99 out of 100 people to opt to pay extra for a cleaning service company. 

People want to use the thing that is designed for the thing. This is perhaps the most heartening sentence a founder of an AI company could possibly hear today.

When you take a step back and look at the three main points addressed above, this is not new advice. These are some of the same core tenants that have made startups successful when building in the long shadows of more technologically-advanced incumbents for decades. 

Being terrified of big players driving your startup to obscurity is a feature of startups, not a bug. It is one of the things on the endless list of things that make building a successful startup really freaking hard.

We all now have an amazing new technology at our fingertips. The natural reaction is to believe that “everything is about to change.” The reality is that some things will change but most will resemble the past. I am willing to bet that one of those things that will persist is the space to create amazing companies and products alongside the headline-grabbing giants.

There will be startup roadkill created by the winners of the foundational model space. I contend that the blast radius will be smaller than most believe today. 

Read more:

The opinions expressed in Fortune.com commentary pieces are solely the views of their authors and do not necessarily reflect the opinions and beliefs of Fortune.