Beyond the Buzz of Generative AI Apps
Are we investing in innovation or short-lived hype? How I feel after examining more than a thousand companies in the last six months.
The generative AI wave hit the startup world massively in the last few years, and we all believe this variety of new solutions will significantly benefit everyone’s life. We all have been impressed by companies like OpenAI and Midjourney because of their immediate usefulness in our day-to-day lives. Besides, observing what humankind has accomplished in such a short time is incredible, and it’s natural to think about where all this can bring us in a few years.
Generative AI requires significant computing power and financial investment to perform tasks comparable or, in many cases, superior to what a human brain can accomplish. This is particularly true when dealing with massive data that needs to be processed and analyzed quickly. And before we can interact with AI models, they must be trained for months. This is what I’ve read recently online:
Training a single generative AI model can require as much energy as several flights around the globe.
And again:
According to a study by researchers from the University of Massachusetts Amherst, training one large generative AI model for natural language processing can emit as much carbon as five cars in their lifetimes.
It’s also true that a fine-tuned AI model can serve billions of tasks a day and theoretically improve human lives in the medium term. People are willing to invest, and founders are fond of experimenting.
What troubles me is how today’s startups use this source of knowledge provided by models like GPT-4, and it makes me wonder if we are really pushing innovation or simply playing around with this cool technology to see what comes out.
Start using different AI tools every day. You can see that they often provide not only good summaries or insightful information but also many redundant answers to questions. However, we are still far from considering them an extension of ourselves. Maybe they are not so “intelligent” today, but they are learning fast. This kind of technology amazed us but is very young. To have tools you can trust with your life or even simply that you can use without questioning their correctness and accuracy will take years.
Nonetheless, one thing we know for sure is that companies providing paid access to their sophisticated generative AI models through an API layer will get most of the immediate advantage and revenue.
But what about all the other AI applications built on top of them?
We analyzed thousands of startup pitches in the last six months, and 99% of them are AI-centric. Most of them may seem helpful but obvious. AI copilots and platforms to talk to documents are everywhere, and I’m sure some will succeed eventually.
But which ones? And, most importantly, are they mature enough to get the job done?
I forecast that 99% of them are likely to fail.
In his WTF VC - Fall 2023 paper, Sam Lessin from Slow Ventures stated that most of the innovation we see in AI companies will be extending innovation and not disruptive:
Two types of players are going to benefit from this extending innovation: First, the big existing platform players. Second, tiny mom and pops.
I agree with Sam on this one!
From what we see on the market, thousands of founders are betting their future to build something that is not a disruptive innovation. Building long-lasting companies that can have a massive impact is what we look for as a VC, and it’s not what I see here. We are all amazed by what you can do by applying AI to niche problems; niches are our bread and butter. But the more I think about AI application companies, the more I believe something must change, and founders should tackle more challenging problems to be called innovation.
The long-lasting test for a newly-born company is more important than quick initial growth, especially when discussing generative AI. Unfortunately, the success of a startup can only be determined retrospectively, based on the alignment between its founding team’s vision and personality, as well as investors’ belief in it.
I’m probably biased because this approach resonates well with our investment thesis, like API-first platforms, infrastructure, developer tools, and open source. And don’t get me wrong, we like growth signals, and I’m not saying that many AI application companies will not monetize well. They will, but I feel it will only last a few years before most of them will be wiped out.
Two trends will be responsible for this massive failure:
- The “smart” incumbents are moving faster than ever with these technologies, and they have capital, R&D brains, and customers to upsell their products.
- The unsustainability of AI costs at scale. That’s why we are eager to invest in solutions like—OpenPipe, for example—to reduce AI consumption costs, making smaller and smarter open-source models the best option for most players.
A wise strategy for these generative AI times is to invest in shovels, pickaxes, and processes to make picking the nuggets efficient and avoid most of the gold rush. We love solutions like Exa, which redefines searches on the Internet, or Texel, which can bust and save money on your video pipeline and make it a viable solution at scale. We invested in both of them. Adopting this strategy is probably wise enough if you, like us, invest at the pre-seed stage, where the chaos is perhaps at its peak, and there is too little data to say if the generative AI app will pass the long-lasting test and is solving an urgent customer problem.
In the world of generative AI applications, a common question frequently arises:
Are these companies meant to last, or is this, for the most part, a phase of experimentation?
Maybe both, but we better look beyond flashy applications and invest in the foundational tools, processes, and cost-effective solutions underpinning this transformative era.
