As with any hype cycle, the huge leaps forward in AI capabilities in recent months have whetted VC appetites. And after OpenAI served up its latest dose of catnip last week, the frenzy is even greater.
But people who’ve built long careers in AI are now worrying about the collateral damage from too much indiscriminate cash being splashed by VCs. Some warn we’re also heading for mass startup failures as a slew of generic “AI startups” pop up in response to VCs’ new obsession, despite the value that generative AI could bring to society.
“I also think that there will be quite a bloodbath,” says Julian Senoner, cofounder of EthonAI, which uses AI to improve manufacturing quality. “Many of these companies will fail because they are now jumping onto this hype but essentially they will struggle to find a real problem to solve.”
An answer in need of a question
Investors are already seeing startups struggling to find that “real problem”.
“Since January [excitement about AI] created a lot of inbound,” says Rasmus Rothe, founder of Berlin-based Merantix. The AI venture studio invests in and advises early-stage machine learning startups.
“You get these application-layer companies that, rather than building a deeptech or AI company that is still hard to build, just use an API… You invest at a very high valuation like with an AI company, but it’s just some people who have built a little tool that got tons of traffic because of Twitter. “
Nathan Benaich, of London-based investment firm Air Street Capital, says the same things about the kinds of companies sliding into his inbox: huge interest has led many people to jump into the sector, often without a clearly distinguishable or unique product.
“I think if there's more than about five companies doing a problem that's already a bit too much,” he says. “I want to walk into a call when I meet a founder to explain to me what they're working on and be like, ‘That's nifty. I didn't even think of that or I haven't seen that before.’”
Benaich adds that he’s already seeing some patterns when it comes to low-hanging fruit, with a lot of startups pitching ideas around content generation, chat, customer support, marketing and legal.
“It's all ‘first order stuff’, which I can't get that excited about. By that, I mean the obvious stuff that doesn't take too many leaps of thinking to view the problem and understand it.”
Products in text generation, in particular, may find it hard to stand out.
“Why would you use a vanilla copywriting tool when you can just subscribe to OpenAI and use their tool, which probably actually works better?” says Stephanie Demetriou, cofounder of AI startup Maekersuite.
Rothe from Merantix adds that he expects to see more casualties in the B2C space.
“If you're a B2C service there's no binding between the customer and the solution, that's why B2B is much more interesting,” he says. “The applications that go deeper into a company — not just like a tool to generate logos or write an email a bit more nicely — but actually integrate much more into enterprise.”
How to sift the wheat from the ChatGPT
So which companies will actually survive? Rothe says one clue is whether a startup is working with any of its own data that others don’t have.
“Is there unique data they're collecting to fine-tune, or unique expertise on the algorithm development side, or even just the integration of the interface to the back end that’s very hard to replicate?”
For Maekersuite, that unique data comes from viral YouTube video transcripts that it uses to produce SEO-optimised videos for business clients.
Felix Quinque, Maekersuite’s head of AI, argues that this is data that’s less available to models like GPT-4, which are trained on content from the internet.
“If a product focuses on writing generic internet content like blog posts, that is almost exactly what GPT-4 is trained on because the training data was scraped from Google results,” he argues. “For us, there's much less training on video scripts because YouTube videos don't necessarily have the transcript readily available on the internet for pre-training.”
Generative AI and other approaches to AI
Other founders say the risk of too much investor focus on generative AI models will be less funding for other approaches to AI, like that used by Senoner’s company EthonAI. The startup doesn’t use generative models in its manufacturing quality software, but longer-standing approaches like classification and regression.
While generative AI is great at spitting out new content, classification algorithms are used for identifying patterns and anomalies in data — useful for spotting things going wrong on a production line. Regression, meanwhile, is used to assess the relationship between independent variables — if you change process “X” in a production line, that will result in higher quality “Y” in the product, for instance.
Senoner adds that investors should be mindful that these kinds of machine learning are often more applicable to valuable sectors like industrial manufacturing — evidenced by the fact that EthonAI’s tech is being used globally by industrial manufacturing giant Siemens.
One reason for this is that generative models have a problem with “explainability”, meaning it’s hard to ascertain how the AI has produced its output.
“If your algorithms aren't explainable to manufacturing experts, they won't use the results because they don't trust it,” he says.
He advises that generalist investors look for startups that focus on the problem they are solving, rather than the tech they’re using to do it.
“We always advertised ourselves as a software tooling company building a platform for quality management, as opposed to being an AI startup,” he says.
Don’t throw the baby out with the bathwater
Merantix’s Rothe also believes there’s a positive impact from the hype — as more specialists from a broad range of sectors are interested in teaming up with a technical cofounder and solving problems with AI.
Meanwhile, he and Benaich agree that generative models have huge potential when it comes to applications like DNA sequencing and finding better medicines.
There’s no doubt that companies like OpenAI and Google are going to change the world with the language models they’re building, as generative AI allows people to interact with machine learning models in more natural ways.
But, for every entrepreneur with an important problem they can now solve with the technology, there are five more who see an opportunity to capitalise on investor herd mentality.
Demetriou from Maekersuite sums it up.
“You can ride the wave of just being like a generative AI startup, but we really don't want to be a business that just exists in the short term and then evaporates because it's just hype.”
