Since ChatGPT was launched in November, a new mini-industry has mushroomed that has defied the broader slump in tech. Not a week goes by without someone unveiling a “generative” artificial intelligence (AI) underpinned by “foundation” models—the large and complex algorithms that give ChatGPT and other AIs like it their intelligence. On February 24th Meta, Facebook’s parent company, released a model called LLaMA. This week it was reported that Elon Musk, the billionaire boss of Tesla and Twitter, wants to create an AI that would be less “woke” than ChatGPT. One catalogue, maintained by Ben Tossell, a British tech entrepreneur, and shared in a newsletter, has recently grown to include, among others, Ask Seneca (which answers questions based on the writings of the stoic philosopher), Pickaxe (which analyses your own documents), and Issac Editor (which helps students write academic papers).
ChatGPT and its fellow chatbots may be much talked about (and talked to: ChatGPT may now have more than 100m users). But Mr Tossell’s newsletter hints that the real action in generative AI is increasingly in all manner of less chatty services enabled by foundation models.
Each model is trained on reams of text, images, sound files or any other heap of data. This allows them to interpret, react to and create statements in natural language, as well as art, music and any other type of content you find on the internet. Even as the venture-capital (VC) industry nurses a giant hangover after the recent tech crash put paid to a bubbly couple of years, entrepreneurs experimenting with generative AI have no trouble attracting investments. In January it was reported that Microsoft poured another $10bn in OpenAI, the startup behind ChatGPT, on top of an earlier investment of $1bn. A spreadsheet maintained by Pete Flint at NfX, a VC firm, now lists 539 generative-AI startups. Not counting OpenAI, they have so far collectively raised more than $11bn in capital (see chart). Mike Volpi of Index Ventures, another VC firm, calls it a “Cambrian explosion”.
Several factors are driving it. Though foundation models have been around for some time, Mr Volpi explains that it took a consumer-facing service such as ChatGPT to capture the world’s—and investors’—imagination. This happened just as venture capitalists disappointed by the cryptocurrency crash and the empty metaverse were on the lookout for the next big thing. In addition, even more than web browsers and smartphones before them, foundation models make it easy to build new services and applications on top of them. “You can open your laptop, get an account and start interacting with the model,” says Steve Loughlin of Accel, yet another VC firm.
The question for venture capitalists is which generative-AI platforms will make the big bucks. For now, this is the subject of much head-scratching in tech circles. “Based on the available data, it’s just not clear if there will be a long-term, winner-take-all dynamic in generative AI,” wrote Martin Casado and colleagues at Andreessen Horowitz, one more VC firm, in a recent blog post. Many startups offer me-too ideas, many of which are a feature rather than a product. In time even the resource-intensive foundation models could end up as a low-margin commodity: although proprietary models such as OpenAI’s GPT-3.5, which powers ChatGPT, are still leading, some open-source ones are not far behind.
Another source of uncertainty is the legal minefield onto which generative AI is tiptoeing. Foundation models often get things wrong. And they can go off the rails. The chatbot which Microsoft is developing based on OpenAI’s models for its Bing search engine has insulted more than one user and professed its love to at least one other (Sydney, as Microsoft’s chatbot is called, has since been reined in). Generative-AI platforms may not enjoy the legal protection from liability that shields social media. Some copyright holders of web-based content on which existing models are being trained willy-nilly, without asking permission or paying compensation, are already up in arms. Getty Images, a repository of photographs, and individual artists have already filed lawsuits against AI art-generators such as Stable Diffusion. News organisations whose articles are plundered for information may do the same.
OpenAI is already trying to manage expectations, downplaying the launch later this year of GPT-4, the highly anticipated new version of the foundation model behind ChatGPT. That is unlikely to temper VC types’ appetite for generative AI. For more risk-averse investors, the safest bet at the moment is on the providers of the ample processing power needed to train and run foundation models. The share price of Nvidia, which designs chips useful for AI applications, is up by 60% so far this year. Cloud-computing services and data-centre landlords are rubbing their hands, too. Whichever AI platform comes out top, you can’t go wrong selling picks and shovels in a gold rush. ■