Many large companies are eager to reap the benefits of generative A.I. but are worried about both the risks—which are numerous—and the costs. In the past few weeks, I’ve had a number of conversations with startups trying to address both of these concerns.
Leif-Nissen Lundbaek is the founder and CEO of Xayn, a six-year-old A.I. company based in Berlin. It specializes in semantic search, the term that refers to techniques that allow people to use natural language to find information, and recommendation engines, which suggest content to customers. Lundbaek tells me that while most people have become fixated on the ultra-large language models, such as OpenAI’s GPT-4 and Google’s PaLM 2, they are often not the best tool for companies to use.
If all you want is to be able to find relevant information, a huge LLM isn’t the most efficient approach in terms of cost, energy efficiency, speed, or data privacy, Lundbaek tells me. Instead, Xayn has pioneered a suite of much smaller models that are better at learning from small amounts of data and surfacing results much faster than a very large language model would. Xayn's models are small enough that they will run on a mobile phone, rather than requiring a connection to a model running in a data center. In a pilot project for German media company ZDF, Xayn’s recommendation software, which the company calls Xaynia, increased the volume of digital content users watched and the click-through rate compared to the media company's previous recommendation model, while reducing energy consumption by 98%, Lundbaek says. He says that compared to OpenAI’s latest model for embedding information, which is called Ada 002, Xaynia offers 40 times better energy performance. It is also about 20 times more energy efficient than using Google’s BERT model.
In a demonstration, Lundbaek also showed me how the model tries to infer what content a user might like based solely on a single search or a single piece of content that a person engages with—in this case, a search for football, which surfaced recommendations about the soccer team FC Bayern Munich, as well as other sports—rather than, as many recommendation engines do, trying to compare a user’s profile with those of similar users it has seen before. Xaynia’s model is based mostly on the content itself. This solves many of the data privacy concerns that companies, particularly in Europe, have about how to personalize content for users without having to store lots of sensitive data about them, he says. “It’s completely individualistic,” he says. “Even if this user looks similar to someone else.”
Another thorny problem for chatbots powered by large language models is their tendency to produce toxic or inappropriate content and to easily jump guardrails. Aligned AI, a tiny startup based in Oxford, England, has developed a technique for content moderation that it says significantly outperforms competing models created by OpenAI. On a content filtration challenge that Google’s Jigsaw division created, OpenAI’s GPT-powered content moderation was only able to accurately filter about 32% of the problematic chatbot responses, while Aligned AI’s scored 97%. On a separate evaluation dataset that OpenAI itself provided, OpenAI’s moderation system scored 79% compared to Aligned AI’s 93%.
Rebecca Gorman, Aligned AI’s cofounder and CEO, tells me that even those kinds of results may not be good enough for many enterprise use cases where a chatbot might engage in tens of thousands or hundreds of thousands or even more conversations. At such scale, missing 3% of toxic interactions would still lead to a lot of bad outcomes, she says. But Aligned AI has at least shown its methods are able to make progress on the problem.
While much of what Aligned AI is doing is proprietary, Gorman says that at its core Aligned AI is working on how to give generative A.I. systems a much more robust understanding of concepts, an area where these systems continue to lag humans, often by a significant margin. “In some ways [large language models] do seem to have a lot of things that seem like human concepts, but they are also very fragile,” Gorman says. “So it’s very easy, whenever someone brings out a new chatbot, to trick it into doing things it’s not supposed to do.” Gorman says that Aligned AI’s intuition is that methods that make chatbots less likely to generate toxic content will also be helpful in making sure that future A.I. systems don’t harm people in other ways. The work on “the alignment problem”—which is the idea of how we align A.I. with human values so it doesn’t kill us all and from which Aligned AI takes its name—could also help address dangers from A.I. that are here today, such as chatbots that produce toxic content, is controversial. Many A.I. ethicists see talk of “the alignment problem,” which is what people who say they work on “A.I. Safety” often say is their focus, as a distraction from the important work of addressing present dangers from A.I.
But Aligned AI’s work is a good demonstration of how the same research methods can help address both risks. Giving A.I. systems a more robust conceptual understanding is something we all should want. A system that understands the concept of racism or self-harm can be better trained not to generate toxic dialogue; a system that understands the concept of avoiding harm and the value of human life, would hopefully be less likely to kill everyone on the planet.
Aligned AI and Xayn are also good examples that there are a lot of promising ideas being produced by smaller companies in the A.I. ecosystem. OpenAI, Microsoft, and Google, while clearly the biggest players in the space, may not have the best technology for every use case.
With that, here’s the rest of this week’s A.I. news.
Jeremy Kahn
@jeremyakahn
jeremy.kahn@fortune.com
