Stay informed with free updates
Sign up on the bus. Artificial intelligence myFT Digest — delivered straight to your inbox.
Artificial intelligence companies that have spent billions building so-called large language models to power AI products are now banking on a new way to boost revenue: small language models.
Apple, Microsoft, Meta and Google have recently released new AI models with fewer “parameters” — the number of variables used to train an AI system and shape its output — but still With powerful abilities.
The initiative is an effort by technology groups to encourage the adoption of AI by enterprises that are concerned about the cost and computing power required to run large language models, the type of technology popularized by chatbots such as OpenAI's ChatGPT advances.
In general, the higher the number of parameters, the better the performance of the AI software and the more complex and fine-grained its tasks can be. OpenAI's latest model GPT-4o and Google's Gemini 1.5 Pro, both announced this week, are estimated to have more than 1tn parameters, and Meta is training a 400bn-parameter version of its open-source Llama model. .
As well as struggling to convince some enterprise customers to pay the large sums required to run generative AI products, there are also concerns over data and copyright liability holding back adoption.
This has led tech groups like Meta and Google to develop small language models with just a few billion parameters as cheap, energy-efficient, customizable alternatives that require less power to train and run. is, which can also color sensitive data.
“By having that much quality at a low cost point, you actually enable a lot of applications for customers to go in and do things that are worth it to them on that investment,” said Eric Boyd. There wasn't enough of a return to really justify doing that,” Eric Boyd said. Corporate vice president of Microsoft's Azure AI platform, which sells AI models to enterprises;
Google, Meta, Microsoft and French startup Mistral have also released miniature models of the language that show advanced capabilities and can better focus on specific applications.
Llama 3's new 8bn parameter model is comparable to GPT-4, said Nick Clegg, Meta's president of global affairs. “I think about every metric you can think of, you see high performance,” he said. Microsoft said its Phi-3-small model, with 7bn parameters, outperformed GPT-3.5, an older version of OpenAI's model.
Smaller models can process tasks locally on a device rather than sending information to the cloud, which may appeal to privacy-conscious users who want to keep information in internal networks.
Charlotte Marshall, managing associate at Edelshaw Goddard, a law firm that advises banks, said “I think one of the challenges that many of our clients face” in adopting generative AI products is data. Regulatory requirements for handling and transfer had to be followed. Smaller models “have given businesses an opportunity to overcome legal and cost concerns,” he said.
Smaller models also allow AI features to run on devices such as mobile phones. Google's “Gemini Nano” model is embedded inside its latest Pixel phone and Samsung's latest S24 smartphone.
Apple has hinted that it is also developing AI models to run on its best-selling iPhone. Last month, the Silicon Valley giant released its OpenELM model, a small model designed to perform text-based tasks.
Microsoft's Boyd said the smaller models would lead to “interesting applications, all the way up to phones and laptops.”
Sam Altman, the head of OpenAI, said in November that the San Francisco-based startup offered customers AI models of different sizes “for different purposes,” and that it would continue to build and sell those options.
“There are some things where smaller models will work really well,” he added. “I'm excited for it.”
However, Altman added that OpenAI's focus will continue to be on building large-scale AI models, including the ability to reason, plan and execute tasks and eventually achieve human-level intelligence.
“There are a lot of times where I think people just want the best model,” he said. “I think that's what people mostly want.”
Additional reporting by George Hammond in San Francisco