Stay up up to now with free updates
Just register Artificial intelligence myFT Digest – delivered straight to your inbox.
Artificial intelligence corporations which have poured billions of dollars into developing so-called large language models for generative AI products at the moment are turning to a brand new solution to increase their revenue: small language models.
Apple, Microsoft, Meta and Google have all recently released recent AI models with fewer “parameters” – the variety of variables used to coach an AI system and shape its output – but still with powerful features.
The moves are an attempt by technology groups to encourage AI adoption amongst corporations concerned about the associated fee and computing power required to run large language models, the variety of technology that underlies popular chatbots like OpenAI's ChatGPT .
In general, the upper the variety of parameters, the higher the performance of the AI software and the more complex and differentiated its tasks could be. OpenAI's latest model GPT-4o and Google's Gemini 1.5 Pro, each announced this week, are estimated to have greater than 1 trillion parameters, and Meta is training a 400 billion-parameter version of its open source Llama model.
Not only is it difficult to persuade some enterprise customers to pay the massive amounts required to run generative AI products, but there are also data and copyright liability concerns hampering adoption.
This has led to tech corporations like Meta and Google offering small language models with just a number of billion parameters as cheaper, energy-efficient and customizable alternatives that require less energy to coach and run and may shield sensitive data.
“By providing a lot quality at a lower cost, you truly allow customers so many more applications to enter and do things that prohibitively didn't provide enough of a return on that investment to justify actually doing it,” said Eric Boyd, corporate vp of Microsoft's Azure AI Platform, which sells AI models to enterprises.
Google, Meta, Microsoft and French startup Mistral have also released small language models that display advanced capabilities and may higher concentrate on specific applications.
Nick Clegg, Meta's president of worldwide affairs, said Llama 3's recent 8 billion parameter model is comparable to GPT-4. “I believe you’ll be able to see superior performance on just about every measurement you’ll be able to consider,” he said. Microsoft said its Phi-3 Small model with 7 billion parameters outperformed GPT-3.5, an earlier version of the OpenAI model.
The small models can process tasks locally on a tool fairly than sending information to the cloud, which could appeal to privacy-conscious customers who need to ensure information is stored inside internal networks.
Charlotte Marshall, managing partner at Addleshaw Goddard, a law firm that advises banks, said that “one in every of the challenges I believe lots of our clients have had” in launching generative AI products was regulatory requirements around processing and transfer of knowledge. She said smaller models “provide a chance for corporations to handle legal and value concerns.”
Smaller models also allow AI functions to run on devices corresponding to mobile phones. Google's “Gemini Nano” model is integrated into its latest Pixel phone and Samsung's latest S24 smartphone.
Apple has hinted that it is usually 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 smaller models would result in “interesting applications, all the best way as much as phones and laptops.”
OpenAI CEO Sam Altman said in November that the San Francisco-based startup offers its customers different sized AI models that “serve different purposes” and that it’s going to proceed to develop and sell those options.
“There are some things where smaller models work rather well,” he added. “I’m looking forward.”
However, Altman added that OpenAI will proceed to concentrate on constructing larger AI models with expanded capabilities, including the flexibility to reason, plan and execute tasks, and ultimately achieve human-level intelligence.
“I often think people just want the very best model,” he said. “I believe that’s what people want most.”
