Malted AI — a startup HQ’d in Edinburgh — has raised £6m to scale up its development of “distilled” GenAI models for businesses. The round was led by Hoxton Ventures and joined by Creator Fund and angel investors.
The company’s distillation process uses a “teacher” system that takes a company’s raw data and feeds it into a string of small, efficient language models, which a company can use to complete specialised tasks.
“It’s quite contrary to where a lot of people are focusing in terms of a bigger, more general [AI model] is better,” says Iain Mackie, Malted AI cofounder and CEO.
The case for small models
Malted AI’s focus on smaller language models mirrors a wider trend going on in generative AI — which for some years was following the logic that bigger models (trained on more data) produced better results. Companies like Paris-based Mistral and Silicon Valley’s Meta are pouring a lot of resources into training models that rely on a fraction of the amount of data that OpenAI’s GPT-4 does.
While general purpose chatbots like ChatGPT, which are expected to respond to all sorts of queries, benefit from a large training dataset, more narrow business applications can get comparable results from smaller, cheaper-to-run models.
“At scale it could easily be 10-100x cheaper [to use smaller models],” says Mackie. “In the long run, really smaller focused models definitely have a cost benefit.”
Malted AI is working on “a couple” of pilots with large enterprises, helping them to build language model-powered apps for things like retrieving complex information from large internal datasets or spotting regulatory issues in text documents. It says that, by using its distillation technology to generate specialised training data, it’s able to develop “smaller, more focused AI models with greater performance at a fraction of the cost.”
A Scottish story
There are lots of companies trying to help businesses make use of generative AI for enterprise use cases, ranging from small startups to foundation model companies that are raising hundreds of millions of dollars in investment.
“What we're focused on is being known as kind of best in class for distillation and making really small focused AI models,” says Mackie.
Malted AI also sets itself apart from the crowd with its Edinburgh base, outside Europe’s leading AI hubs in London and Paris. Scotland, Mackie argues, has a lot of potential because of its leading research universities, even if that’s yet to translate into big AI companies.
“Glasgow is one of the world's largest research areas for information retrievals and search systems. Edinburgh has some of the most talented NLP [natural language processing] people,” he says. “There are really amazing researchers, but right now there aren't many interesting VC-backed startups, so we’re trying to give people really exciting jobs, working on the cutting edge without having to leave for London or Silicon Valley.”
Mackie’s cofounder Carlos Gemmell previously worked in Meta’s language model division. The Malted AI team — currently 10 people — also features experience from DeepMind, Amazon and Bloomberg.
The startup will use the new capital to keep financing the development of its tech, and building out its commercial go-to-market strategy with early partners.