Unlock the Editor’s Digest for free
Roula Khalaf, Editor of the FT, selects her favourite stories in this weekly newsletter.
Does the future belong to a handful of all-powerful, wide-ranging artificial intelligence agents that navigate the world on our behalf — successors to the ChatGPTs, Claudes and Groks that seek to handle almost any task you throw at them? Or will it be populated by a host of specialised digital aides, each trained to take on a narrow task and invoked only when needed?
Some mix of the two seems likely, but the sheer pace of change has left even leaders in the field admitting they have little idea of how things will look a year or two out.
For proponents of the “One AI to rule them all” idea, there have been plenty of encouraging developments. OpenAI, for instance, added a shopping feature to ChatGPT this week that points to how personalised AI agents could reorder the economics of ecommerce. Using a single query to get a chatbot to do your product research and make a buying recommendation threatens to subvert the entire “funnel” that brands have relied on to steer buyers, putting OpenAI very much at the centre.
Advances like these may grab the most attention, but behind the scenes a new generation of more specialised agents is starting to take shape. These promise to be narrowly targeted and — a key consideration — far cheaper, both to build and to run.
Meta’s LlamaCon developer conference this week provided a glimpse of the state of play. The social networking company has placed its bet on the adaptability of its “open weights”, AI models that have a limited form of an open-source structure. This enables others to use and adapt the models, even if they can’t see exactly how they were trained.
One sign that Meta has hit a nerve in the wider tech world is the 1.2bn downloads its “open” Llama models have had in their first two years. The vast majority of these have involved versions of Llama that other developers have adapted for particular uses and then make available for anyone to download.
The techniques for turning these open weights models into useful tools are evolving fast. Distillation, for instance — imbuing small models with some of the intelligence from much larger ones — has become a common technique. Companies with “closed” models, like OpenAI, reserve the right to decide how and by whom their models can be distilled. In the open weights world, by comparison, developers are free to adapt models as they want.
The interest in creating more specialised models has picked up in recent months as more of the focus of AI development has shifted past the data-intensive — and highly expensive — initial training runs for the biggest models. Instead, much of the special sauce in the latest ones is created in the steps that come next — in “post-training”, which often uses a technique known as reinforcement learning to shape the results, and in the so-called test-time phase used by reasoning models to work through a problem.
According to Ali Ghodsi, chief executive of Databricks, one powerful form of post-training that has been catching on involves using a company’s proprietary data to shape models in their reinforcement learning phase, making them far more reliable for business use. Speaking at Meta’s event, he said this is only possible with open models.
Another favourite new trick has been to combine the best parts of different open models. After DeepSeek shocked the AI world with the success of its low-cost R1 reasoning model, for instance, other developers quickly learnt how to copy its reasoning “traces” — the step-by-step patterns of thought that showed how it worked through a problem — and run these on top of Meta’s Llama,
These and other techniques promise a tidal wave of smart agents that require less expensive hardware and consume much less power.
For the model builders, meanwhile, it adds to the risk of commoditisation — that cheaper alternatives will undermine their most expensive and advanced models.
But the biggest winners of all, as the cost of AI falls, could be the users: companies that have the wherewithal to design and embed specialised agents into their day-to-day work processes.