Muse and Meta AI get new models that remember better and keep going

Anyone who gives an AI assistant a job does not want to keep explaining what is going on. And the assistant should not stop at the first setback. Meta now says it is making progress on both points. Since Friday 6 November, the company has been rolling out new models in Muse, its personal AI assistant that carries out tasks on its own, and in the chatbot Meta AI. Chief executive Mark Zuckerberg wrote on X that the improvements are going to all users.

The new models are specialised versions of existing Meta models. They were trained with the help of Muse Wonder, the company's most powerful model, which came out on 12 October. They do not have their own name. In this way, two of Wonder's skills reach many more people: Active Memory and Active Focus.

Remembering what matters

Active Memory is meant to make the assistant remember what is important and recognise when something from earlier becomes useful again. Meta taught the models the difference between temporary and lasting information, and between data that is still correct and data that is out of date.

According to Meta, the new models in Muse forget crucial information 2.4 times less often. In the company's own test, the new model used that information correctly in 90 per cent of tasks, compared with 76 per cent for the previous model.

Keeping going when things go wrong

Active Focus is meant to help Muse keep working towards a goal when something unexpected happens. Meta trained the models on all kinds of obstacles that also occur in real life. In this way they learn to try harder, recover or look for another solution. According to the company, this gives a 67 per cent improvement in finishing everyday tasks in which something unexpected goes wrong.

Meta also says the models are better at sensing when to carry on by themselves and when to check with the user. They can also handle corrections halfway through a task. In Meta AI, the same improvements are meant to make a better conversation partner, one that breaks a problem down into steps and asks sharper follow-up questions.

All the figures come from Meta itself. The next step is for others to measure them independently.

From a powerful model to everyday use

Muse came out on 8 September, at first only in the United States, and also got a free version. According to the tech site RuntimeWire, Muse has now been set up more than 10 million times in North America. At its launch in October, Muse Wonder itself was meant for developers, through Meta's paid developer service and the programming tool Muse Code. Meta announced at the time that it would also use Wonder to make specialised models for everyone who uses Muse. That step has now been taken.

Meta is also working on models that anyone can download and run themselves, such as Muse Spark Open, a large model for programming. The models in this article, by contrast, are meant for Meta's own apps and for ordinary users.

What does this mean for the road to the singularity?

An assistant that carries out tasks by itself only becomes really useful when it remembers what is going on and does not get stuck at the first problem. With this step, Meta shows how a large, powerful model can help bring such skills to models for everyday use. As a result, the strongest abilities of AI reach ordinary users faster. If independent measurements confirm the figures, assistants that finish long tasks on their own come another step closer.

What is the technological singularity? Read the explanation →