Meta has made a huge step in AI, Mark Zuckerberg says

Meta has made a huge step forwards in the AI race, Mark Zuckerberg has said.

The chief executive revealed a new model, named “Muse Wonder”, that it says is as powerful as the latest technology from OpenAI and Anthropic.

For years, Meta has looked to position itself as among the leaders in the development of artificial intelligence. But it has lagged dramatically behind many of its rivals, and has failed to develop a model that has been seen as being at the “frontier” of AI.

In recent months, however, the work of its “Superintelligence Labs” division has made a number of strides. Last month, it released a new “Meta Muse” agent that aims to run your life – and now it says it has caught up with the leading AI companies in the performance of its systems.

The company revealed two new models: Muse Wonder, and Muse Wonder Full. The first is intended to be more versatile, and will do most tasks, while the full version is the “largest and most capable model we have trained”, Meta said.

It is specifically intended to be used as a personal assistant, and Mr Zuckerberg said it was a step towards developing “personal superintelligence for everyone”. Meta has differed from its rivals in a number of ways, including a specific focus on building AI systems that are intended to help people with their lives, providing personal information and doing tasks on their behalf.

The new model will allow the model to undertake long tasks, such as “a cross-country move or a claim their provider denied”, Meta said. Because those tasks involve a number of complicated tasks over many days, less capable models can produce errors as they undertake them.

It also includes “active memory” capabilities that allow it to understand tasks over time. It means the model will be able to remember important context and be able to find relevant details when they are needed.

It is also better at being able to retain focus, Meta said, meaning that it will keep working towards a goal even when it runs into challenges. Meta stressed that the model will “ask for clarification when needed” and check before “taking important actions”, presumably to allay concerns about rogue systems that can become dangerous precisely because they are so focused on their task.

Meta also said that it had focused on safety in building the model. It is particularly able to resist “adversarial instructions” that might try to break the model, for instance, and will be more careful in handling important information.

Using the new models is notably expensive, compared to Meta’s other systems, though is in line with the top models from other companies such as OpenAI and Anthropic. Meta also said that the system will soon be used to power more of the Muse agent, which is free.

Apple is about to launch a new product that could decide its future

Anthropic forced to ban models from the internet to stop them going ‘rogue’

AI torture experiments are testing our understanding of morality