Facebook parent Meta publishes AI model that can argue: “…as AI becomes more and more superhuman…”

Metathe parent company of Facebook and Instagram, recently announced that it is releasing new AI models, including a “Self-Taught Evaluator,” from its research department. Among other things, this AI model uses AI to evaluate other AI models, potentially reducing the need for human involvement in the development process.
Metas Self-taught appraiser is nothing more than a “chain of thought” reasoning technique used by OpenAI's latest model OpenAI o1 and is supposed to provide arguments when answering.

What Meta's new AI model does and how it differs

According to a Reuters report, this approach breaks down complex tasks into smaller, more manageable steps, improving accuracy in areas such as mathematics, scientific analysis and coding.
What's special about Meta's model is that it is trained entirely on AI-generated data, eliminating the reliance on human-labeled data.
In particular, this AI model approach could pave the way for more autonomous AI systems capable of self-improvement. By learning from its own mistakes, AI could become increasingly precise and sophisticated without constant human intervention.
“We hope that as the AI ​​becomes more superhuman, it will get better at checking its work, so that it will actually be better than the average human,” said Jason Weston, one of the researchers.
This can also streamline the development process and potentially lead to more powerful and efficient AI applications.
“The idea of ​​being self-taught and being able to self-assess is fundamentally crucial to the idea of ​​reaching this superhuman level of AI,” he added.

Other AI tools that Meta publishes

In addition to the Self-Taught Evaluator, Meta also released other AI tools, including an update to its Segment Anything image identification model, a tool to speed up response times in large language models, and datasets to aid in the discovery of new inorganic materials.