"Facebook owner Meta
said Friday it is releasing a suite of new artificial intelligence models from
its research division, including a "self-evaluator" that could help
reduce human involvement in the process of building artificial intelligence.
In a paper
published by Meta in August, they introduced the tool, detailing how it relies on the
same "chain of thought" method used by the recently released OpenAI
models, o1, to reliably evaluate model responses.
The method involves
breaking down complex problems into smaller logical steps and appears to
improve the accuracy of answers to complex problems in subjects such as
science, coding and mathematics, according to Reuters.
Meta researchers used
only AI-generated data to train the estimator model, so no human input was used
at this stage either.
Two Meta researchers
working on the project told Reuters that the ability to reliably evaluate AI
opens the door to autonomous AI agents that can learn from their own mistakes.
Many in the field of
artificial intelligence envision such agents as digital assistants that are
intelligent enough to perform many tasks without human intervention.
Self-evolving models
could help avoid the expensive process of learning from human feedback, which
requires a human annotator with specialized knowledge to accurately annotate
data and verify that answers to complex mathematical and written queries are
correct.
"We hope that
artificial intelligence will get better at checking its own work so that it is
actually better than the average human," said one of the researchers,
Jason Weston.
"The idea that
AI can learn and judge itself is fundamentally important to achieving that
superhuman level of AI," he said." [1]
Komentarų nėra:
Rašyti komentarą