Zuckerberg’s Meta Launches AI Model to Check Other AI Models' Accuracy

Mark Zuckerberg’s Meta Platforms released an official statement on Friday that new AI models were released from its research division – Fundamental AI Research (FAIR).

These models have a ‘Self-Taught Evaluator’ that offers the possibility of less human involvement in the entire AI development process and another model that can freely combine text and speech. 

The AI researchers under FAIR said that new releases hold up the company’s goal of achieving advanced machine intelligence while incorporating open science and reproducibility.

Also read: Akash Ambani bats for AI localization at the IMC, 2024 (startuppedia.in)

Meta releases new models

Meta’s newly released models include updated Segment Anything Model 2 for images and videos, Meta Spirit LM, Layer Skip, SALSA, Meta Lingua, OMat24, MEXMA, and Self Taught Evaluator. 

Meta announced in August how these models would rely on the ‘chain of thought’ mechanism. OpenAI used these for its recent o1 models that think before they respond.

Incidentally, bigshot companies such as Google and Anthropic have also published research on the concept of Reinforcement Learning from AI Feedback. However, these models have not yet been released for public use. Meta has declared its new model to be capable of validating other AI models’ works as a “strong generative reward model with synthetic data”.

Big claims about the Self Taught Evaluator 

As per their claims, the Self Taught Evaluator has a new method for generating preference data to train reward models without relying on human interventions. “This approach generates contrasting model outputs and trains an LLM-as-a-Judge to produce reasoning traces for evaluation and final judgments, with an iterative self-improvement scheme,” Meta disclosed in its official blog post. 

As announced by the company, it is a powerful model that performs better than models relying on human-labeled data such as GPT-4.

“Many existing AI voice experiences today use ASR to techniques to process speech before synthesizing with an LLM to generate text — but these approaches compromise the expressive aspects of speech. Using phonetic, pitch and tone tokens, Spirit LM models can overcome these limitations for both inputs and outputs to generate more natural sounding speech while also learning new tasks across ASR, TTS and speech classification,” Meta stated in a recent tweet. 

Also read: Zuckerberg’s Meta Sacks Employees for Misusing $25 Meal Vouchers (startuppedia.in)