What happens when AI starts improving itself? It may sound like science fiction, but it’s beginning to happen. For decades, humans have been laying the foundation of AI research. The recent rise of large language models (LLMs) has accelerated this progress, making AI more accessible to the masses and speeding up development & discovery. While challenges remain, announcements like this bring us one step closer to achieving Artificial General Intelligence (AGI)—a point where machines can match human intelligence across all domains of knowledge. While many remain unaware, or flat out in denial, those that are paying attention are witnessing a moment as powerful as the dawn of the internet.
Meta is making waves in the AI world with its latest innovation, a self-taught evaluator, as reported by Laura Varley. This new AI model is designed to train other AI systems without human input, aiming to boost efficiency and scalability for enterprises using large language models (LLMs). The technology was first introduced in a paper back in August and uses a ‘chain of thought’ method similar to recent Open AI models to generate more reliable responses and judgments.
So, how does this work? Essentially, the self-taught evaluator breaks down complex problems into manageable steps, which helps improve the accuracy of responses across various topics like science, coding, and math. By eliminating the need for human-labeled data, Meta hopes to speed up the evaluation process and cut costs.
There are some clear benefits here. For one, this could significantly reduce the time and expense associated with human involvement in training AI models. Jason Weston, a research scientist at Meta, even suggests that as AI becomes more advanced, it could surpass human capabilities in checking its work, potentially leading to more accurate and efficient AI systems.
However, there are concerns too. The risks associated with reduced human oversight in AI processes are not to be taken lightly. Meta has faced criticism in the past, notably when its Galactica AI tool was pulled after generating biased content. Moreover, there’s a broader industry issue with the lack of robust evaluations and standardization in responsible AI reporting, as highlighted by a report from the AI Index.
For entrepreneurs, this technology opens up several exciting opportunities:
- A startup that uses self-taught evaluators to streamline AI model training for educational technology, making learning tools more adaptive and personalized.
- A business that offers AI-driven coding assistance, using the evaluator to ensure code accuracy and efficiency without human oversight.
- A company that leverages self-taught evaluators to enhance scientific research, providing more reliable data analysis and hypothesis testing.
As we consider the potential of Meta’s self-taught evaluator, one question remains: How can we balance the benefits of reduced human involvement in AI training with the need for accountability and ethical oversight?
Image Credit: DALL-E
—
Want to get the RAIZOR Report with all the latest AI news, tools, and jobs? We even have a daily mini-podcast version for all the news in less than 5 minutes! You can subscribe here.
RAIZOR helps our clients cut costs, save time, and boost revenue with custom AI automations. Book an Exploration Call if you’d like to learn more about how we can help you grow your business.
