Microsoft’s new AoT aims to create more human-like AI

Microsoft teamed up with Virginia Tech to publish a white paper introducing their new "Algorithm of Thoughts" (AoT). The objective? To make language learning models akin to human learning.

https://preview.redd.it/3q8lcq0k96lb1.png?width=2000&format=png&auto=webp&s=fce7a0e3225c64353ad6c51f65e8b490e795feed

If you want to stay on top of the latest trends and insights in AI and tech, look here first.

What's the big idea?

• Microsoft's AoT aims to fuse the accuracy of algorithms with the nuances of human reasoning. A bold aspiration indeed, but not a new one. The goal to empower computers to learn for themselves – akin to human cognition - has been an AI objective since its inception back in the 1950s.
• The AoT could be seen as an attempt to resolve the drawbacks of the "Chain of Thought" (CoT) approach. LLMs following the CoT approach can provide incorrect steps to the right answer, as they base conclusions on precedent.
• With AoT, the model works to evaluate the soundness of initial steps or "thoughts," reducing the risk of one incorrect step leading to disproportionate results.

What could AoT do?

• Mitigate AI "hallucinations:" These funny— but disconcerting — instances of AI outputting false information.
• Enhance the integrity of AI interaction: programmers suggest that improvement in this aspect is crucial for aligning AGI (artificial general intelligence).

The takeaway:

• AI's ability to understand and process information like a human being is a longstanding goal in the field. With AoT, Microsoft seems to be making strides toward achieving it.
• Much remains to be seen on its efficacy: How it will impact the broader AI ecosystem and the user experiences it can create.

P.S. If you like this kind of analysis, I write a free newsletter tracking the most relevant news and research in AI and tech—stay informed in under 3 minutes/day.

(source)

​