This is how Microsoft promises to improve them with this training technique

Sometimes the great advances of artificial intelligence do not happen with hype and cymbal a new product, but by publishing a scholarly paper and waiting a few years to see new knowledge take shape. ChatGPT itself owes its existence to a technique developed by Google scientists in 2017, for example.

Now, it is Microsoft scientists who have just published their own paper (PDF) which, according to some, could transform the landscape of generative artificial intelligence: they have released an innovative training method for AIs called “Algorithm of Thoughts” (AoT)or ‘algorithm of thoughts’, wow.

Its creators point out that it can make large language models, such as ChatGPT, be more efficient and humane in their ‘reasoning’ skillsspecifically of ‘intuitive reasoning’, a characteristic of human thought that seeks to make it compatible with the exhaustiveness of algorithms…

…at the same time that overcomes the limitations of working memory of us humansthus allowing a more complete exploration of ideas.

“[Buscamos] merge these dual facets to increase reasoning capabilities within LLMs”.

“Interestingly, our results suggest that training a model with an algorithm can lead to performance outperforming the algorithm itself.”

A more efficient alternative

The result? A faster, more efficient problem-solving process in terms of resource consumption. In fact, the researchers claim that this technique outperforms previous single-query methods and is on par with multiple-query methods that use extensive tree searches.

The AoT technique represents a significant change in the way language models learn and reason: instead of relying on a “query and answer” approachAoT allows artificial intelligences to internalize a more complete thought process, incorporating this ‘artificial intuition’ to the search for solutions…

…and, by the way, constitutes a response to the limitations of other existing ‘learning in context’ techniquessuch as the “Chain-of-Thought” (CoT) approach, which sometimes produces incorrect intermediate steps of reasoning.

Microsoft can hit the table (if you know how to seize your opportunity)

What can this mean for AI? Well, ChatGPT is already the undisputed leader in the chatbot market… if Microsoft—remember that it has invested heavily in OpenAI and uses ChatGPT as the basis for its own chatbots and assistants—is able to power it up just now that LlaMa 2, Claude and Bard seemed ready to start hot on its heels, it could strengthen its dominant position.

On the other hand, the basis of the research has just been made public with the launch of the ‘paper’: although Microsoft has an advantage when it comes to taking advantage of AoT, nothing guarantees that someone more skilled will not beat him in his own field (as happened to Google at the time).

Image | Marcos Merino through AI

In Genbeta | We have not yet reached ‘general AI’, but according to OpenAI we must think about how we will regulate the next step: ‘superintelligence’