91探花

AI tech news is a constant back and forth between announcing exciting innovative potential and sharing warnings about the concerning capabilities of AI, akin to sci-fi thrillers featuring killer robots 鈥� like a tug of war between the devil on your shoulder and the angel whispering in your ear.

And, because AI tech is changing and progressing at a rate of knots, almost unlike anything we鈥檝e ever experienced before, even prominent, knowledgeable figures from deep within the industry seem to be unable to commit to a single narrative.

Do we like AI or not? Is it changing our lives for the better, or will it lead to the eventual demise of life as we know it and the human race?

A little dramatic, sure, but the point is, there鈥檚 no real consensus or long-lasting agreement on the AI issue.

It鈥檚 healthy (if not necessary), of course, to have experts change their minds after having been presented with new information and evidence 鈥� nobody should be resolutely stuck to a single plan just for the sake of it. However, too much back and forth and contradiction is unsettling, and it feels as though this is becoming the norm in the AI industry.

The latest switch-back in the AI universe is centred on reasoning. Up until recently, the primary objective of AI experts and tech developers was to replicate human reasoning 鈥� that would put us one (big) step closer to achieving generative AI.

Or so we鈥檝e been told.

However, hold on to your britches (or, hold on to your, uh, cargo pants? What are the kids wearing these days?), because as of yesterday, we鈥檙e no longer necessarily all about deep thinking AI. In fact, we鈥檙e being told that having AI do too much reasoning is actually bad news.

The main headline? It鈥檚 expensive.

Not just a few extra pennies expensive 鈥撎� we鈥檙e talking AI companies not only bleeding money, but also consuming way more energy than necessary, making an industry that was already an environmental concern become a greedy, wasteful monstrosity.

听

What Does 鈥淭oo Much Reasoning鈥� Mean?

听

In the past, we鈥檝e been told that 鈥渋ntelligent鈥�, so to speak, processing of AI models is the aim of the game 鈥� that we鈥檙e trying our best to make them as sophisticated as possible. Naturally, a big component of that has been trying to get these models to be able to 鈥渢hink鈥� and 鈥渞eason鈥� in the same way that humans can and do.

So, it鈥檚 no surprise that the most recent revelation that models are doing 鈥渢oo much reasoning鈥� feels like it鈥檚 come out of left field. Why is it bad for AI models to 鈥渢hink鈥� too much? Surely we听want听them to become as sophisticated and 鈥渟mart鈥� as possible?

Well, yes and no.

Yes, we want them to progress in terms of their capabilities to become more sophisticated and capable than ever before.

But, what experts are quickly realising is that while processing is expensive and requires the consumption of energy, having AI models do more advanced 鈥渞easoning鈥� requires even more processing than normal, making it exponentially more expensive and making its energy consumption excessive.

Not only is deep thinking more of a drain on resources, however, the problem is that it鈥檚 also being done unnecessarily.

For instance, Sam Altman of OpenAI recently made headlines by explaining that the simple act of saying 鈥減lease鈥� and 鈥渢hank you鈥� to ChatGPT (or any other AI model, for that matter) costs companies an exorbitant amount of money and requires a shocking amount of energy consumption.

听

听

The Solution, According To Industry Leaders听

听

Much remains to be seen, but Google鈥檚 DeepMind has introduced its own solution to this new 鈥減roblem鈥� 鈥� a dial that allows you to change how much the model reasons. The intention behind this isn鈥檛 to stunt its capabilities and stop it from producing high-quality responses, it鈥檚 intended to stop Gemini from thinking more than it needs to in specific contexts.

For instance, an issue that most experts haven鈥檛 yet been able to solve is the fact that AI models tend to think more deeply than necessary about simple queries, costing way more money and using way more power than it needs to.

In fact, in conversation with MIT Technology Review, Nathan Habib, an engineer at Hugging Face, asserted that this is not the exception, as we may think 鈥� rather, overthinking is more like the rule.

Thus, this new reasoning 鈥渄ial鈥� has been introduced to allow developers (not yet end users) to decide how much they鈥檙e willing to spend on reasoning, which then dictates how much reasoning can take place. The result is that models can鈥檛 just 鈥渢hink鈥� and reason endlessly over basic prompts 鈥� they鈥檙e limited.

The problem that still exists, however, is that it鈥檚 not clear just how much reasoning is a good amount of reasoning for certain tasks. Perhaps in the future, experts will be able to set out specific parameters dictating what kind of prompts require different levels of reasoning, but for now, this is a fairly new issue that AI companies are facing, with potentially serious effects on the environment.