The Economist, in the April 12th-18th 2025 issue, had four articles on the excessive power requirements of LLMs. However, let me start this article with the following extract from an interview with Jeff Bezos of Amazon – the source of the 50 Watts in the title.

<< But you’ve got to remember humans learn in a very different way so its, we don’t yet know how, human are unbelievably data efficient.

We learn these incredibly complex things, you know, you don’t have to drive a million miles to be able to drive a car.

But the way we teach a self-driving car to drive today is we have the algorithms drive a million miles and they are still not as good in certain scenarios as a human would be and so humans are doing all kinds of special things to make that possible.

We are also very power efficient, so you know Alpha Go which is a really impressive achievement, beat the world’s best Go player – as you guys undoubtedly know, but, you know, he’s (the human) operating on about 50 watts.

…If Alpha Go had to be limited to 50 watts, he’d (the human) have creamed it, and so it’s very interesting:-

humans are just doing something fundamentally different from the current way that we do machine learning and machine intelligence. >>

The something different that humans are doing is the subject of my 20 year long research into the mathematics of language and the science of cognition – specifically the science of human-quality conversations. This means conversations that are error-free and it means the creation of persistent memories that capture meaning, it also means the discovery of a verifiable model for the resolution of anaphoric references, which are such an important feature of human languages – although a feature that is unknown in classical computer languages.

There is also one other aspect of my research that in the longer term may be seen as the most economically important – and that is a standardised representation for the essential persistent memories.

IF mainstream research had uncovered a mathematical theory for language, and IF mainstream research had discovered a verifiable theory for the resolution of anaphoric references, then perhaps our interactions with voice assistants would at least approach human-quality. But that would be the world of science — not the alchemy we now have.

If you are unaware of the context
equation, which simply states that the meaning of a sentence is determined when the sentence is mediated by context, then you are unlikely to ever create systems that support human-quality conversations.

It was Edwin Land who made clear that visual perception is a field phenomena and not a point phenomena and now there are a number of famous, but simple images, that demonstrate this truth. In contrast for conversations the field is provided by one’s memories, and as I have discussed elsewhere the meaning of a simple phrase such as ‘so is his wife Kate’ can only be determined from the field in which it occurs i.e. the context – and hence the critical importance of the context
equation.

If we had science rather software, then even without biological machines, we might, with silicon, start to approach the 50 Watt power consumption of the human brain.

Over the years I have performed many simple conversational tests on various so-called AI implementations and perhaps my key discovery has been the critical importance of creating human-quality persistent  memories.

A key point regarding LLM memories is that LLMs do not have a standardised way of representing cognitive information in memory, i.e. those memories that are essential for a  human-quality conversation.

This means that one cannot offload these memories to use elsewhere with specialised, standards-based, tools perhaps for financial, contractual or medical purposes.

I believe that universal standards-based memory is an essential component for long term AI economic success, and it is certainly one area where science + cooperation would seem to be the second most important component, the first being the design of representations that are very simple and also extremely effective.

An on-handset 50W AI must clearly be possible and a such low power solution would then allow Apple to retain its current policy on privacy. The newly discovered mathematics is essential to such success. To see the difference it makes you can watch the YouTube video on the Precise Intelligence channel entitled: Alexa & Siri and Conversational Intelligence.