What AI can do, and what it can’t

A very effective marketing term. The phrase “artificial intelligence” has been used in computing since at least the 1950s. “AI” is now used as shorthand for “machine learning” and Large Language Models (LLMs).

The term “AI” is supposed to signify that these technologies involve machines learning to think. At their core, LLMs are basically a large-scale word guessing game. If I feed ChatGPT the sentence “the cat sat on the ___”, it would look through its enormous database of human communication that it has been trained on, and evaluate that humans are 56% likely to finish that sentence with the word “mat”.

Don’t we also work out what to say based on things we’ve said before? Isn’t that basically learning how to think?

Much of the training that AI bots use is based on human labour. I recently joined thousands of others doing freelance work on an AI industry website. Users are paid nominal sums of cash to evaluate how successful AI chatbot responses are, and write down what the chatbot should say instead. Websites like this are the engine for refining AI models, and they are driven entirely by the correction and evaluation of human brains.

LLMs are not “thinking”, they apply human-developed algorithms to enormous quantities of data. When ChatGPT “answers” a question, what it is actually doing is providing you with a very sophisticated average of what it estimates a human could respond. Notably, humans often lie, and because of that, so does AI.

Is that useful?

Computers have been able to outstrip human intelligence in many fields for a long time. For example, the 1976 Haken-Appel proof of the “four-colour theorem” in maths depends on computer analysis which no human can emulate or even check in detail (rather than outline).

There are enormously useful case studies for technologies that fall under the “AI” umbrella. Machine learning has revolutionised our ability to research how proteins are constructed; and for tasks like coding, which are based on formulaic constructs that can be very easily tested and falsified, AI enormously speeds up the process.

Such genuine uses will continue in everyday life too, like being able to give plain-English instructions to get results from databases. But this is a long way from AGI, and the technofuturist maniacs that fund it.

Hold on, what’s AGI?

Another marketing term, this one less effective. “Artificial General Intelligence” is defined differently by each AI company that uses it, but it’s supposedly the point at which machine learning outstrips human intellectual efforts, and can take over vast swathes of the economy across multiple industries.

LLMs can’t achieve AGI. A more accurate average of human conversation can’t magically transform into original thought. The current period of AI-mania began when developers achieved a big step forward in “human-like” responses by training their models on enormous amounts of data. This made their propaganda around AGI and “thinking machines” seem somewhat plausible to investors.

A recent plateau in the quality of outputs and user base led to OpenAI, the developers of ChatGPT, delaying GPT4.0 for several months while they tried to find something to justify the enormous amounts spent on it. Sam Altman, CEO of OpenAI, switched from speaking non-stop of AGI to saying it “wasn’t that useful a concept”. Last week, Altman announcement that, to drive up user engagement, adult users on ChatGPT would now be allowed to generate porn, a line that they previously promised not to cross.

Is AI going to take my job?

Given its limitations, the number of actual whole jobs that AI could take over is limited. Human labour power is a unique commodity, and one of the most unique things about it is the ability for humans to adapt to new environments and scenarios, and change the labour they do accordingly.

But I can’t get a job right now. Isn’t that AI?

If you’re young or just out of education, there is definitely an employment crisis at the moment. But the data suggests that it’s hitting blue collar, manual jobs more severely than white-collar office jobs and jobs to do with data handling. This is the opposite of what you would expect to see if AI were the cause.

The job market in multiple countries is still running “cold” due to after-effects from the Covid pandemic, inflation being high for a multitude of reasons, and geopolitical uncertainty, most prominently Trump and his tariffs.

So why is everyone talking about AI and the economy?

There is currently a bubble forming around AI – one similar in size to the “dotcom bubble” at the turn of the millennium. The Chair of OpenAI recently acknowledged that company valuations were massively outgrowing their actual value, and former Google boss Eric Schmidt declared “Bubbles are great. May the bubbles continue”.

Much of this valuation is based on speculation about how valuable AI models could be in the future – if AGI is discovered. In the real world, this has all the trappings of a forthcoming crash. Now that user numbers have plateaued, and the companies are scrambling to prove profitability (some have started hiring advertising executives from Google and Facebook, indicating a shift towards new income sources), the clock is ticking.

This is especially the case for companies like Nvidia, a chip producer whose shares have skyrocketed in value since their GPUs have become integral to providing the computing power necessary for LLMs. Given the shelf life for a high-end piece of Nvidia equipment is about three years, investors will be seeking a return on investment much sooner than the promise of the ever-elusive AGI on the horizon.

My friend says AI is killing the planet. Can you tell me why in the style of Frank Sinatra?

I’m a Trotskyist activist, not a performing monkey. But the problems AI expansion pose to the environment are real and dangerous. LLMs require enormous amounts of computing power and data storage. To account for this, there has been an explosion in the building of data centres across the world. An estimated $3 trillion (roughly the size of the French economy) will be spent on building new data centres over the next five years, and there are currently plans for around 100 new data centres in the UK.

These centres use unbelievable amounts of energy. In the US, these centres have been built on sites where electricity is currently cheapest, like Nevada or Utah.

That poses its own problem – those places are some of the driest in the US, and when thousands upon thousands of stacks of computers are run side by side, the heat generated is vast. Water is an essential resource for cooling down the data centres, and the water cost of the average centre is a terrifying prospect for our planet. A water cooling tower in a data centre uses around 19,000 litres of water per minute. Both the water and energy usage far outstrip human usage in the surrounding area — in 2023, indirect water consumption tied to energy use was equivalent to about two million US homes. That will now be higher.

I’d like to write something for Solidarity on AI. ChatGPT has given me some great ideas.

We welcome new contributors to Solidarity, but please, don’t get chatbots to write your articles. The ability to read and collect articles, books and other information, politically evaluate them and produce a coherent, well-written text are all skills that socialists should want to develop themselves. A speech at a union conference will be better if you have put care and a human touch into it, and you will learn how to do that by writing articles yourself and talking to real people.

AI may provide you with options, or perhaps offer some kind of structure to get around writers’ block (personally, I share writer David Simon’s view: “I’d rather put a gun in my mouth”). but the words on the page, and the argument they make, should be yours and yours alone.