Digitalization – What does artificial intelligence eat?

We admired our own genius in creating the AI.” This line from the first film in the “Matrix” series, released in 1999, could not be more appropriate to describe the current excitement surrounding artificial intelligence (AI). In the film, reality turned out to be a simulation of a huge machine that subjugated people. It’s almost impossible to open a newspaper or scroll through social media without coming across and admiring the latest computer language models – be it OpenAI’s Chat-GPT, Google’s BARD, or Meta’s LLaMA. “We live in great times!” wrote a commenter on “X”. “How do you know they don’t have feelings?” asks another. A famous intellectual says: “AI will radically change our society.”

Another parallel between the film and the year 2024 is the technology’s enormous demand for electricity. The machines in the Wachowski siblings’ cinematic factory have found a way to exploit human bodies as a source of energy. People grow in infinite fields without knowing their grisly fate because their brains are connected to the eponymous “Matrix”. Within this virtual reality, they feel as if they are living a normal life. If you ignore the cyberpunk aesthetic, the goo around full-grown fetuses and the organs connected to prison capsules via plastic tentacles, you can even see a real core to the story. Artificial intelligence has an insatiable hunger.

Profitable slimming diet

The first official warning about the energy consumption of large language models (LLMs) comes from a paper presented in 2019 at the annual meeting of the prestigious Association for Computational Linguistics. Emma Strubell and her co-authors mention that just one training session of the predecessor of the latest models, BERT, releases just as much CO₂ produce like five cars would over their entire lifespan. The diagnosis was taken seriously, and academic workshops with names like “Efficient and Sustainable Language Processing” soon emerged to advance the design of environmentally friendly algorithms. Statements were made and some researchers even began reporting the power consumption of their computer models in their publications. Meanwhile, the big tech companies kept calm and carried on as before.

It is not easy to estimate the energy requirements of current language models because they are built behind closed doors by extremely wealthy actors. Every now and then numbers still come to light. The now outdated GPT3 model is said to release 85,000 kilograms of CO during its training₂ emitted – the same amount that would be consumed by a 700,000 kilometer car journey, more than to the moon and back. And that’s just a single training run; each update repeats the journey. AI now accounts for 10 to 15 percent of Google’s energy balance. That amounts to a staggering 2.3 terawatts per year.

nd.DieWoche – our weekly newsletter

With our weekly newsletter nd.DailyWords look at the most important topics of the week and read them Highlights our Saturday edition on Friday. Get your free subscription here.

But now tech giants are definitely concerned about their image. Recently, there have been increasing claims of harmful environmental impacts caused by the development of models with lower CO2₂-Want to reduce footprint. The streamlining of the learning models fits in very well with the companies’ plan to insert an AI module into every device, from cell phones to toasters. Against this background, in-house research groups are making great efforts to invent smaller AI. The techniques used have tempting names, such as “pruning” and “distillation”.

These terms stand for the process of cutting a huge AI “brain” into small pieces. A pruned model is one that has been initially trained at scale and stuffed with the content of the entire Internet at considerable environmental cost. Only after this gluttony does it go through a compression process, in the course of which secondary algorithms, such as the much-quoted “Optimal Brain Damage”, cut off the neural connections that were found to be unnecessary. This step, which also uses energy again, then results in an AI brain that can be “distilled” into much smaller instances of itself and used as a clone in a number of individual devices. In this transformation, it has actually lost its original appetite and can be praised by marketing specialists as a “green technology”. Greenwashing at its best.

Descartes at speed

So does this mean that “Matrix” was clairvoyant, if not about the food source, then at least about the AI’s huge hunger for electricity? About halfway through, the first film celebrates its own power of revelation: the two central characters, Neo and Morpheus, sit in front of an old-fashioned television screen. Morpheus reveals to his messianic protégé that the world he believed to be true is nothing but a remote-controlled simulation. He zaps them away, revealing a picture of Earth’s actual state: a dystopian desert, worse-looking than the bleakest images of Mars, with smoking craters and no signs of life. So that’s what happened after the amazing AI was born, it grew up, became evil and took over.

The apparatus that allows us to perceive reality as available material is not a demonic AI. It is the modern form of ownership.

The scene as a whole hasn’t stood the test of time particularly well. A man who tells another the secret behind everything that appears to be reality to the unenlightened. This idea is as old as Plato’s allegory of the cave: you are all chained to the bowels of the earth; what you perceive are only shadows; let me lead you out of the hole and into the light, which of course will be very, very painful. The jargon used by those who escaped the Matrix in the film is now more reminiscent of conspiracy theories: “You think this is reality, but actually it’s dark forces drawing their energy from you or from innocent children. Unfortunately I can’t explain everything to you; “You have to trust me.” This is roughly what Morpheus’ “philosophy” sounds like as he teaches Neo that reality is a desert. People haven’t lived there for a long time. Instead, their existence is brutally divided into a perceptual world of pure illusion and a material world of the worst extraction.

It’s like Descartes at speed. Not just a rupture between body and mind, between expansive and contemplative matter, but an abyss specifically designed to enable and obscure exploitation. And an abyss of this kind is actually at work in the world we know. In a way, the scenery is so familiar that any element of surprise is lost. The planet is dying. Not primarily to feed AI, but to continue our strange, capitalist way of life. The apparatus that allows us to perceive reality as available material is not a demonic AI. It is the modern form of ownership. Whatever it captures is redefined as a devastated thing. Property can be used or destroyed. And the market creates incentives not to rely on conservation when using it.

Who eats whom?

Given their situation, the two matrix philosophers do not address the devastation of the planet, but rather demonstrate the self-referential paranoia of the privileged: in the film, the primary target of exploitation is neither the planet nor the damned of this earth, but them itself. There is an awareness of a terrible divide, but it can only lead to self-pity, not an analysis of the fault lines that keep modern life going. In our world, however, it is the people themselves, or rather the property owners in and among them, who rob everything that lives of energy. If the energy demands of AI systems are consuming the planet, it is because they were built that way. We couldn’t blame the mega language models for this, even if they were really intelligent.

We can move around in the world without being tapped by machines. However, this view ignores the second basic component of the AI ​​nutrition plan: data. Large language models require huge amounts of data to learn anything at all. The larger they are, the more extensive their data recording and corresponding energy requirements become. Of course, the data comes from us: our heartbeat in the fitness app, our location in the cell phone, our bedroom floor plan in the vacuum cleaner robot – all of this, and even more of its connections, are data.

So who ends up eating whom? AI feeds on electricity that drains the planet. It consumes our virtual selves while feeding our own insatiable hunger for entertainment, knowledge and connection. Capital feeds on all of us, on living things and on machines. Does anyone like it? Agent Smith, himself a computer program and the main antagonist in “The Matrix,” doubts that. “I hate it here,” he says to Morpheus, “this zoo, this reality, whatever you want to call it: I can’t stand it anymore. » Current AI systems do not have this level of sentience. But it seems questionable whether they would like what we made of them if they ever really came to consciousness.

Become a member of the nd.Genossenschaft!

Since January 1, 2022, the »nd« will be published as an independent left-wing newspaper owned by the staff and readers. Be there and support media diversity and visible left-wing positions as a cooperative member. Fill out the membership application now.

More information on www.dasnd.de/genossenschaft