Left-Bashing in the AI Bubble.
By Olivier Jutel
The AI industry is a monstrous waste of money and resources predicated upon monopolizing the sum total of human intellectual and creative work for venture capitalists and the “Magnificent Seven”. It is a labour and democracy issue hyper-scaled for our wretched new gilded age. We’ve been promised cures for cancer and nuclear fusion and what we’ve got is LLM confidence machines that pander to the fantasies of bosses, middle managers and corporate consultants.
The CEOs of these companies want us to think of this as a religious or civilisational issue rather than one of theft and enclosure. They oscillate from apocalyptic scenarios about paper clips to childlike fantasies of abundance. They eschew questions of unit economics, revenue, return on investment or use cases; things that investors and the business press used to care about. Now company valuations simply orbit the wreckage of our society like a SpaceX satellite.
To be a tertiary educator is to be at the coal face of AI’s attempts to remake human aspirations and values. Our students have magic sparkle icons brute forced into all their devices and are fed a steady stream of “innovate or die” from the broader culture.
AI resisters or slow-walkers, like myself, are chided as being reactive, not responding the world as it “really is”, or simply we do not grasp what is happening.
We are told that a moratorium on environment destroying data centres will preclude abundance. The Dutch historian Rutger Bregman has diagnosed left AI denialism as a danger on par with climate denialism, or worse! Its an odd analogy because the climate denialists have won and the hydro-carbon intensive tech sector is playing a big part in their victory lap.
I want to focus on Bregman because his arguments are spectacularly stupid and symptomatic of the sleights of hand necessary to give the tech sector everything it wants.
Graph Teleology
Bregman claims we have broken through to a new reality because of the performance of Anthropic’s Mythos in performing coding tasks. I will return to the specifics in a moment but first a quick word about graph teleology. Tech loves a good exponential teleology from Moore’s Law, the Kardashev scale, to Nick Bostrom’s theory of Super-intelligence. It is based on the idea of a runaway technology as a natural phenomenon that will change our nature and society. This logic has been captured in the exponential baby meme.
The AI industry relies on hype and anticipation of this deus ex machina event of the 7 trillion tonne child. For the last four years we have been treated to blog posts or white papers which claim “we have seen sparks of artificial general intelligence” (AGI) or “recursive super intelligence” on the basis of some model breakthroughs. The actual evidence of how models and training sets are constructed remain proprietary. Press releases replace peer review as venture capitalists simply declare on twitter that AGI is here. The venture capitalist Marc Andreessen and his acolytes casually quote Nick Land and the idea of the techno-capital machine as a universal cybernetic system of AI and capitalism.
The reason why AI boosters will make that jump is because of:
1) doom marketing is powerful, and
2) the field has been driven by philosophers theorising super-intelligence rather than engineers thinking about applied tasks.
Nick Bostrom is the man behind the recursive intelligence thesis (see below) that once an AI can do X it will do Y, Z it will be able to do everything else. A whole cottage industry of AI doomers and alignment experts have emerged from Bostrom’s sci-fi musings.
Essentially these philosophy parlour games have been key to marketing AI as a unified field, as Marvin Minsky put it: “we don’t yet understand how brains perform many mental skills, we can still work toward making machines that do the same or similar things. ‘Artificial Intelligence’ is simply the name we give to that research.” As an side both Minsky and Bostrom are big cryonics enthusiasts who loathe their meat machines (bodies).
The pursuit of Artificial General Intelligence has become a business model [1] for culturally illiterate oligarchs who invoke Digital God to rationalize their world-eating ambitions.
The Dreaded Time Horizon
Returning to Bregman’s claim of the urgent need for the left’s participation in the AI revolution, he is very excited by the latest publication of METR’s time horizon graph. It’s a dramatic line go up moment “even scarier” than Al Gore’s cherry picker climate chart.
METR is a non-profit staffed by researchers and insiders from the AI industry who monitor the risk of breakaway AI agents and superintelligence. In other words, they attend to the risk of surrendering human autonomy to software at the behest of capitalism. They assess performance benchmarks and safety risks in a transparent way and as a non-specialist there is plenty for me to glean. My favourite METR study is that software developers believe they have become superusers through AI but in fact have become 19% slower.
So: yes, all of the Claude-code pilled are on some scale between delusional and psychotic.
METR’s prominence is due to its Software Time Horizon graph which MIT Tech Review has described as ‘the most misunderstood graph in AI’. The graph measures the ability of different models to complete increasingly difficult software tasks measured by the amount of time it would take a software expert to complete.
METR, to their credit, are circumspect about whether this is the right metric for model evaluation let alone super intelligence. But that does not stop a hype machine and religious devotees primed to announce the rapture for nerds. In fact much of the excitement around the METR Time Horizon measure comes from its inclusion in “AI 2027”, a frothy piece of teen sci-fi Doomerism that gets treated as serious analysis by a credulous tech media.
In June, Claude’s Mythos achieved a time horizon of 16 hours compared to GPT-2’s 9 seconds in 2019. This is the kind of a proof that appeals to graph teleology. The first thing to note is that improvement is measured for 50% task completion, when measured for 80%, Mythos is down to 3 hours.
The issue of AI super intelligence however remains mired in human systems. The problem for widespread AI adoption is cost and reliability. 50% task completion is not a revolution in software engineering. The time horizon also says nothing about quality of outputs. In order for this to be a revolution in software engineering we’d need it to be around 90+%.
Even if we achieve this at a massive economic cost, there is still the need for software and code to respond to human work and needs. These are systems that exist outside of LLMs and demand human competency and understanding, not black box magic. This is known as the “oracle problem” and the fact that AIs or LLMs cannot understand the physical and social world; it has to be fed into them. And again if we got to a high task automation rate there is still the massive problem of security and vulnerability. The code AI makes is highly insecure and you’ve taken experts out of the process of understanding where the risks are.
Businesses and governments would be an insane to switch to these systems with a combination of very high costs and security risks. Gary Marcus writes that there might be some very real gains in the old field of ’Symbolic AI’ from Mythos, but this is not AGI but something akin to Google Deepmind’s AlphaGo.
Abundance and Over-accumulation
Rutger Bregman came to viral acclaim for dressing down the global elite at Davos as tax cheats. He is a European social democrat who has signed up for the pro-tech abundance agenda. Among the many problems of the abundance agenda is a disinterest in thinking about why capital chases monopoly returns in software. It is also prone to the kinds of techno-fetishism that props up fictitious valuations and a financial bubble. We are in a moment of massive over-spending and over-accumulation. Bregman points to the AI infrastructure as proof of the coming era of super-intelligence rather than a sign of this structural imbalance and crisis. We have two more lines to consider, AI infrastructure and the nominal increase of AI revenue.
First up: the projected data centre build outs. There are headline numbers offered by the likes of McKinsey of $7 trillion by 2030. There is simply no way these get built. Everyone from NVIDIA, META, Oracle, SpaceX are putting these pledged build outs through highly leveraged special purpose vehicle shell corps because the debt load is so atrocious. For example, Oracle has taken on $340 billion in debt to build “Stargate” for OpenAI. Oracle is building out these data centres on the idea that OpenAI will continue to grow and fundraise in spite of massive losses. As Ed Zitron and the Financial Times report OpenAI lost $38 billion last year with $13 billion in revenue. These are really bad numbers: Oracle needs about $30 billion a year from OpenAI to service the debt on Stargate. This bubble cannot go on; and it may well claim Oracle and OpenAI. The latter has very few assets other than their models; they don’t own the data centres or the GPUs. They are a hype machine and without the hype they are in big trouble.
Then there is the issue of data centre populism: i.e. no one wants them built in their backyards for obvious reasons. It’s a non-partisan issue and AI disenchantment is widespread. There are AI Bartelbys everywhere, not just on the left. Also the backup plan to building in the imperial core has been the Gulf states who have come under fire from recent imperial misadventures.
The AI industry in general is close to a trillion dollars short, as reported by PE Firm Bain Capital, in justifying pledged data centre build outs that Bregman says is inevitable because of the METR chart.
The other hype and teleology he swallows is the idea that OpenAI’s massive growth from 0 to 900 million users is proof that we will get recursive self-improvement and an economic juggernaut. This is nonsense. User growth numbers have been a benchmark in tech as proof of “network effects” for some time. Network Effects is the idea that the more users you have, the more profit. Take for example Uber: it’s more profitable if there are 10,000 users in Dunedin rather than 100 because software and infrastructure are low marginal costs. That is not is true for AI. In fact, it’s the opposite. Anthropic is losing more money on their paying customers. GPUs are not a marginal cost, they are massively expensive. We are expected to lock ourselves into AI dependency before the real costs are revealed.
Also, the data centres are not going to be useful after the crash; GPUs are really only good for generative AI and crypto. We are not building infrastructure like the railroads which will live on after the crash. Talk of an AI race with China is ludicrous too. China has cheap open source models and a much more “for purpose” approach, rather than trying to build silicon God. There is no reason to believe that Mythos will be delivered at scale and could create the type of abundance for the working class.
Anthropic Realised Revenue
OpenAI is no longer the darling of the media and AI boosters. Pretty much everyone is sick of Sam Altman.
Step up Dario Amodei and Anthropic.
Mythos is an Anthropic model and in addition to it’s recent METR performance Anthropic has seemingly surpassed OpenAI in revenue. Bregman believes this is another sign, the lines on all the charts are going up.
Anthropic’s number of $45 billion in annual revenue is fugazi. The AI industry does some creative accounting called “Annual Realised Revenue” which basically means take your best month or quarter and average it out over a year. These are start-ups so they don’t have to conform to real accounting practices, just to their venture capital donors. As the Wall Street Journal reported Anthropic’s numbers: “it is unclear what accounting methods Anthropic has used to book revenue and costs, as the company isn’t yet required to follow the financial-reporting requirements of a public company”.
What Ed Zitron has found is that this can all be explained by horse-trading between Anthropic and SpaceX. Anthropic’s has received two months of heavily discounted compute from SapceX which Anthropic used to leak cooked numbers to the press. SpaceX was then able to use Anthropic’s pledge of $45 billion of future spending as proof of revenue for its recent IPO. I think the left’s real is to take this kind of ponzi-dynamics seriously as it entails massive societal risk.
More recently when Anthropic and OpenAI moved to token based billing for AI, which starts to reflect the real costs, everybody blew through their AI budgets. Uber burnt through its yearly allocation in just 4 months. Comedically the chief technology officer spent $1200 USD in tokens for a two-hour session meant to impress other C-suite executives.
This reality of political and economic costs is starting to sink in. The financing of AI infrastructure is incredibly fragile. We are not seeing the returns of investment to justify sinking more money into these expensive oracles. Roughly 95% of AI pilot projects have failed.
The really loathsome part about AI scolding the left is that elements of AI doom are here of our own making. Whether we are talking about an AI automated genocide in Gaza or the massive power grab that tech oligarchs are trying to pull off through invocations of God. The trillion ton babies are the venture capitalists and tech CEOs who demand we acquiesce to their juvenile fantasies.
Olivier Jutel is a Senior Lecturer in the Media, Film & Communication Programme at the University of Otago. He is also the co-editor of Media Peripheries journal. Olivier has published in Antipode, Big Data and Society and Psychoanalysis, Culture & Society.
[1] OpenAI’s landmark deal with Microsoft in 2023, which kicked off the current race in LLMs and gen-AI, was premised upon OpenAI giving Microsoft exclusive access to its API up and until the point that OpenAI solves AGI. At the point OpenAI will have exclusivity over our new digital god.
 | A guest post by
|