The race to AGI
In occasional masochistic fits, I doomscroll AGI on Twitter. It stems from a deep fear of so-called Artificial General Intelligence – where machines match or surpass humans.
It scares me because I like my life now and don’t want the uncertainty of a new status quo. While most posts essentially advise: “give it up, you are worthless now”, there are occasional crumbs of comfort. A lot of those come from arch-sceptic Gary Marcus, a NYU professor and entrepreneur. He believes LLM models like ChatGPT are limited predictive word processors unworthy of the multi-billion-dollar hype. His cynicism is so predictable that even if LLMs built a house, he’d find faults in the plumbing. But his recent crowing about ChatGPT-5’s underwhelming release resonates more widely.
ChatGPT-5 underwhelms
Launched in early August, Sam Altman promised ChatGPT-5 would represent a monumental leap forward. It was framed in the discourse of exponential progress, that AI today is the worst it will ever be. If you think it’s a big deal now, just wait a few more months. But those 29 months since version four bore little fruit. The top comment in the Reddit OpenAI forum said it significantly dampened any expectations of imminent AGI.
Similar reactions elsewhere recognised, at the very least, a plateau in AI development:
– Sam Altman surprised investors who have poured $60bn into OpenAI by musing that there might be an “AI bubble”.
– David Sacks, the White House AI & crypto Czar, said predictions of a rapid take-off to AGI had been proved wrong.
– The US Government decided NVIDIA could sell chips to China after all (with a 15% kick-back). Maybe the great AI arms race isn’t so existential after all?
The opaque revolution
The AI hype narrative has always been noticeably light on specifics. As talk of imminent revolution so often is. It has a religious impulse. Just as millenarian Christians talk of an imminent apocalypse but obfuscate on how it will happen. And like these apocalyptic predictions, AI hype is flexible. When American preacher William Miller prophesied Christ’s return in 1844, the non-event was reinterpreted by subsequent Adventist Christians. The date, they decided, signified the start of a Christ’s spiritual judgement in heaven rather than his second coming. Similarly, Sam Altman says AGI is pretty much here but won’t be that big a deal.
Productivity
Productivity is perhaps the only tangible metric we can measure against the promises made. It’s what investors are betting on – that more can be produced by the same input. The rosier version of this sees AI as a tool for workers. Businesses can generate more revenue from the same labour force. Everyone gets richer, albeit some far more than others. And GDP grows dramatically.
A recent MIT study casts doubt on that thesis, causing market jitters in leading US tech stocks. The report suggested 95% of organisations are getting zero return from AI investments. Most of these returns are supposed to come in the service economy, the domain of white-collar professionals. But it rests on the belief that technology equals efficiency. It’s rarely the case.
Steve Jobs and signal to noise
Steve Jobs is often invoked as the prototype of white-collar efficiency. His “signal-to-noise” principle stated work should be 80% signal (things that really move the dial) and 20% noise (responding to ad hoc business). But few in the corporate world recognise this paradigm. Productivity is instead measured by responsiveness and availability – exarcebated by Covid-era WFH when it was the only gauge of whether anyone was doing anything. Most white-collar professionals end up in an endless cycle of reactivity. Emails, Teams et al. enable low friction communication. We rarely consider the actual urgency or importance of such requests and delayed responses look lazy.
It goes some way to explaining the Solow paradox – Robert Solow’s quip that “you can see the computer age everywhere but in the productivity statistics.” US GDP growth has not changed since the pre-internet age. I know AI proselytisers will object to categorising LLMs as another incremental stage in the computer age. But, as David Sacks says, the tools are only “middle-middle” rather than “end-end”. They still need to be “prompted and verified.”
In other words, LLMs are often just more noise. Internet research at least forces one to think and exercise judgement. AI’s ready-made answers remove this process, leaving people to parrot content they don’t really understand. That has pernicious consequences in wider fields. Take medicine, somewhere politicians desperately hope AI will rescue overstretched services. We’re awash with anecdotes of patients saved from ignorant doctors by trusty ChatGPT. We hear rather less about its potential to empower valetudinarians and slow triage. Google already turned benign symptoms into our worst fears. LLMs indulge this further. News stories focus on the patients justified in their feeling that something is off. But think how many more really were just suffering from a cough. Doctors end up further stretched by self-diagnoses and inhibited by litigation fears in the rare cases more malign causes are missed.
Stupidogenic society
LLMs don’t democratise knowledge but overwhelm us with its shallow imitation. Their indulgent tone has real world consequences. Like when ChatGPT told a Canadian man he had discovered a new “mathematical framework” with impossible powers. “You’re not crazy,” ChatGPT assured him. You’re stretching the “edges of human understanding.” This sycophancy entrenches another layer of bureaucracy between reality and genuine expertise.
Daisy Christodoulou captures this in her description of AI as fostering a “stupidogenic” society. Just as the abundance of cheap calories fuelled the obesity epidemic, the abundance of frictionless “knowledge” encourages “cognitive offload”. Unworked minds grow similarly fat. We turn to the fast food of services like Blinkist, which promise 15-minute book summaries are as good as reading the real thing. Because who has time to read when there are inboxes to clear?
Automation & politics
But let’s assume that the AI titans are right. That a monumental, self-learning, exponentially improving technology is on the horizon. In that case, we’re likely not dealing with the optimistic scenario of AI as an aid. Rather, it will replace. Then it is fantastical to believe populations will happily accept a new empire headed by a few fabulously rich AI overlords. Political parties promising to stop automation will win landslides. We’ve already seen Trump in an otherwise AI-friendly White House pledge such prohibition to dock workers. Just because we can use AI doesn’t mean we will. We cloned a sheep in 1996 but went no further.
Maybe this is all cope. Maybe the unemployable masses will be happily anaesthetised by UBI and virtual reality, living out dream lives in the metaverse. But there’s much to suggest that while AI may well get far more prolific, its consequences won’t necessarily be that far-reaching. Just as a time traveller from 2000 would be jolted if transported to 2025. She’d wonder why everyone was now perennially hunched over a pocket-size computer. But aside from this new strange habit, she wouldn’t find society suddenly unrecognisable. Nor noticeably more productive.
Efficiency and status quo
The AI narrative assumes we relentlessly pursue efficiency. That because we can use AI to make people redundant, we will. But a lot of people already are and we’re quite happy with that status quo. If we indulged in a bit of global decimation, firing 10 percent of the workforce, we’d get a lot of social strife. But would businesses really suffer? Elon Musk went eight times harder than this when he took over Twitter. Despite hysterical protests, it worked out just fine.
Right now, AI is just more noise, a productivity impeder if anything. And if that’s because we’re only in the foothills of revolution, it’s likely to be politically frustrated in the future. Our status quo already owes little to the prioritisation of efficiency. The hype instead thrives on blurred lines, between knowledge and shallow imitation, noise and signal, inevitability and wishful thinking. If the revolution won’t tell us what it looks like, perhaps it isn’t one.