AGI Debate: ChatGPT Launch Sparks Concerns Over Costs and Job Losses

General AI isn't arriving alone: debates over costs and data

How far are we truly from General AI? The presentation of the new ChatGPT peine this question, along with a catalog of forecasts: some place Artificial General Intelligence (AGI) in 2029 and superintelligence—ASI—before 2040. The tone of some reactions doesn't invite calm: talk is of mass layoffs, hoarding food, and global hunger and poverty. Other, more tempered voices recall that every technological revolution—the train, the assembly line, the internet—brought the same panic and, so far, a different result than announced.

The discussion, in any case, is no longer whether the tool writes or draws, but what collapses when the tool improves.

From 2029's AGI to 2040's Superintelligence

The timeline is on the table. Against the prophecy that general AI and robots arrive "next year," some argue the benchmark is different: AGI by 2029 and, subsequently, a superintelligence that could materialize before 2040. This is the timeframe framework repeated in conversations, with a surprising nuance: the idea that the adjustment won't first wipe out workers, but bosses.

Historical analogy is the other major block. When the train arrived, when assembly-line factories came, when the internet emerged, some announced the end of work. The dominant thesis on that side is that technology destroys jobs and creates others, and that the net balance can be "living better," although the distribution of that better is another matter. The lingering question: what new jobs will this time generate?

A circulating calculation: who wins and who loses with AI

The business, for now, doesn't add up. A circulating calculation argues that all companies involved in AI are losing money hand over fist and that the only ones winning are those renting servers. The most cited case is Stability AI—the firm behind Stable Diffusion—which was attributed a $100 million deficit and imminent bankruptcy. However, the forecast didn't come true: months later, as noted in the thread itself, the bankruptcy hadn't occurred.

The other front is raw material. High-quality datasets, it is argued, have been exhausted; there's nothing left to scrape. The solution being tested is synthetic data, generated by the machine itself, with the evident risk of training a model on invented material. It's the difference between learning from reality and doing so from a mirror.

Energy and water: recurring limits

This is one of the points where several voices coincide. Energy and water appear as limiting factors for continuing to scale AI, and some already speak of a plateau: the technology would reach a point where economies of scale disappear and more power doesn't buy more capacity. What's striking, they say, isn't that this ceiling exists, but how quickly it has been reached.

Translated: the model's arch-enemy isn't competition, it's the electricity bill. Every response coming out of a chat costs energy, and that energy has a price, location, and physical limits. The barrier, finally, isn't philosophical. It's kilowatts.

Which jobs are really on the chopping block?

The risk distribution is unequal. A specific testimony illustrates this: in one company, management accumulates staff that the tool could absorb—three IT specialists, ten administrative employees, a logistics manager, and a marketing head—and only the disconnect between leadership and daily operations prevents the adjustment. The uncomfortable joke is that positions are surplus, but those deciding don't know which ones.

Against this, the defense of trades. Anyone who has worked blue-collar or comes from rural areas, it is maintained, laughs at these forecasts: repairing a truck, hammering sheet metal, or welding aren't easily automated. And a deep technical argument: AI, it is claimed, doesn't have a world model; a child learns to drive with a few hours of practice, while the system, with millions, still doesn't do it equally well. On the other side remains the phrase summarizing the vertigo: the bulk of the population wouldn't know how to do half of what the model already does.

Can AI replace judges and politicians?

Some take the argument to the institutional terrain. If AI managed public decisions, it is proposed, an episode like the DANA (a severe Spanish storm event) wouldn't have gone so far: the dam would have been built and the alarm activated in time. The counterpoint appears in the same sentence: an AI causing hundreds of deaths due to a wrong decision would be unplugged without hesitation.

The reasoning extends to private life: pets, friends, and even partners would be replaced by connected machines, because they don't make messes, don't die, and don't argue. This is one of the darkest derivatives of all this, and also the most speculative. No one is yet billing for the robot dog that stays to care for someone at eighty.

---

While some hoard food and others wait for the boss to be fired, there is a part of the business that, according to the thread, doesn't lose: those who rent servers. Whatever happens.