AI Hype: Elon Musk Echoes the Nonsense Claims from the CEO of Anthropic

In recent weeks, the AI marketing hype has become so intense, and the misinformation so widespread, that well-educated and intelligent people have seriously asked us whether sentient AI might escape from data centres and begin hunting down human beings.

That is not going to happen. There is not a 10% chance of it happening, nor a 1% chance. For all practical purposes, there is zero chance of it happening in the foreseeable future.

AI is software. It cannot escape from a computer as a disembodied being, conjure up a physical body or acquire real-world power merely because someone gives it a malicious prompt. We should point out that the AI CEOs have never claimed this would happen, but they also seem quite happy to let the idea enter the public consciousness.

Where did the latest hype start?

At the start of the month, a few AI staffers, such as Evan Hubinger and Jacob Coxon of Anthropic, claimed there is a very significant chance that “AI could kill all humans”. Hubinger puts that chance at 10% within the next decade (see also “P(doom)”).

The next decade is a handy timescale: it means you can’t be proved wrong for at least 9 years and 11 months. Hubinger and Coxon also don’t provide a mechanism for AI to kill us all, so what exactly could AI do that would wipe out all humanity? After all, there is an easy way to stop AI: switch the power off. It’s simple. It’s even easier than shutting down HAL.

Hubinger and Coxon apparently expect huge leaps in AI capabilities, resulting in super-human ability, driven by AI models improving themselves (“recursive self-improvement”). If that is true, why can’t AI models still count the way a ten-year-old – unless the AI has help from other tools and software? Why do AI models themselves only have a limited working memory (“context window”) that has shown almost zero improvement over the last 18 months? Why can AI not even write a bestselling novel by itself?

What claims have Musk, Amodei and Altman themselves actually made?

Dario Amodei, the CEO of Anthropic, has claimed that within just six to twelve months, a swarm of AI agents might be capable of “taking over the entire Internet”. This is of course the same Amodei who claimed that by now US unemployment might reach 10–20% because of AI. There is no sign of that yet, and while Amodei might claim the colossal economic disruption is still to come, one notices that few, if any, of his predictions have yet happened. But let’s still take Dario Amodei seriously: how will AI actually take over the entire Internet? The Internet is distributed across many providers. We don’t see any plausible way in which AI could take over the entire Internet. And if it did, who would provide the colossal AI footprint needed to do it? There is simply no organisation or country with the necessary resources.

Amodei is short on detail here. So is Musk. Musk simply says “Dario is right”. Meanwhile Altman was a lot less alarmist in his interview with Fortune and measured in his tweet (sic) in support of Amodei, but that did not stop the media claiming that Altman was also calling for an AI slowdown. We don’t think he did.

By the way, Musk, Altman and Amodei all worked together at OpenAI some years ago, when they claimed their AI model (“GPT-2”) might be too dangerous to release, but then released it anyway.

What is the real danger?

Some fear that the real danger is that AI could create a TV sitcom so tedious that we all die of boredom watching it. But there are serious questions about how humans may misuse AI or recklessly connect it to important systems. This, however, is a very different danger from the science-fiction story now being presented as a measurable risk to humanity. Humans have shown they can do some real dumb stuff with bows and arrows, never mind missiles and nuclear arms. It seems entirely possible that some profit-seeking CEO at an energy company could get rid of all their trained staff, put AI in charge of critical infrastructure, and take their bonus payment before the out of control AI shuts down, say, a nation’s power supply. Or that management cuts corners and ends up allowing foreign hackers to use AI to damage a nation.

If anything, the real danger might be to Anthropic’s IPO plans: will their IPO filing have to state that one of their risks is that they might destroy humanity? Or will they miss that bit out of the list of investor risks?