Many of those critical of AI consider that hyperscalers are just doing fear marketing to shift the focus away from the long list of damages AI is already causing today.
“We are so powerful and valuable that we could cause human extinction.”
A lot of friends I feel close to intellectually still argue : “AI is not really that autonomous yet, they’re just saying that to avoid responsibilities. Let’s focus on the real problems.”
Well, clearly something has changed in the last couple of months.
It’s not just the Amodeis or Altmans telling us how disruptive they are.
More and more academics and independent researchers are evidencing that this cannot be summed up to a clever communications strategy.
And it’s not only the Hugging Face / OpenAI hacking by autonomous agents.
METR's (Model Evaluation & Threat Research) has published very relevant reports in the past few weeks, both on OpenAI and Meta.
Among other findings, it emphasises the reality of self-recurring improvements in AI development, in which AI research is increasingly carried out by AI itself.
Anthropic is accelerating the technologies it’s warning us about – and it is doing so faster than the pessimist estimates.
We’re very close to the irreversibly dangerous threshold, at which AI is moving so fast that humans will not understand what’s happening any more, while transparency and oversight will be almost impossible.
Daniel Kokotajl, whose famous prediction about the loss of control over AI have roughly proved to be correct until now, expressed his worries after reading the last METR report on Anthropic:
“If they’re right, that implies an AI-driven acceleration that is 5 months ahead of the AI 2027 scenario. I am scared, sad, and angry.”
We cannot accept this. If you train your dog to bite and it kills someone, you’re facing charges. AI leaders should be held accountable for irresponsibly gambling with our future.
Did you find this reflection interesting?
⮕ Read more articles on LinkedIn.
⮕ Subscribe to the OFF Movement newsletter.
⮕ Request a talk or conference.