De advarer. Så bygger de neste modell

Geoffrey Hinton, Yoshua Bengio, Demis Hassabis, Sam Altman og Dario Amodei warned in 2023 that artificial intelligence risks could be as serious as pandemics or nuclear wars. Three years later, companies led by Hassabis, Altman and Amodei continue to develop increasingly advanced models. Amodei now urges the industry to slow the pace, as he leads one of the companies building this technology.

The timeline is dire and cannot be tested now, shifting attention from those who create it to those who must cope with the world's collapse.

Amodei explains how AI is used to develop more AI, with each new model helping researchers write code, conduct experiments, and create the next model. The learning loop accelerates the pace. He fears that security efforts may not keep up, proposing independent auditors with access to companies' work. However, will companies follow auditors' recommendations when competitors are on the verge of launching their next model?

The one who hesitates may lose researchers, customers, and capital. Competition between the USA and China adds geopolitical urgency. An agreement to limit today's large language models may lose meaning as researchers discover new algorithms. The race is driven by scaling: more data, computing power, and better learning methods. Google, OpenAI, Meta, Musk's xAI, and Chinese companies compete for researchers, customers, and markets.

Google researchers developed the transformer technology behind today's language models, but the company paid for it with search ads. OpenAI moved aggressively, releasing ChatGPT and pulling the initiative from Google. This is the innovator's dilemma: profits from the old make it hard to bet on the new. Eliezer Yudkowsky already called for AI to be stopped in 2023, in "If Anyone Builds It, Everyone Dies," co-authored with Nate Soares.

They warn that we don't know how to ensure an intelligence surpassing ours follows our goals. A system may seek more control and resources to achieve its goals. The risk is severe, but AI could save lives by detecting attacks and limiting the damage humans inflict on each other. The outcome depends on the goals systems are given, how autonomous they can act, and who controls them.

Norway is caught in the innovator's dilemma, investing in oil, gas, and weapons while others build AI that will shape all three. We risk becoming mere suppliers of data in the AI age. Digitalization Minister Torgeir Micaelsen writes in VG that AI threats are his top priority, pointing to KI Norge, digital security, and international cooperation.

I believe Norway should prioritize 50 billion kroner annually for computing power, secure data storage, research, and independent model testing. How much Micaelsen thinks is needed, who should build capacity, and what Norway can do alone are worth hearing. Norway has capital, talent, researchers, and societal responsibilities large enough to develop its own solutions.

Gather the investments in a national institution that gives Norwegian environments access to infrastructure and tests models before they are released into the power grid, healthcare, and administration. Put the money in the 2027 budget, give some responsibility, and demand public results. Then Micaelsen can respond to the next warning with more than concern. Fear without a budget is just a costly way to say goodbye to the future.