I am building an AI company. I still think we should pause AI development.

I have worked with AI for most of my career, first as a computer scientist, then as one of the Nordic leaders of AI work at McKinsey, and now building alice as an application on top of the best language models. We are riding a tailwind where, as the models get better, alice naturally gets better. So it may sound strange that I think we should pause the development of the underlying language models. But I do, because I no longer think model intelligence is what holds us back as a society - and if we continue, we are potentially accepting extremely large risks.

We are no longer held back by how smart the models are

Two years ago, or even just six months ago, I think you could legitimately say that we were held back primarily by how intelligent the models were. Today I do not believe that is the main thing holding us back. What holds us back is how well we use them: implementation, change management, and the time it takes for people, organizations and society at large to adapt.

There may be edge cases. For example, a slightly better model might make us slightly more likely to cure a disease or make some breakthrough in science. But in most cases, it is probably more likely that the intelligence we need is already there, and we have not had time to explore it. And even if a model could tell us exactly how to cure cancer, we would still have to develop the drug, test it and roll it out. That takes years, because we are humans, and humans cannot do most things instantly.

The risks are growing faster than we can handle them

At the same time, the risks are real. The most immediate one is people using AI to cause harm. The International AI Safety Report, written by more than 100 experts and backed by over 30 countries, found that today’s models can give expert-level laboratory instructions relevant to biological and chemical weapons, and that criminal groups and state-linked attackers are already using AI in cyberattacks (International AI Safety Report, 2026). Bill Gates said in September 2026 that AI is “certainly powerful enough to drive events that cause a billion deaths” (Al Jazeera, 2026).

There is also a more distant worry that AI itself could one day act against us. Elon Musk has put the chance of something terrible happening at 10 to 20 percent (Deadline, 2024), and Yuval Noah Harari has warned that we could be looking at “the end of human history” (Fortune, 2023). I do not think we are close to that yet - we are still in control as the human race today.

It is tempting to just keep going. Progress is exciting, and every new model feels like a step forward. But if intelligence is no longer the bottleneck, the next model adds little, while even a small risk of things going extremely wrong could be catastrophic. I do not think that is a trade worth making.

We are on an exponential curve, and humans are bad at understanding exponential development. It is moving so fast that I do not think slowing down is enough. I think we should pause the development of more powerful models until we have better monitoring in place, and rules that make progress controllable enough to protect people. Bill Gates is calling for the first part: “We do need to monitor any sophisticated model and record exactly what’s being done” (NBC News, 2026). I would go one step further and pause until that monitoring is in place. To be clear, I am not against progress. I suggest a pause now, and I am open to moving forward again once we have the right monitoring and feel in control.

This is not a new idea. In 2023, more than 30,000 people, including Elon Musk and Apple co-founder Steve Wozniak, signed an open letter asking the labs to pause for six months (Future of Life Institute, 2023). It did not happen. Instead, the race and the investments behind it have only grown.

A pause would have to be coordinated internationally to work, and I know that is extremely hard. Governments and countries have strong incentives to be the ones in the lead, and nobody wants to stop first. Asked whether the US and China could agree on limits the way they did on nuclear weapons during the Cold War, Gates answered: “This is more difficult than that was” (NBC News, 2026). But if we can find an agreement and enforce it, I think that is the best solution. I think we have to try.

Pause the models, not the use

Because it is no longer model intelligence that holds us back, a pause costs us less than most people think. It does not mean we stop using AI. Companies, schools, public institutions and private individuals can keep using it, and we can keep building applications on top of the models we have. We should just not make the underlying language models more powerful until we are ready for the risks that come with it. It is better to pause now than to walk into a catastrophe and end up putting restrictions in place as a reaction, instead of being proactive.

Critics might say I have an interest in this, since I run an AI company. But this goes beyond alice. It is about what I see, and what I believe is the right choice for us as a society. Precisely because I am exposed to this every day, it is also very clear to me that intelligence is no longer the bottleneck.

If anything, a pause would cost us some of our tailwind at alice. But the models we have today are already more than good enough to give students a personal tutor. The hard part is building it well so it improves learning, and that is what we will keep doing - making visible progress daily. Try what we have built so far.

Several of us at alice are worried about where AI is heading. Let us do something while we are still in control as a society.