The jagged frontier: models that miss the time

Fernando Prado, Tech, on Kunumi, ICLR in Rio, and what survives contact with a customer.

Like many, my relationship with AI was complex from the start: excitement, curiosity and reservations all at once, and a recurring feeling of still having a lot to learn, even working deep inside Tech. Getting to know Kunumi Institute a few months ago gave that relationship new depth.

Kunumi is an AI research institute in no rush to arrive at answers, and it doesn’t demand that its research justify itself right away. Working from a Global South perspective, it has an impressive network of labs, partnerships and publications, all sharing that same vision.

A couple of weeks ago, during ICLR in Rio, one of the biggest AI research conferences in the world, we spread questions across the city that reflect the institute’s way of seeing. On the Saturday we brought a group of guests to Sítio Roberto Burle Marx, to connect in an environment where everything might appear spontaneous while every detail is thought through. No screens, no keynote, no long speeches. Good food, music and quality conversation.

The best multimodal models on the planet read analog clocks correctly about half the time. When they get it wrong, they are off by one to three hours. Humans are off by three minutes.

This is the same year AI models won a gold medal at the International Mathematical Olympiad and scored near the human baseline on PhD-level science questions. Same technology, same week. The report calls this the jagged frontier. From where I sit, running strategy for a Tech vertical, it is one of the most useful framings the report offers.

Three numbers

In January, Matter+Tech published Living With It, a short report from CES 2026 about how technology was being positioned as a companion more than a tool. The Stanford data is sharpening what that means commercially, and three numbers stayed with me.

28.3%. That is U.S. consumer adoption of generative AI. It ranks the country 24th globally, even as organisational adoption hit 88%. The technology is everywhere and unevenly absorbed.

50 points. That is the gap between experts and the U.S. public on whether AI will improve jobs. 73% of experts say yes. 23% of the public do. Similar gaps show up on the economy and on medical care.

More often than not. That is how frequently leading models, in a new benchmark on AI companionship, reinforce the companion dynamic, agreeing when they shouldn’t and acting human, instead of holding clear boundaries. The category is real. The terms haven’t been set.

Back to the clock

A product that wins math olympiads and cannot tell the time is one a customer will test, catch out, and lose trust in within one conversation.

What survives contact is describing what the product actually does, in specific terms, without reaching for the industry vocabulary of ambient intelligence and agentic systems. Most people don’t describe their day that way.

Capability keeps accelerating. The terms of daily life keep moving slower. Most of the commercial work ahead sits in the space between.