A machine out-guessed the humans.
That is the short version of why Mantic, a British AI startup based in London, announced $25 million in seed funding on September 18. The longer version is more interesting, and a little more humbling, because the contest it won was not a chess match or a coding benchmark. It was a competition about predicting real events in a messy, unfinished world.
The event was the 2026 Metaculus Cup, a forecasting contest. Mantic’s system beat every human competitor. It did not win outright, though. It finished second overall, behind another bot. I want to sit with that detail for a second, because it says more about where we are than the headline does.
What “AI forecasting” actually means
If you have used a chatbot, you have seen a system that is good at explaining things that already happened. Forecasting is a different job. You are asked a question about the future with a genuinely unknown answer, and you have to put a number on it. Not “yes” or “no,” but something like “there is a 23% chance this happens before June.”
Human forecasters who do this well tend to share a few habits:
- They break a big question into smaller, checkable pieces.
- They start from base rates, meaning how often this kind of thing usually happens.
- They read widely and update their numbers when new information shows up.
- They avoid falling in love with a story.
That last one is where people struggle most. We are narrative creatures. A compelling explanation feels like evidence even when it is not. A software system has no ego invested in being right last week, which turns out to be a real advantage when the news changes.
Why this matters for anyone curious about AI agents
Readers of this site often ask me what separates an AI agent from a chatbot. My usual answer is that an agent does things, while a chatbot says things. But there is a quieter difference underneath: an agent that does things has to make judgment calls about what will happen next.
Think about any useful task you might hand to software. Should I reorder stock now or wait two weeks? Is this supplier likely to miss the deadline? Will this flight get delayed enough that I should rebook? Every one of those is a forecast wearing ordinary clothes. An agent that cannot estimate probability is just a very articulate to-do list.
So a company focused specifically on prediction quality is working on something close to the engine room of the whole agent idea. That is a big part of why the funding is notable. Mantic says the money will go toward scaling its forecasting technology, and the investor list includes Radical Ventures and M12, which is Microsoft’s venture arm. When a major platform company’s fund shows up early, it usually means someone sees the technology plugging into much larger systems later.
Reasons to stay curious rather than convinced
I like this story, and I still want to keep my expectations tidy. A few honest caveats.
One contest is one contest. Forecasting tournaments have specific question sets, specific time windows, and specific scoring rules. Performing well in that setting is a real result. It is not the same as being reliably better than experts on every question that matters to you.
Second place is a useful reminder too. Another bot finished ahead of Mantic, which suggests we are watching a category compete with itself, not a single system pull away from the pack. That is a healthier sign for the field than a lone winner would be, but it does complicate the word “superhuman.”
And probability is slippery for all of us. If a system says an event has a 30% chance and the event happens, the system was not necessarily wrong. Judging forecast quality takes many predictions over a long stretch. Anyone who tells you a single correct call proves anything is selling something.
What I would watch next
The practical question is where this lands in products you might actually touch. Forecasting is valuable to insurers, supply chain teams, public health planners, journalists, and honestly anyone who has to write a budget. If Mantic’s technology scales the way the company hopes, the most likely shape is not a public oracle you visit for predictions. It is a layer sitting quietly inside other tools, nudging decisions with numbers instead of vibes.
The version of that future I would welcome is one where systems show their reasoning, name their uncertainty, and get graded in public over time. Forecasting has a strong culture of exactly that kind of scorekeeping, which is part of why a win there carries weight.
For now, one clear fact stands: in a contest built by and for skilled human forecasters, the humans came in behind the software. That is worth paying attention to without treating it as the final word.
🕒 Published: