The AI Hype Was Too Cautious

AI Hype Was Too Cautious

For twenty years, Fill The Funnel has run on one simple rule: test the tool and ignore the hype. That rule has saved a lot of sales teams a lot of money. So when someone makes a strong case that the AI hype was actually too cautious, I pay attention.

That is the argument Benjamin Todd makes in a piece published on October 24. He went back and scored 18 months of AI forecasts from superforecasters, AI researchers, the general public, AI company leaders, and himself. His finding is simple: almost everyone guessed too slow.

The usual complaint is that AI hype oversells what the technology can do. Todd’s scorecard says the bigger mistake ran the other way. Here is what he found, and what I think it means for those of us who sell for a living.

The Scorecard

On math and coding, the forecasts were not even close. Todd predicted AI would score 32% on FrontierMath, a very hard math test, by the end of 2026. It reached 40% by the end of 2025 and 94% by this September.

He had expected scores near the top to take until 2030, so AI arrived more than three years early. On a tough coding test called LiveCodeBench Pro, he forecast 23% by the end of 2026. AI passed 53% a full year sooner.

The bigger math story is even more striking. In a late 2024 survey, more than 1,000 AI researchers said AI likely would not crack a major unsolved math problem until around 2054. In September 2026, AI appears to have solved one of the Millennium Prize problems, though the result still needs formal review that could take years.

It was not only math. Researchers expected a two-legged robot to outrun the fastest human over 5 kilometers sometime around 2034. This April, a humanoid robot beat the human record for the half marathon.

Even writing moved. In a New York Times quiz this March, the typical reader preferred AI writing over passages from famous novelists like Cormac McCarthy and Ursula K. Le Guin.

For anyone who watches the money, the revenue numbers tell the story best. Superforecasters expected OpenAI and Anthropic to reach a combined yearly pace of $90 billion by the end of 2026. They passed $100 billion by July, and Todd notes that AI company revenue has been growing about six times a year.

Why the Best Forecasters Missed

Superforecasters are the people with the best record at predicting world events. On AI, they did worse than any other group. Todd traces the pattern back to at least 2022.

Back then, they guessed the biggest AI training run would cost $35 million by 2024 and $100 million by 2030. The real cost likely topped $100 million by 2024, years ahead of their forecast.

Todd’s explanation makes sense to me. The smart way to predict most events is to assume things stay the same or stay on trend. AI keeps breaking the trend, so the safe-sounding guess keeps losing.

He also admits to a math trap he fell into himself. If you think in straight lines, going from 2% to 4% looks like a tiny gain. On a doubling curve, it is the same size step as going from 10% to 20%.

His point is not that experts are useless. AI researchers beat the general public on these forecasts. Their main mistake was failing to expect progress to be even faster than they thought. In other words, the experts were not adding to the AI hype. They were underselling AI.

The AI Hype Crowd Came Closest

Here is the twist. The AI company leaders, the people most often accused of AI hype, were among the few who guessed too high, and only by a little. Todd says that if you took their six-month predictions and assumed six to eighteen months instead, you would have done well.

Sam Altman said in early 2025 that AI agents would join the workforce that year. They trickled in at first and then took off in early 2026.

Dario Amodei said in March 2025 that AI would write 90% of code within three to six months. Anthropic’s own codebase reached 80% after about 14 months. The timing was early, but the direction was right.

Todd’s advice is to take what these companies say seriously and then add some time. The best AI forecasters he knows work this way. Most of them now expect AI that beats humans across the board around 2028 or 2029, and Todd puts his own middle estimate a little later, near 2030.

What This Means for Sales Teams

Thirty years in sales technology taught me to discount every vendor promise. That habit still makes sense for most sales tools. But when it comes to the AI underneath those tools, Todd’s scorecard tells me that tuning out all the AI hype may be the wrong habit.

First, stop planning around today’s limits. I hear it all the time: AI can write a decent email, but it can’t run a real discovery call or read a buyer’s mood. Todd’s record shows that limits people call permanent often fall within months, so build your process and hiring plans for where the tools are heading, not where they sit today.

Second, read roadmaps a new way. When the companies building the core AI models say something is coming, Todd’s rule is to believe the direction and pad the timeline. That is not the same as trusting every sales vendor’s feature list, because a better model does not guarantee a better product built on top of it.

Third, notice the gap. Todd points out that big changes in jobs and the wider economy have not shown up yet, and he did not expect them to by now. The space between what AI can do and what most sales teams actually use is where experienced sellers can get ahead, because the tools are moving faster than the habits.

A Fair Word of Caution

One reader raised a good point in the comments under Todd’s post. Fast wins get scored the moment they happen, while overly bold predictions can’t be marked wrong until their deadline passes. So any scorecard taken today may lean toward “everyone was too slow.”

There is also a difference between passing hard tests and doing real work on its own. Strong scores in math and coding do not prove AI can run a sales territory. Todd himself has nudged his estimate a bit later and now gives real weight to a path where progress slows after 2028.

He also ends on a serious note. In his view, pacing the development of the most powerful AI systems is now the single biggest step for reducing AI risk, and he has updated his list of priorities to put it first.

The Bottom Line

The old rule of discounting hype has served sales teams well for a long time. When it comes to AI hype, the evidence says that rule may now be costing us. That is not a reason to panic or to buy every new tool.

It is a reason to plan for AI that will be better next year than it is today, probably by more than you think. If the people closest to this work keep getting surprised, the rest of us should expect to be surprised too.

Source

Benjamin Todd, AI wasn’t overhyped. Almost everyone underestimated it., October 24, 2026.

Leave a Comment

Your email address will not be published. Required fields are marked *

Scroll to Top