AI Makes Mistakes. That Is Why You Compete on the Number of Attempts: Lessons from 722 Math Manuscripts

You try AI, and it returns a clearly wrong answer. So you conclude, "AI still isn't good enough." I hear this story all the time.

I think that conclusion is a serious mistake. The greatest value of AI is not that it gets the right answer on the first try, but that it runs trial and error at high speed.

722 math manuscripts, and errors in some of them

On October 6, 2026, OpenAI released 722 mathematics research manuscripts written by a new internal model. The results span 17 fields and are organized into 372 groups. According to OpenAI, they came from posing roughly 4,000 problems to the model.

The day after the release, a sign error was found and three manuscripts were withdrawn, and 14 others were revised. Reports say that only about 40 percent of the main results have been formally verified line by line with the software Lean. Results without that verification are not necessarily wrong, but they still need to be checked by people.

You could read this as "AI is amazing" or as "AI is dangerous." I read it as trial and error itself: many attempts, checking, removing errors, fixing, and moving on. If the policy had been to release only perfect work, this volume would never have appeared.

Do not judge AI by a single mistake

When AI produces something wrong, people immediately declare it "imperfect." That is the use of AI I find most wasteful.

When people try something, each round takes time: prepare, build, check, fix. AI makes each round very light. Lighter rounds mean more attempts, and more attempts raise the chance that some of them hit. Misses increase too, but if you cut the number of attempts to avoid misses, you also cut the hits. A first miss is not proof that AI is useless. It is material for the next attempt.

Not everything has to be right for you to move forward

What I have actually seen is using AI to improve features. When a company keeps trying improvements to its internal systems and tools with AI, some of them are actually used on the front line and become part of daily work.

Not all of them succeed. Some are built and never used; some are tried briefly and dropped. That is fine. Release them as prototypes, use them, and decide what to keep and what to discard. Because some were discarded, you could choose what to keep. The more you increase the number of attempts, the more surely you move forward.

Perfectionism, and budgets that do not match expectations

The most common failure is using AI while staying a perfectionist: expecting a finished answer on the first try and declaring AI "useless" otherwise. That wastes AI's strength, which is the speed of trial and error.

Companies where AI is not really being used often share another trait. Their expectations are high, but they have not allocated a budget to match. People have only touched it briefly, there are almost no attempts, and no time is set aside for people to think. They conclude "it wasn't as good as we hoped" without ever drawing out AI's performance.

At the trial stage, produce many attempts without fearing misses. At the adoption stage, check whether something is ready to enter real work. Separating these two matters. Checking is not a way to stop trial and error; it is part of keeping the loop running.

The criterion: did you move forward?

My criterion is simple: did trial and error move you forward?

If you judge by whether you got the right answer, every miss becomes a failure. If you judge by whether you moved forward, even a miss can be progress. Learning that a certain approach does not fit is a step forward. Conversely, if something looks successful but the work has not changed at all, you have not moved forward.

Compete on the volume of trial and error

The main point I want to make is this: compete on how much trial and error you repeat. The quality of what you finally keep comes from choosing among many attempts. Now that AI has made each attempt light, it is better to shift the axis of competition from "accuracy on one try" to "the volume of attempts."

That requires securing budget and time for trying, not blaming people for misses, and evaluating by whether things moved forward. Executives should first align expectations with budget, and they should not judge AI by a single mistake themselves. A single remark from the top can greatly change how much trial and error happens.

What you can do starting tomorrow

First, choose one task to try with AI and decide in advance how many attempts you will make. Do not judge by the first result until you have finished those attempts.

Second, record results not as right or wrong but as "did we move forward?" Did something take hold, and did the next thing to try become clear? Writing these two points briefly is enough.

Third, put what you expect from AI next to the budget and time you are actually spending on it. If they do not match, choose to add budget or lower expectations. Once they are aligned and you try enough times, you can finally see clearly whether AI is truly useful.

▶ Watch the full video (Japanese)
https://www.youtube.com/watch?v=g3gnfzhqnKs