Masayoshi Son disputiert sieben Stunden taeglich mit KI: Plaedoyer fuer adversariales Prompting

Posted on Oct 25 • Originally published at appliedaihub.org

The standard warning about heavy AI use is that it makes you lazy. Hand your writing, coding, and strategy to a model for long enough, the argument goes, and your own judgment quietly wastes away.

There is real data behind that fear. A 2025 study by researchers at Microsoft Research and Carnegie Mellon University surveyed 319 knowledge workers and found a clear pattern: the higher a user's confidence in the AI, the less critical thinking they applied to the output. Most people treat a language model like a vending machine. They put a one-line coin in, take the first plausible answer out, and paste it into their work.

Then you listen to Masayoshi Son. In our timestamped breakdown of his interview on Inside Blackstone, the 69-year-old SoftBank founder admits that he spends three hours a day talking to AI on weekdays and up to seven hours a day on weekends. Yet when the host brings up the fear that heavy AI reliance dulls the human brain, Son immediately pushes back.

Son does not use AI to avoid thinking. He uses it to pick a fight with his own assumptions. Here is how he described his daily habit in the sixth insight card of the interview:

"It's so productive, so exciting. And people say, hey, you know, when AI, if you use, you don't use the brain, it's the opposite. It's totally opposite, okay? It's like having a thought partner. When you start, you know, debating, right? Discussing. If the AI says this, I say, oh, but what about this angle? Can you make it 20% more efficient? Can you do this and get that? You can debate, right? Brainstorm. It's my thought mate now."

Look closely at the verbs he uses: debating, discussing, challenging angles, demanding 20 percent more efficiency. He is not asking ChatGPT to write his emails so he can go play golf. He is running a seven-hour adversarial stress test on his own capital allocation ideas.

That distinction explains why two people can spend the exact same number of hours inside a chat window and walk away with opposite cognitive results. One person's brain atrophies; the other person's brain gets sharper. The variable is not screen time. The variable is whether you use the model for delegation or collision.

To see why collision works so well, you have to look at how language models generate text and how human executives normally make bad decisions.

When you ask a model a neutral question ("Is this product roadmap a good idea?"), RLHF post-training nudges the model toward agreeable, middle-of-the-road support. It lists three polite pros, two mild risks, and wishes you luck. If you accept that output at face value, you are outsourcing your judgment to the statistical average of internet prose. That is cognitive offloading in its worst form: you trade away deep domain reasoning for a smooth summary.

Now consider what happens inside a large company or an investment committee. The higher a founder or executive rises, the harder it becomes to find someone willing to tell them their favorite idea has a fatal flaw. Direct reports hedge. Consultants flatter. Meetings optimize for consensus.

A language model has no career risk. If you instruct it to attack your thesis, demand a 20 percent efficiency cut, or find the exact supply-chain bottleneck that will bankrupt your plan in year three, it will do so relentlessly at midnight on a Sunday. When you push back ("What about this angle?"), the model forces you to articulate the exact mechanism behind your intuition. You cannot wave your hands in a debate with a machine that keeps asking for the numbers.

Author's Comment: Notice that the Microsoft and CMU study found a second, less publicized correlation. Workers who had high confidence in their own domain expertise actually did more critical thinking when using AI, because they treated the model's output as a rough draft to interrogate rather than an oracle to obey. Son's seven-hour sessions are an extreme version of that active interrogation loop.

Son's second contrarian claim in the interview is just as sharp, and it connects directly to how he uses AI. Most engineering leads and startup founders treat long-term planning as a binary choice: either you write a rigid five-year waterfall plan that goes stale in a month, or you throw up your hands, say "the industry is moving too fast to predict," and only plan the next two-week sprint.

Son rejects both. In the seventh insight card of the Inside Blackstone breakdown, he explains how SoftBank handles long-horizon uncertainty:

"Nobody knows exactly what's gonna happen in 10 years. That's why we have to think. You know, people say, oh, 10 years later, nobody knows so let's not think about it. Let's think about next year, okay? I think it's the opposite. It's the opposite. If you don't know what's gonna happen in 10 years, don't give up thinking about it. Think, think, think! Okay?"

Even more striking is the cadence he describes right before that quote. At his meetings every single day, his team debates the 10-year strategy, and the very next day they revise that 10-year plan again because 24 hours of new inputs have arrived.

In quantitative finance, this is simply Bayesian updating applied to a long-horizon target. You do not shorten your time horizon just because variance is high. If you only look 12 months ahead, you optimize for local maxima; as Son puts it in his "70% rule," since human execution yields roughly 70 percent of whatever target you set, a small target guarantees a small outcome. Instead, you lock onto a massive 10-year structural vector and update your path toward it every 24 hours.

This echoes a theme we have seen across other frontier leaders. In our breakdown of Jensen Huang on the BG2 Pod, Huang explains that active reasoning and tree-search exploration require a billion times more compute than simple memorization because you have to test hypotheses and prune dead ends. Son is applying that exact tree-search loop to human strategy, using AI as the exogenous compute engine to re-evaluate his 10-year tree every night.

If Son is right about the 7-hour thought partner and the daily 10-year revision, then 90 percent of builders and knowledge workers are currently configuring their AI workflows backwards.

Most teams spend all their energy automating low-stakes output generation: auto-writing boilerplate tickets, summarizing meeting transcripts nobody reads, or generating first-draft blog posts. Meanwhile, they make their highest-leverage architectural and strategic decisions the old-fashioned way, inside a two-hour Tuesday afternoon meeting where everyone is tired and polite.

When you use AI only to type faster, you hit a ceiling quickly. Worse, you flood your organization with verbose text that someone else now has to read and verify. When you flip the workflow and use AI to stress-test decisions before a single line of code is written, the leverage compounds across the entire project.

You do not need seven spare hours on a Saturday to apply Son's method. You need to stop prompting your model like a secretary and start prompting it like a hostile board member. Here are three concrete protocols you can put into practice tomorrow morning.

Notice how Son specifically asks his AI: "Can you make it 20% more efficient? Can you do this and get that?" Vague critiques produce vague advice. A hard numerical constraint forces the model to cut structural fat.

Before you finalize a system architecture, a cloud infrastructure budget, or a project timeline, paste the spec into Claude or ChatGPT with a strict forced-reduction rule:

Here is our proposed architecture and execution plan: [Paste spec] Assume a hard constraint: we must cut latency, token cost, or engineering complexity by 20% without dropping core reliability. 1. Identify the two weakest or most over-engineered components in this plan. 2. Propose three concrete trade-offs that achieve the 20% reduction. 3. State exactly what breaks if we push that cut to 40%. Enter fullscreen mode Exit fullscreen mode If you want to go deeper on adversarial stress-testing, you can also run your thesis through our Devil's Advocate Viewpoint Challenge prompt or use the AI Blind Spot Mirror to surface the unstated assumptions in your plan.

In a real debate, you do not let the other person deliver a five-page monologue. You go back and forth, one claim at a time.

When you open a brainstorming session, forbid the model from agreeing with you and cap its reply length so you can actually argue with it. We covered why sycophancy ruins single-pass prompts in our guide to the Recursive Reflection prompt framework. In a live chat, you become the reflection loop:

Practical Pitfall Avoidance Guide: Watch out for fake disagreement. If you tell a model to "criticize my idea," it will often invent nitpicks about edge cases while praising your core premise. Force it to attack the load-bearing assumption by asking: "What is the single most likely reason this entire project fails 18 months from now, and what early signal would prove you right within 30 days?"

Instead of treating your long-term technical roadmap as a static document that lives in a dusty Notion folder, keep a compact markdown file of your 3-year or 10-year thesis. Whenever a major new model release, pricing shift, or customer insight lands, feed both the old thesis and the new input into your thought-partner session.

Ask one question: "Given what changed in the last 24 hours, which branch of this long-term plan just became obsolete, and which bottleneck just moved?"

The split in the AI economy will not be between people who use AI and people who refuse to. Everyone will use it. The split will be between passive passengers who let models think for them and active sparring partners who use models to think harder.

Next time you open a chat window to work on a hard problem, resist the urge to ask for the answer. State your best thesis, tell the model to find the flaw, and start the debate.