The people who know the most often sound the least certain

I watched Dwarkesh Patel’s recent conversation with John Schulman, Beren Millidge and Charlie O’Neill this week. For an hour and a half, some of the smartest people working in AI tried to reason, in public, about genuinely difficult questions.

Can AI automate AI research? Can it choose its own research objectives? Does progress hit a wall when the thing we need cannot be turned into a clean reward? What happens when models become better at running experiments but humans are still needed to decide which experiments matter?

These are not small questions. The answers could shape the economy, politics and quite possibly the trajectory of civilisation. But a noticeable share of the reaction was about how often the speakers said “like”. I get it. Once you notice a verbal tic, it can become impossible to stop noticing it. “Like” can make a sentence feel less precise. Used often enough, it can make a brilliant person sound less confident than they are.

But I also found the reaction revealing.

We had people with deep, first-hand knowledge trying to think at the edge of what anyone knows. They corrected themselves. They qualified claims. They searched for examples. They occasionally reached for a filler word while assembling a difficult thought in real time. Meanwhile, much of the public conversation about AI is being led by politicians, executives, and professional commentators who have mastered the opposite skill: delivering simple, polished certainty about things they barely understand.

The people with the least friction from reality often have the least friction in their sentences. That is a problem.

Researchers hedge for a reason. The honest answer to an important technical question is often: it depends on the objective, the data, the environment, the time horizon and what exactly you mean by “intelligence”. That answer is accurate. It is also terrible television.

The less careful speaker has an enormous advantage. They can remove every condition, compress a probability into a prediction and turn a complicated system into a slogan.

All of these travel further than: “There are several plausible mechanisms here, and the result depends on which bottleneck dominates.” Public discourse rewards confidence long before it rewards calibration. This is not unique to AI. But AI makes it unusually dangerous because the gap between what the technology can do, what people think it can do, and what institutions claim it can do is already enormous. Decisions about jobs, education, national security, investment, and regulation are being made inside that gap.

If the people who understand the systems retreat into papers, private labs, and conversations legible only to one another, the vacuum will not remain empty. It will be filled by people whose expertise is language itself: message discipline, emotional hooks, selective statistics, and sentences engineered to survive a news cycle. Language hijinks beat technical truth when technical truth has no storyteller.

This is why formats like Dwarkesh’s matter. He is creating space for experts to reason in public, not merely arrive at an approved conclusion. The disagreement is the content. You hear assumptions being tested, definitions being repaired, and confident positions becoming less confident under pressure. That is what thinking sounds like.

It is also messy.

A polished keynote hides the discarded branches. A live conversation exposes them. Someone begins a sentence before they know exactly where it will end. They use an analogy, realise it is imperfect, and replace it. They say “like” while searching for the bridge between an intuition and a claim.

We should want more of this, not less. But the criticism cannot simply be dismissed either. If experts want to influence public understanding, being correct is not enough. An idea that cannot escape the room loses to a worse idea that can. It is also why I really love the new whiteboard formats. A whiteboard makes the construction of an idea visible. You watch someone draw the model, notice where it does not quite work, erase part of it, and try again. It turns expertise from a polished answer into a process the rest of us can follow.

The answer is not to turn researchers into politicians. It is to help technical people carry more of their intelligence across the gap. I have had to learn this the hard way.

I am a neurodivergent (AuDHD) builder turned CEO. My brain tends to branch.

While I am explaining one idea, I can see three qualifications, two exceptions , 4 steps ahead, and an adjacent problem that suddenly feels important, whilst trying to solve a different problem in my head. If I try to say all of them, the main point disappears. If I suppress them all, I worry that I am being inaccurate. Sales makes this especially unforgiving. A customer does not experience the internal sophistication of your mental model. They experience the sentence that arrives.

Over time, I have found a few things that make speaking easier without forcing me to pretend the world is simpler than it is.

Simple language is not evidence of simple thinking. It is usually evidence that someone has done the compression work. I practise explaining ideas as if the listener were five, then ten, then an intelligent adult outside the field. GPT is useful here, not as a ghostwriter, but as a sparring partner.

Ask it:

Which words in this explanation require prior knowledge?

Explain my idea back to me as if you were a customer hearing it for the first time.

What is the one sentence they are likely to remember?

If the simple version changes the meaning, it is too simple. If it preserves the mechanism, you have probably found the idea.

Silence feels much longer to the speaker than it does to the audience. When I need to think, my instinct is to keep the audio channel occupied: “like”, “um”, “sort of”, “you know”. Barack Obama often did something more effective.

He paused.

A pause can feel authoritative because it signals that the next sentence is being chosen rather than spilled. More importantly, it gives your working memory a moment to catch up. You do not need to eliminate every filler word. Just replace some of them with air.

People rarely repeat your framework accurately. They repeat the story that made the framework make sense. The simplest story structure I know is:

A person wanted something. Something got in the way. Something changed.

That is enough.

In sales, I try not to begin with the architecture of the product. I begin with the customer who was losing leads after 6pm, the employee who came in each morning to a pile of callbacks, and what changed when those conversations were handled immediately. The technology can follow. The story gives it somewhere to land.

Experts often lead with everything that could make them wrong.

“I have not looked at all the data, and there are several possible explanations, but perhaps…”

By the time the claim arrives, the room has mentally left.

Try reversing the order: “The bottleneck is evaluation. There are two cases where that may not hold.”

The uncertainty is still there. It is simply attached to a legible position.

This matters because non-experts will not hedge on your behalf. If you bury your conclusion under seven qualifications, someone with less knowledge will state a worse conclusion in eight words and own the conversation.

Be assertive where the evidence allows it. Hedge where the uncertainty is real. Do not hedge merely to make yourself socially safer.

One idea per sentence is a good default. In a prepared monologue, a longer sentence can work because you control the rhythm and destination. In a live conversation, shorter units are more robust. They survive nerves, interruptions, and the listener’s limited working memory. This does not mean making every idea shallow. It means revealing complexity in layers.

Writing has been the best speaking practice I have found. It forces me to discover whether I have an idea or merely a cloud of associated thoughts. Publishing adds useful pressure. You have to choose the claim, find the story and notice where readers get lost. Substack is particularly good for this because it allows enough space to reason without requiring the polish of a book or the compression of a social post. Write the idea. Tell the story. Publish it. Then explain it aloud without looking at the page. That loop compounds.

The goal is not to make every AI researcher sound like a media-trained CEO. Frankly, we have enough media-trained CEOs. The hesitations, qualifications , and live corrections are evidence that someone is reasoning rather than reciting. We should be careful not to punish every visible sign of thought until only the performers remain.

But technical people cannot complain that public debate is shallow while refusing to learn how ideas travel. We need both: experts willing to think in public, and experts willing to become better storytellers. Keep the caveats that protect the truth. Remove the complexity that merely protects the speaker. Pause instead of filling space. State the claim before qualifying it. Use simple words. Give people one story they can carry into the next room. The future of AI will not be decided only by the people building the systems. It will also be decided by the people who can explain what those systems are, what they are not, and what we should do about them.

If the experts do not learn to tell that story, someone else will.