Read

4 MIN

Date

Category

Sam Altman Admits He Still Works the Old Way.

The man who built Claude Code made sure nobody on his team can. Two companies people put on opposite sides of AI, and the same lesson from both ends: the tool changed, and the job has to change with it.

Sam Altman Admits He Still Works the Old Way.
Champ Smith
Champ Smith

Champ Smith

Operator & AI Cystems Builder

I build the custom AI Cystems that run businesses for the operators who own them — leads routed, content shipped, calls handled. I work from a finca in Málaga, where the intelligence lives in the walls.

The tool changed. The habit did not.

Sam Altman has used a computer the same way for twenty years. He said so himself, and he said it was the question he had been waiting for someone to ask. He has Codex. Everyone at OpenAI has Codex. That should mean he works completely differently. He should not be clicking around, pasting from one messaging app into another, scrolling his inbox for the email that is least painful to answer, keeping a to-do list and grinding through the same rote tasks he has done for decades. He still does all of it. Something in his head is encoded to say this is what work is. He would never claim to like it. By revealed preference, he keeps doing it. His words: it makes no sense.

That is the CEO of OpenAI. The tool changed. The habit did not.

Boris built the tool, then rebuilt the team

Boris Cherny built Claude Code at Anthropic. Ask him what he hires for now, if it is no longer the ability to write code, and his answer is that the split into engineering, design, product, user research and data science is the old way of thinking. On the Claude Code team everyone writes code. Designers, product managers, engineering managers. Not because a memo said so. Because it got easy. His designer does not message him to move a button one pixel. She moves it.

So the roles split the other way, along the life of the product instead of along a person's training. Prototypers, who find the first idea and iterate on it fast. Builders, who take a new idea and get it to market. Maintainers, who keep the software standing once it is at scale. Scalers, who take a product that already fits its market and grow it ten or a hundred times. And sweepers, the one he did not have a good name for yet: the people who polish the product, the infrastructure and the code until the rough edges are gone. Someone offered "perfectors." He said it is one of the most important roles on the team, because polish is what a user actually feels.

The common thread

Put the two side by side. These are the two companies most people place on opposite sides of AI, and the two men are describing the same thing from different ends. The tools moved faster than the way people work. Altman is honest that he has not caught up. Boris made his team catch up, and he did it by changing what a role is, not by handing out a new tool and hoping.

That is the whole point. A person with a skill used to be the unit of work. The skill is now available to anyone, so the question becomes which stage of the work you are best at, and who or what holds the stages you are not in.

I was in Altman's camp

Until recently I worked exactly the way he describes. Too many tabs. Copying between apps. A to-do list I reread more often than I finished. Email sorted by how much I dreaded it. I had agents, and I still did the rote work myself, because doing it felt like working.

What moved me out of that was not a better model. It was watching how Anthropic, and people like Boris, build teams. When the person who wrote the tool tells you his designer ships her own pixel, you stop asking whether you are allowed to hand the work over. You hand it over. I now sit in one of the five roles at a time, mostly the prototyper, and the other four are held by agents. Building runs while I sleep. Maintaining is a scheduled job. Sweeping is a review board that reads every change before it ships, and I never clear my own work.

I am not all the way there. Some days the old way still wins on habit, same as Altman. The difference between his answer and Boris's is the difference between knowing and doing, and it comes down to whether you changed what a role means or just added a tool to the old one.

Why I am backing Boris by name

He built the thing that put code in everyone's hands, watched it change his own team, and then redrew the roles around what actually happened instead of defending the chart he came in with. Most people at the centre of this industry are still selling the tool. He is showing what the work looks like after it. That is worth saying out loud.

What to do with it

Pick the stage you are actually good at. Give the rest away on purpose, to people or to agents. Then count what you still do by hand every week and ask which of it is work and which is habit. Altman's list is a fair audit: clicking around, pasting between apps, email by least pain, the to-do list. If the CEO of OpenAI can name his, you can name yours.

And build the sweeper in. It is the role everyone skips, because polish has no line in the budget until a user hits a rough edge.

Build the Cystem. Watch it work.

Sources: Sam Altman, interviewed by David Senra, on the one question nobody had asked him. Boris Cherny on Bloomberg's Odd Lots, on what the Claude Code team hires for now. Both clips are short and worth watching in full.

Read

4 MIN

Date

Category

I share ideas, lessons, and practical insights from my work.

Related
Articles.

Sep 3, 2026

The Signal Was Already in the Stream

Signal

The Signal Was Already in the Stream

Google Research just built a glucose model with only 0.72 million parameters, and it beats every bigger model that came before it. It reads the continuous glucose monitor people already wear and predicts diabetes risk, insulin resistance and beta-cell dysfunction, with no new blood test. The lesson is bigger than health: the signal you want is usually already in a stream you are already collecting. You are just reading it as spot values.

Aug 30, 2026

The AI That Predicts the World, Not the Next Word

Signal

The AI That Predicts the World, Not the Next Word

Two researchers turned down more than $2 billion tied to Jeff Bezos' new AI venture — then came out of stealth with a model that does not do language at all. It simulates physics in four dimensions. We read it as confirmation of a CYSTEMS conviction: prediction, not conversation, is what intelligence is made of.

Aug 25, 2026

GPT-6, Astra, and Ilya’s Secret Model

Signal

GPT-6, Astra, and Ilya’s Secret Model

Three names are dominating the AI conversation right now: GPT-6, something called Astra, and a mystery model from the most secretive lab on the planet. Depending on who you listen to, one or all of them are landing any day now. Here is what is actually confirmed, what is speculation, and what it means if you are building a business on top of AI.

Aug 24, 2026

The End of the Perfect Prompt

Cystems

The End of the Perfect Prompt

The most important number in AI right now is not a model score. An older, cheaper model wrapped in a loop solved 95 percent of a coding benchmark that the newest model, asked once, solved 67 percent of. Architecture beat intelligence. If you are still hunting for the perfect prompt, you are optimizing the wrong layer.

Sep 3, 2026

The Signal Was Already in the Stream

Signal

The Signal Was Already in the Stream

Google Research just built a glucose model with only 0.72 million parameters, and it beats every bigger model that came before it. It reads the continuous glucose monitor people already wear and predicts diabetes risk, insulin resistance and beta-cell dysfunction, with no new blood test. The lesson is bigger than health: the signal you want is usually already in a stream you are already collecting. You are just reading it as spot values.

Aug 30, 2026

The AI That Predicts the World, Not the Next Word

Signal

The AI That Predicts the World, Not the Next Word

Two researchers turned down more than $2 billion tied to Jeff Bezos' new AI venture — then came out of stealth with a model that does not do language at all. It simulates physics in four dimensions. We read it as confirmation of a CYSTEMS conviction: prediction, not conversation, is what intelligence is made of.

Sep 3, 2026

The Signal Was Already in the Stream

Signal

The Signal Was Already in the Stream

Google Research just built a glucose model with only 0.72 million parameters, and it beats every bigger model that came before it. It reads the continuous glucose monitor people already wear and predicts diabetes risk, insulin resistance and beta-cell dysfunction, with no new blood test. The lesson is bigger than health: the signal you want is usually already in a stream you are already collecting. You are just reading it as spot values.

Aug 30, 2026

The AI That Predicts the World, Not the Next Word

Signal

The AI That Predicts the World, Not the Next Word

Two researchers turned down more than $2 billion tied to Jeff Bezos' new AI venture — then came out of stealth with a model that does not do language at all. It simulates physics in four dimensions. We read it as confirmation of a CYSTEMS conviction: prediction, not conversation, is what intelligence is made of.