The AI labs want to become banks.

This is letter #100.

The first one went out in October 2024, to six people.

When I started, I honestly didn't know what to write about. So I made one simple decision: every week, I would share what I was thinking about from IMPLEMENTING AI in the real world. With real teams, real budgets, real messes. Not from X threads, and not from whatever a founder said on stage that week.

It turns out that "AI in the real world" and "AI on social media" are two different planets right now.

Most people told me to write about hacks. And I did, in a few of the early letters. (One of them was literally called "The Secret AI Hack To Create Stunning Diagrams In Minutes." I still wince.)

But that's not what people were looking for.

So I chose to write honestly about what I think. And that's what brought the best replies, the best discussions, and the people who stayed. That's what made this letter.

Thank you. Really.

I read every reply. I loved every single piece of feedback you sent, and your messages of appreciation are the best reward for this work. This letter only exists because of you.

 

So for #100, I want to do something a bit different.

In AI, a couple of months feels like a century. Everything moves at light speed. What you learned in January is a museum piece by June.

This letter is a recap of my thoughts, and especially the ones that survived all those centuries.

First, a quick macro view of what happened since ChatGPT. Then the handful of things I believe are true about people and companies, whichever model wins next.

Three years in one page

OpenAI led the dance in 2023 and 2024. For a while they were the only lab with a reasoning model, with "o1." On the other side of the Atlantic, Mistral was one of the best teams in the world at mixture-of-experts (MoE), the trick that lets a huge model only wake up the parts it needs.

Then, at the end of 2024 and early 2025, came the DeepSeek moment.

DeepSeek cracked reasoning, found much more efficient ways to do MoE, published everything in papers, and released the models OPEN. That was a huge moment. Suddenly the secret sauce wasn't secret anymore.

The US labs caught up with OpenAI, and the whole race boiled down to compute. Whoever had the most chips and the most power won the next round. That gave Google a brief moment of glory with Gemini.

Then Anthropic took the lead. Their Opus models were incredible. But more importantly, they had the cleverest idea of late 2025 and early 2026: the harness.

It started with Claude Code, and the idea was simple. Take a very capable model. Give it very capable tools: reading and writing files, searching the internet, running code in a sandbox, loading skills. Wrap all of it in clever software that lets the model act, check its own work and keep going.

It gave a tremendous advantage. Most of us can't imagine going back to the "simple chatbots" we had before.

OpenAI recently caught up. They upgraded their harness, and more importantly they shipped a whole new model, GPT-6 Astra, which looks very good at computer use and "visual reasoning." It took the top score on ARC-AGI-3, a benchmark built specifically to resist memorization.

BUT.

AI got better and more agentic, and that has a cost. An agent doesn't answer once. It reads, plans, tries, fails, retries, checks. Every one of those steps burns tokens.

AI bills literally exploded.

The part most people missed

Good news: we already had an alternative.

While all of this was happening, the Chinese labs and the open-weight world kept progressing at an incredible pace. Today they offer an alternative to almost everything frontier AI does: very capable models, harnesses, tooling. Open, and at a fraction of the cost. All you need.

That was a huge game changer.

Thanks to the Chinese (I know this might sound strange to some of you), we live in a world where we are not 100% dependent on closed US AI.

This helped many teams, including many of my clients, divide their AI bills by ten. Sometimes much more.

The money will defend itself

But the story doesn't end here.

The capital invested in closed AI is enormous, and capital defends itself. In my opinion, one of the only ways to do it is regulation and watermarking.

The pitch writes itself: we are the compliant AI, and everything else is "the new dark web." Therefore, trust us, and only us.

Nobody needs to plan this in a room. Just follow the incentives. That is EXACTLY why we are suddenly hearing so much about regulation, cybersecurity claims, watermarking and "pacing" progress. And strangely, all the labs seem to be friends on this one.

(To be clear: I spend a big part of my week on AI governance and trust. Rules are good. Rules written to lock the door behind you are something else.)

A very plausible scenario is that the frontier labs are trying to become the BANKS of the AI market. Regulated, trusted, expensive, and very hard to compete with.

The other scenario is that AI becomes just the hardware market. You run a free model on your phone, on your laptop, on a small box in the office, and nobody pays a toll.

Why AI is not sticky

Here is the deeper problem for anyone who invested billions in closed models.

Compare AI with the previous tech winners. Social media, delivery apps, Amazon. They all have network effects: the more people use the platform, the more valuable it gets for everyone. Leaving means leaving your friends, your restaurants, your purchase history.

AI has none of that. You can switch models in an afternoon.

The labs know it, which is why they are all racing to build "memory." If the model remembers you, maybe you stay.

But memory already lives somewhere else. It lives on your laptop, in markdown files, in prompt libraries, in the folders your harness reads every morning. (I wrote about this in Letter 75: a company in 2030 might run on text files. We're getting there faster than I expected.)

Your context is portable. So there's no stickiness.

What I believe is invariant

That's the tech side. Now the part I care about most: what all of this does to people and companies.

On this, I see a few things that look to me invariably and indisputably true. They were true with GPT-4. They are true with GPT-6 Astra. I expect them to be true with whatever comes after.

1. Your brain is about to need a gym

There's no free lunch in the universe, only transformation.

Cars removed the daily demand for walking. Our bodies paid for it, and a whole fitness industry was built to give us back artificially what the world used to demand for free.

AI just removed the daily demand for thinking.

Every time you offload a thought, you train your brain to expect the offload next time. First the hard memo. Then the easy email. Then the small decision you used to make while waiting for your coffee.

Many of those offloads are good. None of them are free. So you need to think deliberately, on purpose, the way you'd go to a gym. (Letter 86.)

2. AI is only as good as the human operating it

The best AI-generated software comes from the best engineers. The best AI-generated marketing comes from the best marketers.

The best AI-generated anything comes from the best anything.

A model is a prediction engine. Your words are the address. The right term, the right question, the right context pulls the model toward the part of its training where experts live. Without them, it drifts to the middle and hands you the average, with perfect formatting and total confidence.

And better models make this worse. The map gets bigger, the default gets more average, and the mistakes get harder to see. (Letters 70 and 99.)

So it is more and more important to be extremely good at what you DO. There's no way around it.

3. Your best people are now proofreaders, and we stopped making seniors

Making things got almost free. Owning them didn't move at all.

Somebody still has to read the code, understand the memo, defend the decision and pick up the phone when it breaks. So the bill moved from the people who make things to the people who have to own them. Your most expensive people now spend their weeks reviewing. (Letter 94.)

And here's the part that worries me more. The juniors who should become the next generation of reviewers aren't getting the reps. The "grunt work" was the gym where judgment got built: the four hundred boring contracts, the fifteen versions of the same model.

AI handed everyone a forklift and told them to skip leg day. You get the output today. You get the fragility in five to ten years, on the day the model is confidently wrong and the bench is empty. (Letter 85.)

4. You cannot automate a mess

AI doesn't natively have your context: your business, your clients, your exceptions, the rule the CFO changed last week in a Slack message nobody pinned.

Most companies run on organizational folklore. Unwritten habits in the heads of four people. When a human runs a messy process, the mess is contained, because people fill the gaps with judgment. When an agent runs it, the mess runs faster, at scale, with more damage.

If your context is a mess, AI can't help you. It doesn't matter how powerful it is.

The fix is almost always boring. Write down how the work actually runs. In Letter 78, that meant two weeks, twenty-two ugly pages, and the same agents going from chaos to near-zero errors.

The upgrade was in the document.

 

Look at those four again. They have the same shape.

Every jump in AI capability moves more of the value to the human side: the thinking you keep doing, the expertise you bring, the judgment you build and the context you write down.

If you want one thing to do on Monday, take everything you learned about AI this year and sort it into two piles. What expires with the next model (the tricks, the prompt templates, the tool-specific hacks). And what compounds (your expertise, your judgment, your written context). Then move your time to the second pile.

That's also the honest story of this letter. The tool reviews from year one are dead. Grok-3, o1 versus o3, the Midjourney tips: all gone, replaced a dozen times over. The letters about people are still true.

The next 100

Same rule as the first 100. I'll keep writing what I see implementing AI in the real world, and I'll keep telling you what I actually think, including when it's not what the timeline wants to hear.

One favor before you go. Hit reply and tell me which letter changed something in how you work. Or what you want me to dig into next. I read every single one.

Six people read the first one.

Thank you for being here for the hundredth.

Models expire. Judgment compounds.

AI is only as good as the human operating it.

Have a great weekend.

Stay sharp.

— Charafeddine (CM)