What's AI by Louis-François Bouchard

  • Home
  • What's AI by Louis-François Bouchard

What's AI by Louis-François Bouchard Co-Founder @ Towards AI. Making AI accessible through free content, courses, trainings and consultation.

The Claude Code leak revealed something interesting.Good AI agents are not relying on one big context trick.They stack m...
18/06/2026

The Claude Code leak revealed something interesting.

Good AI agents are not relying on one big context trick.

They stack multiple compaction techniques on top of each other.

Things like:
filtering tool outputs,
trimming old context,
compressing past conversations,
removing useless logs,
routing smaller tasks to cheaper models first.

Individually, none of these techniques feel revolutionary.

But together, they make a huge difference in how long an agent stays reliable before the context becomes a mess.

I think a lot of people still focus too much on the model itself.

Meanwhile, a big part of building stable agents is just controlling what the model sees and what it doesn't.

We all know these models are trained on the internet. And the internet is full of spam, duplicates, toxic text, and pers...
17/06/2026

We all know these models are trained on the internet. And the internet is full of spam, duplicates, toxic text, and personal data nobody should be memorizing.

So how does any of that turn into a model that actually works?

It doesn't go in raw. The data runs through a pipeline first, stripping out one problem at a time.

Here's how raw web data becomes a clean dataset 👇

1. Deduplication: copies get removed so the model doesn't overlearn repeated text
2. Quality filtering: low-signal, low-effort text gets cut
3. Toxicity removal: harmful and offensive content is pulled out
4. PII scrubbing: names, emails, and other personal info are stripped so nothing gets memorized
5. Bias audit: the data is checked for overrepresented views and gaps

After this, you've got a clean dataset.

Running alongside it is one more source: synthetic data. When real examples are thin, extra ones are generated, validated, and merged into that same clean set.

A model's capability comes entirely from this data. Clean it well, and the model performs.

I broke down the full data layer in Neo Kim's "The System Design Newsletter" in plain English.

Link in the comments 👇

Loop engineering is not just automation with a new name.An automation is a fixed script.Step 1, step 2, step 3, the same...
16/06/2026

Loop engineering is not just automation with a new name.

An automation is a fixed script.
Step 1, step 2, step 3, the same every time.
One unexpected change and it breaks.

A loop has an agent inside.
It looks at the state, picks the next action, checks the result and decides what to do next.
Retry, roll back, or stop, on its own.

That difference is exactly why this is booming now!
Agents finally got good enough to be the decision-maker inside the loop.

It's why the people building Claude Code and Codex say their job today is writing loops.

Have you tried building your first loop yet?

Last week, I discovered something quite alarming.You can get the GPT or Claude API 90% cheaper.Same models, same code, y...
15/06/2026

Last week, I discovered something quite alarming.
You can get the GPT or Claude API 90% cheaper.

Same models, same code, you change one line, your base URL, and you’re paying a fraction of the official price.

Some are flipping a $20 plan into $400 of API usage and reselling it.

But researchers tested 400 of these dirt-cheap AI API services…

And the alarming part is what they discovered... had to make a video about it: https://youtu.be/sMvIMDWVLmQ

15/06/2026

Your 1 million token context window is lying to you.

The bigger you make the prompt, the worse the model gets at using it. It reads the start, reads the end, and skims the middle.

A longer prompt is not a better prompt. You pay more, wait longer, and usually get a worse answer.

This is why every serious AI agent in 2026 runs context compaction. I broke down 11 of these techniques and the exact order I run them in.

14/06/2026

When this blew up, everyone picked a side. Anthropic good, OpenAI bad.

People pictured a giant stomping on the little guy.

Here's the twist. Anthropic wasn't the little guy.

In the business world, more companies were already choosing Anthropic over OpenAI. Two years earlier it was the other way around. The "underdog" had quietly become the leader.

And the contract everyone was fighting over? 200 million dollars. For a company expecting to make around 18 billion this year, that's small.

Then it went to court, and this is the part most people stopped following.

One judge sided with Anthropic. She said the government punished them just for speaking up, and called it illegal.

A few weeks later, a higher court flipped it and let the ban stand while the case continues.

So one company got the deal. One got the public. One judge said the government broke the law. Another said let it run.

And it's still not over.

I cover a lot of these stories and the drama around AI and the tech world!

13/06/2026

Anthropic just shut down their two most powerful models, 76 hours after launching them.

Not because anything broke. Because the US government told them to.

And it raises a question we've never really had to answer before: who gets to decide when a model is too powerful to stay online, for the whole world?

Broke down what actually happened in the video.

The US government just made Anthropic suspend two of its own models.Fable 5 and Mythos 5, gone for everyone, everywhere....
13/06/2026

The US government just made Anthropic suspend two of its own models.

Fable 5 and Mythos 5, gone for everyone, everywhere. Just three days after launch.

The whole thing started over a "jailbreak."

But when Anthropic looked at it, the jailbreak was basically asking the model to read code and fix bugs. Something devs do every day.

Anthropic is calling it a misunderstanding and complying for now.

The government is treating these models as too risky to leave open.

The 8-step order I use for shipping AI agents in 2026:1. Filter noisy tool outputs2. Load tools only when needed3. Clean...
13/06/2026

The 8-step order I use for shipping AI agents in 2026:

1. Filter noisy tool outputs
2. Load tools only when needed
3. Clean cached history before reusing it
4. Compress long logs and terminal outputs
5. Store memory outside the context window
6. Compact manually around 40%
7. Add retrieval behind the system
8. Keep autocompact as the last resort

What I like about this order is that each step reduces pressure on the context window before the next layer gets added.

The result is not just lower token usage.

The agent stays much more stable during long sessions and keeps track of what it’s actually doing.

That’s becoming a very important skill in agent engineering now.

Not just building agents.

Managing their context properly.

12/06/2026

Yesterday I explained loop engineering. Today, how I actually use it.

Unlike the guys at Anthropic, I don't have infinite tokens.

So I start the loops myself, keep my skills in indexed markdown files, cap my spending, and never let an agent run when I'm not around.
Because at the end of the day, you're responsible for everything it does.

My full setup and the mistakes to avoid, in the video 👇

Address


Website

,

Alerts

Be the first to know and let us send you an email when What's AI by Louis-François Bouchard posts news and promotions. Your email address will not be used for any other purpose, and you can unsubscribe at any time.

Contact The Business

Send a message to What's AI by Louis-François Bouchard:

Shortcuts

Share