This was one of the most consequential weeks in AI in a long time. We just saw three major labs ship flagship models in a single week. At the same time, Amazon committed a total of $13 billion to Anthropic. This is the largest bilateral AI infrastructure deal ever signed: a signal that AI has moved from a shiny new tool to a permanent utility like the power grid.

Here is what is inside this week:

OK, let's get into it!

Last week I mentioned Gemma 4, Google's latest open-weights model family, and a handful of you hit reply asking for more. So let's go a little deeper.

Gemma 4 is different from most model releases you'll hear about: it's not built for the cloud. It's built to run on your device, meaning your data never leaves your machine. For business owners handling client information, financial records, or anything sensitive, that's not a small thing.

The family comes in different sizes depending on what you're working with. The smaller versions are designed for laptops and phones. The larger ones run on a single consumer GPU or a standard workstation. You're not buying server infrastructure to use this. You're running it on hardware you likely already have.

Here's what's been interesting to watch in practice.

The model handles text and images natively, and gives you real control over how it processes visual input. If you need speed, you dial down the visual detail. If you need accuracy on something like reading a dense document or parsing a chart, you dial it up. 

Gemma 4 is also designed to call functions and tools, not just generate text. That means it can fit into an agentic workflow where it's actually doing things, pulling data, triggering actions, completing steps, rather than just answering questions.

Gemma 4 is not going to outperform Claude or ChatGPT on complex reasoning tasks. That's not what it's for. But if you're building a workflow that handles sensitive data, runs repetitive processes, or needs to operate offline, a well-configured Gemma 4 setup can punch significantly above its weight.

I’ve installed their smallest model (Gemma e2b) on my iPhone 17 and I'm continuing to test this and will share a more hands-on breakdown once I have more to show. If you're already experimenting with local AI deployments or curious about where to start, hit reply. I want to hear what you're working on.

A few clients asked me what I use for meeting notes. The answer right now is Granola.

It sits quietly on your computer, no bot joining the call and just transcribes everything in the background. When the meeting ends you get clean, structured notes with action items already pulled out.

What sets it apart from most note-taking tools is the notepad. During the call, you have a live writing space where you can jot down your own thoughts, reactions, and anything you want to flag. When the meeting ends, Granola takes what you wrote and weaves it together with the full transcript into one clean set of notes. So it's not just an AI summary of what was said, it's your perspective combined with everything that happened which is a meaningful difference.

You also stay in control of when it runs. You can choose to have Granola transcribe a call or skip it entirely, which matters when you're jumping into a conversation that doesn't need to be recorded. Not every meeting should be captured and it's good that the tool respects that.

One thing to know before you roll Granola out to your whole team: by default, notes are shareable via link and AI training is opted in. 

Granola is free to start. Business plan is $14 a month. Available on Mac, Windows, and iPhone.

If you've started experimenting with AI agents, you've probably hit this moment: the agent does something impressive, and you wonder how much you can actually hand off to it. It's a good question, and honestly one most people are figuring out in real time.

Here's a framework I've been using with clients inside Claude Cowork that takes the guesswork out of it.

Every AI agent you deploy starts at 20% trust. That number goes up when the agent executes cleanly and anticipates what you need. It goes down when you have to repeat yourself, catch errors, or redo its work. The key is only giving the agent tasks that match its current trust level:

The part that makes this work is the nightly Agent Audit. At end of day, set up a Claude Schedule Task in Cowork with this prompt:

Every morning you get a clear picture of whether your agent earned more autonomy overnight and because it's updating its own instructions, it keeps improving without you rewriting prompts from scratch.

The goal isn't to keep your agents on a short leash forever. It's to build toward real autonomy through a process you can actually feel good about.

Before you go, I want to zoom out for a second.

What used to happen in a quarter is now happening in a week. Three flagship model releases, a $13 billion infrastructure deal, all in seven days. The pace is not slowing down.

Google Cloud Next is this week and Meta's LlamaCon is right around the corner on April 29. I'll be watching both and will have the key takeaways in the next edition.

One question to leave you with: of everything in this edition, what's the one thing you're actually going to try this week?

Cheers,

Julien

PS: If you know someone who would value the information I share each week please forward this to them, or send them a link to subscribe at www.ampra.ai/join-our-newsletter