My future proof AI stack

Google dropped Gemini 3.1 Pro this week and it's a big step up. The reasoning is noticeably better. It handles complex multi-step problems, coding tasks, and long-document analysis way more competently than the previous version. It's 7x cheaper than Claude Opus 4.6 via API with a 1M token context window (Claude's still at 200K for most users). Replit is already using it to power their new animation feature. If you're in the Google stack, this is the best model you've had access to.

Google launched Nano Banana 2 this week, built on Gemini 3.1 Flash Image. It gives you close to Nano Banana Pro quality, but way faster and up to 50% cheaper. And if you're using it inside the Gemini app, Google Search, Lens, or Google Flow, it's completely free with no usage caps.

Here's where NB2 really shines: product photography that is consistent, captures all the details & looks editorial instead of AI-generated, action shots that actually nail the physics of movement, and infographics and diagrams that pull from Google Search so the information and measurements are more accurate than the previous nanobanana pro version.

Justine Moore did an amazing breakdown testing NB Pro vs NB2 side by side. check it here.

This should probably be your go to image generation model now for most things :)

Anthropic had a wild week. THEY HAVE PRETTY MUCH SHIPPED OPENCLAW by adding these features to Claude Code and Cowork. Let's run through the releases.

OpenClaw is literally just a jailbroken Claude Code with schedule triggers, remote access from your phone, and persistent memory, plus a "heartbeat" that wakes the model every 15 minutes to do tasks. Claude is one upgrade away from shipping their own version.

Notion now has three layers of AI. Basic Notion AI. Personal Agent (the one you chat with). And now Custom Agents, which work FOR you. Set a trigger or schedule, and they run autonomously while you sleep or while you're pretending to listen to a meeting.

They connect to Slack, Gmail, Calendar, HubSpot, Linear, Stripe, and more.

My honest take: If you're a heavy Notion user who already runs your ops in there, this is a big unlock. Agents triaging your inbox, prepping your meetings, generating weekly reports - without you touching anything.

The trade off is ease of use for control. It's a super simple setup compared to something like OpenClaw, but you're giving up the customisability you get with a dedicated agent on your own machine. If Notion's already your operating system though, this slots right in.

Jack dorsey (twitter founder) announced he is laying off 4,000 staff today (half of his company).

Within 15 minutes of the announcement the share price was up 25%.

Every CEO & CFO wants to do the same thing, and the market has just told every company that replacing people with Al gets rewarded

I don't mean this in a fear mongering way... but this is your wake up call for what's to come. you need to learn Al. you are still early to it.

I've been building out an AI agent team for a global 3PL company (using openclaw) to run their entire marketing operation, and it's now functional.

It is absolutely insane.

As of right now, the agent creates all the ad creatives, publishes them live in the Meta ad account, sets the campaign structure, and manages budgets. Every morning at 7am the CEO gets a daily report via Telegram. Key metrics, leads, cost per lead, what's working, what's not, recommendations on what to kill and what to scale.

The team shoots all the content weekly from their warehouses and uploads it into a Dropbox folder that the agent has access to. It then uses that footage to create the ads itself.

The reporting is just the start though. It actively creates test variations based on what's working. It looks at which creatives, angles, and messaging are getting results and spins up new tests off the back of them.

It's also plugged into the sales team's Fathom meeting recordings. So it's listening to what objections prospects are throwing out on calls, and turning those objections into ad angles. Real sales conversations feeding the top of funnel in real time.

This is the stuff I get excited about. An AI system running a real marketing function for a real company, end to end.

THIS AI AGENT SHIT ISNT JUST THEORY HYPE ANYMORE. IT F*CKING FUNTIONS!!!!

Something I've been thinking about a lot recently. The future proof AI stack is… markdown files.

OpenClaw laid the foundations for this. Your entire agent is just markdown files in the backend telling it its role, the jobs it does, and how to do them. Then you substitute any AI model in. Your agent can run on Claude, ChatGPT, Gemini, doesn't matter. If Claude goes bankrupt tomorrow, you select ChatGPT instead and everything still runs off your markdown files. If OpenClaw goes to shit, another platform will pop up and you just move your files across. You're never locked in.

And this isn't just for OpenClaw. These same markdown files work in ChatGPT projects, Claude projects, Claude Code, Antigravity, anywhere. All the AI agents I'm building for this 3PL company are literally just markdown files. Markdown telling it our ad creative process. Markdown telling it how we do media buying. Markdown telling it how to use its tools. Your business context as a markdown. Your ideal customer profile as a markdown. Your product list with all your offers, their prices, how they interconnect, as a markdown.

The theme for 2026 is: build your markdown assets.