Memory is a moat.
Google goes agentic, Canva hits $3B in revenue and launches Canva code, plus a not-so-stealth model called Optimus Alpha (hint is in its initials)
Memory could be the moat in the AI race.
As someone who tries out every new model on market, I do find myself going back to ChatGPT for off-the-cuff questions — mostly because it already knows me. The context makes a huge difference.
OpenAI updated its memory feature this week (it can now retrieve context from all prior conversations with you), and it’ll be interesting to see in a few months if this leads to a “lock-in” effect for models, which are otherwise — commoditized.
Two things that would move the needle:
• Giving users more control over how memory is used (e.g., “don’t over-index on this detail”)
• Extending memory to the API, so builders can offer user-specific context in their own apps
Memory is core to a strong user experience that then, reduces churn. The best AI-native products will have to understand how to derive what their users care about, and use that to design a personalized experience.
I enjoyed signull’s prose on memory being the “last dataset” needed.
“Forget training data. This is training humanity”
I’d love your help! Please help me shape the future of The Strange Review.
Even as the world is on fire, AI companies keep shipping. It seems like this is the new norm.
This week, Google made some big moves in the agentic space. They released Agentspace, and an open Agent2Agent protocol similar to MCP, and the Agent Development Kit. They also prematurely releases Firebase Studio, an AI IDE which could have the potential to be a Cursor-killer, if executed well. (Alas right now it appears broken).
Here's why if Google shipped Firebase Studio right, it could pull ahead from the get-go:
- In-built cloud infrastructure: Firebase Studio plugs directly into Google’s full cloud stack (hosting, database, functions), all from a browser-based IDE.
- Access to the most advanced coding model from Gemini, could leverage Gemini Deep Research
- Auth out of the box: You can spin up a full login flow with zero config or boilerplate. This was something several AI IDEs struggled with at the start.
- Built-in web scrapping: most models can do this, but with Google’s infrastructure behind it, I'm assuming this could be a real edge.
Today: the app Firebase Studio deployed was still rough around the edges, i.e. the UI it created sucked, it couldn't find images.
But if Google continues to ship, and turns on agents that can integrate and deploy across the Google ecosystem, they could be well-positioned to be the default OS for the AI-native web.
That’s a big if. I wished they paid more attention to product experience.
Creative tool heavyweight Canva announced two bigs: a whopping $3B in annual revenue and Canva Code, a way to create interactive components and visualizations without touching a single line of code.
Lastly, a stealth open source foundation model named Optimus Alpha is creating waves in the research community for its beating benchmarks in programming tasks.
The not so open secret is that it’s from OpenAI, which would make it a slightly annoying gimmicky leak, ngl.