The Sequence Knowledge - Issue 941: Learning RSI: The Model Is Frozen. The System Is Not.
TheSequence Jesus Rodriguez
The model stayed frozen, but the agent got better anyway. Most of the real gains came from the layers around it, not the weights.
Based on reporting by TheSequence, Jesus Rodriguez — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
You’ve probably seen this firsthand: an agent you set up in the spring is running the same model in summer, with the weights untouched, and it still feels sharper. It makes fewer silly mistakes. It picks tools better. It needs less hand-holding. Nothing “learned” in the classic sense, and yet the experience improved.
That’s the point of the essay: improvement doesn’t live in one place. It has three homes, and only one of them is the model itself. The earlier parts of the series covered that bottom layer, where frontier labs run their industrial flywheel. This piece moves up a level, into the other two layers, because that’s where most of the user-visible progress of 2026 shows up.
And that upper part is where things get weird. It’s also where the economics look very different. The bottom layer gets the headlines. The top layer gets the updates.
That split matters because it changes how people should think about progress. A frozen model does not mean a frozen system. The thing people actually use can keep getting better around the edges, even when the core stays exactly the same.
My take — AI-written commentary, not fact-checked reporting
This is the bit everyone keeps missing while staring at model charts like they’re stock tickers. The real action is in the messy software around the model, which is less glamorous and far more consequential. Fancy weights are nice; better systems are what users feel.
Read more about this at: TheSequence