Gege & Didi: Open-Source + Frontier AI for Small Business

Most planning conversations now bend toward the same two words: token cost. A year ago it was a footnote. Today it reads like an operating manual for anyone running a business on AI.

Most planning conversations I sit in now bend toward the same two words. Token cost.

A year ago that was a footnote, the thing the technical person muttered at the end of a call. Today it sits in the middle of the table. We talk about token hygiene the way a restaurant talks about food cost. What are we spending per task, where is it leaking, what runs hot.

There is a reason it moved to the center. Every serious build now carries an open-source component, and that component is no longer the cheap seat. It sits at near-frontier quality. Just behind, sometimes level.

Over the weekend, Gavin Baker put a frame around it on All-In. He admitted, on air, that GLM 5.2 had moved beliefs he held about the gap between open and closed. Then he said the line operators should keep. The future is composable models. Every enterprise will run what Andrej Karpathy called a council of LLMs. A symphony, frontier models conducting, open source carrying most of the score. His rough split: frontier tokens still take about 90% of the economic value, while open source already runs 80% or more of the tokens.

For a business owner, that is an operating manual.

You stopped compromising

The old trade was simple and painful. Cheap model, lower quality. Good model, higher bill. You picked your poison per task and lived with it.

That trade is gone. When the open model lands a hair behind the frontier one, you are no longer choosing between good and cheap. You are choosing roles.

So the choice changed shape. Last year you picked a brand and routed everything to it. Now you assign jobs. Some models are doers. Some are validators.

Here is the warmer version, and then I will drop it. In Chinese, gege is older brother, didi is younger brother. The younger one does the legwork. The older one checks the homework. Your open model is the younger brother, fast and tireless and cheap. Your frontier model is the older brother, slower, pricier, the one you trust to catch what the first one missed. And there is a small irony worth a smile. The strongest open model on the market right now is Chinese.

From here, plain terms. Frontier and open source.

What that buys a small business

Picture the solo wealth manager buried in document review. The open model reads every statement, pulls every figure, fills the file overnight. The frontier model checks the extraction before a single number reaches the client dashboard. The owner reviews the one thing that needs a human: which client to call first.

Picture the small agency drafting and translating at volume. The open model produces the first pass across forty documents by morning. The frontier model checks tone and accuracy against the standard the client demands. The human signs off on what goes out the door.

Picture the small firm running contracts. The open model does the first read against the template. The frontier model flags what breaks against precedent. The partner still owns the signature.

Eighteen months ago, none of this was affordable. You ran the expensive model on everything, or you ran nothing. Now the volume rides on the open model for almost nothing, and you spend real money only where judgment is on the line. The window that just opened for bespoke work got a second floor.

The seat you should fight for

Baker added a twist worth holding onto. One seat on the council should go to a truth-seeking model. Not for raw intelligence. Because as an owner, your most expensive failure is a council that agrees with you.

You have no board. No partner paid to tell you the plan is wrong. The forecast you do not want, the read on why the deal really stalled, the honest answer about the hire: that input is the one you cannot buy cheap anywhere else. So you build a seat for it. A model whose job is to hand you the truth, especially when the truth is the thing you were hoping to avoid.

A flatterer with a context window is worse than no model at all. It tells you what you already wanted to hear, faster.

The human did not leave the room

Orchestration used to mean agents checking each other. Peers, side by side, catching one another's mistakes. This setup is a hierarchy instead. The frontier model checks the open one. Validation moved up a level, from human to frontier model.

Which sounds like the human finally gets to step out. It is the opposite.

Karpathy has been blunt about it. The human is the bottleneck. Every rung we climb on this ladder moves the human up, never out. The open model does. The frontier model validates it on everything a metric can score. And you validate the frontier model on everything it cannot: the client read, the judgment call, the decision you will answer for when it lands wrong. The output still needs a human before it touches a client. That part has not changed. It just moved up the stairs.

That is the real shift. Not fewer humans. Humans higher up.

The open model does the work. The frontier model checks it. You decide what it means. And you sign.

At Exponential Partners, we build this for owner-run companies. The council, the split, the seat you keep for yourself.