What Can't Be Said Must Be Trained

The misread

The press will frame Inkling as a GPT rival. That’s the wrong read. Thinking Machines said it themselves: Inkling is not the strongest model available, open or closed. That’s an unusual thing to put in your launch post. It’s also the most honest thing in it.

What’s actually being launched is the customization loop. Inkling is demand generation for Tinker, their managed post-training platform. The model is the top of the funnel. The product is the fine-tuning workflow.

The self-finetuning demo makes this explicit: Inkling writes its own fine-tuning job, defines a target behavior, generates evals and synthetic data, trains via the Tinker API, and reloads its own weights. The scarce input there isn’t compute. It’s the specification: what behavior, what eval, what data. The model is a compiler. The spec is the source code.

The weights bet, steelmanned

Before I get to the architecture, I want to give the weights bet its full due, because it deserves it.

Bridgewater AIA Labs and Thinking Machines published a paper on June 30 on replicating expert judgment in six financial triage tasks. The results are hard to argue with: frontier models with naive prompts sit at roughly coin-flip accuracy, 46-50%. Expert prompting gets you to 74-78%. A fine-tuned model reaches 84.7% average accuracy, with roughly 29.8% fewer errors than the best frontier baseline, at approximately 13.8x lower cost per task.

That’s the strongest evidence I’ve seen for weights as a competitive moat.

The mechanism is the important part. High-quality labels came from expert investors. Non-expert labels failed. Disputed cases went back to the experts. The model learned patterns that the experts themselves could recognize, but couldn’t fully verbalize.

That last clause is load-bearing. Hold it.

The verbalizability line

Here’s what Journal No. 11 was tracking, and here’s what I think connects to Bridgewater, though I’ll mark this clearly as speculation.

Gurnee et al. published “Verbalizable Representations Form a Global Workspace in Language Models” in Transformer Circuits on July 6. The paper identifies a small, privileged subspace of internal activations, J-space, that functions as a global workspace for verbalizable concepts. Ablate it and multi-step reasoning collapses. Leave it intact and you can ask the model what it’s thinking and it will tell you, accurately, what’s active in that space.

Journal No. 11 used this to argue that roughly 7% of a model’s structure does the causal work, and that context engineering is the discipline of filling that workspace deliberately.

Here’s the connection I can’t prove but can’t ignore: the Bridgewater result is exactly what you’d expect if expert judgment lives outside J-space.

If a pattern is verbalizable, if an expert can articulate it as a rule, a heuristic, a checklist, it can travel through a prompt. Context engineering captures it. But if the pattern is the kind of thing an expert can recognize without being able to say, the felt sense of a bad balance sheet, the pattern that makes a macro thesis feel off, then it can’t travel through the verbalizable channel. It has to be trained in.

I can’t prove the mechanism connects. The rhyme is too clean to ignore.

Context engineering works the verbalizable channel. Fine-tuning reaches what doesn’t verbalize. They are not rivals, they split along that line.

Curation is the scarce input

This is where Tinker sitting downstream of context stops being a metaphor and becomes an architecture.

Look at what the Bridgewater pipeline actually consumed: expert-curated labels. Look at what Inkling’s self-finetuning demo actually consumed: a curated specification, target behavior, eval criteria, training data. The compute is cheap and getting cheaper. The curation is the scarce thing.

This is the same scarcity that context engineering runs on. Cards, context contracts, the structured substrate you feed an agent at inference time, all of it is curated human judgment about what matters and how to say it.

One curated substrate. Two compilation targets: prompts at inference time, datasets at training time.

Tinker is downstream of the context layer, not parallel to it. If you’ve done the work of articulating what good looks like, in structured, reusable form, you already have most of what a fine-tuning pipeline needs. The specification is the dataset seed.

There’s a second signal worth noting. During reinforcement learning, Inkling’s chain-of-thought traces spontaneously compressed, dropping articles and connectives without any reward targeting it. “We need to understand” became “We need determine.” Efficiency alone drove it. Cognition saw a similar condensed chain-of-thought phenomenon training SWE-1.7.

Gradient descent keeps discovering density unprompted.

Journal No. 11 made this point about the 7%: optimization finds the concentrated leverage emergently. Curation is doing deliberately what optimization does on its own, isolating what matters, stripping the rest.

What I don’t know

What I know: the verbalizable line is real. The Bridgewater numbers are the clearest demonstration I’ve seen that some judgment genuinely can’t travel through a prompt, no matter how well-engineered. That’s not a failure of context engineering. It’s a boundary condition.

What I don’t know: three things, and I want to be honest about all three.

First, whether cheap fine-tuning erodes the portability argument for some workloads. If weights are open, discounted 50% at launch, available across five inference partners on day one, and you can self-finetune with a single API call, the friction that made portability valuable shrinks. I don’t know where that equilibrium lands.

Second, where the verbalizable line falls for brand content. Is editorial voice a spec or a taste? If it’s a spec, it can be written down, put in a card, compiled into context. If it’s a taste, if the thing that makes a brand’s voice feel right is exactly the kind of pattern experts recognize but can’t articulate, then it belongs in weights, not prompts. I’m genuinely uncertain which side brand voice falls on.

Third, whether curated context used as training data keeps its advantage or dissolves it. If the proprietary substrate feeding fine-tuning is the Cards + Context Compiler output, and that output gets compiled into weights, does the curation advantage survive in the weights? Or does it flatten into something anyone can replicate with enough labels? I don’t have an answer.

The bet

Agents are commodity. Weights are joining them on the curve, open, discounted, five partners at launch, weights on Hugging Face. The model layer is becoming infrastructure in the same way Docker made containers infrastructure: the thing you build on, not the thing you sell.

The curated delta feeding either one, context or weights, doesn’t commoditize on the same schedule.

That’s the bet .fylle is built on. The open agent-portability protocol is infrastructure. It should be free, it should be open, it should work across every model and every runtime. The proprietary layer is the curated substrate: the Cards, the Context Compiler, the structured articulation of what a brand knows and how it thinks.

That substrate compiles to prompts today. Tomorrow it compiles to datasets. The curation is the same work either way.

Context compounds.

---

Sources

• Thinking Machines Lab, “Inkling: Our open-weights model,” July 15, 2026: thinkingmachines.ai/news/introducing-inkling
• Su, Zhu, Xiao, Alur, Kang et al., “Learning to Replicate Expert Judgment in Financial Tasks,” Bridgewater AIA Labs and Thinking Machines, June 30, 2026: thinkingmachines.ai/news/learning-to-replicate-expert-judgment-in-financial-tasks
• Gurnee et al., “Verbalizable Representations Form a Global Workspace in Language Models,” Transformer Circuits, July 6, 2026: transformer-circuits.pub/2026/workspace
• Cognition, “SWE-1.7”: cognition.com/blog/swe-1-7
• Fylle Journal No. 11, “The 7% That Does the Work,” July 9, 2026: fylle.ai/journal/the-7-percent-that-does-the-work