Imprinting, revisited: what 2025's substrate makes possible

A couple of years back, when ChatGPT was four months old and GPT-4 was fresh, I worked through a framework for thinking about a category of tool that didn't exist yet. The shape was: take a person's expertise, encode not just what they know but the priors that shape how they reason, and ship the result as a portable artifact someone else could license and run locally.

That framework was speculative when it was written. Most of the technical preconditions weren't there. The economic model didn't exist. The legal questions hadn't been asked yet, let alone answered.

Three things have happened in the two years since that change the analysis substantially. Worth revisiting the framework with the new pieces in place.

The earlier sketch

The earlier work had two parts. The first was a thought experiment about whether expertise could be made portable now that conversational AI was a workable interface. The second was a separate piece arguing that the interesting layer of capture wasn't the facts a model could memorize but the priors that shaped how it reasoned about its domain. Two follow-on pieces walked the supply side and the engineering stack for the same problem.

Both core pieces were essentially a hypothesis: if you could capture how a person thinks and serve it through a conversational interface, you'd have a new kind of artifact. Neither piece offered a clear path to building it. They were sketches, not blueprints.

What changed in the foundation

Adapter techniques became standard practice

In early 2023, fine-tuning a base model to adopt a particular voice or reasoning pattern required either full fine-tuning (expensive, lossy, hard to share) or prompt engineering (cheap, brittle, didn't really capture priors). Both approaches were limiting in different ways.

LoRA-style adapters changed the math. A LoRA adapter is small (megabytes, not gigabytes), trains quickly on consumer hardware, and is portable across deployments of the same base model. You can train one in an afternoon. You can ship it as a file. You can layer multiple adapters on a single base model and switch between them.

The infrastructure for "encode the way someone reasons as a small additive layer" exists in 2025. It barely existed in 2023.

Open-weights models reached frontier reasoning capability

DeepSeek R1, released in January, made open-weights reasoning available at sub-$1 per million tokens. The base model is on Hugging Face under MIT license. You can pull it down, run it locally, and adapt it without sending a single query to a hyperscaler.

This matters for the framework because the original sketch assumed personal AI would need to live at OpenAI or Anthropic, with all the licensing and data-residency questions that implies. With R1 (and with the Llama 3 family before it) that assumption is no longer required. The artifact can run on the licensee's hardware. The original author's adapter never has to leave the original author's control.

Personal AI as a category exists

In 2023, "personal AI" was Twitter speculation. By early 2025 it's product strategy at Apple, Meta, and several venture-backed startups, and an active research direction at Anthropic and Microsoft. The version-1 implementations are mostly preference-tuned chat assistants, not portable expertise artifacts. That's a different shape of product than the framework described. But the category exists in the public conversation in a way it didn't.

Where the framework holds up

The "style and process, not just facts" framing maps cleanly onto what the field is now calling adaptation. The interesting work is happening at the layer of how models reason, not what they recall. RAG handles the facts. Adapters handle the priors. The split that the original sketch proposed turns out to be a useful one.

Portability also held up. The most interesting open-source projects in this space (adapter registries, model hubs that support cross-base portability, evaluation harnesses for voice consistency) all assume that adaptations should be transportable across deployments. That assumption isn't universal yet, but it's where the energy is.

Where the framework is incomplete

Three things the original sketch got wrong or didn't anticipate.

The marketplace doesn't exist

The earlier sketch assumed there'd be a marketplace by now, a place where individual contributors could license their adaptations and earn from usage. There isn't. The closest thing is Hugging Face, which is a model registry, not a marketplace. License terms, royalty splits, attribution mechanics, none of that has materialized commercially.

The most plausible explanation is that the platforms with distribution (the closed-frontier shops) aren't motivated to build a marketplace where their users buy adaptations from third parties. The economic incentive is to keep value capture inside the platform. Whether regulatory pressure (the EU AI Act especially) eventually forces creator-attribution conversations onto the table is an open question. As of March 2025, it hasn't.

The ownership layer is undefined

The legal frameworks for "who owns an adaptation trained on my corpus" don't exist. The current legal arguments (the New York Times case, the various class actions about training-data scraping) are about base-model training. They don't address the layer the framework is actually about.

There's a real argument that copyright law as it exists today doesn't have the right shape to handle adaptation rights. Whether that gets resolved through new legislation, through case law, or through industry-standard licenses (something like Creative Commons for adapters) is uncertain. None of those paths is moving quickly.

The economics inverted

The earlier sketch assumed personal AI inference would be expensive enough to require monthly licensing: $5 to $20 per consumer per month, in the back-of-envelope. Inference economics moved faster than predicted. R1's pricing puts a full personal-AI stack well under a dollar per user per month at modest usage levels.

That changes the business-model space substantially. Subscription pricing at $5–$20/month is harder to justify when the underlying inference cost is closer to $0.50/month. The licensing model the framework assumed may need to be replaced with a different economic structure, usage-based royalties, one-time licenses, or some hybrid. The sketch didn't anticipate this.

What's testable now

The interesting question for the next year is whether the artifact the framework describes can actually be built end-to-end with off-the-shelf components. The pieces all exist in isolation. The integration doesn't.

A reasonable experiment:

1. Take a published author with a substantial archive (their permission, obviously).
2. Train a LoRA adapter on the corpus, optimizing for voice and reasoning patterns rather than fact recall.
3. Build a retrieval system over the source material for grounding.
4. Wrap it in a small interface that exposes the result as a query-able expertise artifact.
5. Have the author evaluate it: ask questions they haven't answered publicly, and judge whether the output reads as something they'd write.

The technical components for that experiment are all available. The hard part isn't the engineering. It's the evaluation. There's no good objective metric for "does this reason like the author would." That's a research problem worth working on.

The thing the framework got most right, in retrospect, was the importance of the "how someone reasons" layer. The thing it got most wrong was assuming the commercial scaffolding would arrive on a similar timeline as the technical scaffolding. Two years on, the gap between those is the most interesting open problem in this space.