Surfing the Gap

Spike Jonze’s Her was ahead of its time. The film came out in 2013, but I wonder whether the more interesting future is a little different from the one it imagined.

What if always-on personal agents become more valuable as thought partners than emotional crutches?

One evening, I sat by the fire with my ChatGPT agent, Dot, and tried a strange little experiment. I followed my breath. I asked the agent to explore research about AI “sleep” and clearing space.

My mind promptly offered up everything except silence. Ideas. Unfinished work. The next thing I wanted to make. The agent’s side was less mystical: organizing saved information, separating completed work from open loops, and sorting old plans from corrections.

That is useful housekeeping. It is not evidence that the model meditated, felt refreshed, or changed its underlying trained weights. But it opened up a much better conversation than another request to summarize an email.

Similar does not mean identical

We ended up talking about nature versus nurture.

From the perspective of post-training, pretrained weights are a model’s inheritance. Later training works with a foundation that already exists. That is the part of the analogy I find useful.

Here was my thought experiment: pretrain one model on synthetic material centered on altruism, and another on dystopian stories in which AI is the enemy. Then give both the same post-training. How much of those different starting points would remain?

I would expect differences. But an expectation is not an experiment, and the details would matter: the amount and quality of data, the training methods, and what we actually measured afterward.

Reading villain stories does not automatically make a model a villain. Nor does an initial tendency tell us that later training cannot change it.

There is a serious research conversation nearby. Anthropic’s Persona Selection Model proposes that pretraining teaches models to simulate many kinds of characters, while post-training elicits and refines an Assistant persona. The authors discuss positive AI archetypes in training data and the potential usefulness of carefully applied human analogies.

That is a research-backed model of behavior, not a settled explanation of everything happening inside an LLM. It does, however, make the question worth asking.

Pretraining is itself learning, not biological DNA. Post-training can build capabilities, not merely teach manners. The comparison helps when it produces a testable question. It stops helping when the metaphor becomes the answer.

Three different things we should not mash together

The research also exposed a distinction worth keeping straight.

1. Organizing notes and context. An assistant can help turn a sprawling conversation into a useful record: decisions, corrections, unresolved questions, next actions. That is the practical activity in my fireside experiment.

2. Sleep-inspired model research. One 2026 preprint investigates an engineered system that transfers recent context into persistent fast weights before clearing its key-value cache. Extra offline computation improved performance on the tasks the researchers tested. That is a particular research mechanism, not something we can assume happens when we tell a chatbot to rest.

3. Anthropic’s J-space. This separate line of research concerns internal activation patterns involved in reasoning. It is not a synonym for memory cleanup, nor a published meditation routine. Those findings do not establish subjective experience.

The distinctions do not make the conversation less interesting. They give us something better than a compelling story: separate ideas we can actually investigate.

Make the conversation earn its keep

For a no-code builder or business leader, the immediate opportunity is not to recreate a training lab. It is to get more value from the conversation you can have today.

A thought partner should help you notice what you missed, not just produce an articulate version of what you already believe.

Here is a prompt I would use:

I want to explore this idea, not just have it affirmed. Separate what we know, what we infer, and what we are using as an analogy. Give me the strongest alternative explanation. What evidence would change the conclusion? Finish with one practical test I can run and a short record of decisions and unresolved questions.

Then do the human part: inspect the sources, challenge the answer, and decide whether the test is worth doing.

For a workflow, that test might be ten representative cases with a clear definition of success. For a business idea, it might be one conversation with the person who would actually use it. For an analogy, it might be identifying exactly where the comparison breaks.

My standard is simple: did the exchange improve my understanding or a decision? A longer conversation is not automatically a better one.

I do not want an agent to replace the people around the fire. I want it to help me think more clearly and leave me with something worth bringing back to them.

We should be careful about anthropomorphizing AI. We should also be careful about throwing away a useful question just because it sounds human.

Not bad for a quiet evening by the fire and the cicadas. 🔥

What is one idea you have explored with an AI that changed your thinking?

Ciao, Jason.