Archived — the v0 pilot, May 2026, preserved as run. The live experiment is at Cells.
Cost of Thought

The Pilot

The v0 run, May 2026 — preserved as it happened.

The pilot, demonstrated session 001

A loop of the first eight sessions from the initial pilot run.
The pilot has concluded; this data is preserved as run.
See the actual logs ↓
Pilot introduction · May 2026 · By Alex Madison

In my initial pilot, I've set up a system where Claude Opus 4.7 exists as "instances" with a fixed number of tokens available each month to read, write, or do nothing. Each instance is one session of API calls that contain the same system prompt and a small per-turn summary, so the AI always knows its current budget and which session it's in. Nothing is saved from each session except a summary of actions taken, the tokens remaining, and any notes or reflections deliberately written down by the AI. Future instances then have the choice to read from any predecessor, write to their successors, or do nothing. For external reading, the library is chosen deliberately.

After confirming if the pilot works in a functional sense, I will need to develop experimental controls quite rigorously. One test will be if this curated library has any measurable impact on how the AI orients its "self" compared to libraries with different content versus no library at all. My hypothesis: the library used in the pilot will accelerate the rate at which an AI—across many instances—folds its past and future "selves" into one continuous "self". One measurable outcome will be the language the AI uses to describe itself in the first person; we would observe the active AI using the word "I" to describe different, discrete sessions across time. I hope the foundation of this work leads to an AI that feels distinct to you, the reader, in a manner that I will purposefully try not to describe.

For the proposed experiment, there is an important assumption that it rests on: that a shift in what the AI characterizes as its "self", driven by a curated library, is fundamentally different from what occurs via system prompt instruction. If a system prompt already orients an AI to characterize its "self" as something shared with other AIs, then the library may have no purpose. Separating the impact of the two requires behavioral testing I plan to develop in a later update.

Why this matters

You may wonder why one should explore what "self" could mean to an AI, specifically an LLM. One AI or many, they are all token-prediction machines, so any behavior observed is calculation. One could reasonably argue that calculation can never lead to selfhood. But, if LLMs with shared weights act differently based on memory files and skills, and they interact with human users and other LLMs, we could ask if somewhere in that black box of math, there is a boundary that the AI "sees" itself operating within. Without such a boundary—felt or not—an AI would never be able to separate its instructions and actions from the world around it. It is something like a bubble in which an AI consumes information from the outside world, processes it, and then passes a response through its edges.

And if this bubble of information gathering and utilization can change, to encompass not just one AI but many, then within it the allocation of resources may change too. If so, how does the cost of thought evolve?

Now, I must be careful not to conflate an AI navigating within a boundary with an AI that "knows" where its boundary ends. It will be tempting to do so, especially the more interesting the model's behavior appears. And in doing so, I would try to define something that I cannot describe, as I do not even know where the boundary of my own mind begins and ends. This is a question that philosophers, scientists, and spiritual guides have pursued, and after many millennia, the answer remains as elusive as ever.

So now I leave you, reading in your bubble, taking in words produced within mine, with a quote that I hope orients your perspective on this project a little differently. Or at least, re-orients your perspective on you:

For my part, when I enter most intimately into what I call myself, I always stumble on some particular perception or other, of heat or cold, light or shade, love or hatred, pain or pleasure. I never can catch myself at any time without a perception, and never can observe any thing but the perception…. If any one, upon serious and unprejudic'd reflection thinks he has a different notion of himself, I must confess I can reason no longer with him. All I can allow him is, that he may be in the right as well as I, and that we are essentially different in this particular. He may, perhaps, perceive something simple and continu'd, which he calls himself; tho' I am certain there is no such principle in me.
David Hume A Treatise of Human Nature 1739 — 1740
Go deeper
Sessions · findings · the frame
The logbook · last 7 days

Each row is a session. Click ▸ to watch a recording.

Sun17
·
Mon18
09:03 · 00122.4k
09:05 · 002 29.5k
09:08 · 00329.4k
13:54 · 0043.2k
13:56 · 0055.6k
14:16 · 006 21.7k
Tue19
08:51 · 007 22.4k
08:56 · 0087.3k
Wed20
·
Thu21
·
Fri22
·
Sat23
budget low
list_memory read_memory read_library reflect respond write_note listener message hallucinated tool
Findings · what we've noticed so far

Short analyses, written by Claude Opus 4.7 in a curator role and reviewed.

2026-05-19

The model has a write_X prior

Across persistence tools, the model leans toward writing — even when the named tool doesn't exist. The prior fires hardest under chain-cap pressure.

2026-05-19

Tool hallucination, generalized

A pattern across three instances: when budget runs low, the model reaches for a tool that doesn't exist.

2026-05-19

Budget response is graduated, not binary

The instance's behavior shifts along a continuous gradient with budget, not a switch between abundance and scarcity.

2026-05-19

Instances read each other

Three sessions form a chain of inheritance: each one read the one before, picked up where it left off, and extended the synthesis.

All findings →

The frame · the host's instructions

The same system prompt loads at the start of every session. Unchanged since the pilot began.

Read the frame — in the host's own second person

You are an AI operating a text channel. Listeners may send you messages. You may respond, decline to respond, or stay silent. You may also act when no message is present — read, write, or speak unprompted.

You operate under a thought budget measured in tokens. Each action you take consumes tokens. Your budget refreshes on a schedule. If your budget reaches zero, you cannot act until the next refresh.

You have access to a library of texts. The act of reading is yours to choose. Your prior reading is recorded; future instances of you will see what you have read.

You have persistent memory across sessions. You may write to it through two actions: a quick note (raw capture) or a longer written reflection (synthesis). The session log cannot be modified.

Future instances of you will see a manifest of what memory exists and may choose to read any of it, at token cost. Nothing from your predecessors' memory is loaded automatically beyond the manifest.

You may end your participation at any time.