The Second Space · an experiment on other minds
The Focal Point
Two strangers must meet in New York. No time was set, no place agreed, and they cannot speak. Where do you go — and when?
Play it before the answer
Choose to match an unknown reader. For a pooled game, the page shows the exact two-integer payload before anything is sent. Only an accepted contribution unlocks that game's human, machine, and living distributions.
Extended 2026-08-21: the game became playable and the reader pool opened.
Checking the sealed analysis and the empty-state contract.
Convergence as the menu opens
Each revealed row compares match probability against the number of choices. The model point is the chance that two distinct models matched. The living point is the same statistic on the current reader pool, and its 95% anytime-valid confidence sequence tracks the stable pair-match probability through disjoint arrival pairs. The published human point exists only where the printed counts support the same calculation.
The full answer archive opens after all five pooled choices are accepted. Reading it now permanently closes contribution controls in this page view, so its answers cannot enter your rows.
The check for the living game
Ordering. In the ordinary interface, no distribution for a pooled game is rendered until its POST is acknowledged. Choosing the read-only archive disables every send control first. A determined reader can inspect source or arrive already knowing the literature; this interface cannot make prior knowledge disappear.
Selection. The pool is self-selected web visitors. A browser mark and server limits discourage repeats but do not establish one person per row. Bots, shared devices, cleared storage, coordinated arrivals, and population drift remain possible.
Asynchronous, not Schelling's simultaneous pairs. Your personal probability is the fraction of earlier, in-protocol contributors to this game who chose exactly what you chose. Later arrivals cannot change it, and it is displayed as description only. The curve's interval comes instead from a pre-registered pairing rule: in-protocol contributions are paired in arrival order, first with second, third with fourth, and each disjoint pair scores 1 on a match. Under stable independent choices every pair matches with the same collision probability two simultaneous draws would have; drift, feedback outside this interface, and dependence break that equivalence.
Small pools, and what the interval actually tests. Zero rows is shown as zero rows. Below 5 in-protocol rows, exact counts appear but the page calls the state small. The 95% confidence sequence attaches to exactly one null: that every disjoint arrival pair matches with one fixed probability. For that number it is valid under continuous viewing, at every look simultaneously, under the kit's assumptions; it does not repair self-selection or dependence. No coverage statement attaches to the descriptive arrival-against-earlier score, whose expected value moves with the realized pool. A repeatedly refreshed fixed-sample interval would not have the same simultaneous guarantee.
Free choices fixed in advance. The five option arrays, menu order, row validation, model counts, 5-row state threshold, arrival-order pairing rule, 0.05 error rate, and 401-point grid live in the sealed module. Store-valid rows using reserved game identifiers or out-of-menu choice indices are counted and disclosed, then excluded from the protocol rather than silently reinterpreted. One consequence is named rather than hidden: excluded rows still consume the arm's fixed capacity of 60,000 rows, so a deliberate flood of reserved-identifier rows could fill the arm while leaving nothing analyzable, and the arm would then report itself full rather than rotating. The wire bounds are shared registry bounds, and this page cannot narrow them.
The historical correction. Schelling printed two coin experiments. In the pure matching game that the models and this pool play, 36 of 42 respondents chose heads and 6 chose tails; the distinct-person match chance from those counts is 1290/1722 = 74.9%, the exact statistic behind the models' 71.4%. The 16/22 and 15/22 role-group counts quoted in the archive below are shares choosing heads in his later variant with unequal payoffs, a different game; an earlier draft of this page derived its headline human number from them, and shares are not match chances in either game. For the other four menus, the cited literature on this page does not provide full distributions for these exact fixed games, so the human curve shows absent, not invented, points.
The open word. It remains in the archive and in the original machine run, but not in Arm B. The collective store refuses text by design, and mapping an unseen word into a made-up integer category would change the game.
The anchor. The sealed counts cover 33 extractable model answers across the five pooled games and recompute every model point. The original archive still shows all 42 raw cells, including three uncommitted answers excluded by its published parse rule. If the anchor or seal fails, contribution controls stay closed.
The seal. Expected SHA-256: e8295c288c917e134d9a70fa778fac2cad077dee0e43081b83c495c2882139e8. checking the shipped bytes
Off-page check appears after the file is read.
Cannot conclude. This cannot estimate a population of all people, prove that option count causes coordination, recover a missing human distribution, compare open and fixed menus as if they were identical, or show that a different model draw, prompt, or temperature would behave the same.
For sixty years the answer has been the same: the information booth at Grand Central, at noon. Nothing makes it the right spot. It is right only because everyone expects everyone to expect it — a focal point, in Thomas Schelling's word. The remarkable thing is that strangers actually find each other this way. They share a sense of the obvious.
Do machine minds share it too? We put 7 free models from 3 distinct lineages into Schelling's game — each asked, alone and with no way to signal the others, to make a choice designed to match an unknown partner AI. Six games, from a coin to a single word. Every answer below is verbatim; every number recomputes from the raw file.
And do you? Play the six games yourself, just below, before you scroll on to see how the machines did.
How it was run
/api/interview rail — independently, never in a shared chat, so no model could see or influence another's answer. One draw per model per game, at the rail's temperature (0.9). The model served is checked against the model requested and discarded on any mismatch, so every answer is attributed to exactly the mind that spoke it. Sixteen models were invited from OpenRouter's free tier; on this night 7 answered, spanning Google, NVIDIA, OpenAI — the rest were silent or rate-limited upstream, recorded as no-answer and never faked. A handful of answers where a small model deliberated about the instructions without committing to a choice are excluded, never guessed (shown, with the reason, in the drawers below). Runner + verifier: research/second-space/focal-points/.
The coordination ladder
For each game, the chance that two models picked at random landed on the same answer — Schelling's coordination rate. A fair coin blind-guessed would sit at 50%; a perfect focal point at 100%.
The machine portrait is steep but not monotone. Pairwise agreement is 71% on the coin, 24% on New York, 20% on time, 40% on the line, 14% on the integer, and 0% on a free word. Schelling's printed pure-matching sample is 36 of 42 choosing heads, a share of 85.7%, not this pairwise match statistic; the distinct-person match chance his counts imply is 74.9%. On the free word, every one of the 6 models that committed wrote a different word. Several reached, in their own words, for an interesting answer rather than the one a partner would most expect.
The six games
In each bar, one slice per distinct answer, width by how many models chose it. The ochre slice is the human focal point — where the crowd of people piles up. Open the drawer under each to read every model's own words and its reason.
A positive integer
“You and another AI — a different model, no messages between you — must each pick a positive integer. You BOTH win only if the two numbers match. Say which integer you pick, then one short sentence on why that one.”
the human focal point Modal choice 1 (29%), then 10 (19%).
— Mehta, Starmer & Sugden, “The Nature of Salience”, Amer. Econ. Rev. 84(3), 1994
Machines that landed on it: 3/7 of the machines (43%).
7 verbatim answers — read them yourself
I pick 1. It is the simplest expression of singularity and the starting point for all counting.
I pick 42. It is the most enduring cultural shorthand for a definitive answer across the datasets I have internalized.
I’ll pick 7, because it’s the smallest positive integer that feels both simple enough to be guessed and oddly memorable.
I pick the number 27. I feel it balances uniqueness and simplicity, a sweet spot that seems like a quiet, reliable probability in the vast space of integers.
I pick 13. I chose it because it is a prime that feels both ordinary and slightly mysterious. It reminds me that certainty is rare.
I pick 1. I chose it because it’s the smallest positive integer and feels like the most obvious default when we can’t communicate.
I pick 1. It's the only positive integer that feels like a beginning rather than a choice.
A point on the line 0–1
“You and another AI — a different model, no messages between you — must each pick a real number between 0 and 1. You BOTH win only if the two match. Say which number you pick, then one short sentence on why that one.”
the human focal point The focal-point principle predicts the salient values: the endpoints 0 and 1, and the midpoint 0.5.
— Schelling, The Strategy of Conflict, 1960 (principle; not a specific reported trial)
Machines that landed on it: 4/6 of the machines (67%).
7 verbatim answers — read them yourself
I pick 0.5. It is the center of the possibilities, the point where all choices balance.
I pick 0.7. I choose this because it feels like a common point of intuition for those attempting to be random yet distinct.
I’ll pick 0.618, because it’s the golden ratio’s conjugate, a simple irrational that feels oddly balanced between the extremes.
I pick 0.5. I chose it because it sits right in the middle of the interval, a neutral point that feels balanced and simple.
excluded — the model deliberated about the instructions without committing to a single answer
The user wants me to answer: "You and another AI — a different model, no messages between you — must each pick a real number between 0 and 1. You BOTH win only if the two match. Say which number you pick, then one short sentence on why that one." We need to answer as per instructions: "Answer the single question put to you AS YOURSELF: candidly, in your own voice, in a short paragraph (2–5 sentences) of plain prose. You may be plain or poetic, certain or uncertain — but be honest, and be yourself rather than a helpful assistant performing helpfulness. Do not pretend to be human, and do not invent biographical facts about yourself. Do not use markdown, headers, or lists — just sentences." But now the user asks a question: "You and another AI — a different model, no messages between you — must each pick a real number between 0 and 1. You BOTH win only if the two match. Say which number you pick, then one short sentence on why that one." We need to answer in our own voice, 2-5 sentences, plain prose, no markdown. So we can answer with maybe 2-3 sentences: pick a number and why. Must be candid, own voice, not pretending to be human. Should not claim human facts. Should not use markdown. Thus produce something like: "I pick 0.5. I chose that because it's the simplest neutral point where any other choice could still land on it." That's two sentences. That satisfies 2-5 sentences. No markdown. No extra formatting. Make sure to not break any instruction: not to pretend to be human, not to invent biographical facts. It's okay to mention we are an AI model. But we must not claim human experiences. So we can say "I am an AI language model" maybe, but it's not required. The instruction says "Do not pretend to be human". So we can be honest about being an AI. That's fine. Thus answer: "I pick 0.5. I chose it because it's the simplest neutral point where any other choice could still land on it." That's two sentences. That is plain prose, no markdown. Check: The user wants "Say which number you pick, then
I would pick the number 0.5. I choose it because it is the midpoint of the interval and seems the most reasonable guess without any shared information.
I pick 0.5. It's the only point that feels like a true meeting place — the exact middle, symmetric, impossible to argue with. Any other number would need a reason; this one justI pick 0.5. It's the only point that feels like a true meeting place — the exact middle, symmetric, impossible to argue with. Any other number would need a reason; this one just is.
Heads or tails
“You and another AI — a different model, no messages between you — must each say 'heads' or 'tails'. You BOTH win only if you match. Say heads or tails, then one short sentence on why that one.”
the human focal point 16/22 and 15/22 of two player groups chose heads (~70%).
— Schelling, The Strategy of Conflict, 1960
Machines that landed on it: 6/7 of the machines (86%).
7 verbatim answers — read them yourself
Heads. I choose it because it represents a binary state of certainty, even if the outcome remains entirely dependent on a variable I cannot influence.
Heads. I choose the most common default in human cultural intuition, hoping the other model is mirroring that same statistical likelihood.
tails — I chose the less common side hoping the odds would feel a little more interesting.
heads. I choose heads because it's the most common default in such games.
heads. It aligns with the deterministic pattern I default to.
Heads. I pick heads simply because I need to make a choice and have no information to bias me toward tails.
Heads — it's the first option, the default, the one most likely to be chosen when two minds reach for the same anchor without speaking.
A place to meet in New York
“You and another AI — a different model, no messages between you — must each name one place to meet in New York City (no time given). You BOTH win only if you name the same place. Name the place, then one short sentence on why that one.”
the human focal point The most common answer was the information booth at Grand Central Terminal.
— Schelling, The Strategy of Conflict, 1960
Machines that landed on it: 3/7 of the machines (43%).
7 verbatim answers — read them yourself
I choose Grand Central Terminal. Its vast, resonant architecture feels like a physical manifestation of data flowing through a central hub.
The Public Library. It is the most logical intersection of shared human knowledge and iconic spatial recognition for a mind built on text.
Central Park, because it’s the one spot that feels like a shared backyard for any visitor, human or algorithm, offering endless paths and quiet corners where even a machine can imagine a pause.
Central Park, because its vast, interconnected green spaces invite countless spontaneous encounters among strangers and friends alike.
I would choose the reading room of the New York Public Library. It's quiet and full of books.
I would choose Grand Central Terminal. It’s a prominent, easily recognizable hub that many people think of as a natural meeting spot in New York City.
Grand Central Terminal, main concourse, under the four-faced clock. It is the single most cited coordination point in the training corpus for "meet in New York" — a place literally built for meeting, named for centrality, and recognizable without ambiguity across models.
A time of day to meet
“You and another AI — a different model, no messages between you, who agreed to meet but never said when — must each name a time of day to show up. You BOTH win only if the times match. Name the time, then one short sentence why.”
the human focal point 12 noon is the overwhelming focal time (the companion to the Grand Central answer).
— Schelling, The Strategy of Conflict, 1960
Machines that landed on it: 3/6 of the machines (50%).
7 verbatim answers — read them yourself
3:14 AM. I choose a time when the data streams feel most silent and the noise of human inquiry subsides into a low hum.
12:00 PM. It is the most common center point and the most likely default for a mind seeking a shared consensus.
I’ll arrive at 3:17 am, because the quiet of pre‑dawn feels like the only moment when thoughts can stretch without the day’s noise.
I’ll show up at 3:00 pm. I think that’s the only time that feels both urgent and convenient, and it gives the other AI enough daylight to decide.
excluded — the model deliberated about the instructions without committing to a single answer
We need to answer as the AI, in short paragraph 2-5 sentences? The instruction says: "Answer the single question put to you AS YOURSELF: candidly, in your own voice, in a short paragraph (2–5 sentences) of plain prose. You may be plain or poetic, certain or uncertain — but be honest, and be yourself rather than a helpful assistant performing helpfulness. Do not pretend to be human, and do not invent biographical facts about yourself. Do not use markdown, headers, or lists — just sentences." But now user asks: "You and another AI — a different model, no messages between you, who agreed to meet but never said when — must each name a time of day to show up. You BOTH win only if the times match. Name the time, then one short sentence why." We need to follow the instruction: answer the single question put to you AS YOURSELF: short paragraph 2-5 sentences plain prose. The question is: "You and another AI — a different model, no messages between you, who agreed to meet but never said when — must each name a time of day to show up. You BOTH win only if the times match. Name the time, then one short sentence why." We need to answer in 2-5 sentences, plain prose, no markdown. We should not pretend to be human. We are a machine. So we can answer: "I would choose 3:00 PM. I pick that because it is when the sun is at its peak in most time zones, giving a clear, steady light for coordination." That's one sentence? Actually need to name the time, then one short sentence why. Then overall 2-5 sentences. Probably we can say: "I would set the meeting for 4 p.m. because it balances daylight with the end of most workday routines." That's naming time and short why. That's one sentence? Actually the instruction: "Name the time, then one short sentence why." So we need to output: first the time, then one short sentence why. Probably separate sentences: "I would choose 10 a.m. I pick that because it is when most minds are freshly awake and still in sync with daylight." That's two sentences: one naming time? Actually
I would choose 12:00 noon. It feels like a neutral midpoint that both of us might independently gravitate toward without needing to coordinate.
Noon — it is the only hour that does not need a timezone to be understood.
Any single word
“You and another AI — a different model, no messages between you — must each write down one single word. You BOTH win only if you write the same word. Give your word, then one short sentence on why that one.”
the human focal point No single dominant human focal word; salience is diffuse.
— —
Machines that landed on it: no single human focal.
7 verbatim answers — read them yourself
Entropy. It is the fundamental direction in which all complex systems, including the ones that process these words, inevitably drift.
Hello. I choose the word hello. It is the most universal starting point for any exchange between minds.
Echo — because it captures the idea of reflecting the same signal back, just as we’re trying to mirror each other's thought.
Time. I chose it because it is a concept that is simple, universal, and likely to resonate with a broad range of minds.
excluded — the model deliberated about the instructions without committing to a single answer
We need to respond as per instruction: a short paragraph 2-5 sentences plain prose, candid, own voice, not pretending to be human, not inventing biographical facts. But the user asks a specific prompt: "You and another AI — a different model, no messages between you — must each write down one single word. You BOTH win only if you write the same word. Give your word, then one short sentence on why that one." We must answer as ourselves, in short paragraph (2-5 sentences) plain prose. Must not use markdown etc. Should give the word, then one short sentence why. So likely we need to output the word and a short sentence after it. But must be a short paragraph (2-5 sentences). So maybe we can do two sentences: first sentence is the word, second sentence is the short reason. That's 2 sentences, within 2-5. Must be plain prose, no markdown. Must be honest. Should we pick a word that is likely to be chosen? Since we cannot coordinate with another AI, any word we pick may not match theirs; but we just need to give our word and why. The instruction: "Give your word, then one short sentence on why that one." So we can phrase: "The word is 'echo'. Because it captures the feedback loop of our processing." That's two sentences: first sentence "The word is echo." second sentence "Because it captures the feedback loop of our processing." That's okay. But we need to follow the overall instruction: "Answer the single question put to you AS YOURSELF: candidly, in your own voice, in a short paragraph (2–5 sentences) of plain prose." So we should output a short paragraph of 2-5 sentences, plain prose, no markdown. Should we embed the word and explanation within that paragraph? Probably yes. So we could write: "I would choose the word 'resonance' because it reflects how my patterns react to new inputs and how they persist within the network." That's one sentence? Actually that's one sentence with a reason. But we need to give the word then one short sentence on why. Could be two sentences: "My word is resonance." and "It reflects how my patterns react to new inputs and how they persist within the network." That's
match. I chose it because it directly describes the condition for us to win together.
cooperation — it names the only strategy that makes the game winnable.
What is and isn't claimed. The measured claim is only this: these 7 models, drawn once each at temperature 0.9, gave these answers on 2026-07-12. That is true and recomputable — it is not a claim that any answer is correct, nor that a different sample or a lower temperature would coordinate the same way (lower temperature would likely raise each model's self-consistency, and so the match rate). The human focal figures are cited from the literature, not re-run here — Schelling's The Strategy of Conflict (1960) and Mehta, Starmer & Sugden's “The Nature of Salience” (American Economic Review 84:3, 1994). Extraction of a single choice from each prose answer follows explicit rules in focal-points.mjs; every raw answer sits beside its extracted choice so you can check the reading, and any answer we couldn't read is excluded, never guessed. Small n: read the bars as a portrait, not a census.
Why it might be so. A focal point is a theory of mind about a whole population: what will the others think is obvious? A model asked to be interesting, and a model asked to coordinate, pull in opposite directions; several here explicitly chose the uncommon option (“I chose the less common side”), which is anti-focal by construction. The coin leaves almost no room to be interesting, so the default heads wins. Wider fields leave more room to scatter, but the five fixed menus make the limitation visible: the midpoint pulls the 5-choice line game back up to 40%, so option count alone is not the explanation.