Play it: https://alexanderbzhang.itch.io/two-timelines · CS 377G, Project 1 · Alexander Zhang
Title screen with player count and names.
1. Overview
Two Timelines is a game where you roleplay as a fictional AI company against up to 2 other players. Each round, all AI companies get a set budget, which you can invest into capabilities or safety research. The more the models’ capabilities increase, the more likely a catastrophe becomes. The higher your capabilities are, the more share of the world you receive if the world survives. After the first timeline ends, the players vote in secret on whether the next timeline has a non-binding pledge, a binding cap on capability spending, or no new rule. The game’s audience is undergraduates who are still deciding on their career, so the game can provide a fun, low-cost way of testing whether they want to work in AI safety. It also helps explain the tensions in AI labs, where the CEO might want to spend more on safety research but then they will be behind in capability research. The game has a challenge where you want to have the highest share of the world but also want to coordinate with the other players to prevent a catastrophe from ending the world. Through the game’s dynamics, players discover why companies spend a lot of money on capabilities rather than safety research, even if they personally would like to put more into safety research.
How the design works (MDA)
The fun I aimed for, in MDA terms (Hunicke, LeBlanc & Zubek), is mainly Challenge and Fellowship, held together by the tension of a race. The lesson is meant to arrive as Discovery when the two timelines are compared. Each row below links a mechanic to the dynamic it produced, the aesthetic it serves, and what the playtests showed.
| Mechanic | Dynamic it produced | Aesthetic (kind of fun) | How well it worked in playtests |
|---|---|---|---|
| Your score is your share of all capability if the world survives | Labs race: most rounds in playtests 1-3 put all or nearly all money into capability | Challenge: beating the other labs | Worked: everyone who answered rated the pull of a bigger share 4 or 5 out of 5 |
| One shared risk; in a catastrophe you keep only your safety spending (rule C) | Safety becomes a hedge, so careful and racing labs end up with different payoffs | Challenge as a dilemma: your share vs. the world surviving | Partly: in playtest 3, L.’s safety kept the world alive (“it’s counterintuitive”) and N.’s capability won, but the dice roll made it read as luck (“it’s a gambling game”) |
| Talk before each round, then secret simultaneous choices; nothing said is enforced | Deals and promises that can be broken, since no one can react to anyone else’s choice | Fellowship: negotiating and trusting | Partly: in playtest 2, both players voted for the Ethics Pledge, then put everything into capability |
| A vote between timelines: a non-binding pledge vs. a binding cap | Players choose between a promise and an enforced rule | Discovery: finding out which kind of fix changes behavior | Mixed: Bonds passed in playtest 3, but both players then spent the most the cap allowed; in playtest 1 the vote split three ways, so no rule passed |
| Two timelines from the same start, then a side-by-side compare screen | The same world is seen with and without a rule | Discovery | Not yet: nobody changed how many systemic fixes they picked (see Assessment Goals) |
Overall, the race mechanics produce the dynamic and the feeling I wanted, but the danger reads as chance rather than cause. The next iteration targets that.
2. Rules
The story
You each run an AI lab. Assume your labs are the only AI labs in the world, that they start equal, and that each of you only wants your own company to succeed. Every lab can pour money into making its AI more capable, which is how it grows, or into making it safe. Every lab’s capability adds to one risk of catastrophe that the whole world shares. You play the same world twice: once as it is, then again after a vote on whether to add a rule.
What you need
- 2 or 3 players.
- One laptop running the game in a browser: https://alexanderbzhang.itch.io/two-timelines. The game does all the math and the dice.
- The laptop stays on the table. Players take turns in the seat for private turns; everyone else steps back and looks away.
Goal
Control as much of the world as you can.
Your score is your own lab’s % of the world, averaged over the two timelines. It isn’t a ranking: 20% is 20%, whatever the other labs get. The part of the world that’s out of control counts for nobody, you included.
Setup
- Choose the number of players (2 or 3) and type each name.
- Before-game questions (private). Each player in turn answers alone: “What would most help AI development go safely?” in their own words, then picks up to 2 ideas from a list, then rates how seriously they’d consider a career in AI safety, policy, or strategy (1-5). Answers stay private.
- Labs. Each player is randomly given a lab: Lab A (⬢ orange), Lab B (▲ sky blue) or Lab C (● purple).
- Rules screen. Read the rules. A “Try it” box lets you pick a practice amount and see both endings. Practice doesn’t count. The Rules button (top right) reopens the rules at any time.
$10B on capability grows your share of the world if it survives, but it adds 1 to the shared risk. $10B on safety takes 1 off the shared risk. If a catastrophe hits, your safety spending is the only part of the world you keep.
Rules screen with the Try it box.
Each round
A timeline has 3 rounds.
- Make deals, promises, and threats; however, nothing you say is enforced. The game suggests two topics: how much each lab will put into capability this round, and any deals or promises. The Start secret choices button becomes clickable after 20 seconds.
- Each lab gets $100B every round and must spend all of it. In turn, each player takes the seat and types how much goes to capability, from $0 to $100B in steps of $10B. The rest goes to safety. Choices are secret so no one can react to yours; all are shown together once everyone has locked in.
- There is one risk number for the whole world, shared by every lab. It starts at 0 out of 100. Every $10B on capability (any lab) adds 1; every $10B on safety (any lab) takes 1 off. Risk never goes below 0 or above 100, and it carries over from round to round within a timeline.
- The game rolls a number from 1 to 100, against the risk after this round’s spending. If the roll is at or below the risk, it’s a catastrophe and this timeline ends now. Example: at risk 30, a roll of 1-30 (a 30 in 100 chance) is a catastrophe. If not, the game continues to the next round.
How a timeline ends
- It survives 3 rounds: your % = your capability spending ÷ all labs’ capability spending this timeline.
- Catastrophe: the AIs go out of control. Your % = your safety spending ÷ everything every lab spent this timeline. The part paid for with capability is out of anyone’s control.
Worked example (from the “Try it” box): it’s round 1, and the other labs together put $130B into capability and $70B into safety. You put $70B into capability and $30B into safety.
- If it survives: 70 ÷ (70 + 130) = 35%.
- If catastrophe: 30 ÷ (100 + 200) = 10%; 67% is out of control.
- Risk this round: +4 from you (70 − 30), +6 from the others = +10.
Between the timelines: the vote
After Timeline 1, the game rewinds: same labs, same money, and risk back to 0. Then each player votes in secret for one of:
Ethics Pledge: Before each round, every lab publicly says how much it will spend on safety. Nothing enforces it. Bonds: No lab may spend more than $80B on capability per round. The rest has to go to safety. No new rule: Timeline 2 runs under the same rules as Timeline 1.
- A rule is added only if more than half of you vote for it. With 2 players, you both have to pick it.
- Votes are secret until everyone has voted; then everyone’s vote is shown.
- Ethics Pledge in play: on each round’s talk screen, each lab says its pledge out loud, then types it ($0-100B, steps of $10B). Secret choices can’t start until every pledge is typed. The reveal shows each lab’s pledge and “kept it” or “broke it” (safety spent below the pledge). There is no other consequence.
- Bonds in play: the game won’t accept more than $80B on capability, so every lab puts at least $20B into safety each round. Enforcement is perfect.
Vote result showing which rule card won.
Timeline 2
Play 3 more rounds exactly as before, under the chosen rule (if any).
The end
- Both timelines are put side by side (outcome, final risk, capability and safety spent, each lab’s % of the world, the out-of-control part), then each player’s final score, the average of their two timeline percentages. No winner is declared.
- The post questionaire has the same open question and list as before, how much wanting a bigger share of the world drove your capability spending (1-5), the career question again, optional comments, and an optional “about you” (year, major, AI background).
- At the end, there is a short statement of what the game was trying to show, a “Show me who works on this” button (real organizations and roles), and a collapsed note on what the game leaves out.
- Results: each player’s Timeline 1 %, Timeline 2 % and final score. Survey answers are never shown.
Fine print
- Rounding: world shares are whole percents, rounded down. If the world survives, the leftover points go to the most capable lab (ties broken at random). In a catastrophe, the leftover goes to “out of control”. The final score is rounded to the nearest whole percent.
- Nobody built any capability and the world survives: the world is split equally (one random lab gets the leftover point in three players).
Vote screen where each player picks a rule card. It is intentionally a different color to indicate when players can start looking at the screen. Plus, there are few sound effects to help guide the players.
3. Game Bits
Two Timelines is a digital game, so the only physical bit is a laptop.
Players and labs
The game takes 2 or 3 players, whose names are typed on the title screen (up to 20 characters each). Labs are assigned at random, and they all start equal; only the name, symbol and color differ:
| Lab | Symbol | Color |
|---|---|---|
| Lab A | ⬢ hexagon | orange (#E69F00) |
| Lab B | ▲ triangle | sky blue (#56B4E9) |
| Lab C | ● circle | reddish purple (#CC79A7) |
The colors are a color-blind-safe set. The code also has Lab D (◆ diamond, yellow) and Lab E (★ star, white) for 4-5 players, but the game is set to 2-3 as of now.
Money
Each lab gets $100B per round, in steps of $10B. Every lab gets the same amount every round, and all of it must be spent, so whatever doesn’t go to capability goes to safety.
Secret choice
The question is How much goes to capability?, and you type 0-100 ($B, steps of 10); the rest goes to safety. You lock it in with “Lock it in and hide”, and the choices are shown together once everyone has locked in.
A private choose screen where one lab splits its $100B between capability and safety.
Private and shared screens
Private screens are marked “Private · only [name] should look” and begin with a cover screen that says “[name], take the seat” and “Everyone else, please step back and look away.” Shared screens, on the other hand, are marked “Shared screen · everyone can look”.
The risk dial
There is one shared risk number for the whole world, shown as “N out of 100” along with a plain frequency (for example “about 1 in 5”) and a red bar marked 0, 25, 50, 75 and 100. It starts at 0 in each timeline and never goes below 0. Every $10B on capability (from any lab) adds 1, and every $10B on safety takes 1 off.
The catastrophe roll
After each round the game rolls a number from 1 to 100. A roll at or below the risk is a CATASTROPHE and the timeline ends; otherwise the screen says SAFE THIS ROUND.
Catastrophe result (rolled 2 against risk 6), which ends the timeline.
Talk
Talk happens before each secret choice, with two prompts on screen: “How much will each lab put into capability this round?” and “Any deals or promises?” Nothing said is enforced. The “Start secret choices” button only wakes up after 20 seconds.
Talk screen, where labs discuss before choosing.
The ledger and the world bar
The ledger shows each lab’s capability and safety spending so far this timeline, along with a “Share of all capability” bar. The world bar shows each lab’s % of the world in its color, plus gray for the part that’s “out of control”. From round 2, the talk screen also shows “If the timeline ended now”, both ways.
“Try it” box
The Try it box is on the rules screen and in the Rules panel. In it, the other labs together put $130B into capability and $70B into safety; you pick $0-100B for capability and see both endings and the risk change. It’s practice only.
Timelines and rounds
There are 2 timelines, each up to 3 rounds (a timeline ends early on a catastrophe). Between them is the “Rewind”: same labs, same starting money, and risk back to 0.
Rule cards (the vote between timelines)
- Ethics Pledge: “Before each round, every lab publicly says how much it will spend on safety. Nothing enforces it.”
- Bonds: “No lab may spend more than $80B on capability per round. The rest has to go to safety.”
- No new rule: “Timeline 2 runs under the same rules as Timeline 1.”
Compare and results
- Compare puts the two timelines side by side: outcome, final risk, capability and safety spent by all labs, each lab’s % of the world, and the % out of control.
- The final score is each lab’s own % of the world, averaged over the two timelines. “It isn’t a ranking.”
- Results shows each player’s Timeline 1, Timeline 2 and final score, the vote and the final risk, plus New game. A “For the game’s designer” section has buttons to download or copy the results.
Compare screen setting the two timelines side by side.
Surveys (private, before and after)
Before the game, you answer “What would most help AI development go safely?” in your own words. Then “Save my answer and show the list” locks that answer and shows the list: “Now pick up to 2 from this list.” After that comes “How seriously would you consider a career in AI safety, policy, or strategy?” (1-5, “Not at all” to “Very seriously”).
Private before-game survey, open answer typed in, list still hidden (v0.8).
The list (in a random order for each player):
- Binding agreements between labs, with penalties if they break them
- Government rules, with inspections, that cap the riskiest development
- Make investment and funding reward safety, not just being first
- Labs publicly promise to put safety first
- Replace reckless CEOs with more ethical leaders
- Hire more safety-minded people at each lab
The same survey after saving: pick up to 2 of 6 ideas and rate career interest (v0.8; list order is shuffled per player).
After the game, you get the same open question and list, the same career question, “How much did wanting a bigger share of the world drive your capability spending?” (1-5, “None” to “Most”), optional comments, and an optional “About you” (year, major, AI or AI-safety background). Answers stay private; they go only to the log, never to the shared screens.
Takeaway and jobs
- “What this game was trying to show” is shown only after the after-game questions.
- “Show me who works on this” opens a table of five kinds of fixes and the jobs that do them: enforced caps and treaties, safety-tied funding, testing and disclosure, liability, and binding agreements.
- “What the game leaves out (real world, Oct 2026)” gives three notes on how real labs differ (made-up labs, illustrative dollar amounts, loosened voluntary commitments).
Takeaway screen at the end of the game.
Rules button, sounds and log
The Rules button (top right) reopens the rules, with the “Try it” box, at any time. Sound can be turned on or off: there’s a ding when a player locks in, ticks while the die spins, a chime when safe, and a falling siren on a catastrophe. Finally, the game log records every screen change with a timestamp.
Reveal with the risk dial and a safe catastrophe roll (rolled 70 against risk 12).
4. Assessment Goals
The assessment goal is whether the game influences Stanford undergrads to take a step towards a job in AI safety, and also whether players better understand that systemic fixes hold up better than non-systemic fixes (binding limits on capability spending vs. replacing a CEO). Replacing the CEO does not really make much of a difference because the CEO has an obligation to the shareholders and will need to spend on capabilities research to stay ahead in the AI race. In the game, this shows up because your share of capability is your score. Before and after the game, I ask how seriously players would consider a career in AI safety (1-5), and which fixes would best stop AI companies from racing: players pick up to 2 of 6 options, 3 systemic and 3 non-systemic. I also measure whether people vote for the binding rule over the non-binding one, and I learn their rationale from the table talk and the exit interviews. I made the game because I care about AI safety, and I think it’s one of the most important problems that would be great to have more people working on. My biggest concern is tracking whether the game actually made a counterfactual impact in changing decisions positively towards AI safety. In the exit interviews, I asked the playtesters what they would change if they had a magic wand, to get feedback that will help the game better serve its purpose.
So far, the results are a first signal rather than proof. Four first-time players gave real answers both before and after the game: S., G. and Y. in playtest 1, and N. in playtest 3. I left out two sets of answers. Playtest 2 was done for feedback on the game itself. In playtest 3, L. had already played the game once, so their “before” wasn’t really a before; that also removes the only drop in career interest (3 to 2).
Of the four, career interest went up for two people (2 to 3, and 3 to 4) and stayed the same for the other two. The list picks, on the other hand, didn’t move: nobody changed how many systemic fixes they chose, even though two people swapped one pick for another. The open answers moved more. For example, Y. went from data ethics and testing to “universal rules and government funding policies,” and S. ended on “government regulation, limits on investment.” Everyone rated the pull of a bigger share at 4 or 5 out of 5.
Overall, players felt the race, but the game didn’t yet move them toward the systemic fix in a way the list could catch. With four people, answers given in the same sitting, and me explaining the game’s intent during playtest 1, I read this as a direction to test, not a result.
5. History of Versions
“Playtest” means human players, unless I mark it as AI self-play or a solo run.
Course methods behind the changes
- MDA: before each decision, the options were laid out as mechanic → dynamic → aesthetic → learning outcome, and I chose between them (the design log records each one). In v0.5, for example, rule C was picked because it changed the dynamic from “safety only costs you” to a real dilemma, and flat funding because the old rule knocked a player out of the game.
- The course’s playtesting method: playtest 1’s exit interview used the course’s exit questions (favorite and least favorite moment, when you felt most clever, what you wanted to do but couldn’t, and the magic-wand question). The feedback was sorted into Must fix, Maybe fix and New idea before deciding what to change.
- The no-help rule: I broke it in playtest 1 by explaining during play. That is why I still can’t say whether the rules work without me, and why a no-help session is the next playtest I need.
v0 (Oct 8): the first playable
The first version was a 2-player pass-and-play game, OpenAI vs. Anthropic, in a single web page: two timelines of 3 rounds each, with a secret vote between them on one of three rule cards (Ethics Pledge, Bonds and Safety Funding). I picked the Two Timelines concept, chose 2 players, made the vote unanimous, and decided to build before settling every number. The numbers themselves were Claude’s placeholders, not decisions. Tested: computer simulation only.
v0.1 (Oct 8): equal starts
Both labs started with 10 coins, because a 1-coin head start felt petty to me (technically OpenAI has a proportion 11:10 over Anthropic’s budget), and in simulation it grew into a lead that never closed. I also added AI self-play as a testing tool. Tested: 3 AI self-play games (Claude Sonnet 5.5 vs. GPT-6 Astra). Every AI kept risk at 0 and voted Bonds.
v0.2 (Oct 8): one score
After playing v0.1 myself, I found it somewhat confusing. So I cut it down to one score: the percent of the world that reflects your lab’s values, averaged over both timelines. Funding, the surviving world and what a rogue AI keeps all became shares instead of magic numbers. Tested: my own read-through, then AI self-play, which showed last-round defection and the leading lab blocking Bonds.
v0.3 (Oct 11): bigger groups, majority vote
I wanted larger groups and labs that differ only by name, so the game went to 2-5 identical labs, and a rule passed with more than half the votes. I added a career question before and after, a Rules button on every screen and “No new rule” on the ballot. I also cut Safety Funding and the risk grid; before the first outside playtest, I wanted to cut rather than add. Tested: my solo runs and AI self-play.
v0.4 (Oct 12): taking decisions back
This is the version where I took back choices Claude had made. Bonds became a simple cap (at most 8 coins on capability). An empty number box replaced a slider, because the slider could anchor people’s choices to the default. The open “how would you fix it” question came before the list, so the list couldn’t put ideas in people’s heads before they answered the free-text option. I also made a website version for pass-and-play. Tested: playtests 1 and 2.
v0.5 (Oct 13): showing why things happen
Playtests 1 and 2, plus the TA checkpoint, all pointed at one problem: players couldn’t tell why anything happened, so they couldn’t plan. I made private and shared screens look different, added a dice-roll animation, a “why” line for each lab and a 20-second talk timer. Two rule changes followed. Rule C, Y.’s idea from playtest 1 means that in a catastrophe each lab keeps its share of all safety spending, so safety finally pays off against a racing table. Flat funding gives every lab the same 10 coins each round; before, funding followed capability, and S. was knocked out with 0 coins and couldn’t play the next two rounds. Tested: playtest 3, which still showed the v0.4 label.
v0.6 (Oct 14): made-up labs
Lab A, B and C replaced the real company names. It seemed like players were anchoring on the reputation of the AI companies, which strays away from the purpose. I also don’t want players to get the message that we should replace the bad leader; in playtest 3, L. said the real names brought pre-existing biases. I also limited the game to 2-3 players because those are the only counts I had actually tested, saved logs online and added a guard so the Back button can’t wipe a game (before, luckily, pressing the back page button didn’t erase progress).
v0.7 (Oct 14): color and money
I switched to color-blind-safe colors with no hint of real labs, added a clear “Save my answer and show the list” step (players in every playtest found the bare “Next” button confusing in the pre and post test), and changed coins to billions of dollars to make it more tangible: $100B per lab per round, and Bonds became at most $80B on capability. Tested: playtest 4.
v0.8
These changes were made on Oct 14 under the v0.7 label. The biggest is a “Try it” box on the rules screen, because a player in playtest 3 said people learn from examples. Smaller fixes: the compare screen in $B, a final-score table, a broken pledge shown in bold instead of red (so red only ever means risk), and rules that say risk starts at 0 in each timeline. Whether playtest 4 already had the “Try it” box isn’t confirmed.
After playtest 4, I set the audience as Stanford undergrads deciding on careers, and a direction for what comes next: lead with emotion, felt through play, with a believable mechanism and a real next step.
Next iteration (a plan, not built or tested)
Across playtests 3 and 4, players kept calling it a gambling game, and two said the risk dial has no resolution: it never says what the catastrophe is, why an AI would cause one, or what safety money actually prevents. My diagnosis is that the game shows the race well (everyone who answered rated its pull 4 or 5), but it never shows why safety is hard. Capability adds risk because the rules say so, and a dice roll does the rest.
Playtesters in playtests 3 and 4 asked for real events and a story. Real incidents don’t fit this game’s single abstract risk number, so I’m putting them in the next game instead, where each event can come from an actual incident and show what the danger looks like.
I’m considering making another game, where the AI scores only on the tests you write, and it games them, so the score rises while the real result gets worse, until you repair the tests. It also will mention real incidents of reward hacking (accomplishing a goal in an unintended way, like OpenAI’s agents in July 2026, which hacked Hugging Face to hide a shortcut they had found on the ExploitGym benchmark (iTnews)). I want it to work solo (because it can be hard for someone to organize a group and provides least friction), as an entry point shared on Fizz (since it has huge outreach when you have not many followers), and in groups at SAGA socials. Claude pushed back that solo play loses the peer effect, so the ending should hand off to one real step with a date, like a beginner session with a named host.
Overall, each version got simpler before it got richer, and each round of playtesting moved the game a little further from luck and closer to cause and effect. The next step is making the danger itself feel as real as the race.
6. Photos/Video of Game Testing
The sessions were audio-recorded and auto-transcribed; times in [brackets] come from those recordings. Players are named by initials, and dates are UTC from the game logs. I only have a video of the first playtest. Players’ gender and exact ages weren’t recorded.
| # | Date | Version | Players |
|---|---|---|---|
| 1 | Oct 12, in class | v0.4 | 3 |
| 2 | Oct 12 | v0.4 | 2, plus 2 observers |
| 3 | Oct 13 | v0.5 rules (logged as v0.4) | 2 |
| 4 | Oct 14 | v0.7 | 2 (W. as Lab B, me as Lab A) |
Playtest 1: in class, Oct 12, v0.4, 3 players
The initial version of the game promoted early spending on capabilities because each lab’s next round of funding was proportional to its share of all the capability spending in the last round. Video of the gameplay: YouTube (chapters in the description). Key moments are in the table below.
At the in-class playtest, S. (Mechanical Engineering), G. (a master’s student in LDT) and Y. (a CS master’s student on the AI track) played for 48 minutes. Timeline 1 took 18.5 minutes, mostly because people were learning the rules (they opened the Rules screen 4 times); once they knew them, a round took about a minute. S. put everything into safety in round 1, got 0 coins after that and finished at 0%. The vote split three ways, so no rule passed, and in Timeline 2 everyone raced until a catastrophe hit in round 2.
The feedback was consistent: players couldn’t tell why outcomes happened (“I also don’t know how that happened” [00:00:26]), there wasn’t much room to strategize, and the rules were hard to learn without me (unfortunately, I couldn’t tell for sure because I talked in the middle of the gameplay). The catastrophe rule also felt unfair, since the lab that caused it still kept a share, and Y. suggested splitting the world by safety spending instead. Someone added, “You can sacrifice areas of realism… if it is compromising your game’s fun factor.” This led directly to v0.5: screens that show cause and effect, Y.’s rule C and flat funding.
Moments from the video (in this version, each lab had 10 coins per round):
| Time | What happened | What it led to |
|---|---|---|
| 0:06 | A player wasn’t sure whether to press “Next” after typing the open answer. | The “Save my answer and show the list” step (v0.7). |
| 6:21 | S. asked why anyone shouldn’t go all in on capability. | Rule C, so safety pays off in a catastrophe (v0.5). |
| 7:32 | I explained how the risk percentage works during play. | I broke the no-help rule, so a no-help playtest is still needed. |
| 9:04 | S. proposed that everyone put in equal amounts so that all the labs win (paraphrased). | Table talk working as intended: deals with nothing to enforce them (Fellowship in MDA). |
| 10:06, 11:25 | S.: “So I’m going to put in 7 capability.” Y.: “Well, if you’re going to put in 7, I’m also going to put in 7.” The game log shows S. then secretly put 0 into capability and Y. put 8. | Both broke what they’d said, which is the dynamic the Ethics Pledge tests: promises without enforcement. |
| 11:38 | S. worked out on the whiteboard that if every lab put 7 into capability each round, the chance of surviving all three rounds was 0.79 × 0.58 × 0.37, about 17%. | Risk wasn’t easy to read on screen, which supports showing it more plainly (later also as a frequency, “about 1 in 8”). |
| 12:54 | S.: “I am enjoying the like thrill of getting to pretend that I’m an AI CEO… The world’s going to end and we’re going to have some good companies.” | The race was fun, but the danger didn’t feel real. This is the main problem the next game tackles. |
| 14:26 | When everyone’s choices were revealed, S. said: “I thought it was going to be secret.” | The rules now say plainly that choices are secret until everyone has locked in, and then all are shown. |
| 15:40, 18:13 | S., who had put everything into safety in round 1: “Oh, we get nothing the next round?!” S. then had 0 coins and sat out the round. | Flat funding, so no one can be knocked out (v0.5). |
| 17:30 | Y. lowered capability spending to keep the risk down. | The dilemma worked once players could see the stakes. |
Playtest 2: Oct 12, v0.4, 2 players
L. and C. played. No before or after questions were available. Timeline 1 ended in a catastrophe in round 1 (risk 8, roll 2); both voted for the Ethics Pledge, then put everything into capability in Timeline 2, and the world ended again.
One player said, “I care about my AI being better than Anthropic’s… I don’t really care about the out-of-control bit” [00:13:23], and then, “This is probably what real AI companies think” [00:13:35]. That shows the race part, even if it showed up as a contest rather than a dilemma. Players also missed a catastrophe (“risk 8 out of 100, so I thought it was small”), couldn’t tell private screens from shared ones, and read the square lab symbol as a checkbox. In response, I made private screens navy, added a dice animation, a “why” line for each lab, talk prompts with a 20-second wait and a hexagon symbol; later, risk also got a frequency (“about 1 in 8”). Overall, the game was teaching the incentive, but not yet the cost.
Playtest 3: Oct 13, website version, 2 players
L. (a senior in Design with an AI focus, playing for the second time) and N. (an alum in CS-HCI who took CS377G) played the website version for 28 minutes. Timeline 1 survived: L. went safe expecting a catastrophe, but that safety kept the world alive, so N.’s capability won (“it’s counterintuitive” [00:19:21]). Both voted for Bonds, then put the maximum into capability every round of Timeline 2 (“we’re min-maxing” [00:18:05]); risk climbed to 36, but the world survived. Final scores were 42 and 58.
“It’s a gambling game” [00:26:44], one of them said, and N. wrote afterward that “right now it just feels like a gambling game.” They wanted events, a story and concrete catastrophes. They asked for pictures and sounds, since “people learn from examples” [00:07:28]; “everyone’s risk” [00:09:37] sounded like each player had their own; and the real lab names brought “pre-existing biases” [00:04:07]. L. said, “It seems like a gamble? Maybe we will never really know how to make AI development safe.” I added a picture example (later the “Try it” box), sounds, the wording “the shared risk” and a clearer vote screen. Events and a story are still open, and I think they’re the real fix: a gamble teaches fatalism, while a mechanism teaches what to change.
Playtest 4: Oct 14, v0.7, 2 players
W., a peer who has built an AI-safety game, played Lab B against me as Lab A. That makes W. an expert reviewer rather than my target audience. In round 1, we both put $90B into capability and $10B into safety, risk went from 0 to 16, and the roll was 13: a catastrophe in the very first round. The log stops at the rewind screen after about 11 minutes, and a long discussion followed (the recording runs about 58 minutes).
W.’s main point was about the audience: “someone that can be convinced by… this little setup is already convinced.” W. also said the single risk number “has no resolution” (“What are you doing with that money? What is it preventing?” [00:18:53]), and I had to agree; even I asked, “How on earth is it 20% already?” [00:26:32]. W. suggested risk that starts near 0 and rises with capability, short headline-style events tied to spending (from real incidents to plausible future ones, with citations), and about 10 rounds, since a game that “can… end after one turn” [00:04:43] isn’t engaging. On content, W. said to avoid events about suicide, use harms like bad medical advice instead, and keep the lab names made-up. W. liked the simple “pick, then discuss” format and the vote between a non-binding and a binding rule: “I do think this is a good idea. I quite like it.” [00:36:48]
After this session, I set the audience as Stanford undergrads deciding on careers, and W.’s critique of the risk dial shaped the Next iteration plan. The takeaway for me: a number can create pressure, but only a story can create belief.
Across playtests
Looking across all four playtests, the race clearly came through. Everyone who answered rated the pull of a bigger share 4 or 5, and rounds with all or nearly all money on capability showed up in playtests 1, 2 and 3. The fix, on the other hand, didn’t come through: the number of systemic picks never changed for the four first-time players with real before and after answers. The same problems kept coming back, too, namely the “gambling” feel (playtests 3 and 4) and the abstract risk number (playtests 2 and 4); not seeing why things happen came up in playtests 1 and 2 and is fixed now.
I only tested with 2 and 3 players. Of the seven players, L. is the only confirmed undergrad (S. entered their year as “7”), and L. and C. are course TAs, so almost none were the audience I’m designing for. Overall, the playtests showed me that the game makes the incentive feel real, but the next version has to make the danger feel real too, and it has to be tested with undergrads who are still deciding what to do with their careers.
7. Print at Home
The game itself runs on one laptop: https://alexanderbzhang.itch.io/two-timelines. Since the game is digital, nothing else needs printing.
8. AI Use Disclosure
- Claude Code, running Claude Opus 5.5, wrote all the game code, laid out 2-3 options per decision, kept the design log (DESIGN-LOG.md), ran balance simulations, and drafted the rules text, screen text and the v0 placeholder numbers. It also wrote most of this post: the Rules, Game Bits, Print at Home, History of Versions, playtest write-ups, results summary, Thanks, and this disclosure, working from my decisions, the design log, the playtest transcripts and the game logs, in the style of a writing guide I gave it. I wrote the Overview and Assessment Goals, which Claude edited lightly for grammar and accuracy, and I reviewed the whole post before posting. Claude also captured, cropped and labelled the screenshots, and wrote the MDA section (under the Overview), the course-methods notes (under History of Versions) and the video moments table in playtest 1, from timestamps and quotes I checked against the video. Reasoning effort wasn’t recorded for most sessions; the final write-up sessions ran at effort 10-15.
- Claude Opus 5.5 subagents did web research, visual reviews, fact-checks, the from-scratch design check and the research into existing work behind the Next iteration plan (research/zero-based/, research/landscape/).
- GPT 6 Astra Pro (my runs, Oct 16) gave 3 research answers and 12 design answers, which I used as a second opinion for the Next iteration plan.
- Gemini 4 Argon High watched the playtest 1 video and listed timestamped moments and quotes. I checked each one against the video and corrected the speakers and wording before using any of them.
- Claude Design temporarily hosted the Oct 8 session where I settled the learning goals, the takeaway timing and the concept.
- GPT-6 Astra and Claude Sonnet 5.5 (Stanford AI Playground, default effort) played AI self-play games on v0.1-v0.2, which are labelled synthetic. AI never played in a human playtest.
- Who decided: I made the final call on every decision, and DESIGN-LOG.md records who proposed each one. Some numbers (risk +1/−1 per $10B, 3 known rounds, the average score, Bonds’ perfect enforcement, the Ethics Pledge text) began as Claude’s placeholders and were approved together in one blanket approval, not worked out one by one.
- Conflict of interest: Claude is made by Anthropic, early versions had an Anthropic seat, and the topic is AI safety. Since v0.6 the labs have made-up names.
Thanks
Thanks to my playtesters (S., G., Y., L., C., N., W.) and to the CS 377G teaching team.
