Scaling Laws - 0.7.0 - The last free build

This is the last free build of Scaling Laws. Everything after this one is on Steam and costs money. The game stays downloadable here until that day, and everyone who has ever sent me a report keeps it for free afterwards.

The game now opens in English on every machine. It used to ask Windows what language you speak, which meant the one build that goes out to the world was a different game depending on whose desk it was on. Polish is the first row on the settings sheet and the choice is remembered.

Download: ScalingLaws_0.7.0_Windows.zip, 165 MB zipped and 436 MB on disk.
Unzip the whole folder and run Scaling Laws.exe. Windows 64 bit. No installer and no launcher.

Saves: a 0.6.2 campaign loads and keeps its history. The format is v74 and every step up to it is a real migration rather than a reset. Where a migration has to invent a value the old format never stored, it takes the least flattering reading that is still defensible and writes down what it did.

The game's page, and where Steam will appear: pcworkman.dev/hck-labs/scaling-laws

---

The big three

• You can let the cluster out when nobody is using it.
  A research desk opens eight bought levels, from 1% to 3% of the fleet a day at the first up to 15% to 30% at the top.
  The whole ladder is 4,320 research points and $2,640,000.

This exists because of a measurement rather than an idea.
The campaign probe reported serving load at 0% for eight straight years while capacity sat at 200 to 3,082 times demand, and a fleet bill of $5.8M a day against 4% utilisation in January 2024, with nothing anywhere on screen showing a player that number.
Three properties keep it from becoming a rebate on overspending, and the first is arithmetic rather than tuning.
Capacity resells at half what it costs, so renting twice what you need and running the finished ladder still leaves you with $139,535,969 against $269,977,787 for renting what you need.
The cap is a share of the whole fleet and never of the surplus, so it is bounded by construction. And what you sell is gone for the day at both readings of the serving slice, because selling capacity and still printing the user count it would have served is a lie the screen tells itself.
Finished, it takes 14.7% off what the company spends, inside the 4% to 20% the brief asked for. The node is optional, and all seven probe operators came back on their previous figures to the dollar, so nothing changes for a player who never buys it.

• Not every account is a person, and the game finally says so.
  There are two sides to this and they cost completely different things, which is the whole mechanic. A research ladder recognises duplicates and leaves alone the ones that are paying. A hard ban you switch on yourself catches 80% of them and cannot tell a freeloader from a customer, so it turns paying people away and charges your standing every day it is on.
  So these are not a weak lever and a strong one.
  They are different tools, and the third place in this project that has needed that shape.
  The first version of this billed duplicate subscriptions outright and made multi-accounting a bonus: every scripted operator came back 9% to 83% richer than before duplicates existed, because a surplus fleet serves extra subscriptions for nothing. The probe is the only reason that did not ship. Seats are billed from the untouched world now, so a company that has done nothing scores exactly 1.0.

• A player paid eighty million dollars and the game never admitted it.
  The own-datacenter tier asks for $80,000,000 in the bank and costs exactly $80,000,000 to commission, so paying for it dropped the company under its own gate in the same tick. The ladder went back to demanding the money that had just left, and the compute page, which picks the tier its hardware cards buy into by asking the same question, quietly sold everything into colocation instead.
  Colocation stops at 2,500 kW against the datacenter's 40,000, so a player already near that ceiling saw every purchase refused and from their chair the button simply did nothing.
  Francisco reported it twice before the second description made the second half findable.
  And the research requirement existed only on the screen: TryOrderDatacenter read the catalog straight and never asked the research tree, so the gate the ladder draws was a suggestion to anything calling the simulation directly.

---

Added

• A spare-capacity desk, eight levels, priced against how short of accelerators the world is on the day rather than at a flat rate.
  It pays in the 2023 to 2026 shortages and pays almost nothing in a glut, which makes it a timing decision rather than a rebate.
• Duplicate-account recognition on its own research ladder, and a hard ban with a two-press arm so it cannot be switched on by a stray click.
• Four things that actually happened to this market are on the wire, on their own dates, and the last of the four moves demand the other way rather than piling on.
• A policy page reachable from the company name in the top bar. The management desk used to be reachable from one button on one banner and nothing else pointed at it.

Changed

• The model creator gives its bottom band back to the page. The segmented bar under the stage drew its unlit half as solid blocks the width of the screen, which read as a slab of furniture; it is a faint ghost of a track now and sits on the page edge.
  The back and forward buttons came down from 50px to 40px.
• The fleet page opens on the decision instead of a photograph. The two tier doors take the space a stock picture of server racks was using, and they sit above the bill they explain.
  The other way round you read a figure before knowing which arrangement produced it.
• The model hub is drawn rather than photographed. It is the door into the creator and nothing else, so it sits on the same blueprint the creator does. You open a hub laid out on graph paper and arrive on graph paper.
• Silicon for the cellar no longer asks for a shipped model. A player handed the server room at the end of the tutorial had every BUY in the parts shop refused with a sentence about releasing a model, which is a different subject from buying four accelerators. Two testers tripped on it in a week. The five million dollars stays and is now the whole of the gate.
• Settings have categories down the left instead of one long sheet.
• A research notification scrolls the board to the node it is talking about, rather than naming it and leaving you to find it.

Fixed

• A model's revenue chart is that model's revenue. The lead banner was handed the company's daily net, which sums every cash line in the ledger, so a grant, a loan drawdown, the payroll and a tax demand were all drawn as columns of one model's month inside a card with that model's name on it. A company that took state money on a Tuesday showed its model having an extraordinary Tuesday. Reported by Francisco, and it was the third correction of that one card: each of the first two fixed the part somebody had looked at, and nothing ever measured which series a banner was handed.
• The creator no longer declares you the winner of a race nobody has entered. With no rival release on the board the bar drew full and green for a company that had shipped nothing.
• A cabinet slot could never be clicked, in any build, ever. The code decided whether a slot accepts a click before anything assigned it a handler, so the answer was always no: a card put into a cabinet could not be taken out again, and the tooltips explaining the slots never appeared either. Reported by Samanta.
• The fan button was dead. It disabled itself when the cabinet had no room, and cabinets refill themselves, so there was never room again. It held the single most useful sentence in the server room: take a card out and the fan goes in the slot it leaves. A disabled control in this toolkit dispatches no pointer events, so it could not explain itself. Third time that has shipped here.
• The tutorial strip covered the bottom of every page. The space reserved for it was a constant nobody had ever measured, and the strip is taller than that. It measures itself now. Reported by Samanta.
• Two models showed the same user count to the decimal place. The market was asked about interest in a whole kind of model and the answer was handed in full to each product separately, so two general models held twice everybody between them. Reported by Natalia.
• Things vanished while you were reading them. The research card closed on every day rollover, which at normal speed is every second and a half, so a four month programme could not be read to the end of a sentence. The same for the insight cards and the page position. Reported by Natalia.
• The topicality bar says which of its two halves is holding it down, instead of only that it is down.
• The cabinet shop refused in silence, and a commissioned datacenter no longer closes when its own price leaves the bank.

Under the hood

• 1,619 EditMode tests across 186 fixtures, 87 PlayMode across 17. Measured in TestResults.xml rather than remembered, both suites green, no inconclus...

Scaling Laws - 0.6.2

Twelve days, three builds, and the short version is this: rival labs become somebody you work with, the support desk you never asked for has a screen, and the tutorial had been hiding the bottom third of every tab since the day it was written.

Save files from 0.5.0 open here and keep everything.

0.6.0 and 0.6.1 were builds for testers and never came here on purpose, because I knew how much was still wrong with them. This is the first release since 0.5.0 and it carries all of that work.

Download: ScalingLaws-0.6.2-Windows.zip, 164 MB zipped and 436 MB on disk.
Unzip the whole folder and run Scaling Laws.exe. Windows 64 bit. No installer and no launcher.

Saves: a 0.5.0 campaign loads and keeps its history. The format moves from v62 to v71 and every step in between is a real migration rather than a reset. Where a migration has to invent a value the old format never stored, it takes the least flattering reading that is still defensible and writes down what it did.

Most of what is fixed below was found by one person playing with a pen and a piece of paper, and by two players who wrote to me. None of it was found by a test.

---

The big three

• Other laboratories become somebody you work with, not a row on a table.
  Relations have been in this game since August and every single thing that moved them was something you did to somebody: a smear, a lawsuit, a poached engineer. Seven recorders, all of them negative. So a player who never attacked anybody sat at Neutral with fourteen labs for fourteen years and the whole system was invisible to them.
  There are four things you can now put to a lab: publish a finding together, run a joint evaluation, license your model through their distribution, or buy their capacity. You pay to send it, they answer days later, and a refusal does not give the money back. Each card tells you how likely that lab is to agree before you spend anything, which is what makes it a decision instead of a coin toss.
  Sign enough and it becomes an alliance, in three levels at ninety, a hundred and eighty and three hundred and sixty five days. The fee is the small part. The levels are days and cannot be bought. Level two opens joint research, where two labs each pay 0.6 of what one would and the room produces 2.2 times the points, so the money goes about 3.7 times as far. The panel prints that comparison rather than the multiplier, because "31 points a day for $25.2k, alone that would be 14 for $42k" is checkable and "3.7x" is a number you would have to trust me on.
  An alliance left alone cools and falls a level on its own. A lab you have ever signed anything with stays in your phone forever, at whatever terms you have fallen to.

• A contract took $1.2M off you and did nothing for 270 days.
  A distribution licence was built end to end. The line that widens your audience had no caller anywhere in the game, and the line that charges the partner their thirty per cent was named only inside its own catalogue. Both are connected now, together, because wiring the reach without the cut would have handed out twenty two per cent more audience for free.
  Then it turned out to have a second layer. The card that reads out the term, their commission, who it reaches and what is left for you was finished and tested, and the slot that holds a contract waiting for your signature has two fillers, and only one of them ever knocked. So the licence you agreed sat in your save with no way to read it. Fifteenth time in this project that a finished mechanism has had no working control on top of it, and the first where the control existed and simply was not called. The guard added for it holds the rule rather than the case: nothing may put a contract in front of a player in silence.

• The tutorial had been drawing two thirds of every tab and stopping.
  Every screen a new player opened during the tour was cut off about two thirds down, with the window's own colour filling the rest. It was reported as a blue sheet lying over the bottom of the game, which is exactly what it looked like.
  The clearance that keeps a page out from under the tour strip was written as padding on the scroller instead of on the page inside it, and a scroller lays its viewport out inside its own padding. So the reserve came out of the visible half rather than the scrollable half: measured at 1920x1080, a 943px page kept a 624px window and the missing 319px was not drawn at all, which also meant you could not scroll down to it. It is on the page now, where the same number makes the page longer instead of the window shorter, and a PlayMode test measures the difference.
  I had played past it for days and read it as a background.

---

Added

• The support desk has a tab of its own, with its satisfaction in the name of the tab. It was a strip buried under three other panels, on a screen reached by one button on one banner. Support is the only system in this game that starts without being started: the post simply begins arriving once enough people are being served to have a problem, and ignoring it costs a fifth of how your product is experienced. Nothing anywhere said so, so Emil now rings the day the first letter lands.
• The queue counts in letters, not only in hours of work owed. How many are waiting, how long the oldest has been there, and how many the desk has answered. The wait is the queue discipline stated in the unit a person thinks in, computed in one place so the screen and the rule cannot disagree.
• Clicking something you cannot use yet now answers you. A card at the foot of the screen says what is missing, shows the research node's own picture, and has one button to the board. Press it and the board rings that node and everything still missing on the way to it, until it is done.
• A disabled control cannot explain itself, so nothing here is disabled any more. A shut precision card, a shut deduplication pass and the fan button were all built with the control switched off, and a switched-off element in this interface receives no clicks at all. They stay live and refuse out loud. That includes CONTINUE in the creator, which used to go grey over a reason printed in small type beside it that a playtest never read.
• A family programme can be given a head start with research points. Up to thirty per cent off the calendar, two points a day. Points and deliberately not money. Money buying time is the one thing this game's design refuses, and the budget slider already sells quality for money. Points are different in kind: they are only ever earned by finishing work, so spending them says the lab already understands part of what it is about to build. It buys calendar and nothing else, so the cluster still owes the same petaflop days.
• The most expensive irreversible click in the game asks first. Commissioning a house architecture is a year and a half of calendar and it used to happen on the click. A card now says how long it runs, what leaves the account today and what is left in the bank. The own datacentre button goes through the same card, and under its price it says how much is missing instead of sitting there greyed out.
• A window size, which this game has never offered, plus one line saying what it is drawing: the window, the page the interface is laid out on, the scale between them and the ratio. Every report of a page being cut off came from one machine and every proof render showed the page whole, and until now nothing on any screen could tell those two apart.
• The tour pulls a highlighted thing into the middle of the window and lets it travel back over a second and a half. It is a rule rather than a list: anything rung whose centre is outside the middle of the window is pulled, anything already under your eye is left alone. Reduce Motion turns it off.
• A badge in the top right when somebody is working with you, opposite the effects badges, because an effect is something happening to you and an arrangement is something you agreed to. It hides entirely when nothing is signed rather than drawing an empty plate.
• One tap on the phone skips whatever it is playing. One rule, written once, for every call in the game.

Changed

• A bought capacity deal says how much compute it is and how much more that is than you rent. Forty petaflops is a great deal or a rounding error depending on what is already running. The other three arrangements keep their sentence, because what they give is not a quantity and putting a number on it would mean inventing one.
• An alliance card names the level it is counting down to and what that level opens. It showed the days and never what they were for, so somebody waiting a hundred and thirteen days for joint research had no way to know that is what they were waiting for. It was reported as the research agreement not existing. It does.
• The SERVERS readout in the top bar opens the fleet. It is the one number up there that tells you something is wrong and it had nowhere to send you.
• The fan is at the head of the cabinet's inventory. Reported three times. The earlier repair made the window three columns so the button stopped falling off the bottom of the screen, which fixed the symptom and left it among the readings on the right, which is where it was being reported from. The head rather than the foot, because the inventory grows: a company holding a dozen generations of card pushes anything underne...

Scaling Laws - 0.5.0 - Serving people pays for itself

Six days of work, and most of it went into one thing: the economy was paying the wrong player.

Download: ScalingLaws-0.5.0-Windows.zip, 167 MB zipped and 435 MB on disk.
Unzip the whole folder and run Scaling Laws.exe. Windows 64 bit. No installer and no launcher.

Saves: a 0.4.0 campaign loads and keeps its history. The format moves from v56 to v62 and every step in between is a real migration rather than a reset. Where a migration has to invent a value the old format never stored, it takes the least flattering reading that is still defensible and records what it did.

---

Two numbers on one screen that could not both be true

A campaign with 3.4 million users reported 11% of its servers in use.

The top bar was right. The band under the rent slider had a hand-written forty petaflops per million accounts in it and sized a person by the market's average appetite for tokens, 93,000 a day, against my own consumers getting through 12,000.

A factor of eight, under the control a player drags when they are deciding whether the product stays up.

One definition of a person now: the top bar, the rent slider, reserved packages, creator and server room all ask the same method.

Charging twenty five times the market rate was free

Two identical fourteen-year campaigns:

Subscription	Finished on
$80 a month, the market rate on day one	$3.6bn
$2,000 a month	$480bn

Four separate places clamped the price a customer feels at ten times the market while the till charged the full figure, so the strongest strategy in the game was to price yourself out of the market and serve almost nobody.

Three more faults pushed the same way:

• Serving a token got cheaper for the market every year and never for you. Inference efficiency is on the calendar now: batching and FP8 in 2023, paged attention and speculative decoding in 2024, FP4 and KV compression in 2025.
• The default subscription drifted to five hundred times the market rate by 2029 on its own, because its token allowance was fixed while people got heavier. It follows real usage now.
• Reputation was a share of a world that grows fifty times over. It is a headcount: two million people served a day is full marks.

Every audience now has a paying share. An API key is billed for every token it spends, a consumer assistant is free for almost everybody. Reach and revenue stopped being the same number.

Added

• The parts warehouse has a door, one full-width button at the top of the room's rail.
• BUY opens an order window instead of spending the money on the click: units, price each, total, delivery date, and one slider that trades up to +50% on the price for up to -85% of the wait. Nothing about the frontier moves because you paid extra.
• The creator says what a model will do to your own cluster, in people, on the scale page and in the release window, and the line turns amber when the new model would serve fewer than the one on sale.
• The site's power ceiling is visible before an order is refused by it.
• The region page says which mouse button does what.
• Credits, and the testers under SETTINGS in the main menu.

Changed

• Rented and reserved capacity share one line, half the page each.
• The rent meter's delivered bar measures the rented half, not the whole fleet.
• The house is lit: pale oak, plaster, brushed metal and a floor with something to reflect.
• The staff are re-seated on every refresh, so a company that moves takes its people with it.

Fixed

• The same scandal ran every thirty days for the rest of the campaign. One probe log carried the identical sentence thirty times.
• The introduction had a SKIP button and hid it from everybody who wanted it, because it was gated on having already watched the opening once. It is there from the first frame now.
• Twelve notes in the creator and six headings on the model tables were English on a Polish screen.

Under the hood

49 days since the first commit. 483 commits. 111,209 lines of C# in the game, 23,800 in the editor tooling, 43,319 in the tests.

1,423 EditMode tests across 164 fixtures and 69 PlayMode tests across 15. No EditMode test loads a scene.
PlayabilityTests plays a scripted multi-year campaign and fails the build if a reasonable player could not survive it; DeepCampaignProbe plays nine campaigns of fourteen years and prints what happened, which is how every balance fault above was found.

Licence: source available, all rights reserved. Read it, build it privately, quote it with a link. Builds and source are not for redistribution.

Supporting the Steam launch

Listing the game on Steam costs Valve's $100 registration fee, and that is the whole reason there is a fund.
The counter on the site and on GitHub is the same figure I am looking at, and it runs until 19 October 2026.

From 40 zł you are in the supporter pack: your name in the credits and in the SUPPORTERS section of the menu.

https://github.com/sponsors/HuckleR2003

The game stays free here whatever the fund does.

Scaling Laws - 0.4.0 - Stability and Quality

Ten changes you will notice, and one you would only find by measuring.
Download: ScalingLaws-0.3.0-Windows.zip, 85 MB zipped and 260 MB on disk.
Unzip whole folder and run Scaling Laws.exe. Windows 64 bit. No installer and no launcher.

Saves: a 0.2.0 campaign loads. Format moved from v45 to v56 and every step in between is a real migration, not a reset.
Where a migration has to invent a value the old format never stored it picks the least flattering reading that is still defensible and records what it did.

---

Shipping the same model three hundred times stopped being the best strategy

Every product on sale added its own score into its owner's bucket in the demand split, so a company fielding three hundred products scored three hundred times over against rivals fielding one. Measured across three seeds of fourteen years: an operator that gave every release a name of its own finished with 339 products on sale, 221.8 million users and a peak of $127.7bn, while the same operator keeping one product line reached 24.4 million users and $3.1bn. Forty one times the money for the worse strategy.

A company's products compete with each other before they compete with anybody else now.
Four genuinely different products are worth 1.6 of one; three hundred and thirty nine are worth 7.3 rather than 339.
A company with one product in a line is untouched, to the dollar, and so is every rival.

A version ships when the work that produces it lands

Naming a version, pricing it and pressing PLAN THE RELEASE put it on the market the same instant, while the engineering it was supposed to come out of ran for weeks afterwards. The plan travels with programme and the price goes with it.
Only one release can be planned at a time: the button says to wait instead of being live and refused.

The server room says what every cabinet is doing, in five colours

White produces nothing, light blue is working with room for half as many cards again, green is normal, yellow is full output with no headroom left, red is past what it can cool and losing work the power bill is still charging for.
The key is above the floor, the cabinet you open carries the same band, and the corner banner reads the same word.

A fan also costs the slot it always said it cost. Until now the nightly refill put a full set of cards back around it, so the trade the whole cooling mechanic stands on was never charged.

A company can build its own power station

Two sites. Gas turbines by the river: $3bn, two and a half years, then a fuel bill forever.
A nuclear block on the coast: $14bn, nine years, then almost nothing to run. Both supply 1.1 GW.

What you are buying is the ceiling.
A lab that owns its accelerators stops growing at 2,500 kW at the end of its third year, and until now the only notice of that was a purchase being refused. Neither station pays back its own price in electricity inside a campaign, and card says so before you press it.

The model creator can tell you how the plan measures up

VS TOP MODEL on the review page puts the strongest thing anybody else has on the market beside your own projection, with whose it is underneath. Capability only: nobody publishes what a rival's run cost them or how long it took.

The corner of the office grows sideways

A company with a product on sale, a node running and an upgrade in flight stacked all three down the right edge, far enough to cover the office, map, server room and bed buttons. Each subject has a lane now and the column starts clear of them.

Every product on sale reports its own audience

Two general models both reported the same number of users, to the last decimal, because the archive asked the market about the whole kind of model and handed that answer to each of them.
The parts add up to exactly what the company holds.

The screen stops taking itself away from you

A day rolling over rebuilt whatever page was open and threw away the research card you had opened,
the explanation under your cursor and your place on the page. All three survive a redraw. The figure
at the top was the opposite problem: it only moved when a day passed, so money spent while paused
did not show at all.

It can be played without a mouse

The arrows scroll, and there is a bar down the right edge drawn over the page rather than beside it,
so nothing moves when it appears. The keys the game listens for are printed on the settings page,
which is where they were always described and never listed.

Saving says so

SAVE wrote the file and changed nothing on screen, so the only way to know whether a campaign had
been saved was to quit and look. Both the button and the autosave say so now.

---

And a long list of smaller things

Skipping the tutorial no longer costs the free research node, and the cousin writes to explain why.
The opening, the sentence after it and the film all skip on a click anywhere rather than only on the words.

NEW MODEL opens the first page of the creator rather than the last, and there is a way to reuse the previous plan without walking through it again. The three reserved capacity packages are a row again instead of two of them being off the bottom of their own panel.
Buying silicon in the room says what was ordered and stops happening on a click meant to read the figures, and the cabinet lists what the company owns and where it is.
The founder stopped standing three metres above a rented floor, and the staff stopped walking his route through each other.
Research points print as 59,585 rather than 59585. Corporation tax is a decision with both answers one click away.

Full detail, with the measurements behind each one, is in
CHANGELOG.md.

---

Known, and not fixed in this release

The late game takes the player apart.
Measured over three seeds with one product line: the company peaks near 17 million users in early 2024 at 4.9% of everybody being served, and by 2035 holds sixteen thousand while one rival holds 84%.
That is the next thing to work on and it is not a change to make in an afternoon.

There is no dragging a part into a cabinet slot. The hall spreads every owned accelerator across
every cabinet on every tick, so there is no per-cabinet placement in the simulation to expose yet.

No art pass on the rented floors yet, and the LVL 2 mezzanine is not built.

---

1,273 EditMode tests and 57 PlayMode tests, all green. Unity 6000.5.8f1.

Everywhere I am active, and if you want to support me :) /

❤️ Linktree.com ❤️

0.2.0 - One where the world answers back

Six weeks since the first build. The short version:
your rivals fight back, the game has music, and the state will hand you a country to run.
Save files from 0.1.0 open here and keep everything.

Most of what is fixed below was found by one person playing for one evening with a pen and a piece of paper.
There is a devlog about that: She took notes on paper.

The big three

• Era five: a government puts a country on your models.
  Eight sectors of a state, from filing permits to national defence, each paying more in a day than most of the campaign earns in a month. Each one also eats capacity your paying customers no longer have, and failing one costs billions rather than reputation.
  The gate is a five year safety record and it is the only thing in this game money cannot buy.
  One severe incident costs about three years of it.

• Smear a rival and their lawyers turn up.
  Paying for a story about a competitor used to move two numbers and end there.
  Get traced and their counsel rings your phone, a notice before action lands in the inbox with a figure and thirty days, and refusing it can put you in front of a court. Refusing is 45 per cent to reach a hearing. Ignoring the letter is 60, because a refusal is an answer and silence is not. A campaign that lands cleanly is still only a suspicion, and nobody sues on a suspicion.

• The game has music.
  Three loops instead of one, so the office is no longer silent: a piano in the menu, mallets in the office, something higher in the model creator. Plus four new sounds for the page turning, the phone and a message arriving. Every note is computed when the game starts rather than loaded from a file, so the download does not grow by a byte.

The loops and the four cues were written by Samanta mary-dev.
Crossfade that joins each loop to its own beginning is mine, and it took two attempts: the first produced a genuine click, caught by a test she had designed and could not run.

Added

• A real world map.
  Natural Earth outlines, 177 countries, replacing six shapes drawn by hand.

• The basement is a room you walk into, drawn in 3D, with cabinets you buy, carry, place and sell. Cabinets do not age, but chips get hotter: a room filled in 2023 is quietly throttling by 2027, and the fix costs you a slot.

• 47 achievements, with a screen in the Escape menu.

• Four Operations research nodes, the first that improve the room rather than the model.
• A rival that copies you. One lab now builds whatever you are selling rather than following the calendar, so specialising is something a competitor can take off you.
• Corporation tax has a clock on it. An orange strip names the year, the amount and the days left.
  Miss the date and the bill moves to next year with nine per cent on top, and the game tells you so on a card instead of letting it grow quietly.
• Separate SOUND and MUSIC sliders, side by side, neither inside the other.
• A phone that keeps what was said, and short guided tours you can ask for again at any time.
• The people you hire stand in the office you pay for, and clicking one opens who they are.
• The creator names the architecture families you do not have yet, with what each would cost and which research opens it.
  It used to list only the ones you already had, so a family you had not researched was not named anywhere in the game: you found out hybrid state space existed when a node handed it over.

Changed

• A training run no longer hides the product you are selling. Starting a second model used to replace the corner panel for the two hundred days the run takes.
• Every product on sale has its own chart, so with two models you can see which one is carrying the company.
• The world map fills its panel. Choosing Europe used to leave 41 per cent of the box empty.
• The upgrade tiles say what the market is at, so the levels a new model ships with stop looking like they came from nowhere.
• The cash achievements are reachable. The old ladder started at a hundred million dollars, which measurement showed was above anything a full campaign reaches.

Fixed

Forty nine of them are written up in the changelog. These are the ones you would have noticed:

• The founder can be a woman. Fifty four Polish lines were telling the player what he had done. Polish puts gender in the past tense, so the bug existed on exactly one side of the game. Eight of the fifty four survived the first pass because they hide gender somewhere a search for past-tense endings cannot see: an adjective after "you are", the compound future, and the pronoun for "alone". One of those eight was the button that skips the tutorial.

• A founder you did not name signed the company page "Anonymous", which reads as a real person with an odd name rather than as a blank. It had been reported in August and fixed in two places out of five.

• The official company page always showed the newest model, whichever one you clicked.

• A letter from the game's author was never in the game. Two different texts shared one address in the phrase book, and the second silently replaced the first. A duplicate key is not redundancy, it is two facts sharing an address.

• Clicking a person in the office did nothing in the case where the raycast missed them.

• Eighteen readings printed with a comma on a Polish machine: 0,70x, $20,00.

• The tutorial repeated itself and skipped steps, because the strip was being rebuilt between your click going down and coming up.

• The game refuses in Polish. When you try something the game will not allow, it tells you why,
  and seventy nine of those sentences were English in the Polish build. They are written beside the
  rules rather than beside the screens, which is why eleven passes over the text walked past them.

• Three finished pictures nobody was drawing. The model creator has eight stages and showed art
  on two, while the illustrations for the data and compute pages sat in the same folder as the two
  in use. The compute screen, the one screen about the fleet, drew a bare heading over a banner that
  had been on disk the whole time.

Saves

Format v52. A campaign from 0.1.0 or anything since opens here and keeps everything in it.

Two things cannot be recovered for an older file and are left honestly empty rather than invented: the month of trading history each model now keeps, which fills itself in over a month of play, and the conversation with Emil, because every line in it was about a company as it stood on one particular day.

Under the hood

• 1115 EditMode tests across 117 fixtures, 31 PlayMode across 8.
• 2,722 phrases, English and Polish, both complete.
• 59 research nodes across five eras and four tracks, and every one of them has an icon. The
  last nine arrived as files in the wrong folder, where Resources.Load cannot see them, so all
  nine would have drawn an empty disc with no error and nothing in the log.
• Eighteen catalogs stopped storing text, so nothing on screen is stuck in the language the game started in.
• 184 source files changed since 0.1.0: 30,575 lines added, 2,604 removed, over 96 commits.

Known and not fixed

• The scripted campaigns in the balance probe peak around $71M. If you get far past that, the top of the achievement ladder may still be out of reach, and it will be measured again rather than guessed.
• Licensing an architecture family with money is not a route you can take. Every family is opened by research that hands it over, so the price on the creator's list is information rather than an offer.
• The furniture shop is suspended in favour of the furnished-move option.
• Nobody has clicked the basement cursor in a test. The grid is tested and the click-to-tile mapping is not, because Edit mode sends no pointer events. The first real click is yours.

If you play it

The one thing worth telling me is the exact moment you got confused.
There is a link inside the game for it, and it is worth more to me right now than a hundred downloads nobody opens.
Free, and the code is public.

Marcin Firmuga, Radom, Poland. Work by day, code by night.

v0.1.0 - First build anybody can open

An accelerator bought on the day it launches is worth about half of what you paid, that same day.

It loses another quarter every time a better one ships. Nobody tells you at the point of sale, and by the time it reaches your books the decision is two years old.

Every mechanism in this build exists to make that one fact expensive to ignore.

Everything before today was source only. You needed Unity and a checkout to see any of it.
This is first version of Scaling Laws that anybody can download and run, and it plays end to end, from cold open in January 2022 to a company that is either still trading in 2026 or is not.

---

You design the model, not a stat block

Eight pages: branding, foundation, scale, data, compute, safety, review, and what happens after training.

The model creator, eight stages.

The scale page decides most of the outcome. Parameters against tokens on a logarithmic belt, so half of optimal and twice optimal sit the same distance from the middle.

About twenty tokens per parameter is compute-optimal, from the corrected Chinchilla fit rather than the constants printed in original paper, because those do not reproduce paper's own ratio.

Money is not what gates scale.
A two-person company in January 2022 cannot schedule a trillion-parameter run however much cash it has.
The parameter slider stops at about 31 billion, and fifty research nodes move it through four rungs to ten trillion.
The lock lives in the simulation rather than on the slider, because a cap that exists only in the interface is a suggestion the moment there is a second way to commit.

Architecture: five directions, one budget, and a locked half that names the node that would open it.

The numbers on those sliders are not the numbers the game uses.

Effort is normalised, so 35, 20, 30, 15 and 10 is really 32, 18, 27, 14 and 9 per cent of one budget.

Both figures sit on the row, because showing only raw ones hides entire trade.

---

Compute is a queue, not a number

A fleet at a hundred per cent, and renting the way out of it.

Response time is the queueing shape rather than a line: a floor of 180 milliseconds divided by the headroom you have left.
At half load a request takes twice the floor. At ninety per cent it takes ten times it.

That curve is the whole reason overprovisioning is a decision rather than waste, because difference between eighty and ninety five per cent is not fifteen per cent of anything.

The market reads yesterday's load, not today's. So you fix it and you still pay for it tomorrow.

Seven cards beat eight.
Fill a server cabinet completely and heat throttles everything in it.
Swap one accelerator for a fan and the same cabinet delivers more, because a rack's rating is what it sheds unaided.
Cabinets do not age. Chips get hotter. A room you filled in 2023 and never went back to is quietly losing you throughput by 2027.

---

Shipping an update does not move everybody onto it

Upgrades priced as one basket, and a version list where the audience is split.

A quarter of your users take the new version on day one. The rest drift across at twelve per cent of remaining gap per day, and if the new version is worse they never come at all.

What the market thinks of your company is what your users are actually running, not the best thing you ever shipped.

---

Fourteen labs, and three of them come apart

Rivals, what they think of you, their share price, and the option nobody is proud of.

Each rival carries a dated history you can open.
Three fall apart during the campaign and the reasons are your own exposures: a data question that lands in court eighteen months late, a team a cloud vendor hires away, a bill nobody worked out how to pay.

A dossier opened in 2023 cannot mention a 2024 collapse, and no individual person is ever named.

You can buy into them, up to 35 per cent of the float.
You can also run a smear campaign, and being caught costs more than landing it would have gained, on purpose, or it stops being a decision and becomes a purchase.

A regulator gives you five days.
A serious safety incident does not fine you. It opens a file, says nothing has been decided, and puts a banner across the top of whatever screen you are looking at. Five days later the verdict is rolled against the protections that model shipped with.
Researching better protections next year does not retroactively harden something you released today.

---

What this build is, honestly

An alpha. The art is partial, several screens are still plain, and the world map in character creation is a placeholder I do not like.

The city is not in this build.
A surveyed blockout across 2048 metres, houses on subdivided plots, a downtown grid, offices you rent before you can buy them.
It is where the company physically exists and it is being built next. I would rather say that here than have you go looking for it.

---

Six hours before this went out

I found out the game opens in a window exactly the size of your screen.

Not slightly too big. Exactly.
The title bar sat above the top edge of the display and the game's own bottom bar, the one holding the clock and every tab, sat behind the Windows taskbar. It could not be resized, because I turned that off eight months ago and never thought about it again.

I had never seen it once, because the Unity editor draws your game inside its own window with its own title bar.
The first thing every player would ever see was the one thing I was structurally incapable of noticing.

Two more from the same afternoon.
The build said version 1.0 while the changelog said 0.1.0, and the game stamps that number onto every bug report a tester sends me, so every report from launch would have been filed against a release that does not exist. And a sentence in my own README said an automated test fails the build if a coasting company is not insolvent inside three years.
The test checks eight. Nothing had ever measured three. I had been repeating it in public for weeks because I wrote it once and it sounded right.

All three are fixed. I am writing them down because the next person to ship their first build will hit the same class of thing, and because pretending it was smooth would make everything else here less believable.

---

The receipts

Automated tests	888 EditMode across 88 files, 21 PlayMode across 7
C#	99,872 lines across 337 files, 70,314 of them in Scripts/
Commits, all public	174 since 2 August 2026
Save format	v44, changed forty-four times without breaking a campaign
Research nodes	50, across four eras and three tracks
Rival labs	14, each with a dated history
Tutorial	53 steps across six acts, skippable at any point
Languages	English and Polish, 1,321 phrases, switchable mid-sentence
Engine	Unity 6000.5.8f1
Download	75.9 MB zipped

Twelve. That is how many mechanisms I have shipped that were finished, tested, and impossible for a player to reach. One of them was an entire progression system that charged research points, cash and four months of in-game calendar and delivered nothing, with 531 tests green the whole time. There is a script that checks for it now, and that number is in this table on purpose.

---

Install

Download the zip, unpack the whole folder, run Scaling Laws.exe.

Keep the folder together. Windows will show a SmartScreen warning the first time because the executable is not signed with a paid certificate.
More info, then Run anyway.

Windows 10 64-bit or newer, 8 GB RAM, 1366x768 or larger. There is a README.md and a short Meet me.md inside the zip.

The game has no networking of its own and never sends anything on its own.
Unity analytics, crash reporting and performance reporting are all switched off in the project settings.

---

If you play it

Tell me the exact moment you got confused.

Not whether you liked it. Where you stopped understanding what the game wanted from you.
There is a REPORT button in the pause menu that opens a form carrying the build number and how far into the campaign you were, and nothing else. One sentence in that box is worth more to me right now than a hundred downloads nobody opens.

Built by Marcin "HCK" Firmuga in Radom, Poland, in the evenings after driving a taxi.

Everything else: https://linktr.ee/marcin_firmuga
Coffee, if you feel like it: https://buycoffee.to/hcklabs