Scaling Laws 0.3.0: I found the exploit in my own game by measuring it

Post news Report RSS Scaling Laws 0.3.0: I found the exploit in my own game by measuring it

The best way to play my AI lab tycoon was to ship the same model three hundred times under different names. It was worth 41x the money. 0.3.0 closes it, adds power stations, and colours every cabinet in the server room by what it is actually doing.

Posted by HuckleR on Sep 13th, 2026

You run a frontier AI lab, starting in January 2022, with $12M.

You rent compute before you can own it.
You pick a model size you can actually supervise rather than one you can afford. And decide whether three weeks of safety evaluation is worth delaying a release, and a regulator will judge you on what you had switched on the day you shipped, not the day they knocked.
Fourteen rival labs ship alongside you on the dates the real ones shipped.

The thing I found this week

For weeks the strongest way to play was to ship the same model over and over under different names.

Every product on sale added its own score into the demand split, so a company fielding three hundred products scored three hundred times over against rivals fielding one. Nothing in the game told me.

Every test passed. I found it by running my scripted operator twice over fourteen years, three seeds, once naming every release separately and once keeping a single product line, and putting the two results side by side.

a new name every release 339 products on sale 221.8M users $127.7bn peak one product line 1 product on sale 24.4M users $3.1bn peak

Forty one times the money, for what is obviously the worse company.

The repair is a nested logit, which is the standard way of saying that two things can be closer substitutes for each other than for the rest of the field.
Within one company and one kind of model, products now share an audience instead of each bringing their own. Four genuinely different products are worth 1.6 of one. Three hundred and thirty nine are worth 7.3.

What made it safe to change the centre of the economy the night before a release is a property of the arithmetic: a company with one product in a line, and every rival in the game, comes out with exactly the number it had before. I re-ran all nine campaigns afterwards and the disciplined operator finished identical to the dollar. Only the one standing in the hole moved.

Then it broke my own balance test, because that test had been standing in the hole too.

The rest of 0.3.0

A version ships when the work that produces it lands.
Naming a version and pressing PLAN THE RELEASE used to put it on the market the same instant while the engineering ran for weeks behind it. A tester reported it in one sentence and he was right.

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. A cabinet with nothing in it used to be drawn the same green as one working happily, so a half empty room looked like a room that was fine.

A fan finally costs the slot it always claimed to cost. One method charged it and the nightly refill did not, so the floor held as many accelerators with fans as without, and the trade the whole cooling system is built on was never made.

VS TOP MODEL in the creator puts the strongest thing anybody else has on the market next to your own projection. Capability only.

Nobody publishes what a rival's training run cost them or how long it took, and I am not going to invent a number on the screen where you commit a hundred million dollars.

What is still wrong, because you will find it

The late game takes the player apart. Measured over three seeds with one product line:
company peaks near seventeen million users in early 2024 at 4.9% of everybody being served, and by 2035 it holds sixteen thousand while one rival holds 84%.

That is next real piece of work. It is not a patch I can do in an evening and I would rather say so than let you find it and think I do not know.

Who is writing this

I am Marcin, 22, from Radom in Poland. I write this after taxi shifts.

Fifteen months ago I started building a desktop assistant called PC Workman, mostly because I needed to prove to myself that I could still finish something. A year later it was on the Microsoft Store.

It is my flagship project and it is open source and free, like everything else I make.

The honest arithmetic of that: fifteen months of work, and it has earned me about €16.

Scaling Laws is 42 days old. 330 commits. 145,863 lines of C#.
1,273 automated tests and 57 more that load a scene, render a page to a PNG and let me look at it without opening the editor, because every layout fault in this project has been found by looking rather than by a test.

The save format is at v56.
That means the data model has changed under a live game fifty six times and every one of them is a real migration, not a reset. When a migration has to invent a value the old format never stored, it picks the least flattering reading that is still defensible and writes down what it did.

Where to get it

Free. Windows, 85 MB zipped. English and Polish.

• Download and source: Github.com
• Everything else: Pcworkman.dev

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

❤️ Linktree.com ❤️

If something breaks, or you get stuck somewhere and cannot tell whether it is you or the game, that is the thing I want to hear. There is a feedback form in the game under the last tab.

Post a comment "); var textarea = $("#reply"+post); var mentions = $("#commentform .mentions-input").length > 0; if(mentions) { textarea.mentionsInput({source: "https://www.moddb.com/html/scripts/autocomplete.php?a=mentions&p=home&l=6&u=666"}); } $("#cancelreply"+post).click(function() { $("#"+post+" div.replypostholder").remove(); }); $("#savereply"+post).click(function() { postcontent = mentions ? textarea.mentionsInput("getValue") : textarea.val(); if(postcontent) { $(this).remove(); $("#cancelreply"+post).remove(); textarea.after("Saving your reply"); if(mentions) { $("#commentsummary").mentionsInput("setValue", postcontent); } else { $("#commentsummary").val(postcontent); } $("#commentform input[name=replyid]").val(post); $("#commentsubmit").click(); } else { $(this).text("Please enter a comment"); } }); } return false; }); $(".games92058watch").click(function(){ $.achtung({message: "Only members can follow games. Sign up today and follow your favorites.", className: "achtungFail", timeout:5}); return false; }); $(".c76625d26ff7f0c55142db2c328da119watch").click(function(){ $.achtung({message: "Only members can follow games. Sign up today and follow your favorites.", className: "achtungFail", timeout:5}); if($(this).hasClass("button")) { $(this).addClass("buttonon"); } return false; }); $(".togglertxinfo").click(function(e){ if($(this).attr("href") == "#rtxremix") { localStorage.setItem("rtxtoggled", true); $("#rtxremix").show(); return true; } else { localStorage.removeItem("rtxtoggled"); $("#rtxremix").hide(); return false; } }); function cookieScript(e) { var status = CookieScript.instance.currentState().categories.includes("targeting") ? "false" : "true"; var cookies = CookieScript.instance.currentState().categories.includes("performance") ? "true" : "false"; var cookie_data = { npa : {status:status,cookies:cookies,version:"1.0.13",chosen:"true"}, expires : 365 }; Cookies.set("dfp-npa", cookie_data.npa, { expires: cookie_data.expires, secure: true }); if (typeof gtag === "function") { gtag("consent", "update", {"ad_storage": cookie_data.npa.status == "false" ? "granted" : "denied", "ad_user_data": cookie_data.npa.status == "false" ? "granted" : "denied", "ad_personalization": cookie_data.npa.status == "false" ? "granted" : "denied", "analytics_storage": cookie_data.npa.cookies == "true" ? "granted" : "denied"}); } if(cookie_data.npa.status == "false") { // personalize if (typeof permutive === "object") { permutive.consent({ "opt_in": true, "token": "Allow All" }); } } else { // don't personalize if (typeof permutive === "object") { permutive.consent({ "opt_in": false }); } Cookies.remove("permutive-id", { domain: "moddb.com", secure: true }); } } if(!Cookies.get("dfp-npa")) { Cookies.set("dfp-npa", {status:"false",cookies:"true",version:"1.0.13",chosen:"false"}, { expires: 60, secure: true }); } window.addEventListener("CookieScriptAcceptAll", cookieScript); window.addEventListener("CookieScriptAccept", cookieScript); window.addEventListener("CookieScriptReject", cookieScript); window.addEventListener("CookieScriptClose", cookieScript); $("a.thickbox").click(function() { $.colorbox({href:$(this).attr("href"), rel:$(this).attr("rel"), open:true, onOpen:function(){ $(".nextmediabox").hide(); }, onClosed:function(){ $(".nextmediabox").show(); }}); return false; }); });