A puzzle app spent $220 on Google app ads — 60% of the installs were robots

A one-dev indie studio just published the cleanest writeup we've seen of why "installs" is the most dangerous number in mobile advertising: a bot farm found his $220 budget, Google billed him for it, and the ad algorithm rewarded the fraud at every step.

Nick Abe, the developer behind the puzzle app Dayzle, ran a Google Ads campaign for Android at CA$40 a day with the goal set to installs — and when he removed his cost-per-install target after a slow start, the campaign immediately spent double his daily budget and reported 21 installs in a day while his own admin panel counted one. Digging into the raw analytics, he found 20 of those 21 devices were running an old version of the app the Play Store had stopped serving days earlier. Every one of them claimed Google Play as the installer, yet none could have gotten the app from Play — they installed from a saved copy of the app file instead, opened it once, spent zero seconds on any screen, and never came back. Twenty-eight phone models across nineteen US states, all doing exactly the same thing.

Over the full two weeks the split was stark: 56 installs billed, 33 matching the bot pattern, 7 more from countries the campaign wasn't even targeting — and 13 actual humans, who finished 92 games between them. The mechanics of the fraud loop are the part every advertiser should internalize. The farm would watch the shortest video in the ad group without clicking, then install the app from the cached file because it's faster and Play might notice. Google counts a view followed by an install as a conversion, so every fake install made the campaign look better, which sent more of the budget to the farm, which installed more. The optimization target and the fraud were the same signal.

The fix is instructive precisely because it's a goal change, not a fraud filter: the campaign now optimizes for "won a puzzle" instead of "opened the app," on the theory that it's cheap to script opening an app and clicking around, and expensive to script solving a Sudoku. Abe is waiting on Google's answer to an invalid-traffic claim, and his advice generalizes — if a bot farm can find a CA$40-a-day budget from a tiny puzzle app, it can find yours, and larger apps almost certainly see more of this than they know. For anyone buying AI-adjacent app installs right now — and the market is full of "AI companion" and "AI assistant" apps doing exactly this — an install count that isn't validated against your own analytics is a number a machine is writing to you.

Why it matters: install-goal campaigns are the default for every small app developer, and this is a first-person, receipts-included demonstration that the default quietly funds fraud. Ad-fraud bots keep getting more autonomous — watching ads, installing, mimicking a device farm's worth of diversity — and the platform's own optimization loop is what feeds them.

What to watch: whether Google refunds the invalid-traffic claim, and whether Abe's "won a puzzle" goal actually starves the farm.

Has an ad platform ever billed you for traffic you could prove was fake — and did you ever get the money back? Tell us in the comments.

Houmo moves compute inside memory — and A20 enters production

Beijing chipmaker Houmo used its partner conference this week to lay out two compute-in-memory routes at once — and to say the first chip built on the newer one is already in production. Houmo's A20, its first chip built on a 3D compute-in-memory architecture, has entered production. The part supports up to 16 chips fully interconnected, claims a 50% improvement in compute utilization over the previous generation, and covers models from 3B to 30B parameters, according to co-founder and product

Kling 4.0 goes live with 30-second native video at 10-bit HDR

Kuaishou's Kling AI has spent two years becoming the video model people actually finish projects with. Its new flagship, Kling 4.0, went fully live on October 15 — and this round is less about prettier frames than about the two things that make a generated shot usable in a real edit: length and control. Kling 4.0 generates up to 30 seconds in a single native pass, outputs 10-bit HDR at 4K, 1080p and 21:9, and takes up to 15 reference assets in one task. That budget splits across ten images, fi

Brain organoids classified Braille at 83% — in a trio, not alone

Two stories today: human brain tissue on a chip decoded a tactile signal, and the data-centre company behind some of the largest AI contracts in Europe turns out to be six months behind on its flagship site. Human forebrain organoids grown on a microelectrode array classified Braille letters with 61 percent accuracy on their own and 83 percent when the responses of three organoids were combined, in a peer-reviewed study from the University of Bristol. The team took event-based tactile readings

Suleyman says Microsoft's base model beat OpenAI's on a quarter compute

Microsoft finished pre-training its biggest in-house model yet — then used it to make a point about the company it bankrolls. Microsoft AI completed pre-training of MAI-Base-2, a mixture-of-experts foundation model with 80 billion active and 2.8 trillion total parameters, and says the recipe is 4.2–12.3x more pre-training FLOPs-efficient than DeepSeek-V4-Pro-Base. The run itself is the number worth keeping: 16,000 Nvidia GB200 GPUs for 38 days across August and September, at 21% average model

Read more

Houmo moves compute inside memory — and A20 enters production

Kling 4.0 goes live with 30-second native video at 10-bit HDR

Brain organoids classified Braille at 83% — in a trio, not alone

Suleyman says Microsoft's base model beat OpenAI's on a quarter compute