putting 100% local AI to good use.
Remote
Mario Zechner
@badlogic@mastodon.gamedev.place
mastodon 4.7.3libGDX, Beginning Android Games, RoboVM, Spine. Member of the council of Warios.
5448 Followers
486 Following
50 Posts
Joined November 13, 2017
github:
twitter:
the wall street journal put out an interview with me, basically saying vibe slop will kill our software industry.
here is my mom giving the exact number of fucks go give, when i told her about it.
zero.
Introducing Masto Reader. Like Thread Reader App, but for Mastodon.
How does it work? Mention @mastoreaderio@mastodon.social(@mastodon.social) anywhere in a thread and say "unroll" and the little bot will do the rest.
Here's an example (that takes a bit to load for reasons)
https://mastoreader.io/?url=https%3A%2F%2Fmastodon.gamedev.place%2F%40badlogic%2F111071396799790275
You can share those links with others.
Links to the crappy source code can be found in the bot's profile.
Everything runs in your browser. The server doesn't log anything either.
Enjoy and spread the word.
Replying to
The whole thing runs client-site. The server fetches the latest data from the stores once a day. All data fits into 5mb of gzipped JSON. Small enough for the client to do anything. The server just serves 8 static files. It can handle serve all of Austria easily and could be scaled trivially. It's just static files.
Being the idiot I am, I also made it open-source:
https://github.com/badlogic/heissepreise
And as usual, people flocked to it and contributed. In no time we had all stores in Austria in there.
Replying to
I scrambled to integrate his data into my platform. I added analytics tools. And then I ran my first few analyses. And my jaw dropped.
"Well, that's a bit to much of a price increase even given higher energy prices."
So I started to dig. And boy did I find a lot of things...
back to my old bullshit
Replying to
Then we also got German and Slovenian stores. Then we normalized product categories across stores and added some light data science techniques to match the same or similar products across stores to make prices more easily comparable. You know, iterative improvements.
And then some anomymous guy in Twitter send me the data he crawled for the two biggest chains. Starting in 2017. And that's when thinga really got interesting...
Replying to
E.g. given the historical data, I could see price movements for a product across the two chains. And you won't believe what I found (well, you know what's coming...)
Them fine grocery chains changed the prices of the self-branded low cost products with one to two days, or even on the same day. And they both came up with the exact same price.
Replying to
If I were trying to describe it in more flowerly terms: It's asymmetric information war fare.
The stores tell you they are good and benevolent and only have your interest at heart, so here are discounts. Discounts for everyone. They even gamified the whole thing with stickers. I shit you not. People collect stickers they put on the products in the convery belt at the register. There's also apps, which will give them all info on you
In reality it makes it impossible to know how much things cost
Replying to
This wasn't only happening in the low-price chain-brand segment. It also happened in the mid-range segment of self-branded goods.
And it all started happening when inflation went through the roof.
Clearly, something was up. My guess was: tacit collusion, meaning, oligopolic price coordination without explicit coordination.
Meanwhile, others have build platforms like I did as well. And they too saw these patterns.
There were more.
always have a plan
Replying to
Even more interestingly, products exclusively produced in Austria cost less outside of Austria.
Billa is the Austrian version of REWEDE.
Even fucking Red Bull, an Austrian brand, costs more in Austria when it is discounted here, than it costs normally without discount in Germany.
WTF.
Replying to
Of course you do, cause you are a smart cookie.
While their claim that they decreased the regular price is correct, they also increased the discounted price that comes into play every other X weeks/days.
So they are again technically correct: the regular price was decreased.
But on average, a consumer pays more if they buy the product every week, as the discounted price has been increased. The average is higher than before.
Sneaky.
Replying to
My first analysis actually happened before I build the platform. I was manually comparing prices of products the stores themselves offer in the lowest price segment. Things like grocer store brand milk or flour.
I compared 40 product pairs across the two biggest chains. And lo and behold: their prices matched exactly to the cent!
An NGO picked this up on Twitter and did the analysis for 600 product pairs. Same picture.
With my platform in place, I could do more advanced stuff.
Replying to
We could also show that the exact same product cost up to 40% less in Germany, a country with higher mean income and higher cost of living.
Replying to
The grocery chains got a little iffy about all that somewhat negative media coverage, some of which was spurred by my continued analyses.
They started to put out these things in the store. It basically says "We've already lowered the prices of 450 products for you this year". With a sortiment of 22000.
They were also dumb enough to put out a machine readable PDF with all the products they lowered the price for.
With a little data science magic, I was able to match those with my database...
Replying to
We could show shrinkflation, meaning products with less content are sold for the same or even higher price.
Examplified by e.g. laundry detergent.
Replying to
And lo and behold. There was fun to be had.
There are products that are cyclic in their price changes. E.g. this axe shower gel, which they listed as having a lower price now.
Yeah, you lowered the price from 3.99 to 2.99. But that follows the exact pattern this product's price had over the last couple of years.
Technically correct. But not a permanent price decrease.
Second picture is another example of that.
Replying to
Then I looked at an aspect pretty unique to Austria: discounts.
You see, in a normal country, with a competitive grocery market, you usually have about 10%-20% of products that get discounted on average.
In Austria, that rate is 40%. It's a fantastic way to obfuscate the actual price of a product. As a customer, you'll never know what you'll pay on that day until you see the current discounts directly in the store.
The chains are very generous and will send you discount leaflets via mail.
Replying to
making off.
step 1: buy a discount €9 shitty robot toy. use violence.
Replying to
The spot check showed that their claims were true on the surface.
But I'm a stickler for data, so I looked a bit closer.
Replying to
Given the historical data I had, I was able to also check for patterns in the discounts they give. How often, how high.
Replying to
But there's a more "nefarious" kind of price decrease.
As I said, Austria is a country of insane amounts of cyclic discounts. Many products will be sold for their "regular" price for one week and a discount price the other.
The real price for the consumer is the average of the regular and discounted price.
Given this knowledge, do you notice something with the prices for this product the grocery chain claims to have decreased the regular price on?
Absolutely fantastic long read by @pluralistic@mamot.fr
Recommended reading!
https://pluralistic.net/2025/12/05/pop-that-bubble/#u-washington
(I also seem to have graduated from kid to guy)
Replying to
I would likely not have done this project manually, as it is journalistically niche (at least that's what my journalist friend told me). She's pretty happy now, that it exists and has already published the first article on the matter.
I share many of the reservations about AI, specifically wrt exploitation of IP, low wage workers, AI in education and politics and so forth.
But I also think we as a community should exploit these tools to create things for the public good if possible.
I've integrated Mistral AI models in my little coding agent harness pi and am happy to report that Devstral 2.0 is a good agentic model. Somewhere between Sonnet 4.0 and 4.5. It's also OSS and self-hostable (though you need a small cluster for the bigger model, e.g. via DataCrunch)
Europe is back, baby.
Everything is rotting.
Replying to
Things are never black and white.
Final note: you can open a repository named "nazi-radio" on GitHub without issue.
New blog post, wherein I tell you what I learned building my own opinionated and radically minimal coding agent.
With way to many words. Ask an LLM to summarize it for you.
Recommended reading.
I wrote up all the things I did outside of family and business this year. It's a bit ... long. May it entertain you.
https://mariozechner.at/posts/2025-12-22-year-in-review-2025/
Replying to
Notably, this project uses AI in two ways.
Speech to text transcription is done via Whisper (through Groq API).
All the scraping and bot detection evasion code was iteratively created by a coding agent, with me in the loop to give instructions what to try to poke through protections.
The whole project took about 2 hours in total and cost €12 in API credits for the coding agent. Based on previous experience, it would have taken me about 2-3 nights by hand.
Replying to
If you want a minimal coding agent, with many creature comforts, you can give pi a try:
https://github.com/badlogic/pi-mono
Supports any model you can throw at it, including local models or models you "self-host" on providers like EU-hosted and GDPR-compliant DataCrunch. In case you do not want to share all your data with big tech.
Recommended reading with lots of good thoughts and stories by @grumpygamer@mastodon.gamedev.place
Replying to
The lawyers are going to be filthy rich
Replying to
@hackaday@hackaday.social hey maybe this one fits your style!
Replying to
@BratschCbra@sueden.social Haha, ja
Replying to
@ashdnazg@fosstodon.org fair. The flood gates have been opened tho, and I posit that automated flagging will not really work, leading to a lot of finger pointing and suspicions, which hurts the scientific process.
Replying to
@_benui@mastodon.gamedev.place @mattburgess@infosec.exchange or very dumb depending on if and when i get surd :D
Replying to
@Wandel@mastodon.gamedev.place euch ausgenommen. Aber das war schon sehr late in the game :)
Replying to
@jake4480@c.im @standev@mastodon.online @mattburgess@infosec.exchange I'm currently finishing a rewrite which should make it easier for folks in other countries to plug in scrapers for local stores and do i18n. I'll post about it here once it's ready.
Replying to
step 2: figure out the pcb, find out it's a mystery soic, and an h-bridge. the advertisefld "matrix LED display" is 2 segmented color LEDs behind a plastic screen.
Replying to
step 3: ignore everything and rewire the h-bridge to GPIOs on an ft232h, so the phone can later drive the single motor that can either rotate the head left, or move the robot forward in the direction it's heading in. love the mechanism!
Replying to
@_benui@mastodon.gamedev.place @mattburgess@infosec.exchange it is open-source. https://github.com/badlogic/heissepreise
Currently rewriting it so it's easier to get going in other countries.
Replying to
@cenobyte@mastodon.thirring.org @mattburgess@infosec.exchange I know of two forks that are trying to get this to work for Canada. We'll see how that works out.