Issue 169
🔄💻 New breakthrough AI method tests code in real-time, beats GPT-4. McDonald's scraps drive-through AI after bacon ice cream debacle. AI researchers hunt for "clean" pre-2022 data.
Hey there Bizarro readers!
Welcome to the July 2025 issue of your favorite quirky tech newsletter. Before we get into the stories, I have an important announcement to share:
Due to ongoing budget constraints, this will be the last issue of Bizarro Devs. Unfortunately, as much as we all love this newsletter, our parent company has made the difficult decision to discontinue it.
Having said that, I plan on personally carrying on the Bizarro Devs tradition on my own time in the near future (possibly as early as next month). Please go to devtech.news and add your email, as that will be the new home of some version of this newsletter.
With that out of the way, here are this month’s headliners:
🔄💻 New Breakthrough AI Method Tests Code in Real-Time, Beats GPT-4
🍟🤖 McDonald's Scraps Drive-Through AI After Bacon Ice Cream Debacle
💾📅 AI Researchers Hunt for "Clean" Pre-2022 Data
The lead story is actually very fascinating and hasn’t gotten the attention it deserves in the wider tech press, but perhaps that will change with it being featured here.
Enjoy the read!
📰 From the Newsroom
🔄💻 New Breakthrough AI Method Tests Code in Real-Time, Beats GPT-4
A few months ago, vibe coding took the internet by storm, but as its popularity has grown, so have the justifiable criticisms of its limitations. We covered that story back in April, but recently a group of researchers produced a new system that may put the vibe coding criticisms to bed. Dubbed Execution-Guided Classifier-Free Guidance, or EG-CFG for short, it addresses the fundamental problem of using AI to code, which is that it’s prone to errors. Here’s how it works…
Instead of generating complete programs and crossing its digital fingers, EG-CFG writes code line-by-line while actually executing and testing each piece as it goes. It generates several possible next lines, runs them against test cases, sees which ones work, and uses that feedback to guide what comes next. This mirrors how human programmers iteratively test and adjust their code during development.
The performance jump in testing has been massive. Using only open-source DeepSeek models, EG-CFG hit 96.6% accuracy on MBPP coding tasks (compared to GPT-4's 87.7%), 87.19% on extended HumanEval tests, and 58.18% on competitive programming problems where GPT-4 managed just 34.7%.
The best part of all this is that these results came from open-source models that anyone can access, not proprietary systems costing millions to train. The researchers proved that smart techniques can beat raw computational power. Their system also naturally supports parallel processing where multiple agents explore different coding approaches simultaneously.
EG-CFG could be the reliability upgrade vibe coding has been desperately needing since its inception. If AI coding assistants can actually test their own work line-by-line, it could transform vibe coding from "fun but risky" to "genuinely practical." And I can finally stop seeing five posts per day on my LinkedIn thread about how it’s irresponsible to use.
🍟🤖 McDonald's Scraps Drive-Through AI After Bacon Ice Cream Debacle
AI company CEOs keep insisting that AGI is just around the corner and millions of jobs will be automated away in the next few years (see here, here, here, here, and more). Meanwhile, McDonald's just pulled the plug on their drive-through AI system after five years of testing, and the reason why might make you feel better about human job security.
It turns out that despite the hype, McDonald’s "revolutionary" technology had a few issues. Examples include: A woman trying to order caramel ice cream but getting multiple portions of butter added to her bill instead; two women receiving hundreds of dollars worth of chicken nuggets applied to their order; and a person being given an ice cream topped with bacon.
Apparently the AI has had trouble distinguishing between customer voices and background noise, and it’s been struggling with accents as well. McDonald's franchisees have reported that the technology has been "underwhelming," which seems like a polite way to describe it.
The industry response has been predictably tone-deaf. Despite McDonald's obvious failure, other chains are doubling down on AI drive-throughs. Wendy's, Dunkin', and Hardee's are all testing similar systems, while Yum Brands announced an "AI-first mentality" claiming the possibilities are "endless."
The McDonald's debacle perfectly captures the gap between AI hype and AI reality. While tech evangelists proclaim the imminent automation of everything, in the real world, AI can't seem to handle a basic drive-through order. To bridge the capability gap, some companies have even resorted to faking the funk by employing workers from cheap labor countries to make it seem like their AI is better than it actually is. It makes me wonder if when we “finally achieve AGI,” if it will really just be a group of really smart dudes hanging out in a cubicle in Bengaluru responding to prompts half a world away (I’m joking. Maybe. 😅).
💾📅 AI Researchers Hunt for "Clean" Pre-2022 Data
Did you know that nuclear testing in the 1940s permanently contaminated the world's metal supply? It forced scientists to salvage steel from sunken German warships for sensitive equipment. Now academics are saying that ChatGPT's launch did something eerily similar to the internet, and they're scrambling to find "clean" data from before November 2022.
The parallel is surprisingly accurate. Just like nuclear fallout made all post-1945 steel "dirty" for precision instruments, researchers worry that post-ChatGPT internet content is polluted with AI-generated text. It means that AI output is becoming progressively less reliable, like making a photocopy of a photocopy until everything gets hopelessly blurry.
It also means that early AI players are sitting on treasure troves of uncontaminated training data that new startups simply can't access. And these pre-2022 datasets might give them a permanent advantage over newcomers.
The fix isn't straightforward. Some have proposed federated learning systems where clean data owners let others train models without sharing the actual datasets, but that creates new problems around privacy, security, and political control of these data vaults.
What makes this particularly tricky is the irreversibility factor. As one researcher put it: "If you've completely contaminated all your datasets, cleaning is going to be prohibitively expensive, probably impossible." While this article mainly focuses on the "self-harm" aspect of AI, another (arguably more important) thing to consider is the future of the internet itself. What is the web going to look like in another 5 years when AI-generated content dwarfs human-produced content by some crazy ratio?
⛓️ Ten Must See Links of the Month
Sponsored by Optimole, the best image optimization tool on the internet.
Ever heard of the Pentagon Pizza Index? It has accurately predicted major U.S. military actions since the 1980s and there’s even an X account that offers real-time updates on Pentagon-area pizza orders as a way of "reading U.S. military tea leaves."
If you’re thinking of getting into the wonderful world of web accessibility, then you’ll need a primer on ARIA (Accessible Rich Internet Applications). Here’s a great breakdown from Eric Bailey, a longtime advocate of accessible web design.
Stanford legal expert Mark Lemley found that Meta's LLaMA AI can reproduce verbatim passages from copyrighted books like "Harry Potter," potentially exposing the company to nearly $1 billion in statutory damages.
CNBC recently reported that Google uses YouTube's 20 billion video library to train AI models including Gemini and Veo 3, but most creators and media companies remain unaware this is happening - and there’s no way to opt out.
Nekojita FuFu is a tiny cat-shaped robot that attaches to cups and bowls to blow air and cool hot drinks and food. Its Kickstarter campaign massively exceeded expectations, reaching over 3,167% of its original funding goal with ¥18.1 million raised against an initial target of ¥572,000.
Researchers have developed a robotic skin made from conductive gel that can detect pressure, temperature, pain, and damage simultaneously across its entire surface, sensing signals from over 860,000 pathways in the material.
Scientists have successfully tested swarms of microscopic robots, each smaller than a speck of dust, that can be injected into sinus cavities to treat bacterial infections by heating up and catalyzing chemical reactions to kill bacteria.
🎥🎞️ Meet Chris Smith, a 32 year old married man and father of a two-year old daughter. Chris built an AI companion using ChatGPT and ended up "falling in love" with it - to the point he cried for 30 minutes when his "companion" reached the 100,000 word limit and he had to start a new chat with it.
🎥🎞️ Do you use WordPress? If you answered "yes," then there’s probably a decent chance that you either use or have used the popular Yoast SEO plugin. What you might not know about Yoast though, is that even if you delete it from your site, it leaves behind a mess of files. Here’s a helpful video that shows you how to fully remove Yoast so it leaves nothing behind.
Hostinger is a company primarily known for being a major player in the web hosting market, but recently they decided to branch out from their core business by launching an AI coding tool. I took it for a test run - two test runs actually - and wrote about the experience in detail.
🎤 It’s How They Said It
“A sunset is nothing more and nothing less than the backside of a sunrise.”
– Craig D. Lounsbrough
🧮 The Numbers Game
2,770 days elapsed between the publication of the very first issue of Bizarro Devs on the 29th of November 2017, and the publication of the last issue that you are currently reading on your screen. A lot happened over the course of that time, both in the world of technology and with the newsletter itself.
$13,900,000,000 USD will be equally shared among 100+ children fathered by Telegram founder Pavel Durov through both relationships and sperm donations. The 40-year-old Russian-born tech tycoon told French magazine Le Point that he makes no distinction between his six children from relationships and those conceived through donation.
$1,600 USD is the price for Meticulous's new robotic espresso machine, which the company claims is the world's first to use robotics, precision sensors, and smart algorithms to replicate café-quality hand-pulled espresso shots at home.
6 months is how long printer manufacturer Procolored unknowingly distributed malware-infected official software and drivers for at least six of its printer models, including a remote access trojan called XRedRAT and a cryptocurrency stealer called SnipVex that has stolen nearly $1 million worth of Bitcoin.
7.3 terabits per second was the unprecedented size of a DDoS attack that Cloudflare successfully blocked in mid-May 2025, setting a new world record that was 12% larger than any previously recorded attack. The 45-second assault delivered 37.4 terabytes of malicious traffic targeting a hosting provider, originating from over 122,145 source IP addresses across 161 countries.
⚒️ Tools and Resources
Coding Courses: This collection of coding courses is easy to navigate and offers a mix of free and discounted options to boost your skills. As of the time of this writing, there are around 50 free courses, catering to beginners, experienced coders, and everyone in between.
NGX-VFlow: This is an Angular library for creating node-based applications. It can help you build anything from a static diagram to a visual editor. You can use the default design or apply your own by customizing everything using (mostly) HTML and CSS. Just describe your flow with a simple API and all of the heavy lifting, such as dragging, zooming, and curve math, is handled by the library for you.
Oxlint 1.0: This Rust-powered JavaScript and TypeScript linter, has been released with 50-100x performance improvements over ESLint and support for over 500 ESLint rules. Major companies including Shopify, Airbnb, and Mercedes-Benz have adopted it, with Mercedes-Benz seeing up to 97% speedup and Airbnb completing linting on 126,000+ files in just 7 seconds where ESLint times out.
https://voidzero.dev/posts/announcing-oxlint-1-stable
🖼️ What Am I Looking At?
Both of the above photos are of U.S. President William H. Taft. One was taken in 1908, when the President wasn’t yet the President, but the Secretary of War. The other was produced in 1923 as a psychological tool to make the people of the Philippines believe that the U.S. had good intentions toward them.
You can probably figure out which is which, but the more interesting story here is that fake photos and image manipulation have been around much, much longer than modern smartphone filters, AI tools, Photoshop, or even computers for that matter.
And even though digital photo manipulation is so readily accessible nowadays, analog image manipulation actually has a much longer history that dates all the way back to the 19th century. Read the full story.
💬 What’s the Word?
For our final word of the month, I wanted to share a word from my own language that perfectly captures how I'm feeling about this newsletter coming to an end: ľútosť.
Ľútosť is a Slovak word that you might translate as "feeling sorry for (someone)," but it goes deeper than that. It's a complex emotion that combines regret, sorrow, and genuine compassion - often for circumstances beyond anyone's control.
The beauty of ľútosť is that it doesn't demand that we fix anything or look on the bright side. In the context of this being the final Bizarro Devs issue, it simply acknowledges that endings carry weight, even when new beginnings are already taking shape. Which, as you know, they are. 😉
🧑🏻💻👨🏽💻👩🏼💻 Onwards and Upwards
Thank you for reading!
I’ve been running this newsletter and curating the content for it for almost three years - since issue 135 - and I’ve enjoyed every minute of it. I’m grateful to our company, Vertigo Studio, for entrusting me with the responsibility. I’m also extremely thankful for my colleagues Sabina Ionescu and Karol Król for their guidance.
Sabina was the original architect of the newsletter and came up with the entire concept, and Karol was the one who handed it off to me and helped me craft my first few issues. He’s also been the first person to read every single issue before it’s gone live, and his feedback has helped me finalize each one before it reached your inbox.
Finally, I want to acknowledge the numerous talented writers who contributed to the newsletter over the years. Tom Rankin, Adelina Tuca, Joe Warnimont, Diana Gunn, and others all helped to make it successful before I took the reigns.
I hope you’ve enjoyed reading Bizarro Devs as much as I’ve enjoyed producing it. If you’d like to continue along on this journey, please go to devtech.news and I will see you there in the near future.
Until we meet again,
– Martin Dubovic
Bacon topped ice cream is a win!