ReHacked vol. 343: LLMs Can Get "Brain Rot", Smartphones manipulate our emotions and trigger our reflexes and more
Support ReHacked newsletter with one time donation. Thank you very much!
"In my experience, each failure contains the seeds of your next success—if you are willing to learn from it." --Paul Allen
In this work, we introduced and empirically validated the LLM Brain Rot Hypothesis, demonstrating that continual exposure to junk data—defined as engaging (fragmentary and popular) or semantically low-quality (sensationalist) content—induces systematic cognitive decline in large language models. The decline includes worse reasoning, poorer long-context understanding, diminished ethical norms, and emergent socially undesirable personalities.
Fine-grained analysis shows that the damage is multifaceted in changing the reasoning patterns and is persistent against large-scale post-hoc tuning. These results call for a re-examination of current data collection from the Internet and continual pre-training practices. As LLMs scale and ingest ever-larger corpora of web data, careful curation and quality control will be essential to prevent cumulative harms.
Make a donation - support Ukraine. Щира подяка. Разом до перемоги!
Like what you read? Subscribe now! Please share if you like what you read here, subscribe (if not yet) and leave a comment. Any form of feedback is very important. Thank you very much!
RSS feed available if you don’t want to clutter your inbox.
You can also support ReHacked newsletter with one time donation.
Thank you for being a part of the community. Together, let's continue fostering a culture of knowledge-sharing and making a positive difference in the digital landscape.
I’m drowning in AI features I never asked for and I absolutely hate it #ai #technology #copyrights
The problem with generative AI goes far deeper than annoying features. It starts with how these models are trained. Large language models learn by analyzing massive amounts of data scraped from the internet, like books, articles, photos, videos, and social media posts — all taken without asking permission. What it really means is that someone's original work is being used to build a commercial product they'll never profit from.
Every creative field has been affected. Writers, artists, and photographers have found their work inside datasets used to train AI tools that can now mimic their style. In most cases, there's no credit, no payment, and no consent involved. When an AI tool generates an article, an image, or even a song, it's built on the collective effort of thousands of people who never agreed to be part of the process.
If you engage with a post that features an AI-generated video, your view still counts as engagement. That interaction feeds money and visibility to whoever uploaded it, not to the original artists whose work the model learned from. The system rewards the wrong side of creativity while quietly taking more data from users to fuel itself.
Smartphones manipulate our emotions and trigger our reflexes — no wonder we’re addicted #psychology #health #society #technology
Short of powering off or walking away, what can we do to manage this dependency? We can access device settings and activate only those features we truly require, adjusting them now and again as our habits and lifestyles change.
Turning on geolocation only when we need navigation support, for example, increases privacy and helps break the belief that a phone and a user are an inseparable pair. Limiting sound and haptic alerts can gain us some independence, while opting for a passcode over facial recognition locks reminds us the device is a machine and not a friend. This may also make it harder for others to access the device.
So-called “dumb phones” limit what a user can do with their devices, though that’s a tough sell when 24/7 connectivity is becoming an expectation.
Manufacturers can do their part by placing more invasive device settings in the “off” position in the factory and being more transparent about their potential uses and data liabilities. That’s not likely to happen, however, without stronger government regulation that puts users and their data first.
In the meantime, at a minimum, we should broaden our public discussions of dependency beyond social media, gaming and artificial intelligence to acknowledge how phones, in themselves, can capture our attention and cultivate our loyalty.
A new national study led by researchers from the University of Virginia, University of Pennsylvania and the American Institutes for Research found that Montessori preschool programs (ages 3 to 6) in public schools deliver stronger early learning outcomes for children—and at a sharply lower cost to school districts and taxpayers.
The first randomized controlled trial of its kind, published in Proceedings of the National Academy of Sciences, tracked nearly 600 children across 24 public Montessori programs nationwide.
By the end of kindergarten, children who won a random lottery to attend public Montessori preschools outperformed their peers in reading, executive function, short-term memory, and social understanding—all while costing approximately $13,000 less per child than traditional preschool programs.
Those costs do not include anticipated savings from improved teacher morale and retention, a dynamic demonstrated in other data.
4 ways to fix 'tech neck,' according to a physical therapist | Popular Science #health #fitness
- Wall angels
- Thoracic spine wall rotation
- Row variations
- Prone I’s, T’s and Y’s
More details inside.
The Internet's Biggest Annoyance:Why Cookie Laws Should Target Browsers, Not Websites | NEDNEX #internet #privacy
How do you want to handle your data?
- Essential Only: "Only allow data necessary for websites to function (e.g., keeping me logged in, remembering my shopping cart)."
- Performance & Analytics: "Help creators improve their sites by letting them see anonymous data about how I use them."
- Personalized Experience: "Allow sites to use my data for personalized content and relevant advertising."
- Custom: Fine-tune your settings for specific data types. You make your choice once. Set it and forget it.
Doomsday Scoreboard #fun #futurism
Democracy and the open internet die in daylight #ai #privacy
...you get free access to news if you download PerplexityAI’s proprietary browser, which is built entirely around agentic AI.
grinds teeth into powder
Leaving aside the idea of access to any form of content being conditional on the use of a proprietary browser, which is a particularly horrid 1990s throwback, I’m going to call this day 0 of an experiment in shifting the funding model of journalism from adtech to agentic AI.
Oman’s plan to become a space powerhouse by building a launchpad - Rest of World #space #economy #longread
Oman’s dream for a spaceport sits in a swath of desert overlooking the Arabian Sea. If all goes according to plan, three separate complexes here will launch everything from small suborbital rockets for scientific experiments to superheavy behemoths into space, bringing payloads of satellites to the stars. There will be a large mission-control building, warehouses for rocket assembly and testing, and a business park. If the spaceport succeeds, by 2027 Oman will join the small number of countries — just a dozen to date — that possess the facilities to launch objects into space.
Tesla is heading into multi-billion-dollar iceberg of its own making | Electrek #automotive
Tesla’s ‘Full Self-Driving Supervised’ expansion is back firing as it exposes its shortcomings. Customers left without promised features are growing discontent and demanding to be compensated.
In 2016, Tesla proudly announced that all its vehicles produced onward are equipped with “all the hardware for full self-driving,” which would be delivered through future software updates.
The automaker turned out to be significantly wrong about that.
At the time, it was producing its electric vehicles with a hardware suite known as HW2, which it had to upgrade to HW3 because it couldn’t support self-driving (FSD) capability.
HW3 was produced in vehicles from 2019 to 2023 and Tesla switched to HW4 in 2024.
The catalyst is Tesla’s current FSD expansion in international markets. Previously, Tesla’s FSD was limited to North America, but over the last year, the automaker has been expanding FSD to China and now Australia and New Zealand.
However, the expansion is back-firing as HW3 owners are starting to realize that they will never get what they paid for.
In Australia and NZ, Tesla only launched FSD on HW4 vehicles with no clear plan for HW3, which the automaker already admitted won’t support unsupervised self-driving. The automaker appears to have only adapted its latest version of FSD for HW4 to the Australian market.
To add to the insult, with the launch of FSD in Australia, Tesla started to offer FSD subcriptions for $149 AUD a month for both HW3 and HW3 cars despite the software not being available for HW3.
Foreign hackers breached a US nuclear weapons plant via SharePoint flaws | CSO Online #security
A foreign threat actor infiltrated the Kansas City National Security Campus (KCNSC), a key manufacturing site within the National Nuclear Security Administration (NNSA), exploiting unpatched Microsoft SharePoint vulnerabilities, according to a source involved in an August incident response at the facility.
The breach targeted a plant that produces the vast majority of critical non-nuclear components for US nuclear weapons under the NNSA, a semi-autonomous agency within the Department of Energy (DOE) that oversees the design, production, and maintenance of the nation’s nuclear weapons. Honeywell Federal Manufacturing & Technologies (FM&T) manages the Kansas City campus under contract to the NNSA.
Wikipedia says traffic is falling due to AI search summaries and social video | TechCrunch #internet #ai
Wikipedia is often described as the last good website on an internet increasingly filled with toxic social media and AI slop. But it seems the online encyclopedia is not completely immune to broader trends, with human page views falling 8% year-over-year, according to a new blog post from Marshall Miller of the Wikimedia Foundation.
The foundation works to distinguish between traffic from humans and bots, and Miller writes that the decline “over the past few months” was revealed after an update to Wikipedia’s bot-detection systems appeared to show that “much of the unusually high traffic for the period of May and June was coming from bots that were built to evade detection.”
If you would like to propose any interesting article for the next ReHacked issue, just hit reply or “Leave a comment” link below. It’s a nice way to start a discussion.
Thanks for reading this digest and remember: we can make it better together, just leave your opinion or suggestions after pressing this button above or simply hit the reply in your e-mail and don’t forget - sharing is caring ;) Have a great week!
Dainius