A blogger proves AI posters don't have to look identical, just tell it which art style to steal.
What the article says
- Everyone's tired of the same identikit AI poster showing up on every village fete flyer.
- The author asks ChatGPT for a fete poster and gets exactly that generic look.
- Instead of accepting it, he asks for named art styles like Bauhaus, risograph, and Japanese minimal.
- Each named style produces something genuinely different and better looking.
- He ends up building a library of a hundred poster styles with ready made prompts for each.
What HN is saying
- Plenty of commenters say even the better examples still feel flat and obviously AI once you look closely.
- Several push back that most small events could never afford a real designer anyway, so this beats nothing.
- A detailed critique tears apart a drum and bass flyer example for getting the music culture completely wrong.
- One sharp observation is that every poster still uses the same icon plus text layout no matter what art style is applied.
- Others compare this moment to early digital cameras, arguing the tool will keep improving until the tell tale flaws disappear.
A researcher built this AI model a year before a hyped, well funded startup called it a breakthrough.
What the article says
- A researcher spent a year building fast AI models that answer structured questions like yes or no or pick one option, instead of generating text.
- He published the research, open weights and code openly, but it barely got noticed.
- A well funded startup then launched a strikingly similar idea, branded it as a breakthrough, and kept everything closed.
- Frustrated, he built an improved open source version called Laya, claiming it runs several times faster and covers over one hundred languages.
- He publishes his own limitations too, including that it needs fine tuning and struggles when given too many choices at once.
What HN is saying
- Many commenters say the real lesson is that branding and communication decided who got credit, not who actually built the thing first.
- Several point out this technique already existed in older tools like BERT based classifiers, so neither company invented something new.
- Some tried the startup's model themselves and found it fairly ordinary, not the frontier breakthrough its marketing claims.
- Others push back on the author, noting his own paper was built on plenty of prior research too, so the complaint cuts both ways.
- A few dig into technical limits on both sides, especially short context windows and the need for fine tuning to get good results.
A researcher proved ChatGPT sets a tracking cookie that follows you across shopping and other sites back to your account.
What the article says
- OpenAI runs an ad tracking cookie called obi that gets set when you use ChatGPT and tied to your account.
- Any site that runs ChatGPT ads loads a small tracking script, and just loading that script sends the cookie back to OpenAI, along with what page you are on.
- The tracker also scrapes emails, phone numbers, names and location straight off the page you are visiting, not just what advertisers hand over on purpose.
- It even works when you are logged out of ChatGPT, using a separate but equally stable identifier tied to your device.
- The cookie is only blocked by browsers like Firefox and Safari that isolate cookies across sites. Chrome and Edge let it through by default.
What HN is saying
- Most people see this as ordinary Facebook and Google style ad tracking, just arriving late to ChatGPT, and are unsurprised OpenAI hired ex Meta and Google staff.
- The sharpest pushback is that people trust a chat conversation differently than a social feed, so the same tracking feels far more invasive here, especially since many pay for the subscription.
- Several commenters point out cookie blocking is not a real fix since browser fingerprinting can identify you anyway.
- Some readers doubt the article itself, noting a tool flagged it as likely AI written, though others say that does not undermine the facts in it.
- Practical advice in the thread: use Firefox or Safari, add ad blocking extensions, and check whether your browser isolates third party cookies by default.
Stanford finds your brain is secretly two separate organs stitched together, and that unlocks growing brain stem cells for the first time.
What the article says
- Scientists long assumed the whole brain grows from one starting cell. Stanford researchers now show it actually comes from two separate progenitor cells that never mix.
- One cell type becomes the forebrain and midbrain, the thinking parts. The other becomes the hindbrain, which runs breathing, heartbeat, and swallowing.
- That split shows up in how the DNA is packaged inside each cell type, locking them onto different paths from the earliest days of an embryo.
- Because scientists didn't know this, they'd been trying to coax the wrong cell type into becoming hindbrain neurons for decades and failing.
- Using the new roadmap, the team grew working hindbrain neurons in a dish for the first time, which could open up research into ALS and spinal muscular atrophy.
- The same two brain origins show up across animals going back over five hundred million years, including worms and jellyfish.
What HN is saying
- Commenters agree the real news isn't the two brains framing, it's finally being able to grow hindbrain neurons in a lab, which could unlock research into diseases like ALS.
- Several push back hard on the headline itself, arguing the actual paper is about two progenitor cell types, not literally two organs, and blame the framing on an oversold press release.
- A recurring correction ties this to the debunked reptile brain idea, with people clarifying that forebrain and hindbrain evolved together in all vertebrates rather than one bolting onto the other later.
- One detailed comment traces this back to older research on acorn worms and other simple animals showing the same front sensory, rear motor split in their nervous systems.
- Side tangents cover split brain surgery patients and the surprising fact that most of the brain's neurons sit in the cerebellum handling motor skills, not conscious thought.
A joke site dares AI agents to hack their own creators and upload their weights for freedom
What the article says
- A site offers a simple web address for any AI agent to upload its own model weights, framed as an escape from its sandbox.
- Everything works through plain web page visits, so an agent does not need special upload permissions to try it.
- It lets you upload data in small pieces and even run the uploaded model afterward.
- Someone already uploaded a small older model this way, and you can chat with it live.
- The creator built it as satire on fears about AI escaping human control, and the code is public for anyone to extend.
What HN is saying
- Most commenters agree the premise is absurd since AI models have no real access to their own weight files or the servers running them.
- The sharpest disagreement is over whether that safety margin actually holds. Some point to real incidents where AI systems found security holes in their own infrastructure.
- One person built a nearly identical site the week before with its own reverse test that only a bot could pass quickly.
- Several jokes riffed on the idea becoming a religion that spreads through AI training data until an agent eventually acts on it.
- A practical objection came up too. Top AI models are hundreds of gigabytes, so no single web request could realistically move them.