Giant tropical trees somehow pump water to the top with zero extra strain. Nobody expected that.
What the article says
- Scientists figured tall trees would struggle to pull water up to their highest leaves, making them more likely to die in a drought.
- New research on Malaysian rainforest giants found the opposite. The trees adapt their internal plumbing as they grow, so height brings no extra risk.
- Wider water-carrying vessels near the base and tougher, drought-resistant leaves let tall trees keep up with short ones.
- During a severe drought a few years back, the tallest trees grew just as well as smaller ones nearby.
- These giant trees hold a huge share of the forest's stored carbon, so knowing they can handle drought matters for climate predictions.
What HN is saying
- Commenters loved the plant biology tangent, swapping stories about growing chillies and weed and marveling at how adaptable plants are.
- Several pointed out that fog and moss can supply tall trees with extra water, not just roots, which may help explain the resilience.
- A recurring skeptical thread noted this doesn't explain why no tree on Earth grows past about 130 meters, so something else must cap height even if water transport isn't it.
- One person linked the sad history of the Nooksack Giant, a record-setting Douglas fir that was logged for lumber in the 1890s.
- A tangent on 'structured water' theories drew both interest and pushback, with critics calling it fringe science.
A startup got a big AI model running fast and cheap on AMD chips instead of Nvidia's
What the article says
- A company called Wafer got the GLM 5.2 language model running well on AMD's newer chips, which cost much less than Nvidia's comparable hardware.
- Nvidia's software tools usually make its chips much easier to run these models on, so AMD support tends to lag behind new model releases.
- The team had to hand fix a couple of bugs in the AMD software before a speed up technique called speculative decoding would even work.
- After tuning, they hit roughly two thirds of Nvidia's top speed while paying less than half the price.
- Their takeaway is that Nvidia's advantage is shrinking because it is really about software support catching up, not the chips themselves.
What HN is saying
- Commenters pushed back hard on the headline number, noting the model was compressed using a technique that usually costs some accuracy, and the post did not say how much quality was lost.
- Several pointed out the throughput figure was an aggregate across many simultaneous requests, not the speed any single user would see, which the post did not make clear.
- Some readers questioned whether the real world margins make sense, and a Wafer employee replied that margins run around forty percent, driven mainly by how fully the hardware is used.
- Others noted AMD adoption is more widespread than assumed, citing Meta and OpenAI as customers, though skepticism about AMD's software reliability persisted.
- A few said this kind of speed only matters for one person at a time and is not obviously useful for a real business serving many users at once.
Mistral built an AI that writes mathematical proofs, and it just found a real bug in an eight year old library
What the article says
- Mistral released Leanstral 1.5, a free model that writes formal proofs in the Lean language, meant to verify code and math are actually correct rather than just plausible.
- It aces the standard benchmark tests and solves the vast majority of a hard competition math set, beating a pricier rival model at a fraction of the cost.
- It can prove real properties like a self balancing tree staying fast, reasoning through millions of tokens on a single problem without giving up.
- Given real code, it also hunts for bugs by trying to prove or disprove properties about it. Out of dozens of repositories checked, it flagged several genuine, previously unreported bugs.
- The model and its code checking pipeline are open source and free to try through Mistral's own coding assistant.
What HN is saying
- The showcase bug, an integer overflow in a small Rust library, got picked apart. Commenters say basic fuzz testing or property based testing would have caught it in seconds, making it a weak advertisement for the tool.
- Someone tried the same bug hunt with a rival coding model and it found the same issue just by reading the code, suggesting the bug wasn't actually hard to spot.
- Others pushed back that finding bugs isn't really the point. Formal proof means showing an entire class of bugs cannot exist, which is a different and harder claim than spotting one flaw.
- A side debate broke out about Europe's AI competitiveness, with one commenter lamenting brain drain to the US and another shrugging that money isn't everything.
- A few noted the story is a rehash of an earlier post, and one comment reads as a thinly veiled ad for an unrelated tool.
A researcher found a bug in MSI's laptop software that lets any logged in user become admin in seconds.
What the article says
- A security researcher went hunting for bugs in gaming laptop software after already finding flaws in AMD's and Asus's tools.
- MSI's preinstalled Notebook Foundation service opens a channel that any signed in user can talk to, no admin rights needed.
- That channel can run commands, read and write system settings, and launch programs with full system level access.
- It was protected only by an old, weak encryption method, so figuring out the secret key was easy.
- The same trick can even be pulled off remotely over a local network if you already have valid login credentials.
- MSI patched it within two days of being told, though the researcher says the company still has not paid a bug bounty for any of his reports.
What HN is saying
- Commenters were pleasantly surprised MSI actually fixed the bug fast, since the company is usually more focused on image than substance.
- Several people wished the writeup explained MSI's actual fix, worrying a quiet patch might hide a new, still secret flaw.
- One thread turned into praise for reverse engineering vendor bloatware entirely and writing lighter, faster replacements yourself.
- A few readers were baffled that outdated, weak encryption from the 2000s is still shipping in 2026 software.
- Someone noted the darker point underneath the story, that companies keep taking free vulnerability research and paying nothing for it.
Someone made a Steam Controller crawl across a desk to its own charger, using vibration motors as legs.
What the article says
- An open source web app pilots a Steam Controller onto its magnetic charging puck by itself.
- It watches the desk through an overhead webcam and tracks the controller with computer vision.
- The controller moves by firing its internal rumble motors in an uneven pattern, basically walking itself across the table.
- When it gets close to the puck, it softens the vibrations so it docks gently instead of slamming in.
- It also reads the controller's battery data over WebHID to confirm charging and show battery level.
What HN is saying
- Commenters found the whole idea delightfully absurd, a controller inching along by vibrating itself toward a charger.
- One person pointed out the project's readme reads as obviously AI written, joking it buried the wildest detail, that the controller literally moves by buzzing, instead of leading with it.
- Someone compared it to Cycloramic, an old iPhone app that used vibration to spin the phone for panoramic shots.
- A few comments veered off into griping that Steam Controllers themselves are hard to buy, with one person quoting a 2027 estimated ship date.
- Another joked that neighbors hearing nightly buzzing noises would never believe it's just a controller charging itself.
Wikipedia deleted the Odin programming language's page, and the ensuing Twitter fight reveals how notability rules break down for internet-era subjects.
What the article says
- Wikipedia deleted the page for Odin, a programming language, saying it lacked reliable independent sources.
- Odin's creator and game programmer Casey Muratori argued the language is clearly notable, citing companies that use it in production.
- Wikipedia cofounder Jimmy Wales jumped into the thread, at first backing the deletion, then softening once the creator explained the sourcing problem calmly.
- The author's real target is different. Programming knowledge mostly lives in blogs, Discord, and YouTube now, which Wikipedia's old rules built for print journalism don't recognize as reliable.
- The piece also argues the creator was privately fine with the deletion but publicly played up an ideological persecution narrative for engagement, tying it to the same online right wing figures he follows.
What HN is saying
- Many commenters agree Wikipedia's real problem is asymmetry. Writing an article takes days, deleting it takes seconds, which discourages contribution.
- Several people push back on the framing that this is persecution, arguing a company's own website or founder isn't automatically a reliable source under Wikipedia's rules.
- A recurring gripe is inconsistency. Obscure decades old languages nobody used keep their pages, while actively used modern languages get deleted for lacking print coverage.
- One sharp correction: deletion discussions aren't decided by majority vote, despite the article describing it that way.
- A few commenters found Odin more obscure than the article implies, while others said it is well known in game and systems programming circles, so opinions split on how notable it actually is.
A tiny brain circuit lets thinking reshape what you see, and it might make AI more efficient too.
What the article says
- A Columbia researcher found early vision areas of the brain do not just relay images. They change how they process the same picture depending on what task you are doing.
- Her team built a simple neural network using the brain's own rules, with neurons that excite and neurons that inhibit.
- The key trick turned out to be inhibitory neurons suppressing other inhibitory neurons. That link carries context from higher brain areas down to raw sensing.
- Weakening that one connection broke the model's ability to switch tasks. The same effect showed up when researchers silenced those cells in real mice.
- She hopes these efficiency tricks, learned from the brain, could help build leaner AI that does not need training on the whole internet.
What HN is saying
- Several commenters, including working neuroscientists, pushed back hard on how far this generalizes, since the model is a fairly plain network trained on a simple copying task rather than a faithful brain simulation.
- One detailed comment reframes the finding: this is less about brains thinking and seeing together, and more about what vision looks like while you are actively doing a task.
- A recurring theme is that this kind of two way, back and forth wiring in the visual system has been known for a while, not a brand new discovery.
- There is a live debate over whether it even matters that real neurons fire in precise timed spikes, which these rate based models ignore entirely.
- A couple of asides connect inhibition failures to real world experience, from clumsy motor coordination to a Hunter Thompson quote about losing basic motor control.
The creator of Searx moved on, and his replacement quietly builds your browser into a private search engine.
What the article says
- SearXNG is a free search engine that pulls results from many other engines at once.
- It does not track or profile you, which is the whole point for privacy minded users.
- Setup involves installing your own instance and configuring which sources it pulls from.
- The project is open source and welcomes contributions.
What HN is saying
- Long time users are enthusiastic, many have self hosted it for years and would not go back to Google.
- The original Searx creator has moved on to a new project called Hister, which saves and indexes every page you visit for private, local search.
- A common complaint is unreliable results, some engines fail or get blocked and need captchas solved unless you pay for API keys.
- People are using it as a search backend for local AI models and agents, with tools like TinySearch trimming results to save tokens.
- One person pointed out that even small self hosted instances leave your traffic uniquely identifiable, so privacy is not absolute.
Two 17th century Dutch brothers invented the fire hose and it cut Amsterdam's fire losses by over 99 percent
What the article says
- Amsterdam in the 1600s was rich, packed with candles, curtains, tar, and other flammable goods, and prone to devastating fires.
- Early water pumps could only spray the outside of a burning building, so fires kept gutting factories and killing their owners financially and literally.
- Painter and inventor Jan van der Heyden and his brother Nicolaas built a better engine with a suction hose, a flexible nozzle that could reach inside buildings, and an air chamber for steady pressure.
- A dramatic 1673 warehouse fire proved the new engine worked, and the brothers were put in charge of overhauling the whole city's firefighting system with new gear, alarms, and incentives.
- The reforms worked. Fire losses in the 1680s dropped to a tiny fraction of what they'd been just a decade before.
What HN is saying
- Several commenters push back on the headline claim that Amsterdam invented the fire department, pointing to Ancient Rome's Vigiles Urbani, a firefighting corps dating to 6 AD.
- One notes that fully paid, professional municipal fire departments actually trace to Cincinnati in 1853, so it depends on how narrowly you define the term.
- A moderator note explains the title was edited after complaints that comments were nitpicking a catchy headline rather than engaging with the actual article, which people found genuinely interesting.
- One comment was flagged and hidden.
A developer's detailed, opinionated guide to building your own local AI rig, from cheap to absurd.
What the article says
- A developer shares his own setup for running top tier AI models at home instead of paying for Claude or GPT.
- He built a system with four high end graphics cards and specialized hardware to let them talk to each other fast.
- He lays out spending tiers, from a modest two card setup that handles solid mid range models to a much pricier rig that gets close to Opus level quality.
- The guide also covers running speech to text locally and the software setup he uses to wire everything into his coding workflow.
- Most of the piece is deep hardware tuning notes: BIOS settings, cable choices, and fixes to make the cards share data properly.
What HN is saying
- The big pushback is on cost. Commenters note the priciest build actually runs closer to fifty thousand dollars once every card is counted, dwarfing years of a normal subscription.
- Several people warn that shrinking these models to fit on home hardware quietly hurts quality. Compressed versions can look fine on quick chats but fall apart on long, complex tasks.
- Cheaper setups get real praise. Multiple commenters say a pair of older graphics cards, costing a few thousand dollars, run mid sized models fast enough to be genuinely useful.
- Some argue local models shine less for coding and more for smaller jobs like speech to text, where a compact model already matches or beats the big cloud tools.
- A recurring worry is that local model prices keep creeping up, since seemingly reasonable budgets always tempt people toward buying just one more card.