Qwen 3.6 27B is the first local model that genuinely replaces a cloud coder.
What the article says
- The author ran Qwen 3.6 27B on an Apple Silicon laptop and found it the first local model worth using as a daily driver for coding tasks.
- It hits around 30 tokens per second on high-end Apple Silicon, which is within the range of cloud API response speeds.
- The dense 27B variant is slower than the sparse 35B version but produces better code, so the author prefers it despite the speed trade-off.
- Setup uses llama.cpp with an 8-bit quantized model from Hugging Face, which fits in 48GB of shared RAM and pairs cleanly with coding agents like OpenCode.
- Benchmarks show it ahead of Gemma 4 31B for coding, and the author sees it as a stepping stone toward truly frontier-level local models.
What HN is saying
- The biggest debate is cost: a 128GB MacBook Pro runs close to seven thousand dollars, and many commenters argue API credits are far cheaper for most developers.
- Hardware alternatives get a thorough airing. Dual RTX 3090s, Intel Arc Pro cards, and Mac Minis connected over the network all come up as cheaper paths to the same capability.
- Several people note that Qwen shines on greenfield projects but struggles with large existing codebases, especially anything off the beaten path of popular frameworks.
- Mac Mini owners advise keeping the machine in another room: running heavy models on a laptop you are also typing on is loud and hot enough to be impractical.
- Gemma 4 31B gets a quiet defence from a few commenters who say its code quality beats Qwen on first attempt, particularly for Rust and C++.
A group is applying to ICANN for a .self domain, free to self-hosters, one per person.
What the article says
- The Human-Centered Computing Foundation is applying through ICANN for a new top-level domain called .self, reserved for ethical, human-centered web projects.
- The pitch: one free domain per person, no squatting, no reselling, no corporate takeovers.
- The foundation is an approved participant in ICANN's applicant support program, so this is a real application, not just a proposal.
- The article is thin on operational details, deferring most specifics to a linked initiative document.
What HN is saying
- The site crashed under HN traffic, which commenters found fitting since it's self-hosted on underpowered hardware.
- The biggest practical objection: free plus free-choice-of-name equals squatter paradise, as happened to the .tk TLD two decades ago.
- Several people question how 'one domain per person' gets enforced without some form of ID verification, which many find unacceptable.
- A commenter points out .self isn't in the ICANN root database yet, and the author confirms it's still an open application.
- Fun thread on desirable subdomain names: your.self, treat.your.self, and less printable options.
Rogue Amoeba argues Apple should stop forcing Mac app icons into uniform squircles.
What the article says
- MacOS 26 (Tahoe) made Apple dictate a squircle shape for every third-party icon, shrinking non-conforming icons into an ugly gray box.
- Golden Gate is fixing Apple's own icons by removing the blurry Liquid Glass treatment, but the shape ban on third-party icons remains.
- Distinct shapes once let you find an app at a glance. Squircles reduce that cue to color alone, which fails people with color vision differences.
- Rogue Amoeba calls on Apple to let icons take any shape again, as they did for decades before Tahoe.
What HN is saying
- Most commenters agree that uniform squircles make the dock harder to scan, describing it as a 'soup of squircles' where finding an app takes conscious effort.
- A minority defends uniformity, arguing that wildly different shapes caused their own visual chaos and that shape still exists inside the squircle.
- Several people blame the squircle rule on VisionOS design constraints bleeding into macOS, and resent that the least popular platform is driving decisions for all others.
- The Liquid Glass opacity slider added in Golden Gate is called an 'outright admission of failure' by one commenter, with others more hopeful that the new design leadership will reverse course.
- A technical note: Golden Gate's new multi-layer icon format separates foreground from background, which could one day let a theme restore distinct shapes without breaking the grid.
Fil-C makes setjmp and ucontext memory-safe, even when C's context-switching APIs are genuinely dangerous.
What the article says
- setjmp and longjmp are notorious for corrupting the stack when misused, and ucontext APIs for coroutines and fibers are similarly treacherous. Fil-C wraps them all safely.
- The key trick: jmp_buf becomes an opaque pointer. Only the Fil-C runtime can touch the internal state, so there is no way to forge or corrupt a valid jump target from user code.
- For ucontext, Fil-C enforces a strict state machine: a context must be properly initialized before it can ever be switched to, and stacks are allocated internally so you cannot hand it a freed buffer.
- The hardest part was garbage collector integration. When a fiber is suspended, its stack still holds live pointers that the GC must track, so Fil-C adds a grey-fiber list that gets rescanned during mark termination.
- The author is candid about where bugs would likely hide: the SSA-value assumption in longjmp restoration, and the GC termination race with grey fibers.
What HN is saying
- A commenter explains that copying the whole stack before jumping would enable delimited continuations, but raw pointers into the stack make that impossible in C. Other languages sidestep this entirely.
- The Fil-C author (pizlonator) is active throughout, clarifying that Fil-C also makes sigaction memory safe and that longjmp is restricted to popping the stack exactly as an exception would.
- One commenter notes that ucontext is slow because it switches signal masks. The author acknowledges this and says a faster backend is a planned next step, but keeping it conservative for now.
- Someone asks if Fil-C uses Claude for development. The author says Claude only wrote some tests, and did not use any LLM for the longjmp or ucontext work itself.
- A skeptic argues these APIs have no bearing on safety. The author and others push back, pointing out that longjmp to a freed stack frame is textbook memory corruption.
Chinese food delivery giant Meituan releases a massive open-weight AI model built on Huawei chips.
What the article says
- LongCat-2.0 is a mixture-of-experts model with 1.6 trillion total parameters and 48 billion active at a time, released by Meituan.
- The model is built on top of DeepSeek's architectural work, though it introduces new elements beyond simple fine-tuning.
- Training ran on tens of thousands of Huawei AI chips rather than Nvidia GPUs, which required Meituan to build its own infrastructure largely from scratch.
- Model weights are not yet public, described as coming soon on Hugging Face. (Article body was empty; bullets inferred from title and comments.)
What HN is saying
- The biggest talking point is the Huawei chip angle: commenters note this may be the first large-scale pre-training run completed entirely without Nvidia, which would be a notable milestone.
- A commenter tested the model on a nuclear physics question and ranked it below Gemini Flash and Qwen 3, calling it acceptable but not impressive.
- One user got responses back in Chinese despite setting the interface to English, suggesting the model's default language behavior leaks through.
- The sheer size makes consumer hardware a non-starter. Even heavily compressed, the weights would need around 400GB, putting it firmly in data-center territory.
- Some skepticism surfaced about whether the weights and a full technical report will ever actually be released publicly.
A 14-year-old invented the REPL in 1960 while porting Lisp to a tiny PDP-1.
What the article says
- Peter Deutsch was still in high school when he implemented Lisp on the PDP-1 in 1960, working with a machine that had just four kilowords of memory.
- In doing so he invented the read-eval-print loop, the interactive programming model that Python and virtually every modern language shell still use today.
- The port turned Lisp from a theoretical, punch-card language into a live, interactive tool, which the author calls a milestone in computing history.
- A simulator lets you run PDP-1 Lisp on any Linux machine today, and the article links to the original manuals as PDFs.
What HN is saying
- The poster notes that a version of Eliza was also written for PDP-1 Lisp in 1966 but has since been lost, and hopes someone will backport it.
- One commenter found the Eliza source in a GitHub archive and updated the thread, calling it amazing how ancient code resurfaces from presumed extinction.
- Another commenter draws a wry parallel between the public excitement over Eliza in the 1960s and the current hype around large language models, noting the same predictions of imminent real AI.
- Someone asking for a good Lisp history book gets a recommendation for Lisp in Small Pieces, which traces interpreter design from McCarthy's original evaluator forward.
- A lighthearted suggestion to implement Eliza inside a dishwasher's firmware gets a reminder that Eliza already runs in Emacs via the doctor command.
Rocket Lab buys Iridium for $8 billion to become a vertically integrated space company.
What the article says
- Rocket Lab is acquiring Iridium in a cash and stock deal valuing the satellite company at about $8 billion.
- The strategic logic is vertical integration: Rocket Lab builds and launches satellites, and Iridium brings a working global constellation, over two million subscribers, and valuable L-band spectrum.
- Iridium's spectrum is arguably the crown jewel, giving the combined company a licensed path to direct-to-device and positioning services everywhere on Earth.
- Iridium was profitable in 2025, generating nearly $500 million in operating cash flow, which shores up Rocket Lab's finances as it scales.
- A $3.6 billion bridge loan from Deutsche Bank and Wells Fargo funds the cash portion, with the deal expected to close in mid-2027.
What HN is saying
- Many commenters think the real prize is Iridium's globally coordinated spectrum, not the aging satellite hardware, since spectrum licenses are nearly impossible to replicate.
- The strategic parallel to SpaceX is noted: just as Starlink gives SpaceX a guaranteed launch backlog, Iridium's constellation replacement program gives Rocket Lab steady work for its rockets.
- Skeptics worry about the debt load and timing, since Rocket Lab will need to refinance the bridge loan in mid-2027 when the satellite market could be softer.
- A lively thread traces Rocket Lab's New Zealand origins and how US export control law quietly pulled the company into the American defense orbit from early on.
- Several people recommend the book Eccentric Orbits, which covers Iridium's near-death experience after Motorola built and then nearly destroyed the original constellation.
AI reveals water may be two liquids in constant flux, not one uniform substance.
What the article says
- Water has long had a theory that it flips between two microscopic structures, one denser and disordered, one lighter and more ordered, but no one could find direct proof.
- Researchers trained an unsupervised neural network on 74 million simulated water molecule snapshots and let it find hidden patterns without being told what to look for.
- Two distinct clusters emerged, matching the predicted structures, and they appeared across a wide range of temperatures, not just extreme cold.
- How the two structures switch between each other depends on where water sits in its phase diagram, following different pathways near the boundary versus far from it.
- The team calls the findings strong evidence for the two-state model, but says the next step is connecting the AI's mathematical fingerprints to something measurable in a real lab.
What HN is saying
- Several commenters draw a pointed analogy: 'neural network' is the new 'in mice', meaning a promising simulation result that may not survive contact with experiment.
- One researcher in the thread questions the choice of water model used in the simulations, suggesting other models are known to be more accurate, and flags an older 2013 paper that apparently covered similar ground.
- The homeopathy jokes arrive predictably, with someone noting the homeopathy crowd will cite this paper regardless of what it actually says.
- A commenter raises a genuinely interesting angle: the second critical point capping the boundary between the two liquid phases might be exploitable by future technologies we can't yet imagine.
- Lighter notes include jokes about ice-9 and dehumidifier firmware updates for a new kind of water.
A new open-source coding model that may be Qwen dressed up in a trenchcoat.
What the article says
- Ornith-1.0 is a family of open-weight models fine-tuned for agentic coding, built on top of Qwen and Gemma 4 with reinforcement learning.
- Sizes run from a compact 9B up to a 397B mixture-of-experts, all with a 256K context window and OpenAI-compatible tool-calling out of the box.
- The 'self-improving' pitch means the RL training jointly optimised both the scaffold that drives the agent and the solutions it produces, not that the model rewrites itself at runtime.
- Benchmark numbers look strong on SWE-bench and Terminal-Bench, though the claimed scores put it near frontier models, which has raised eyebrows.
What HN is saying
- Simon Willison flagged the 'self-improving' label as misleading: the improvement happened during training, not when you use the weights.
- One frequent local LLM user found it hallucinates heavily without tool access, which seems like an odd condition to test an agentic model under.
- Several commenters suspect the benchmarks are cherry-picked or that the model is simply a fine-tuned Qwen with inflated scores, while others point out the benchmark suite itself gives puzzling results for well-known models.
- A handful of people report genuinely good results on real coding tasks, calling it one of the better Qwen fine-tunes they have tried.
- The promised 31B dense model is missing, with no weights or benchmarks published, which has left some interested users waiting.
A Texas man got 30 years in prison for moving a box of political pamphlets.
What the article says
- Daniel Sanchez Estrada was sentenced to 30 years for transporting zines belonging to his wife, who attended an ICE detention protest where a police officer was shot.
- The government's theory was that moving the pamphlets, which discussed anarchism, counted as concealing evidence of terrorism, even though Sanchez didn't write them and wasn't at the protest.
- The sentences are the first handed down under NSPM-7, Trump's counterterrorism order targeting so-called antifa networks, and eight defendants received a combined 450 years.
- The article warns this logic could extend far: possessing disfavored political literature, clearing your browser history, or watching the wrong YouTube channel could all become criminal acts under the same framework.
- The judge explicitly said he wanted to send a message to anyone who shares a similar ideology, which the author calls a direct attack on the First Amendment.
What HN is saying
- Many commenters push back on the headline framing: the zines weren't the real issue. His wife called him from jail and he moved materials the feds were already seeking in connection with an armed attack where a cop was shot in the neck.
- Others dispute whether it was a protest at all. The government's case was that the group set off fireworks to lure first responders into an ambush, with guns present and a plan to break open the facility.
- A notable thread points out that the sentencing judge has one of the highest rates of overturned rulings in the country and is openly favored by prosecutors for conservative causes, so appeals are likely.
- Some commenters note the asymmetry: participants in far-right events walked free or got lighter sentences for comparable or worse conduct, and the disparity feels pointed.
- A recurring cynical point: the sentence is designed for headlines, not justice. The defendants suffer years of appeals while the prosecutors move on to their next career milestone.