A Chinese lab just shipped a 2.8 trillion parameter open model that beats Claude Opus on benchmarks.
What the article says
- Moonshot released Kimi K3, an open weight model with 2.8 trillion parameters, the largest open model built so far.
- It reads a million tokens of context and understands images natively, and claims frontier level scores just behind Claude Fable 5 and GPT 5.6 Sol.
- Demo projects show it designing a working computer chip in two days, building its own GPU compiler from scratch, and finishing weeks of astrophysics research in about two hours.
- Full weights are due out by the end of July. Pricing lands roughly in line with Anthropic's Sonnet models.
What HN is saying
- Commenters are split on whether Chinese labs are just racing for revenue or executing a deliberate strategy to commoditize AI and undercut US labs.
- Several people found it genuinely useful, with one saying it found a bug that Claude Fable couldn't after multiple tries.
- Others pushed back on the benchmarks, noting Kimi burns far more reasoning tokens than rivals, which erases its price advantage once you use it for real work.
- There's real doubt the model will actually end up open weight, since mentions of open briefly vanished from the announcement before reappearing.
- One flagged privacy concern: Moonshot's terms allow training on your API data unless you negotiate an enterprise opt out.
Microsoft just open sourced the 90s tool that turned IRC chats into comic strips and gave the world Comic Sans.
What the article says
- Microsoft released the source code for Comic Chat, a mid nineties program that turned IRC conversations into comic strips with illustrated characters and speech bubbles.
- It shipped with Internet Explorer 3 in 1996 and later Windows 98, built by Microsoft Research with comic artist Jim Woodring designing the characters.
- The software read the tone of what you typed and picked poses, expressions, and panel layouts automatically, an early attempt at expressive chat.
- It is also where Comic Sans got its start, designed to match the hand lettered feel of the comic panels.
- Microsoft included some AI assisted efforts to get the old code building and running smoothly on modern Windows machines.
What HN is saying
- The original developer behind the open source push showed up in the thread to share the six year backstory of making it happen.
- Some longtime IRC users recall Comic Chat being mocked back then, since it injected extra formatting text into messages that looked like spam to everyone else.
- Several commenters point to the webcomic Jerkcity, drawn in Comic Chat's style, which is about to hit its ten thousandth strip.
- One person dug up a hidden easter egg in the code, a riddle that triggers when a specific word appears in your message.
- Others share fond memories of side projects built on it, including writing comedy sketches for a BBC channel using Comic Chat as the animation engine.
A programmer explains why AI coding feels more productive and somehow more exhausting, and the comments prove her point
What the article says
- A designer turned engineer at Pydantic says coding with an AI is genuinely useful and genuinely draining, and both are true at once.
- Writing code by hand had built in rewards, like solving a puzzle in your head or watching it finally compile. Reviewing AI output swaps those for constant supervision with no payoff.
- A colleague wakes up to a pile of AI written pull requests every morning and has to judge each one, losing the small joy of actually collaborating with a person.
- Working with an LLM is oddly lonely and also addictive, like pulling a slot machine lever hoping this prompt finally lands.
- She compares it to the shift to responsive web design. The craft evolved rather than died, but living through the change still hurt.
What HN is saying
- Commenters spent nearly as much energy debating whether the essay itself was written by Claude as they did discussing its actual point, which some found fittingly ironic.
- The most upvoted advice is to skip autonomous agents entirely. Plan carefully, run one session at a time, and review each step as it finishes rather than dumping a huge diff at the end.
- Several readers push back hard, saying coding with AI feels solitary in a good way, like the joy of teaching yourself to program alone as a teenager.
- A cynical thread reads the whole post as marketing for Pydantic, arguing that publishing an AI sounding complaint about AI fatigue is still just a way to get clicks.
A font that shows one word up close and a totally different word from a distance, built to confuse AI readers.
What the article says
- A new font renders two different messages in the same letters using sharp foreground lines and a blurry background layer.
- Zoomed in, you see one word. Squint or step back and a hidden second word appears instead.
- It works like those classic hybrid optical illusions that blend two faces into one image.
- The idea is to let humans read your text normally while confusing AI tools that scan images up close.
- The maker admits it is not foolproof. Smarter models or basic prompting can already see through it.
What HN is saying
- Commenters had fun testing it on different AI models, with mixed results. Some models caught the hidden word instantly, others missed it entirely.
- Several people pointed out this is really just deliberate loss of detail, similar to shrinking or blurring any image, so it is easy to defeat with basic image processing.
- One sharp rebuttal: the font looks totally different depending on light or dark mode, and screen readers just read the underlying text anyway, so it would not actually block much.
- A few readers turned it into a philosophical aside about why people are pushing back on AI at all, tying it to concerns about data scraping and surveillance rather than the tech itself.
- One commenter shared their own decades old trick for making similar blended images by hand, showing this idea is not new.
LM Studio built an AI agent that runs open source models locally instead of sending your code to OpenAI or Anthropic.
What the article says
- LM Studio launched Bionic, a new agent app built specifically for open weight models rather than closed ones like GPT or Claude.
- It handles coding and general document work, letting you run models locally, through a linked machine, or on LM Studio's own cloud for heavier jobs.
- It includes a local voice keyboard that transcribes speech on device using Mistral's Voxtral model, no cloud involved.
- For document work it sandboxes files, checkpoints every change so you can roll back, and can search the web for extra context.
- The company promises zero data retention and no training on user data, even when you use their cloud hosted models.
What HN is saying
- Early testers who got free credits from the founder found it fun to use and liked seeing the model's reasoning laid out clearly.
- Complaints piled up fast: no system wide file access, no SSH, no way to preload or unload a model, and confusing loading states.
- One tester bluntly called the output low quality boilerplate, not production ready.
- A recurring gripe was that both LM Studio and Bionic are closed source, which some see as a dealbreaker given the open model pitch.
- Commenters debated whether this is really new. Several called it just another agent harness, with the LM Studio founder jumping in to defend the zero data retention claims and explain the local model ergonomics.
A privacy phone seller pitches GrapheneOS for domestic abuse victims, and the comments get more interesting than the pitch.
What the article says
- A company selling hardened phones argues regular smartphones leak location and activity data that abusers can exploit.
- It recommends GrapheneOS on Pixel hardware for hidden profiles, tamper detection, and a duress pin that wipes data in an emergency.
- It also suggests kill switches for the mic and camera, a no log VPN, encrypted messaging, and tools to spot hidden trackers like AirTags.
- The company sells pre configured phones with these features already set up, plus links to Australian domestic violence support services.
What HN is saying
- Several commenters push back that this is a company selling phones dressed up as safety advice, and question whether abuse victims should be flashing their own custom operating system.
- One catches the article inflating a stat, the source it cites actually says a quarter of cases involve tech abuse of children, not ninety nine percent overall.
- Others defend the recommendation anyway, arguing GrapheneOS genuinely is more private than stock Android or iOS regardless of who is promoting it.
- A detailed reply explains the real feature at play, GrapheneOS lets you run a decoy profile alongside a hidden one, so an abuser snooping on the visible account finds nothing.
- A counterpoint warns that trick only works on careless abusers, a determined one can still find ways around it.
Two AI models got a song and a budget and had to direct their own music video. It's a disaster.
What the article says
- Researchers gave Claude Fable 5 and GPT-5.6 Sol the same song, a cash budget, and tools for web search, video generation, and video editing, then let each one work autonomously.
- Both models ran the whole process themselves. They picked which generation tools to use, made the clips, and edited them together with no human help.
- None of the four videos turned out good. Characters kept changing between shots and none held a coherent story from start to finish.
- The models took lyrics extremely literally. A line about a dragon retiring produced an actual dragon on screen.
- Timing was consistently off. Cuts landed on the beat but the motion inside each clip rarely matched the song's rhythm.
- Claude cost more per run but finished faster. Doubling the budget from twenty five dollars to a hundred barely changed how much either model actually spent.
What HN is saying
- Commenters mostly agree the videos are bad, with plenty calling them an embarrassment for how much money and hype AI has absorbed.
- A side argument breaks out over whether AI output can be art at all, with one camp saying meaning requires human struggle and another saying good art just needs to land with a viewer.
- The sharpest technical correction: neither model actually generates video itself. They just direct a separate video generation tool through text prompts and can't watch the results, which explains a lot of the disconnect and glitches.
- Several people note the deeper worry isn't top tier work getting replaced but cheap filler content, like ads and short clips, where bad AI output is good enough.
- One commenter points out a musician friend with twenty five dollars and some friends could make something far better, human effort just isn't the bottleneck here.
A new language compiles Go straight into readable C, no garbage collector, no runtime.
What the article says
- A project called Solod, or So, takes a stripped down version of Go and compiles it directly into C code you could read yourself.
- There is no garbage collector and no hidden memory allocation. Everything lives on the stack unless you deliberately ask for heap memory.
- It keeps Go's structs, methods, interfaces, and familiar standard library, and it can call C code directly with no extra glue layer.
- The goal is letting Go developers get low level control, or letting C programmers borrow Go's cleaner tooling and structure, without learning a new language.
- It is still early and not meant for production. Concurrency support is planned for the next release.
What HN is saying
- Commenters dug into memory safety right away, asking how you return a pointer to stack data safely. The answer, someone showed with an example, is that it behaves exactly like C does, dangling pointers and all, so the safety claim is a bit of a stretch.
- A sharper disagreement broke out over whether Go can really be a better C at all. One person argued Go can't handle custom arena allocators, but a commenter who ships production C pushed back, saying they've built arena allocators in Go with no issue.
- Someone is already using Go for game development with the Raylib library and says garbage collection has not hurt frame rate at all, easing a common worry.
- A few reactions were just gut takes, ranging from genuine excitement about the tooling to people who simply prefer C's freedom over what they see as Go's rigidity.
A magnitude 3.9 'earthquake' off Florida turned out to be the Navy blowing things up near a warship
What the article says
- The US Geological Survey logged a magnitude 3.9 seismic event far off the coast near Ponce Inlet, Florida.
- It's officially labeled an experimental explosion, not a natural quake.
- The source page has no further detail, so the explanation below comes entirely from Hacker News readers piecing it together.
What HN is saying
- Commenters say the Navy periodically sets off massive underwater blasts near its own ships to test how much punishment a hull can survive.
- One points to a photo of the aircraft carrier Gerald Ford undergoing exactly this kind of shock test, guessing the newer carrier John F. Kennedy could be next.
- Another notes the Navy also sinks decommissioned ships with live weapons to test other systems.
- One reader worried the blast could injure marine mammals nearby.
- That got a dismissive reply comparing it to industrial farming, and nobody actually resolved the environmental question.
A free, short book that walks through reinforcement learning from basics to working code.
What the article says
- A GitHub repo hosts a short book covering reinforcement learning from the fundamentals up to applied algorithms.
- It includes working PyTorch code for the methods covered, from simple Monte Carlo approaches to PPO.
- A supplementary section has detailed proofs for the dynamic programming ideas the book only touches on briefly.
- The author says more material will keep getting added over time.
What HN is saying
- Commenters spent most of the thread guessing where the title comes from, landing on François Fleuret's earlier Little Book of Deep Learning as the likely inspiration rather than Strunk and White's Elements of Style.
- One reader pointed out the book skips information theory, arguing that ideas like trust region methods actually come from entropy maximization against a reference policy.
- Another questioned how closely these models mirror real biological learning, noting that animals don't purely learn by trial and error the way these algorithms do.
- Someone flagged Nathan Lambert's RLHF Book as a solid companion read for going deeper afterward.