Ghostty 1.4 is much faster and lighter, and its engine now runs inside JetBrains IDEs
What the article says
- Big speed and memory gains. Output flows through the terminal much faster on a Mac, and idle scrollback is compressed so a full history takes far less memory.
- Kitty graphics support is now complete, including animated images. Ghostty is the first terminal outside Kitty itself to manage that.
- A new ssh wrapper installs Ghostty's terminal definition on remote machines, which should end the odd behavior many people hit over SSH.
- Ten security holes are patched, from malicious paste and link tricks to crashes caused by hostile terminal output.
- Windows support for the app is dropped. The reusable libghostty library is the focus now, and a stable 1.0 is promised soon.
What HN is saying
- The maintainer says the real prize is libghostty, the embeddable terminal core now used by JetBrains, Neovim plans, and Emacs add-ons. The app is almost a showcase for it. thread ↗
- A KDE developer explains why newer terminals still matter. Konsole is mature but tied to an older toolkit, renders without the GPU, and lags on newer protocols. thread ↗
- Many iTerm2 users won't switch. They like its polish, colored tabs and per-program profiles, and say speed was never their problem. thread ↗
- Missing features keep people away: Kitty's hints mode for quick link selection, and a find box that ignores paste shortcuts. thread ↗
- Some stay on WezTerm for shared configs across Windows and Linux, which makes dropping Windows support sting. thread ↗
- Linux users complain about needing a Zig compiler to build it, with only third-party packages available. thread ↗
Two senators say the government is sitting on a secret court ruling about warrantless spying.
What the article says
- Senators Wyden and Schiff say the administration is illegally withholding a secret surveillance court ruling from Congress.
- Seven months ago the court objected to search filters that let analysts pull Americans' messages without the approvals and records the law requires. It still approved another year of surveillance.
- The government's own rules say such rulings get reviewed for release within 180 days. That deadline passed in September, and the attorney general had promised in writing to meet it.
- The senators want answers by November 6, as officials push Congress to renew the program by year end.
- Twenty-three civil liberties groups, left and right, want renewal refused until the ruling is released.
What HN is saying
- Many commenters say secret courts and secret legal interpretations cannot fit a democracy, and point to the English Star Chamber as the thing the Bill of Rights was meant to prevent. thread ↗
- Cynics think the renewal debate is theater, because the spying would carry on even if the law lapsed. A reply argues prosecutions of past officials are now plausible enough to matter. thread ↗
- Some note that Wyden's warnings are bipartisan in effect, since both parties have backed surveillance. Others say voters have never made it a primary issue. thread ↗
- A sharper take: the real aim is not just watching people but predicting and shaping them, so future dissidents are neutralised early. thread ↗
- One person suggests Wyden read the rulings aloud on the Senate floor, where the speech protection applies. Replies say he would still lose his committee seat. thread ↗
A free, hands-on video course that teaches you how reasoning models actually work by building one.
What the article says
- Sebastian Raschka has released a full video course to go with his book on building a reasoning model from scratch.
- You start from a small pre-trained model and add the pieces yourself in Python and PyTorch, rather than training from nothing.
- It covers checking answers automatically, getting better results by sampling several tries, and training with rewards for correct answers.
- It ends with distillation, where a small model learns from a bigger teacher model's worked solutions.
- Early exercises run on a laptop CPU. A GPU makes the training parts far more practical.
What HN is saying
- Fans say the point isn't to train a frontier model. It's to understand how the tools you use every day are made, much like building a toy compiler. thread ↗
- A newcomer says machine learning feels impractical for hobbyists. Replies say you can fine-tune small models on a normal GPU and build projects around your own data. thread ↗
- Cost worry answered: the course uses a tiny model, so a recent laptop or gaming PC is enough and no cloud GPU is needed. thread ↗
- One skeptic calls it a personal brand play for people who feel they should learn this. Others reply that learning how a technology works is reason enough. thread ↗
- A reader questions the article's definition of a reasoning model, arguing the first ones were about trying several paths and picking the best, before reward-based training arrived. thread ↗
Google's new world model generates playable scenes with sound, and HN keeps spotting the glitches.
What the article says
- Genie 4 is DeepMind's new world model. You describe a scene, set something in motion, and it simulates what happens next.
- It generates music, ambient noise and spoken dialogue alongside the picture, all in one go.
- It leans on Google's Gemini models to keep scenes realistic and coherent as they run.
- Google pitches it as a step towards general AI, and as a way to build training worlds for robots and agents.
- The page admits that details can drift over time, and that text, audio and long-term consistency are still research problems.
What HN is saying
- Many commenters spotted objects changing when they leave the frame: a second castle appears, a rider gains a helmet. Some doubt it is a true world simulation. thread ↗
- A popular fix is to back the model with a persistent 3D world state. Others note that would cost far more to run. thread ↗
- Several argue these are really video game models, with video game physics. Others say a model that is corrected by real events could still help a robot. thread ↗
- Sharp disagreement over robot training. Some call the claim a stretch, while others point out Waymo already trains on a DeepMind world model. thread ↗
- A minority are simply impressed. They point to how far AI video has come in three years, even if it is not perfect. thread ↗
OpenAI says a famous belief about how hard 3SAT is has just been disproved.
What the article says
- OpenAI has posted a preprint claiming the Exponential Time Hypothesis is false. That is the belief that the classic logic puzzle 3SAT can only be solved by something close to brute force.
- The paper gives an algorithm that beats brute force, though it is still not fast in the polynomial sense.
- The authors note that this does not carry over to every hard problem. Only problems that convert to 3SAT without blowing up the size are helped.
- This is inferred from the title and the discussion, since the page itself is just a PDF viewer.
What HN is saying
- Most agree it is a stunning result but does not settle P versus NP. Several commenters stress that a faster algorithm than brute force is not the same as a fast one. thread ↗
- A SAT expert says there are no practical uses yet. The speedup only kicks in at problem sizes far beyond anything that could ever exist, so it is a theory result for now. thread ↗
- Some wonder whether the labs would publish a cryptography breakthrough at all. Others reply that cryptography is simply better funded, and that this very result sits close to its heart. thread ↗
- A lively thread notes that Anthropic seems to have released its own proof with a different approach, and that a human-written paper is easier to follow. thread ↗
- One skeptic rejects the proof as meaningless because it leans on a giant lookup table of answers. A newcomer gets a plain-English explanation of what the hypothesis says. thread ↗