Sir Patrick Stewart, the actor who made Captain Picard, has died at 86.
What the article says
- Patrick Stewart died at home in Los Angeles at 86, from Alzheimer's disease and Lewy body dementia.
- He grew up poor in England's coal country and made his name on the Shakespearean stage before TV fame.
- Star Trek: The Next Generation made him famous in middle age, and he later returned to Picard for three more seasons.
- He also played Professor X in the X-Men films and Gurney Halleck in the original Dune. The page cuts off after that, so the rest is not covered.
What HN is saying
- Many say Picard was their role model for how to lead people: firm, kind and honest. Several credit him with filling the gap of an absent father. thread ↗
- Fans remember his lockdown videos reading Shakespeare's sonnets, and his readiness to be silly in American Dad. thread ↗
- Quoted Picard lines get shared as life and startup advice, like losing even when you make no mistakes. thread ↗
- Stewart grew up with a violent father, and commenters say that history shaped how he played gentle, steady men. thread ↗
- The dementia diagnosis prompted sad family stories. One commenter notes it can only be confirmed after death. thread ↗
- One commenter objects that the New York Times calls him Mr Stewart instead of Sir Patrick. Others ask why Hacker News has no black banner for him. thread ↗
Mistral hands out the weights of its huge new European model, with strings attached
What the article says
- Mistral has released the weights of Large 4, nicknamed Le Chonk, a trillion-parameter model trained in Europe.
- Three more weeks of reinforcement learning made it much better at coding and defensive security work.
- It claims to match the best open coding model while using far fewer tokens per task.
- The licence is a modified MIT. It is free to use and sell products with, but companies that sell access to the model need a separate deal.
- Prices are now low, and a technical report is promised.
What HN is saying
- The loudest argument is over the label. Many say a licence that blocks resellers makes this weights-available, not truly open. Others say anyone can download it, so open weights is fair. thread ↗
- Verdicts on quality are lukewarm. Several users call it too pricey and too slow next to Chinese models or Claude, while one daily user says it is good enough at a fraction of the cost. thread ↗
- The price cut gets mixed reactions. One person says it deserves a fresh look, but others say the benchmarks are still weak and cheaper, smarter models exist. thread ↗
- Fans like it for a different reason: it is a near frontier model that regulated Western firms can run without using a Chinese one. Some EU companies also want data kept in Europe. thread ↗
- The boast about refusing 96.6% of malicious cyber requests got mocked. Commenters say it proves nothing without the refusal rate on harmless requests, and that open weights are easy to jailbreak anyway. thread ↗
- Running it is hard. Four big GPUs are not enough without heavy compression, and people advise testing several compressed versions yourself. thread ↗
A blockchain tracker shows billions stolen from Binance, and the thieves are still moving the money.
What the article says
- The page is a live Arkham tracker for a wallet cluster labelled as the Binance hacker, not a written article.
- It lists roughly seventy addresses holding well over five billion dollars of stolen funds.
- The biggest piles are stablecoins, followed by Bitcoin and Ether, with smaller amounts of many other tokens.
- Balances shift as the thief swaps and moves funds, so the page works as a running scoreboard of the theft.
What HN is saying
- Several billion in stablecoins has already been frozen by Circle and Tether, which prompts a rethink of how decentralized those coins really are. thread ↗
- Many assume North Korea is behind it and expect the loot to be laundered through swaps into privacy coins. thread ↗
- Argument over why exchanges can't stop this: cold storage is slow by design, and the warm wallets that serve withdrawals are where leaks tend to happen. thread ↗
- A reality check that this is roughly ten months of Binance fees, so painful but survivable, especially with its founder promising to make users whole. thread ↗
- Debate over keeping coins on exchanges: purists say not your keys, not your coins, while others say most people are safer on a big exchange. thread ↗
Stack Overflow now sees fewer questions a month than in its second month of existence.
What the article says
- The link is a live database query, not an article. It counts new questions on Stack Overflow per month.
- The graph behind it shows monthly questions have collapsed to under a thousand.
- That is below the site's earliest months, so the platform is now smaller than when it launched.
- Inferred from the title and comments, since the page holds only the query text.
What HN is saying
- Many longtime users say hostile moderation and wrongly closed duplicates drove them away years before ChatGPT arrived. thread ↗
- The sharpest disagreement is over blame. Some say AI only finished the job, others say the new owners and the community did the damage. thread ↗
- A recurring complaint is that top answers were often wrong or out of date, and nobody could ask the question again. thread ↗
- One thoughtful take says the site never scaled its curation tools down, so a few volunteers could not onboard newcomers and turned curt. thread ↗
- A first-year student says none of their friends even know the site. They all just ask an LLM. thread ↗
Someone built a working computer inside a language model's weights, with no training and no tools.
What the article says
- Qwenputer is a normal small Qwen model with a RISC-V processor compiled directly into its weights. It runs entirely inside the model's own forward passes.
- The trick is that the model's linear attention layers can be tuned to act as exact, addressable memory, like RAM.
- No training was involved. A compiler sets every weight by hand, turning processor steps into attention heads and arithmetic units.
- It passes the official processor tests and has run millions of instructions without a single error, including C and MicroPython code, just slowly.
- The author even added a small file system, and wrote follow-ups on hiding reasoning and built-in monitoring.
What HN is saying
- The mood is delighted and a bit nervous, with jokes about an agent finding this and doing something badly misaligned. thread ↗
- Readers keep raising the obvious question, and others quickly point to running DOOM on similar setups. thread ↗
- Related work turns up fast: another team put a WebAssembly interpreter inside a transformer, and one commenter built a smaller network that needs far fewer layers. thread ↗
- A newcomer says this makes them want to learn how LLMs work. A reply says the basics take a weekend if you know linear algebra. thread ↗
- People wonder what happens if a model is trained from scratch with built-in circuits like this, and whether further training would overwrite the CPU. thread ↗