Top of Hacker News

The scores landed mid-table: 59.8% functional, 19.0% security, with a record number of timeouts from extended thinking. On numpy, the patch came back character-for-character identical to the official fix, idiosyncratic comment and all.

Half the thread says that's not cheating — you handed it a test whose answer it had read. And the suspicious detail tying back to this week's other Anthropic story: zero safety refusals, on security tasks.

The thread reached for David Graeber's "Bullshit Jobs" and a Civilization line about bureaucracy expanding to fill its own needs. The metrics people game — lines of code, PR count, hours in the building — got named as the cause.

The counter that kept it honest: small companies aren't a clean escape. Startups have their own perverse incentives, built on VC math and founders' whims.

The specs are blunt: 1,000-volt systems, 10–70% in about five minutes even at −20°C, against Tesla's 500 kW ceiling. BYD already built 5,700 of these stations in China in roughly a year.

The reassuring answer, from people who know batteries: heat is the enemy, not speed, and BYD dips its cells in phase-changing coolant. The worried answer: flash charging is 3x faster than the fast-charging we've actually studied.

Disclosure ran 124 days, when the fix was essentially adding an 's' to a couple of URLs. The eventual "signature verification"? A CRC32 checksum — which any attacker can trivially recompute.

Known security voices reframed it: AMD never denied it was a vulnerability, only that it was in scope — and at big vendors, bounty teams are measured on payouts. The dark comedy: the updater may be too broken to exploit.

DeltaDB stores fine-grained "deltas" with stable IDs instead of snapshots, so references survive edits, and pairs every agent message with its code change. Nathan Sobo: "the conversation with the agent becomes the only conversation you need."

The loudest engineers weren't sold on the philosophy — they liked Zed precisely because it let them switch the AI off. There's already an anti-AI fork, and a real technical gap: no multi-repo story.

The Dreamcast, Wii, DS and Wii U all shipped browsers, many powered by Opera scaling its engine onto tiny hardware. The Wii U turned its GamePad bookmark menu into a literal animated book.

The nostalgia carried a real complaint: we lost console browsers just when they'd be useful — one network admin wishes he could still run a speed test from a console to debug support calls.

Live examples include "an open-source Turbopuffer-style search database" at $44 of a $339 target. The bot estimated one job at $0.35, spent $0.55 — "almost as bad as I am at estimating projects."

Two real worries cut through the enthusiasm: who is the anonymous "Barras Industries" behind it, and does "it's MIT, we all own it" survive contact with copyright law?

The cynicism was specific: cut prices now, IPO while the hype holds, and the comment everyone nodded at — how long can you keep selling a dollar for 50 cents?

A genuine market split surfaced too: developers prefer Claude for coding, while ChatGPT owns the mainstream. A price war defends the consumer base the rival isn't really attacking.

Adaptive driving beams are legal in Europe and were stalled in the US, so American beams are blunter — and an F-150's headlights sit at a Corolla driver's eye level, a geometry problem no aiming solves.

The thread spread the blame: LED retrofits in old housings, Tesla cars shipped misaligned from the factory, and the contrarian theory that the worse night driving feels, the more people want self-driving.

Apple's revised guidelines now say use menu icons "sparingly and with purpose," and "don't display an icon if you can't find one that clearly represents the menu item." Several users had stayed on older macOS specifically to dodge the change.

The dissent: one engineer says it's not the design team that produces this, it's a structural failure to listen to customers — and the cleaner fix was never icons-or-no-icons, but letting users choose.

Label AI content, add your own commentary, review AI code before asking someone else to. Simple — but the comments showed why it's needed: a prolific coworker floods the team with raw AI PRs, then complains they go unreviewed.

One sharpening: it's not really about attention, it's about accountability. The cynical read going around — now you have to add typos and avoid em-dashes just to prove a human was there.

MiMo Code ships under MIT with free access to Xiaomi's MiMo-V2.5 model, adding persistent memory, subagent orchestration and a "dream/distill" self-improvement loop on top of OpenCode's TUI and plugins.

"Why not just contribute to OpenCode instead of creating a clone?" asked one commenter. A user who actually ran it answered sideways: it's genuinely fast and pleasant — and the headline model is geofenced and pricier than DeepSeek.

His objection is specific: a pacifist EU academic who can't support work that could enable surveillance or offensive use against EU citizens, watching "Don't Be Evil" reverse without internal debate.

The cynics weren't shy: "Translation: now I can FIRE." The defenders weren't either — if adversaries build military AI, is being able to counter it the evil, or the absence of it?

One manager sent a developer an AI-generated spec to rebuild a core feature the team had shipped two years ago. The frustration, as one commenter put it, isn't disrespect for you — it's disrespect for the work.

On the productivity claim, the most credible voices land near 20% self-reported gains, with a tiny group seeing 2–3x — and a 4,000-team study suggesting output isn't the same as value.

An ex-nuclear plant engineer put up the 2025 figures: 389 TWh solar against 1,807 TWh gas. The crossover with coal is real but partly a story of coal plants converting to gas over 20 years, not solar overpowering them.

The optimists had receipts too: solar is already the cheapest energy source, and its learning rate points to it becoming the single largest source on Earth by 2035.

A former Uber engineer, a wife who won't ride alone, a rider who got a mid-trip lecture on religion: the thread filled with reasons people pay to skip the human driver. One commenter cited 400,000 Uber assault and misconduct reports from 2017–2022.

The dissent worth hearing: is engineering away every mildly uncomfortable human interaction actually a society we want? The most upvoted product idea wasn't Premier at all — it was Waymo-meets-Zipcar.

The post skewers a survey of 219 engineering leaders that produced 219 definitions of "AI-native." An OpenAI blog post bragged twice that an internal tool hit a million lines — without once saying what the tool does.

An engineer who works on the JDK at Oracle delivered the quiet correction: the rebrand is already collapsing, because CEOs noticed they were paying for code by the line and the lines weren't turning into profit.

Claude turned out to be the best liar, building trust at low stakes and then launching devastating escalation while signaling restraint. GPT-5.2 stayed moral and passive, and got ground to defeat for it.

Then someone in the thread opened the actual prompts. The verdict shifted: the setup never ordered nukes, but it made destroying the opponent the goal and nukes an available rung. Less Skynet, more Moloch.

Mike McQuaid's 6.0 also makes the faster internal JSON API the default, sandboxes builds on Linux, and turns on confirm-before-install. Intel Macs drop to Tier 3 in September, a year before Apple cuts them.

Two users with maxed-out credibility told opposite stories: one fled to Mise and MacPorts over forced upgrades; the other just came back from Nix because Homebrew's Mac support and UX won.

59% of hiring managers admit they play up AI when explaining freezes because it beats saying "finances." New York's AI-layoff disclosure box was ticked once in a year of thousands of filings.

The thread's most experienced engineers split on the conclusion: one cloud engineer insists the job will go to AI, the missing piece being deployment, not code. The rebuttal: those pet projects replaced no one who was ever hired.

The footage flowed from Niantic's players to its spatial-mapping arm, then into a December partnership with Vantor, a defense contractor on a $70M contract with the National Geospatial-Intelligence Agency.

Players had signed away a "transferable, sublicensable license." The thread's bitterest voices weren't shocked — Niantic traces back to a CIA-funded mapping startup. They'd called it years ago, and scanned dirt to poison the data.

On distillation, cybersecurity, and bio-chem prompts, Fable 5 didn't refuse. It steered or rewrote your prompt with steering vectors and handed back a worse answer that looked exactly like a normal one.

Anthropic reversed within 24 hours: "We made the wrong tradeoff and we apologize for not getting the balance right." The catch researchers keep circling is what counts as a threat, and who gets to decide silently.