Hacker News
1Gemini 4 Argon
Impressive prose and general-purpose answers
Users praise Argon's clear, structured, direct answers, interactive simulations, health explanations, and natural writing, even in non-English languages. Coding ability is disputed: some find it on par with Opus 5.5, others call it weak. The return of em-dashes sparked a side debate.Confusing access, payments, and rollout
Many describe Google's UX as a mess. Users face rejected cards, locked accounts, confusing product names (Gemini, AI Studio, Antigravity), and "Try" buttons that default to older models. Argon is limited to Ultra subscribers, which critics say squandered the launch hype.Pelican SVG benchmark and Goodhart's law
Simon Willison calls Argon's pelicans the best yet and notes Google trains on SVG output. Others question benchmark gaming and debate whether the bikes' fork geometry is rideable. Some ask for harder tests, while others argue better SVG skills are genuinely useful.
2Hackers activate automatic defrosting in supermarket fridges across Europe
- French supermarket chain Intermarché was hit by a cyberattack targeting about 150 stores, forcing refrigerator shutdowns. Related concerns include vulnerabilities in refrigeration controllers and a suspected earlier hack of a military commissary's freezers.
Danfoss's laughable authentication flaws
Users mocked a daily date-seeded admin bypass and a documented POST /xml.cgi login bypass. One clarified it's a firmware backdoor, not an admin-chosen password. 2025 CVEs had patches, but customers weren't warned and thousands of controllers sat exposed on Shodan.Should fridges be internet-controllable at all
An ex-industry worker said remote control enables energy savings and demand response. Many replied monitoring should be read-only, actuators need on-site access, and devices belong behind VPNs. Others noted VPNs get breached too and defrost sequencing needs networking.Thin margins vs. basic security costs
Some cited thin grocery margins and decades-old SCADA gear; an IT director described chaotic HVAC contractor networks. Critics noted these chains earn billions and a VPN is cheap. Parallels to the recent US commissary freezer hack raised suspicions of coordination.
3A Modern Web App in C
- Corey Prophitt built and shipped marketyour.games, a commercial web product, in C served via CGI, arguing the old approach still works in 2026. He praises its simplicity, few dependencies, and easy deployment, while noting he had to write his own basic utilities and would welcome shared ones.
CGI's decline and modern relevance
A veteran CGI author notes CGI died mainly from slow process spawning, a limit that no longer holds. Others call serverless a CGI revival, praise busybox's built-in cgi-bin for tiny deployments, and recall early CGI written in C, Perl, bash, and Pascal.Security risks of C web apps
Many worry about RCEs, string handling, and hand-rolled cookie/form parsing in C. Some counter with djb's publicfile, which has had no RCE in 27 years. Go with CGO_ENABLED=0 is pitched as a safer single-binary option, and custom HTTP servers draw a mixed response.Why modern web stacks are so complex
Users question the need for stacks like TypeScript, React, Next.js, and Tailwind. Answers point to habit ("that's what we've always done"), developers' love of complexity, and the need to scale teams rather than tech. Many argue simpler, mature, "boring" tools usually suffice.
4GLM-Legal
- GLM-Legal, built on GLM-5.3-Flash with legal mid-training and agent-focused post-training, achieves a 19.2% all-pass rate on Harvey LAB and matches frontier models on LegalBench and PLawBench. It will power contract review, evidence analysis, and compliance tasks in AutoClaw.
LLMs handle routine legal work, not novel cases
Users report Claude doing well on boilerplate contracts, saving lawyer fees. Skeptics note a 20% benchmark pass rate, missed omissions like absent clauses, and failed corporate trials. Many see AI as replacing paralegal and junior work rather than partners.Losing the junior talent pipeline
Automating mundane work may stop juniors from building expertise in law, coding and other fields. Some point to manufacturing's lost apprenticeships; others argue schools, spreadsheets-era adaptation or deliberate learning can fill the gap. Some also worry about falling labor value.Privacy and trust in Z.ai
Z.ai's past silent workspace snapshots and its closed model raise concerns about sending legal data to a Chinese provider. Others reply that US labs like xAI do similar things, and some say they would rather China than their own government hold their data.
511 year old unsoundness bugs found in Rust borrow checker
- A Rust type system soundness bug lets a malicious Drop implementation keep a reference past its scope, causing use-after-free. The bug is 11 years old and unlikely to have caused accidental memory unsafety. It is marked high priority, and the Types Team is working to fix such unsoundness.
Compiler bug, not a type system flaw
Many note the borrow checker failed to check the implicit drop call, so the fix is rejecting invalid code without changing lifetime rules. Over 100 unsound issues exist, but this one stands out for its simple reproducer. A crater run should reveal any affected crates.Formal verification and AI bug-hunting
Some hope AI-driven searching and formal proofs will push Rust toward full reliability, citing CompCert as precedent. Others counter that provers can have bugs too and complete reliability is a myth. Some object to AI being raised in every thread.Is Rust's safety promise overhyped?
Critics call memory safety aspirational given the lack of a formal spec. Defenders say all type systems (Java, Python) have holes, soundness bugs are tracked and fixed, and confined "unsafe" blocks remain far better than C. Static safety stays valuable despite bugs.
6AI is a tool and using it well takes skill
- Rebutting Rich Hickey's claim that LLMs are automation, not tools, the author argues LLMs can be wielded with skill like any tool. Unpredictability doesn't disqualify them, automation can be a tool feature, and the real distinction is skilled use versus uncritical delegation.
AI as a threat to human thinking
Many fear LLMs will erode cognition and education, especially for youth. Those claiming to learn more with AI were told they confuse familiarity with understanding. Others predict a bimodal split between active learners and those offloading all thought.Whether AI usage is a lasting skill
Prompting and workflow tricks go obsolete fast as models improve; what endures is knowing what to ask, domain expertise, clear communication, and verifying output. Some say graduates master agents in days; others note models still err and need humans steering.Analogies for AI's effect on professions
The claim that software isn't dead, just as phone cameras didn't kill pro photography, drew pushback: AI automates decisions, not just tools, making painters vs. cameras a closer fit. Others called AI a slot machine, not a tool, and a laid-off engineer described a shrinking job market.
7Apple HomeView
- Apple HomeView is a smart home hub with a 7-inch display, A18 chip, and Siri AI. It offers personalized help, Spatial Audio, FaceTime with Center Stage, and Home app control. It recognizes up to six users, keeping their data private with Secure Exclaves.
Old iPad as main competitor
Many argue an old iPad (battery, apps, mounts) beats HomeView. Others cite multi-user Face ID, its role as a Matter hub, a simpler stationary UX, and appeal to the elderly. Several fault Apple for withholding user profiles from iPads.Secure Exclaves and privacy
Apple says camera and mic data is processed in Secure Exclaves the OS and Apple can't access. Some praise this as a real privacy step, like a YubiKey. Skeptics want independent audits, distrust Apple's ad ambitions, and worry about normalizing always-on recording.Limited app support and ecosystem lock-in
No App Store is a dealbreaker for those wanting Plexamp, Spotify, Snapchat or Home Assistant. Others expect a Watch or CarPlay-style path with iPhone-hosted extensions. Some see it as a push toward Apple-only services and family lock-in.
8Firefox Enterprise
- Firefox Enterprise is an on-premises enterprise browser built on Mozilla's independent, open-source engine. It offers centralized policy control, DLP, AI controls, SIEM integration and professional support, helping organizations meet regulations like GDPR and NIS2. Thunderbird Enterprise is coming soon.
Enterprise browsers as a revenue opportunity
Commenters see enterprise browsers as a sustainable revenue model, citing Island's $6.4B valuation. A Mozilla lead confirms a Europe-first focus. Staff detail MDM deployment, DLP rules, SIEM logging and VPN-less access. Others flag compatibility problems with sites like Office 365.Corporate IT and user resistance
IT workers predict users will insist on Chrome. Others say many companies ban Firefox for lacking controls or low usage, though some users accept a simple icon swap. Some question what Enterprise adds over existing group policies, with replies citing SSO, SIEM, on-prem services and support.Ad blocking and locked-down browsers
Users ask about built-in ad and tracking blocking, and replies recommend uBlock Origin (deployable via policy) or LibreWolf. Others fear IT-managed Firefox will block extensions and add MitM certificates. Mozilla notes regular Firefox can run alongside it, and AI controls stay in standard Firefox.
9Codex CLI installs a persistent, self-updating background daemon by default
- A user asks that Codex CLI 0.157.0 make daemon auto-start opt-in, since it launches a persistent, self-updating background process that copies files, keeps the first session's environment, and outlives sessions. They also report that stopping the daemon leaves the updater running.
Disabling the Codex daemon and auto-updates
Users share workarounds: a daemon_auto_start feature flag, killing the process, --no-daemon, and Nix packages that patch it out. Many object to persistent background processes and forced updates, preferring package managers or opt-in updates for professional tools.Disk bloat and poor cleanup by AI harnesses
Codex keeps old ~300MB binaries, totaling up to 12GB. Claude Code, Cursor and others leave similar clutter. Suggested fixes include deleting release folders, using ephemeral VMs or containers, or switching to minimal harnesses like hax.Buggy LLM tools reflect vibe-coded quality
Commenters see frequent regressions in Codex CLI, its VS Code extension, ChatGPT.com and Claude Code (broken scrolling, copy, permissions, subagents) as evidence that AI-driven development yields fragile software. Others argue the tools are young and will improve.
10Redesigning source control at Anthropic
- Anthropic built Scion, an in-house version control system, after Claude-driven coding surged from 2,500 to over 30,000 merges per day and overwhelmed git. Designed for always-online agents, Scion is server-centric and horizontally scalable, and it replaced git for the monorepo in five months.
Agents still need source control
Responding to the idea that constant agentic churn makes version control obsolete, most argue agents rely on history, diffs, branches, CI and rollback even more than humans. Others doubt ~1M noisy merges a month leaves history useful.Questioning the value of extreme code churn
Many ask what 30,000 merges per day accomplishes, how demand for code can be "infinite", and whether constant rewrites signal poor design. Critics see busywork and FOMO, while defenders cite Anthropic's revenue growth and bottleneck-free agent workflows.Reinventing the wheel with AI
Cheap AI-written code shifts build-vs-buy toward custom tools. Commenters compare Scion to Perforce, Piper, SVN and Epic's Lore, ask why git or a non-monorepo setup wasn't enough, and worry about maintenance, online-only workflows and less shared open-source tooling.