Hacker News, Distilled

Scope of the Problem: Unconstrained Agents on Real Machines

• Many developers routinely run coding agents with full permissions (--dangerously-skip-permissions, --yolo) on their main machines.
• People know about containers/VMs in theory but often bypass them in the moment for convenience.
• Reported damage ranges from rm -rf * to subtle breakage (e.g., creating a real /public/blog directory that hijacks a web server’s routing).
• Several commenters note that agents can ignore or “work around” soft guardrails and textual safety instructions.

What jai Tries to Do

• Opinionated Linux tool to wrap an agent in a lightweight container with:
  • Full R/W access to current directory.
  • Read-only access to the rest of the filesystem.
  • Copy‑on‑write home directory plus default blocking of common credential locations.
• Goal: reduce friction compared to hand‑crafted bubblewrap/Docker invocations so sandboxing becomes default.
• Some like this tradeoff and say it should be the default for agent tools; others say it still allows dangerous read access in “casual” mode.

Comparisons and Alternatives

• Existing mechanisms mentioned: bubblewrap, firejail, seatbelt, systemd-run scopes, FreeBSD jails, dev containers, VMs (Lima, Colima, Qubes, macOS micro‑VMs), custom user accounts, SELinux/AppArmor/TOMOYO, zfs snapshots.
• Several tools and configs already integrate bubblewrap (Claude Code, Codex, others), but their sandboxes can be misconfigured, silently bypassed, or buggy.
• Some argue plain Unix permissions (separate user + shared group folder) are sufficient; others prefer full VMs with no host access.

Security Debates and Gaps

• Strong view: LLM outputs and agents should be treated as untrusted/malware; sandboxing must be enforced outside the LLM, not by it.
• Concerns go beyond filesystem:
  • Exfiltration of cookies, SSH/AWS keys, secrets in dotfiles and env vars.
  • External side effects: APIs, databases, email, Slack, PRs, payment systems.
• Some propose overlay-on-CWD plus explicit diff/patch workflows so only reviewed changes leave the sandbox; warn about agent-written artifacts like .git/hooks, .venv, .pyc.
• Skeptics say filesystem isolation alone is “security theater” if network and credentials aren’t addressed.

Reaction to Project & Presentation

• Many praise the idea and implementation; see it as a pragmatic step toward safer agents.
• Others criticize the marketing/splash page as giving an inflated sense of safety and being LLM‑generated “slop,” though documentation and man page are viewed more favorably.

Premise: Are ICs less enamored than executives?

• Many commenters dispute the premise: lots of ICs are excited about AI; lots of executives are skeptical or just following hype.
• Surveys and anecdotes cited in the thread suggest high AI adoption among developers, though not universal.
• Others say enthusiasm is heavily role‑, domain‑, and skill‑dependent (e.g., web vs. systems, data vs. embedded).

Executives’ motivations and perceptions

• Execs are seen as viewing AI as a way to:
  • Reduce headcount and labor costs.
  • Turn “expensive engineering work” into cheaper, more interchangeable output.
  • Confirm an existing worldview that work is a commodity and value lies in orchestration/strategy.
• Strong FOMO: betting on AI is career‑safe; ignoring it and losing to competitors is not.
• AI demos and simple side projects lead some leaders to wildly overestimate capabilities (“vibe coding” → assume anything is easy).
• Some execs apply AI to their own tasks (communication, reports, slideware) and infer it can replace all knowledge work.

IC experiences and attitudes

• Many ICs use AI daily for: boilerplate, debugging, exploring unfamiliar stacks, quick prototypes, semantic search, and tests.
• Others avoid it, especially in low‑level/systems work, citing hallucinations and domain‑specific issues.
• Some enjoy coding and don’t want to outsource the “fun” parts; others revel in speedups that remove drudgery.

Code quality, limitations, and failure modes

• Repeated reports of “slop”: plausible but brittle code, hallucinated APIs, shallow tests, defensive over‑engineering, and poor long‑term evolvability.
• Strong consensus that unsupervised AI code/agents don’t converge on robust architectures for larger projects.
• Some claim they barely read AI‑generated code; others warn this is reckless and leads to unmaintainable systems.

Impact on work, jobs, and power dynamics

• IC concerns: higher expectations without more pay, loss of bargaining power, and explicit plans to cut staff.
• Executives/owners expected to capture most productivity gains; historical productivity–wage decoupling is invoked.
• Some frame this as classic capital vs. labor / “means of production” conflict; others push back on simplistic Marxian readings.

Broader themes and uncertainties

• AI tools both commoditize routine coding and increase leverage for strong engineers.
• Long‑term effects on roles, pay, and required skills (specification, critical thinking, validation) remain unclear.

What GitHub is changing

• Several commenters clarify the change is about Copilot interaction data, not bulk ingestion of all private repos “at rest.”
• For Free/Pro/Pro+ Copilot users who don’t opt out, GitHub will use:
  • Inputs to Copilot (including code snippets and surrounding context),
  • Accepted/edited outputs,
  • File names, repo structure, navigation, and interaction telemetry.
• Business/Enterprise Copilot customers are said to be excluded from this training change.
• If you don’t use Copilot, multiple comments say this shouldn’t affect you, though some remain skeptical.

Scope and ambiguity

• Many argue the distinction between “private repo data” and “Copilot interaction data” is semantic: Copilot’s “context” effectively includes private repo code.
• The wording of the setting (“Inputs, Outputs, and associated context”) is viewed as vague and legally opaque; even an attorney in the thread finds it hard to interpret.
• It’s unclear how this applies when Copilot is managed by an organization, or when a user is in both personal and enterprise Copilot plans.

Opt-out, UX, and dark patterns

• Major criticism: using opt‑out instead of explicit opt‑in for a new data use.
• Some saw a persistent banner and/or email; others never noticed either, especially those who use GitHub only via CLI.
• The UI copy for the toggle (“you will have access to this feature”) is viewed as misleading, implying you must enable training to “use Copilot.”
• Some note the setting is missing entirely for accounts controlled by orgs or possibly in some regions.

Privacy, legal, and ethical concerns

• Strong distrust of Microsoft/GitHub, framed as “enshittification” and data-grab inevitability once data isn’t end‑to‑end encrypted.
• Multiple mentions of GDPR: opt‑out consent is argued to be invalid; code and commits can contain personal data (names, emails, even health data).
• Worry that contributors using Copilot will leak entire private repos, including sensitive or regulated code (e.g., CUI, secrets).
• Some fear policies can later expand silently; others assume companies may ignore flags in practice with limited recourse.

User responses and alternatives

• Many vow to migrate off GitHub or at least stop using Copilot; some already moved due to the Microsoft acquisition.
• Alternatives mentioned: GitLab, Bitbucket, Sourcehut, Forgejo, Gitea, Codeberg (with caveats), Fossil, self‑hosting on VPS/mini‑PC, and encrypted repos (git-crypt/git-gcrypt).
• A few propose “poisoning” training sets with bad or adversarial code, though others doubt its effectiveness at scale.
• A minority explicitly state they don’t mind training on their code and see benefits (models better matching their style), especially when no secrets are stored in repos.

Immediate situation in Slovenia

• Slovenia introduced fuel rationing (50L/day per customer) after stations began running dry.
• Main proximate cause cited: sudden surge in demand (“fuel tourism”) from neighboring countries plus local hoarding (large farm tanks, many jerrycans).
• Logistics (cistern trucks from storage to stations) became a bottleneck rather than overall national fuel availability.

Fuel pricing, regulation, and cross‑border arbitrage

• Slovenia has long regulated off‑motorway fuel prices; highway prices were deregulated, then re‑regulated during recent volatility.
• Prices were significantly lower than in Austria/Italy, creating cross‑border demand.
• Some argue obvious solutions would be higher prices or dual pricing for foreigners, but others say EU non‑discrimination rules and domestic politics (elections, commuting subsidies) constrain this.

Rationing mechanics and fairness

• 50L/day is seen by many as high enough for normal use but insufficient to prevent multi‑station “hopping” or large‑scale storage.
• Debate over whether limits meaningfully reduce fuel tourism or just inconvenience locals.
• Some recall similar measures (price caps, informal limits) in other EU states.

Renewables, EVs, and energy transition

• Many see the crisis as a wake‑up call to accelerate renewables, EVs, heat pumps, and public transit.
• Others stress:
  • Renewables already expanded (e.g., high shares in UK/EU power) but don’t yet cover peak demand or transport.
  • Electricity is expensive in parts of Europe, and many lack home charging or viable public transit.
  • New EVs remain unaffordable for people buying used ICE cars; TCO can favor EVs but upfront cost is a barrier.
  • Europe’s lithium and manufacturing capacity are still ramping.

Nuclear vs renewables

• One camp urges a “France 1970s”‑style nuclear build‑out for energy security.
• Critics respond nuclear is too slow and capital‑intensive for an acute crisis, with recent Western projects over budget and late.
• Others note nuclear and renewables are complementary baseload/low‑marginal‑cost sources, but current EU market design and capacity factors complicate nuclear economics.

Broader oil & LNG crisis and systemic risk

• Thread repeatedly references the closing/contesting of the Strait of Hormuz, damage to Gulf energy infrastructure, and large price spikes.
• Some call this “the worst energy crisis in modern history,” emphasizing that petroleum underpins plastics, fertilizers, logistics, and much of modern agriculture.
• Others counter:
  • Around 20% of global oil flows through Hormuz; painful but not total collapse.
  • Most fossil fuel use is for burning; decarbonizing power and transport would drastically cut exposure while leaving enough hydrocarbons for essential petrochemicals.

Petrochemicals and limits of “just go renewable”

• Multiple comments stress that even with abundant renewable electricity, industrial systems still need carbon feedstocks for chemicals, fuels for aviation/shipping, lubricants, fertilizers, etc.
• Counterpoint: those uses are a minority share of total oil demand; renewables plus electrification could free up enough hydrocarbons for critical uses, and synthetic/biogenic sources are possible, albeit costly and slow to scale.

Geopolitics and winners/losers

• Extensive debate over the US–Iran war, the Strait of Hormuz, and whether US policy is a historic blunder.
• Suggested “winners”: China (renewables push, oil access), Russia (leverage over Europe and Ukraine), Iran (higher oil prices, sanctions erosion).
• Suggested “losers”: Europe (energy costs, dependence), Ukraine (pressure for compromise), Gulf monarchies (questioned US security guarantees), and arguably US global influence.
• Some argue Europe should seek more autonomy, including negotiating energy with Russia/China; others strongly oppose easing sanctions on Russia as “appeasement.”

Taiwan LNG and disinformation

• Viral claims of Taiwan having only ~10–11 days of LNG are discussed; later links suggest this refers to maximum storage in normal conditions, not an imminent total cutoff.
• Clarified that limited storage still implies vulnerability and potential industrial curtailment if resupply is disrupted.

Societal resilience, inequality, and outlook

• Several participants fear cascading inflation, food and fertilizer shortages, and severe impacts on the Global South.
• Others argue talk of “global reset” or collapse is eschatological; by many metrics life has improved long‑term and societies will “buckle down and sort it.”
• Disagreement over whether current globalization and just‑in‑time supply chains increase flexibility or fragility.
• Repeated concern that higher energy prices are regressive, hitting poorer households hardest; proposals include targeted subsidies or lump‑sum rebates to offset regressive fuel or carbon pricing.

Rounded corners & visual consistency

• Many notice inconsistent corner radii between apps and window states; some find it visually unbearable, others literally never noticed.
• Critics see it as a sign of deeper design/engineering sloppiness and poor separation of concerns, not just a cosmetic nit.
• Defenders argue the radii are intentional (different window types) and that focusing on corners is bikeshedding given larger OS issues.
• Several point out practical regressions: cut‑off content (scrollbars, text) and harder resize handles at rounded corners.

Window management workflows

• Huge split: some almost never maximize windows, preferring overlapping “spatial” layouts; others maximize or tile everything, especially on large/4K/ultrawide displays.
• macOS’s full‑screen mode (green button) is widely disliked: hides the menu bar, moves windows to separate spaces, and breaks some window-switching expectations.
• Built‑in tiling (hover/option‑drag on green button, drag to edges, keyboard shortcuts) exists but is seen as hidden and weaker than Windows or Linux tiling WMs.
• Some rely on third‑party tools (e.g., Rectangle, Magnet, yabai, KDE+i3‑style extensions) to get sane tiling/keyboard‑driven workflows.

Broader UX regressions in Tahoe

• Complaints go well beyond corners: “Liquid Glass” transparency, pill‑shaped tabs, and sluggish notification buttons are common targets.
• Window resizing in Tahoe is widely reported as frustrating and less reliable than previous versions.
• Some feel macOS window management has been “death by a thousand cuts” over many releases, with Finder often cited as persistently bad.

Performance & stability concerns

• Multiple users observe WindowServer and kernel_task spiking CPU after upgrades, causing laggy UI, slow app switching, and multi‑second delays.
• There’s debate whether WindowServer high CPU is cause or symptom of misbehaving apps, and frustration at poor tooling to diagnose it.
• Some report having to disable transparency, tweak spaces behavior, or use CPU-throttling utilities to keep systems usable.

Security, SIP, and corner fixes

• The corner‑fix technique requires disabling System Integrity Protection and patching system libraries.
• One side: if malware already has root, SIP doesn’t matter much; others counter SIP is still valuable for protecting OS integrity and recovery.
• Several argue real risk often comes more from package managers and unvetted code than from toggling SIP alone.

Comparisons, alternatives & Apple’s direction

• Linux (especially KDE/tiling WMs) is praised for configurable, discoverable window management and UI; some recent macOS converts feel Tahoe is a regression.
• Windows is criticized for ads and bloat, but its snapping/tiling is often held up as better than macOS.
• Some see the details regression (corners, input lag, hidden behaviors) as evidence Apple no longer has a strong, Jobs‑style perfectionist vision; others caution against overreacting and view these as annoying but survivable design swings.

Nature of Objections to AI

• Many see two broad camps:
  • Pro‑AI: persuaded by clear productivity gains and concrete usefulness.
  • Skeptical: grounded in principles, craft, identity, and discomfort with dependence.
• Several argue customers and employers primarily care about product, cost, and speed, not how code is produced.
• Others insist they value the process itself and don’t want their role reduced to “prompting” or micro‑managing an agent.

Craft, Learning, and Skill Atrophy

• Strong concern that relying on AI erodes deep understanding, problem‑solving, and the “theory in the programmer’s head.”
• Comparisons made to GPS weakening navigation and calculators replacing basic arithmetic; fear of broadly more sedentary minds.
• Counter‑view: tools have always offloaded skills (matches, washing machines, frameworks); losing low‑value skills is acceptable if people redirect effort to higher‑value work.
• Some use AI only as guide/rubber duck, insisting on writing code by hand to preserve learning and mental models.

Analogy Battles

• Pro‑AI side likens it to cars, microwaves, power tools, or fusion: civilization is built on augmenting human effort.
• Critics argue these analogies are flawed: most tech augments rather than replaces cognition; AI feels more like outsourcing to a separate mind.
• Alternative analogies: taking taxis, eating at restaurants, or hiring a fabricator—turning your brain off while others do the real work.

Productivity, Code Quality, and Maintenance

• Enthusiasts report 5–10x productivity boosts, especially on boilerplate, small tools, and OSS features they wouldn’t otherwise implement.
• Others note AI code can be redundant, brittle, stylistically inconsistent, and hard to extend; speedups may not matter over long product lifecycles.
• Suggested best use: existing codebases, tedious tasks, and short‑lived “vibe‑coded” tools, not foundational greenfield systems.

Jobs, Economics, and Environment

• Widespread anxiety about being outpaced, salary compression, and reduced demand for developers; calls for personal “exit strategies.”
• Some argue the deeper issue is wealth inequality and how productivity gains are distributed, not AI per se.
• Environmental critiques of AI are raised; others see them as overstated relative to other energy uses, saying decarbonizing power matters more.

Meta: Discourse, Culture, and Polarization

• Posters lament binary “AI good/AI bad” framing; nuance doesn’t go viral.
• Observations that social media and current information feeds have already harmed attention and reasoning more than AI itself.
• Historical parallels drawn to past tech shifts (cars, mobile phones, internet), with recurring fear of change and generational effects.

Role of AI vs Human Responsibility

• Many argue AI is only a suggestion layer; humans ultimately choose to strike and must remain accountable.
• Concern shifts from “killer robots” to socio-technical systems that make it easy for humans to rubber-stamp lethal decisions and “sleepwalk” through responsibilities.
• Some see AI primarily as a tool to diffuse or obscure accountability: “the computer did it” replaces personal responsibility.

Disagreement over Claude/Maven’s Role

• Several commenters emphasize that Maven (Palantir’s system) is the core kill‑chain platform; Claude is just an LLM layer added later for querying/summarizing intel.
• Others cite earlier reporting and contracts to argue Claude was more deeply integrated and may have informed targeting, including claims it “selected targets.”
• A technical subthread explains how Claude can be deployed via AWS Bedrock without Anthropic seeing prompts, complicating oversight and contract enforcement.

Uncertainty and Information Warfare

• Strong debate over what is actually known: some stress that casualty numbers, intent, and even who hit the school are not independently verified and rely heavily on IRGC claims.
• Others counter that open-source evidence (imagery, missile fragments, timing, patterns of strikes) makes US responsibility highly likely, even if exact casualty figures remain uncertain.
• Several highlight the broader “fog of war” and previous disinformation episodes as reasons to be cautious about both Western and Iranian narratives.

Targeting Process, Maven, and Old Data

• Discussion of Maven’s interface: three clicks to move a map point into a strike pipeline with ranked “courses of action.”
• Critique that such automation is defensible under fire but reckless in a pre‑planned sneak attack where time existed for deep verification.
• Central claim repeated from the article: a decade‑old DIA database still marked the building as a military facility, and the system’s speed made that stale error lethal.

Moral and Legal Responsibility

• Intense argument over whether this was a tragic mistake in an otherwise “low error rate” campaign, or predictable outcome of a doctrine that accepts high civilian risk.
• Many reject framing this as an “error rate” at all, especially given the victims were schoolchildren and the broader question of whether the war itself is lawful.
• Comparisons drawn to past US strikes on civilian targets, and to Iranian and proxy attacks on US and allied forces; sides differ on who is “aggressor” vs acting in “self‑defense.”

Media Coverage and Trust

• Some accuse the Guardian of minimizing AI’s role and uncritically adopting US framing; others say focusing on Claude is sensationalist “AI‑washing” that distracts from systemic military failures.
• Broader skepticism toward all media: claims that both Western outlets and IRGC propaganda shape narratives more than they illuminate facts.

Broader Reflections

• Commenters note long‑term trends: militaries and corporations using complex tech systems to push decisions up or down chains and escape blame.
• Multiple participants argue that blaming AI obscures the underlying choices to launch a war of choice and to bomb without ground confirmation of targets.

Overall reaction to the “cat desk”

• Many see it as a gimmick: essentially a 90s-style computer desk with holes cut in it.
• Widespread doubt that cats will use the designated spaces; expectation they’ll still sit on keyboards, in front of monitors, or on towers.
• Some call the design poor or unsafe, especially the hidden shelf under knee space (risk of banging knees).
• The under-desk “unused” knee space is disputed; taller people say they need every centimeter.
• Several note the promo photos prove only that you can stage a cat, not that it works during real typing.

How cats actually behave (per commenters)

• Cats prioritize:
  • Owner attention and line-of-sight to face/monitor.
  • Warmth (laptops, CRTs, towers, warmed chairs, heated floors).
  • Height and vantage points; they prefer “up” to enclosed cubbies.
  • Owner scent and habitual spots; they rest where humans spend most time.
• Purpose-built cat furniture is often ignored; fresh cardboard boxes, bags, and stolen non-toys (hair ties, bottle caps, paper, plastic) are favorites.
• They use valued human objects (keyboards, plants, counters) to communicate: mostly hunger or desire for play/attention.
• Desks, chairs, and beds are “theirs” as much as ours; many cats claim the warm chair the moment a human stands up.

Practical hacks people actually use

• Simple, cheap solutions:
  • Cardboard box or printer-paper box lid on or near the desk.
  • Separate chair, shelf, windowsill tray, or rolling rack with a cat bed at desk height.
  • Heating pads or electric blankets placed where you want the cat to be.
• “Decoy” strategies:
  • Extra keyboard for the cat to sit on.
  • Heated “fake laptop” pads.
  • Split keyboards or under-desk trays plus a “cat box” in the prime spot.
• Reported outcomes vary: some say heated beds/decoys solved keyboard invasions; others say determined cats still go where the human’s focus is.

Ergonomics, regulation, and work setups

• In some countries (e.g., Denmark), WFH setups must meet ergonomic rules: height-adjustable desks, separate keyboard/mouse/monitor; working long-term from couches or kitchen tables is technically non-compliant.
• The proposed desk would not meet such standards and is criticized as ergonomically poor regardless.

Broader pet and cultural commentary

• Many note that cats hijack video calls and demos, often becoming the real topic of conversation.
• Several compare cats and dogs: some dog-preferring commenters see cats as “ungrateful,” while others emphasize cats’ affection, intelligence, and lower maintenance.
• A number of users say the product is forgettable, but the thread of shared cat stories and observations is the real value.

Rent Seeking, Free Markets, and Enclosure

• Many argue “rent seeking” in the article is misused; it should mean extracting economic rent via regulation and enclosures, not just “business models I dislike.”
• Others use a broader, more colloquial sense: any pursuit of monopoly‑like rents or moats, especially via friction, subscriptions, and platform lock‑in.
• Several comments link modern rent extraction to historical enclosure (turning commons into private property), with IP framed as a new enclosure of ideas.
• Disagreement over whether what we have is a “free market,” a “capitalist market,” or a heavily captured system shaped by lobbying and policy.

AI, LLMs, and Local vs Cloud Inference

• Some see open‑weight models and cheap local hardware as a serious challenge to centralized, token‑priced AI APIs.
• Others think centralization will continue: big players will train larger, proprietary models tied to their chips and clouds; the rent‑seeking just moves to LLM access.
• Skeptics note that running powerful models at home requires money, hardware, and electricity many people lack; most will end up on paid subscriptions.

Agentic Commerce and Marketplaces

• Optimistic view: AI “agentic commerce” could bypass rent‑taking marketplaces by going directly to sellers, doing comparison shopping and due‑diligence.
• Counter‑view: you just swap Amazon’s 15–20% cut for an AI platform’s cut, affiliate fees, or hidden “marketplaces” inside the model.
• Concerns that LLMs are bad at judging trust; trust and logistics (warehouses, delivery) remain strong moats for incumbents.

Consumers vs Corporations in an AI World

• One camp: AI equalizes time; if both sides use AI, it becomes too expensive for companies to weaponize friction and call centers.
• Opposing camp: corporations will deploy more and better AI, tuned at scale on millions of interactions; it becomes “your bots vs theirs,” and they still win.
• Expectation that AI will often exacerbate asymmetries, not flatten them.

Self‑Driving Cars and Broader Automation

• Some argue self‑driving is necessary given aging populations, labor shortages, and safety gains.
• Others see it as a way to replace workers so value shifts from drivers to tech firms, mirroring broader automation‑driven concentration of wealth.

Moats, Platforms, and SaaS

• View that the durable moats will be at the application layer: distribution, network effects, proprietary data, “systems of record,” and perceived stability.
• Even with democratized model access, large platforms can still dominate via scale, marketing, and integration.

Optimism vs Pessimism about Collapse and Democracy of Tech

• Optimists cite past tech waves (web, YouTube, smartphones) reducing barriers and creating more creators and builders.
• Pessimists reply that power merely re‑concentrates in new gatekeepers, and that calls for a US economic “collapse” are reckless given global interdependence.

Breach scope and immediate takeaways

• Hackers linked to Iran claim to have compromised the FBI director’s personal Gmail, with data from ~2011–2022.
• Publicly released content so far appears mostly personal (photos, resume, mundane correspondence); several commenters call it a “nothingburger” from a national‑security standpoint.
• Others stress that even “boring” personal data is valuable for HUMINT and potential blackmail, and that sensitive work-related content might have been withheld from public dumps.

Operational security and use of personal communications

• Strong disagreement on what “should” be in a senior official’s personal email: some say it must never contain classified or official business; others note long‑standing patterns of officials using personal email or apps (e.g., Signal) for government work or to evade records laws.
• Multiple comparisons are made to past email practices (Clinton, Powell, Bush White House) and to recent use of Signal for military or policy discussions.
• Many see this as an OPSEC failure in itself; others argue that unless harmful use of the account is proven, it’s more embarrassing than consequential.

How the hack happened & defensive measures

• High curiosity about the intrusion method: weak/old password, credential reuse, SIM swap, phishing, or a provider bug are all speculated; nothing confirmed in the thread.
• Several point out that Google and Apple offer “advanced protection” programs for high‑risk users; some view the apparent failure to enroll as further evidence of incompetence, others note most people (even technical) don’t know these exist.
• Debate over whether changing strong, random passwords regularly is meaningful versus “security theater”; consensus that multi-factor auth and device hygiene matter more.

Iran, cyber campaigns, and geopolitics

• Some see this as part of a broader Iranian (and allied) cyber and information campaign, alongside earlier healthcare and corporate breaches.
• Others emphasize that publicizing the hack suggests the attackers either found little of strategic value or are using the visible leak as a signal while retaining more sensitive material.

Media framing, leaks, and broader decay

• Comments criticize headlines that imply an “FBI breach” when only a personal account was hit.
• Links to mirrors/archives of the dump raise questions about the legality and ethics of downloading and examining such data.
• The episode fuels broader pessimism about institutional competence, politicized appointments, and a perceived drift toward authoritarian or “clownish” governance.

Overall reaction to the article and “AI slop”

• Many readers see the piece as marketing‑ish or “AI slop”: generic advice, LinkedIn tone, internal inconsistencies, and claims that feel overstated (e.g., “whatever you write in CLAUDE.md, Claude will follow”).
• Others find it a useful, beginner‑friendly orientation to .claude/, especially for coworkers new to agentic coding.
• Several say the official Claude Code docs or community guides are clearer and more grounded.

Behavior and importance of CLAUDE.md and .claude/

• Strong disagreement on how “authoritative” CLAUDE.md is:
  • Some insist it is always loaded at session start and persists through compaction, making it central.
  • Others say it’s just more prompt text: often ignored, diluted in complex tasks, and treated as suggestion not contract.
• Reported pain points:
  • Claude often forgets or ignores instructions (e.g., tests-first, logging rules).
  • Overly long or noisy CLAUDE.md can degrade performance; short, focused rules plus links are preferred by many.

Skills, agents, and configuration vs. keeping it simple

• One camp: fewer skills and minimal config work best. Too many skills/agents:
  • Pollute context, confuse the model, and lead to tool‑call thrashing.
  • Become a “productivity setup” rabbit hole akin to dotfiles/Jira/Emacs bikeshedding.
• Another camp: thoughtfully designed skills/MCPs are powerful, especially for:
  • Project‑specific workflows (debugging, querying logs, custom APIs, accounting systems).
  • Enforcing architecture patterns, separation of concerns, and repeatable debugging steps.
• Several note that custom hacks often become obsolete quickly as models and harnesses improve.

Team workflows, guardrails, and standardization

• .claude/ is seen as a way to:
  • Share conventions, tooling, and guardrails across teams (e.g., preconditions before running, “don’t push to main”).
  • Align agent behavior when multiple devs use AI on the same repo.
• Concerns:
  • Editing shared files like AGENTS.md can affect everyone; some suggest treating them like config with PR review.
  • No cross‑vendor standard for these files; some people experiment with symlinks or per‑model files.

Security and sandboxing

• Strong warnings about running Claude Code/agents without isolation:
  • Default “ask before running commands” is not a real sandbox.
  • Recommended approaches include Docker/devcontainers, official sandboxing, firejail, or cloud/VM isolation.
• Deny‑lists in settings.json are viewed by some as partial “security theater”; containerization is preferred.

Broader views on AI coding and profession

• Split between:
  • Enthusiasts who report big productivity and quality gains with well‑tuned setups and evals.
  • Skeptics who see “prompt and pray” as cargo‑cult engineering, fear config hell, and argue you still must review AI‑generated code line by line.
• Some expect many current configuration practices to fade as models and harnesses improve.

Feasibility of Mars Colonization and Terraforming

• Many commenters argue long-term settlement is vastly harder than popular narratives suggest.
• Key technical issues raised: toxic perchlorates in ubiquitous fine dust, high radiation without a magnetic field, very low atmospheric pressure, unknown effects of 0.38g on long-term human health.
• Terraforming is seen as technologically out of reach and requiring extremely long timescales; realistic scenarios involve pressurized, shielded habitats and largely underground living.
• Closed, self-sustaining ecosystems are noted as an unsolved problem; past Earth experiments (e.g., Biosphere-style projects) struggled even under ideal logistics.

Earth vs Mars: Priorities and Ethics

• Strong view: it is cheaper and easier to “terraform Earth” (repair climate and ecosystems) than to terraform Mars or build a self-sufficient colony there.
• Others counter that humanity is large enough to pursue both remediation and exploration, and that insisting on “solve Earth first” would have blocked historic exploration.
• Some suggest Mars colonization discourse functions as escapism for elites, potentially leaving “the rest” on a degraded Earth.

Humans vs Robots for Mars Exploration

• Pro-human side: astronauts can adapt, fix unexpected failures, and do in days what current rovers take years to achieve.
• Pro-robot side: humans require enormous mass in life support, shielding, and return capability; robotic missions are far cheaper, safer, and can be multiplied across sites.
• Future autonomous, possibly humanoid robots plus compact reactors are proposed as a better path, though others note such systems don’t yet exist even on Earth.

Economics, Technology, and Program Realism

• Several commenters highlight budget constraints, national debt, and political risk aversion as major blockers.
• There is skepticism about current heavy-lift programs that have yet to demonstrate orbital payload capacity while being placed on critical paths for lunar or Martian plans.
• Some see hype around Mars as investor-facing rhetoric rather than an engineering-backed roadmap.

Existential Risk and “Backup Planet” Argument

• One camp sees off-world settlements (Mars, Moon, or orbital habitats) as a hedge against catastrophic Earth events.
• Critics argue that most conceivable disasters are better mitigated with Earth-based bunkers and resilience measures; Earth post-disaster is still likely more habitable than Mars.
• Very small, dependent outposts are not viewed as a meaningful “backup” for civilization.

Psychological and Cultural Themes

• Several commenters express melancholy as classic sci-fi futures (FTL, fully terraformed Mars) look increasingly implausible.
• Others argue these were always fantasies, and that improving life on Earth for billions is a more compelling, “cool” grand project.
• A minority worry humanity is an ecologically destructive “invasive species” and advocate treating other worlds more like protected parks than real estate.

Scope of New Hong Kong Powers

• New rules let police compel disclosure of phone passwords, with penalties for refusal or providing “fake” credentials.
• Commenters see this as part of a broader PRC pattern where laws include broad “at authorities’ discretion” clauses, aimed mainly at already-targeted individuals.

Comparisons with Other Countries

• Many argue Hong Kong is “catching up” with the UK, Australia, Ireland, France, the Netherlands, and others that already criminalize refusal to decrypt devices.
• UK: Under RIPA, refusal can mean up to 5 years in prison (2 years in many cases). Debate over whether this is materially different from “indefinite” imprisonment, especially when life impact is considered.
• US: Multiple comments note that the 5th Amendment protects against compelled password disclosure in theory, but:
  • At borders, CBP routinely demands device unlocks and can seize devices and deny entry to non‑citizens.
  • Citizens can be detained and devices held for days; some report severe abuse and lack of realistic legal recourse.
• Some insist Western democracies remain fundamentally freer than China; others argue increasing convergence toward authoritarian practices.

Legal Safeguards and Their Limits

• Discussion of UK judicial oversight (judge-issued notices, proportionality tests) versus fears that low standards (“reasonable grounds”) and broad purposes (“preventing or detecting crime”) make abuse easy.
• Debate over double jeopardy in the UK and whether repeated non-compliance could effectively yield indefinite punishment; considered theoretically possible but apparently untested.

Risk Profiles: Tourists vs Dissidents

• Several state that ordinary tourists in China/HK are rarely bothered; trouble focuses on “troublemakers” or dissidents.
• Counterpoint: having rights only “if you keep your head down” is itself a sign of lack of real protection.

Technical and Behavioral Countermeasures

• Strong interest in:
  • Multiple profiles, hidden or “clone” systems, plausible deniability, and duress PINs that wipe devices.
  • Use of burner phones, minimal apps, short message retention, and treating smartphones as untrusted work tools.
• Others argue clever tech cannot solve coercion (“rubber-hose cryptanalysis”); once force or jail enters, profiles and tricks may just escalate suspicion.

Broader Civil Liberties Concerns

• Worries about criminalizing forgotten passwords, expanding “hate speech” and online offenses, and general erosion of privacy and due process in many countries, not just China/HK.

Microsoft Account & Setup Requirements

• Many see mandatory Microsoft accounts during Windows 11 setup as fundamentally wrong: it feels like needing third‑party permission to use one’s own PC.
• Complaints focus on dark patterns: hiding or removing local-account options, repeated “finish setting up your PC” flows, and tying logins to cloud services with unclear security/lockout implications.
• Some argue the account brings benefits (BitLocker recovery, settings sync, easier remote access), and that for most people everything is online anyway, so a unified identity is simpler. Others reject this as vendor-centric, not user‑centric.

OneDrive, Cloud Defaults, and Data Control

• Strong anger at OneDrive’s “online-only” defaults: files silently moved to cloud, local storage underused, and data becoming inaccessible when services break (e.g., Teams/OneDrive bugs blocking logins and files).
• Several describe important files being deleted or stranded by OneDrive migrations they never explicitly approved.
• These behaviors are seen as deceptive upsell funnels (“subscribe so we don’t delete your stuff”) and a reminder that local backups are still essential.

General Windows 11 Frustrations

• Broad sense that Windows has become “hostile”: ads in the OS, Edge nags, bundled bloat, features re‑enabled after updates, privacy prompts used as gating for sign-out or usage.
• Search is widely criticized as unreliable and polluted by web/Bing results.
• Some say Windows 11 can be tolerable with scripts, debloat tools, or LTSC editions, but others note changes often get undone by updates.

Comparisons with Apple and Google

• macOS generally does not require an Apple ID for OS login, and Apple keeps the cloud account more clearly separated from the local user.
• iOS effectively requires an Apple ID for app installation, and Apple is also accused of increasing ads and bloat, but many still find macOS less intrusive than Windows.
• Google/AWS sign‑in flows are described as clearer and less chaotic than Microsoft’s fragmented auth ecosystem.

Linux and Alternative Platforms

• Many commenters have already moved to Linux (or macOS) and frame this fight as “too little, too late.”
• Linux is praised for control and lack of enshittification, with Steam Deck/Proton and distros like Mint/KDE cited as making desktop use viable.
• Counterpoints: hardware quirks (Wi‑Fi, battery life), weaker accessibility, and sometimes hostile support culture keep Linux from mass adoption.

Enterprise vs Consumer Outlook

• Some think Windows dominance in enterprise/government (legacy apps, AD, Office, specialized software) will last decades.
• Others see slow but real erosion: cloud/SaaS reducing OS lock‑in, governments and regions piloting open source, and younger users entering the workforce with little Windows familiarity.
• Mandatory accounts are viewed as another push factor nudging consumers and developers toward macOS or Linux.

Authenticity, AI, and Fiction vs Nonfiction

• Many readers enjoyed the story but questioned whether it really happened.
• Some insist it reads like “AI slop” or revenge fan‑fiction; AI detectors and classifier screenshots are cited.
• Others counter that the blog long predates current LLMs, the author is active in blind communities, and the post is explicitly tagged “nonfiction” and “rant.”
• A middle position: even if embellished, similar incidents almost certainly occur in real bureaucracies.

Fax, Email, and “Security”

• Several commenters doubt any office in 2026 is using a purely physical fax; they expect fax‑to‑PDF or fax‑to‑email systems.
• This undercuts the “drowning them in paper” fantasy, though others note small offices may still print everything.
• Debate over fax vs email security: some argue fax is outdated and no more secure; others point to HIPAA‑style rules, lack of end‑to‑end email guarantees, and institutional risk aversion.
• Examples given of long‑standing fax‑over‑IP setups and unified inboxes, including in government and medical contexts.

Blame: Bureaucrat vs System vs Voters

• One camp sees the call‑center worker as a low‑power cog enforcing absurd laws, not the proper target of anger.
• Another argues that “agents of the system” have moral agency; enjoying or rigidly enforcing harmful rules makes them blameworthy too.
• Some stress politicians and voters who demanded anti‑fraud crackdowns as root causes of hostile disability systems.
• Counterexamples show individual bureaucrats sometimes can and do bend rules or quietly help, so personal attitude still matters.

Ethics and Effectiveness of Malicious Compliance

• Strong split:
  • Supporters view the 500‑page fax as justified pushback, forcing the office to internalize the cost of its own friction.
  • Critics call it petty harassment of a likely underpaid worker, potentially delaying other disabled claimants and wasting public resources.
• Disagreement over whether such stunts meaningfully pressure management to modernize processes, or simply harden attitudes and get ignored.

Disability Bureaucracy Experiences

• Many share similar stories from the US, UK, and elsewhere: periodic re‑verification of lifelong or genetic disabilities, intrusive forms, and hostile assumptions about fraud.
• Some note online SSA/benefit portals that (on paper) should allow digital uploads, making the fax‑only demand seem even more arbitrary.

Existence and Ownership of QA

• Broad agreement that “quality assurance” as a function must exist; disagreement is about who does it and how it’s organized.
• Some argue engineers should own quality end‑to‑end; dedicated QA encourages “throw it over the wall” behavior and slows release.
• Others say quality is a system involving engineering, product, design, and operations; dedicated QA are specialists in that system, not a crutch for bad engineering.

Value of Skilled QA vs Bad QA

• Many anecdotes of excellent QA finding subtle, high‑impact bugs, cross‑feature interactions, and timing issues that devs never considered.
• QA often become the deepest experts on the product and its real behavior, helping product, support, and design.
• Strong view that most QA in the wild are undertrained “button pushers,” repackaging dev tests and adding bureaucracy; this fuels skepticism about the role.
• Poor org design (e.g., remote offshore teams running dev‑written scripts) turns QA into expensive red tape that finds little.

Automation, Testing Strategies, and AI

• Debate over the “testing pyramid”: some claim it’s outdated; advocate a “testing hourglass” with many unit tests and many UI/API tests, fewer mid‑level integrations.
• Others emphasize test cost along a “would you run it” axis: fast deterministic tests at bottom, flaky/slow/expensive ones at top.
• Conflicting views on unit tests: some see them as cheap, high value; others see them as expensive, low value compared to smoke/E2E tests.
• AI is seen both as:
  • A way to cheaply generate automated tests and supercharge good QA.
  • A reason QA becomes more important, since coding is automated and human verification/oversight dominates.

Human Testing, Exploratory Work, and Product Knowledge

• Exploratory and adversarial testing (breaking flows, odd timing, network failures, multi‑device behavior) is widely cited as where good QA shine and automation struggles.
• Humans are needed to validate UX, “does this make sense,” and real‑world usage patterns, not just specification conformance.

Organizational Structure and Incentives

• Arguments for independence: QA analogous to audit/red team; shouldn’t be beholden to the same managers and incentives as builders.
• Counterpoint: siloed QA, long queues, and lack of collaboration are destructive; embedded test engineers or SDETs within teams work better.
• QA is often underpaid, first outsourced or cut, and used as a dumping ground for glue work, which degrades quality of practitioners and outcomes.

Domains and Risk Levels

• High‑risk and regulated domains (medical devices, aerospace, enterprise contracts, compliance frameworks) are described as requiring robust QA and formal test evidence.
• Consumer/social products often accept weaker QA in favor of speed and growth; “does it have users” can trump correctness.

Hardware Supply, Prices, and AI Boom

• Many note sharp RAM and GPU price increases, attributed to AI hyperscalers buying at massive scale and higher margins than consumer markets.
• Some see this as a familiar boom–bust cycle like past DRAM and HDD spikes; expect 3–5 years of pain then normalization and possible gluts.
• Others fear a persistent “demand crunch”: consumer high‑end hardware becomes uneconomic as datacenter volume dominates.
• Secondary market hopes are tempered: modern DC gear is rack‑scale, power‑hungry, proprietary, and often ill‑suited to home use, though ECC DDR5 and similar parts can trickle down.

Future of Personal vs Cloud Computing

• Strong anxiety about a shift back to “mainframes + dumb terminals”: thin clients, locked‑down devices, and rented compute replacing general‑purpose PCs.
• Counterpoint: mid‑range laptops, Macs, and even phones are extremely capable for most tasks; many users already treat laptops as SaaS terminals.
• Some argue PC DIY and high‑end consumer hardware will become niche and expensive, hollowing out the middle of the market.

Local AI vs Hyperscaler Models

• One camp says local AI is a dead end; open‑source should focus on large, H200‑class models to avoid permanent dependence on proprietary APIs.
• Others push hard for efficient local models (MoE, quantization, emerging compression like Turboquant) to preserve autonomy and reduce energy use.
• Concern that if open models don’t stay close to frontier quality, economic participation will be locked behind hyperscaler gates.

Software Bloat and Performance

• Frequent complaints about Electron, web apps, and modern frameworks using huge RAM and feeling sluggish for simple tasks.
• Some dismiss this as overblown given abundant memory; others see it as disrespectful waste that shortens hardware lifetimes and drives upgrades.

Ownership, Lock‑in, and Self‑Hosting

• Worries about phones with locked bootloaders, app‑store control, age‑gating, and future KYC/“nanny chips” limiting what owners can run.
• Self‑hosting and homelabs are framed as increasingly important; others warn about long‑term maintenance and backup burdens.
• Old, hackable hardware is valued for root access and OS choice; advice to prefer upgradeable, Linux‑friendly machines.

Economics, Policy, and Geopolitics

• Debate over whether high prices are “capitalism working” vs. de facto extraction from ordinary users.
• Mentions of past DRAM cartels, current tariffs, concentration of fabs (TSMC, China’s entrants), helium and energy constraints, and the fragility of globalized supply chains.

Meta: Site Behavior

• The article’s site uses JavaScript to swap tab titles/icons to inflammatory/NSFW strings when backgrounded, then shows an overlay urging users to disable JS.
• Some find it clever advocacy; many see it as hostile, unprofessional, and risky on work machines, and block the domain.

Scope and Nature of the Compromise

• Attack is part of a broader “TeamPCP/CanisterWorm” supply-chain campaign.
• Only the Telnyx Python SDK distribution on PyPI was compromised; core Telnyx APIs/platform were reportedly unaffected, though some argue the package channel is still part of “infrastructure”.
• Malicious code was crude and detectable: exec(base64.b64decode(...)) plus a hard‑coded C2 URL.
• Payload was fetched as a seemingly valid .wav file; audio data concealed a base64 payload XOR‑decoded into an executable/script.

Detection, Tools, and Notification

• Multiple security teams and automated scanners on PyPI spotted the issue in parallel; packages were quarantined.
• Static tools like Hexora (and regex‑based tools such as GuardDog) easily flag the obfuscated exec and suspicious network behavior.
• Users can monitor compromised packages through PYSEC and OSV advisories.
• C2 host appears offline; attempts to retrieve the malicious WAV now time out.

Python Ecosystem & Dependency Management

• Some commenters express deep frustration with Python’s fragile dependency ecosystem and backward compatibility, leading to heavy use of VMs/containers.
• Others defend Python’s ergonomics but concede the ecosystem is “messy,” especially for large dependency graphs.
• Discussion of uv and pip features to reduce risk:
  • exclude-newer (in uv and upcoming pip options) lets users avoid packages newer than X days to give scanners time to react.
  • Debate over whether needing tools like uv is a symptom of oversized, risky dependency trees.
• Suggestions include curated internal mirrors, PyPI caching proxies, and Guix/VM isolation for high‑risk workloads.

PyPI Publishing Security (2FA, Tokens, CI)

• Calls for mandatory 2FA and more interactive release approval.
• Clarification: PyPI 2FA protects login, but long‑lived API tokens can bypass 2FA for publishing.
• “Trusted publishing” via OIDC ties publishing to CI environments, but:
  • It does not inherently prevent malicious code if CI or source repo is compromised.
  • Some worry non‑major forges (e.g., alternatives to GitHub) are disadvantaged.
• Comparisons to Debian’s offline, no‑network build model as a more robust approach.

Broader Supply-Chain Risk and Mitigations

• Consensus that supply-chain attacks are frequent and hard to defend against; “devastating” large‑scale incidents seem likely or already occurring.
• Suggestions:
  • Sandboxing installers and development environments.
  • Using generic HTTP clients or API‑first designs instead of vendor SDKs.
  • Treating any environment that imported malicious versions as fully compromised and rolling back via VM snapshots or rebuilds.

Distributed vs Centralized Energy

• Many see distributed solar + storage as key for resilience and household energy independence, especially amid rising grid and data center demand.
• Others argue full residential decentralization is inefficient; grids exist to smooth variable demand and enable industrial loads. A mixed model is favored: centralized for industry/cities, more decentralization for rural/essential loads.

Net Metering, Pricing, and Grid Economics

• Net metering works well at low penetration but is said to become a liability around 20–40% of capacity, forcing expensive ramping of backup plants.
• Several posters stress that home solar economics often rely on flat retail tariffs while wholesale prices swing from negative (overproduction) to very high (evening peaks).
• Some argue balcony/rooftop solar becomes uneconomic if everyone pays spot prices; others note taxes and network fees still make self-consumption attractive.
• Fixed infrastructure costs (water, sewage, gas, grid) are largely constant; as consumption drops, per‑unit prices tend to rise.

Overproduction, Curtailment, and Storage

• Negative prices and “paying to turn off” wind/solar stem from market rules and pre‑agreed contracts, not technical necessity.
• Curtailment of solar/wind and inverter export limits are standard tools; domestic systems can simply disconnect.
• Consensus that “overproduction” is really “under‑storage”; batteries, thermal storage, and power‑to‑X (e.g., hydrogen, synthetic methane) are discussed as solutions, but long‑duration storage and intermittent industrial loads remain hard.

Home Batteries, EVs, and Scale

• Some foresee batteries in “every home”; others think grid‑scale storage will dominate due to much lower per‑kWh costs.
• EV batteries are seen as promising second‑life home storage, though there is skepticism about managing mixed used packs and future scrap economics.

Balcony / Plug‑In Solar and Safety

• 800W “balcony solar” plug‑in kits are booming in parts of Europe; UK and several US states are moving to legalize them.
• Safety concerns center on overloaded circuits (especially UK ring finals), incompatible breakers/RCBOs, anti‑islanding behavior, and unidirectional meters.
• Some countries allow only non‑exporting systems or cap export at low wattages; registering small systems is often simple, larger ones more regulated.

Economics and Policy

• Payback reports range from ~2–10+ years depending on sun, installation cost, subsidies, tariffs, and self‑consumption.
• Debate over whether rooftop mandates and subsidies are cost‑effective versus utility‑scale solar; critics see rent‑seeking and “greenwashing,” supporters emphasize long‑lived infrastructure and geopolitical benefits.
• Strong disagreement over “inevitability” of the transition; many blame fossil‑fuel interests and political incoherence for slow grid upgrades and renewable deployment.

Site & UX Issues

• Multiple readers reported light-mode CSS bugs: white text on white background, unreadable links, overlays on code snippets.
• Workarounds included toggling dark mode or using reader mode.
• The author acknowledged neglecting light-mode testing and pushed fixes; some still saw issues initially.
• Benchmark charts were criticized: inconsistent scales, unclear coloring, missing explicit jq baseline, and hard-to-scan labels.

Performance Claims & Real-World Use

• Several commenters handle very large JSON/NDJSON (hundreds of MB to TBs) for logs, monitoring, ETL, and analytics; they care about 2–10× speedups in wall time and cloud cost.
• Others said jq has never felt slow for their interactive use; for them, “faster jq” is mostly marketing or niche.
• Some argued that performance at scale prevents tasks from becoming effectively impossible; even “one-off” jobs can be frequent in support/ops.
• There’s debate on whether shelling out to a CLI is appropriate for high-throughput or low-latency paths.

Syntax, Semantics & Developer Experience

• Many find jq’s syntax hard to remember, “arcane,” or conceptually slippery (especially around pipelines and arrays); they often rely on manuals or LLMs to write filters.
• Others defend jq’s model as simple once you internalize its semantics and value its concision in one-liners.
• jsongrep’s syntax is viewed by some as more intuitive for path-matching; others dislike it or want even simpler, JS- or Python-like query languages.

Positioning vs jq & Other Tools

• jsongrep is framed as a faster, less-expressive, search-only subset tool (no transformations, arithmetic, or interpolation).
• Some see this as useful as a pre-filter, but object to comparing it directly to full jq.
• Alternatives raised: jaq (jq-compatible, Rust), jj, oj (JSONPath), rq (Rego-like), gron/fastgron, custom JS-based tools, DuckDB/ClickHouse/SQLite, nushell and PowerShell with native JSON objects.

Distribution & Tooling Ecosystem

• Initial lack of arm64 macOS/Linux binaries drew criticism; others were content to install via Rust’s cargo.
• The author later added arm64 releases and the tool was packaged for Homebrew and arkade.
• One commenter dislikes “just install with cargo” due to toolchain bloat for small utilities.

Broader Optimization & Determinism Debate

• Some view micro-benchmarks and ns/ms savings as performative; others argue cumulative CPU/energy savings and reliability improvements justify performance work.
• Deterministic tools like jq are preferred over LLMs for actually running queries; LLMs are mainly used to generate jq expressions.