Hacker News, Distilled

Context: Freetube, Invidious, and Legal Pressure

• Freetube can use Invidious backends but, by default, appears to pull directly from YouTube.
• Discussion ties this to YouTube’s cease‑and‑desist against Invidious and the general “cat-and-mouse” dynamic around alternative frontends.
• Some argue non‑official clients and ad-skipping are DMCA risk or “theft of service”; others dismiss these as overreactions or legally dubious claims.

User Experience & Features

• Praised for: built‑in thumbnail “de-clickbaiting” (e.g., random frames, DeArrow‑style), privacy, and removing distracting recommendation algorithms.
• Criticized for: slow startup (Electron/x86 on Apple Silicon), laggy searches, missing 4K, casting, playback shortcuts, sync, and membership perks; some say it feels like “just a worse browser.”
• Several users report bugs, crashes, and issues on multi‑monitor setups.

Ethics of Ads, Tracking, and the “Social Contract”

• One camp: blocking ads/trackers on YouTube is freeloading and ethically akin to piracy; users should either watch ads, pay Premium, or not use the service.
• Opposing camp: modern adtech is invasive “malware”; users have full moral and technical right to control their own client and block ads/tracking.
• Strong disagreements over whether adblocking is petty theft vs. legitimate self‑defense; analogies include unmanned fruit stands, museums with gift shops, and TV ad-skipping.

Business Models, Costs, and Alternatives

• Some creators say they’d pay monthly hosting fees to avoid ads; others respond that realistic storage/egress costs for large catalogs and popular channels are far higher than creators expect.
• Debate over how cheap bandwidth really is at scale and whether Google’s costs are overstated.
• Alternatives mentioned: Vimeo, Wistia, PeerTube, Nebula; but network effects and YouTube’s search/discovery dominance make them hard substitutes.

Google’s Role and Antitrust

• Some view Google as providing invaluable free infrastructure and keeping “the internet alive.”
• Others describe Google as a surveillance-driven monopoly and “parasite,” arguing that draining it via adblocking is ethically positive.
• Breaking up Google/YouTube is discussed; several participants question how meaningful or effective structural breakups would actually be.

Other Clients and Tools

• Alternatives cited: NewPipe (Android), SmartTubeNext (TV), Invidious instances, other alternative frontends.
• Some label NewPipe and similar apps as “piracy” tools; others use them happily for privacy and UX reasons.

AI / LLM Desires

• Some wish for LLM tools to auto‑summarize and de‑duplicate topic videos so they don’t waste time on repeated, clickbaity content.

Overall reception & learning value

• Many commenters praise the workshop as clear, practical, and a good way to revisit fundamentals of transformers and LLMs.
• Several say it hits a “just right” level for people already comfortable with deep learning/PyTorch and who don’t want ultra-low-level autograd-from-scratch material.
• Others share additional resources (e.g., other “GPT from scratch” writeups/videos) that complement this, each emphasizing different aspects (training vs inference, numpy-level math vs framework use).

Data cleaning, instruction following, and real-world models

• Some ask for more detail on how major models clean and structure training data, suggesting this is where long-term differentiation will lie.
• Commenters point to sections in large model papers (e.g., “steerability” / instruction tuning) as partial answers.
• One thread stresses that unstructured pretraining alone yields a babbling model; instruction-following behavior requires additional structured training with human feedback.

“From scratch” and abstraction level debate

• Significant discussion centers on whether building an LLM “from the ground up” should use PyTorch or go lower-level (numpy, custom autograd, or even C/assembly).
• One camp: PyTorch nn is “low level enough” for understanding transformers; going deeper is mostly for framework/hardware developers.
• Another camp: “from scratch” should avoid major dependencies and expose more of the mechanics; they cite bottom-up tutorials (e.g., autograd by hand) as more educational.
• Some propose a pedagogical progression: basic programming → text processing → n‑grams/Markov chains → then transformers.

Should people build their own LLMs?

• Skeptical voices argue most individuals can’t train competitive models and should focus on building applications on top of existing LLMs.
• Others counter that educational value, intuition-building, and niche/small models on modest hardware still justify learning to build and train models.

Alternative/simple language models and terminology

• A long subthread debates a non-LLM “transformer” project based on n‑grams/Markov chains plus rules.
• Critics say calling it a “transformer” is misleading in today’s NLP context, where that term refers to a specific architecture.
• The author defends the broader mathematical meaning of “transform/transformer” and argues that n‑grams, POS tagging, and embeddings are intertwined in modern systems.
• Multiple commenters push back that terminology in ML has become specialized and reuse of core terms can confuse users.

Platform and tooling notes

• Some Windows users wonder about compatibility; others recommend WSL2 with CUDA as a practical route.
• A separate guide is mentioned for training nanoGPT on cloud GPUs for relatively low cost, though its practical utility is described as mostly educational.

Language around “coding”

• A minor tangent discusses dislike for the term “coding” versus “programming” or “software engineering.”
• Views differ by culture and personal taste; some see “coder” as less professional, others embrace it as long-standing slang.

OpenAI’s Moat and Competitive Position

• Many argue individual models are transient (12–18 months) and will be obsolete within a few years; no lasting moat at model level.
• Others see moats in products, ecosystem, branding, and especially data and RLHF from 200M weekly users.
• Upcoming internal models (Strawberry, Orion, Q*) are rumored to use synthetic data and advanced reasoning methods. Some think this could keep OpenAI ahead; others say competitors are doing similar work, so advantage may be modest.
• OpenAI is seen as ahead in multimodality (text, image, audio, some video), but slow, partial productization (e.g., Sora, GPT‑4o voice/screen sharing) fuels skepticism that they have anything dramatically better “hidden.”

Models vs. Wrappers and UX

• Many participants think “AI wrappers” (tools with strong UX built on top of LLMs) may have more durable value than the base models, since a lot of usage is simple (tagging, extraction, etc.) and doesn’t need the very best model.
• Others counter that wrappers are easy to copy and OpenAI itself has decent UX and an API that’s straightforward to integrate.
• Switching models via API is technically easy, but prompt migration and behavior drift create friction, which some see as a soft moat.

Infrastructure, Costs, and Hardware

• Hardware and GPU access are viewed as a major structural moat; training frontier models appears mostly “elastic with capital.”
• CUDA dominance is cited as a barrier to AMD and others, even when alternatives are competitive on raw performance.

Search, Perplexity, and Google

• Perplexity and OpenAI’s SearchGPT-style offerings impress some users, who see them as Google-threatening.
• Others stress Google’s data advantage (fresh index, maps, shopping) and ad business; AI search quality and cost per query may not yet beat traditional search.
• Some note AI search can be biased or safety-constrained (examples around criticizing religions).

Data, Feedback Loops, and Reliability

• Free ChatGPT is widely seen as a data acquisition engine: conversations, thumbs up/down, and multi-turn dialogs provide exclusive training data and “experience flywheel.”
• Some are skeptical this interaction data cleanly separates good from bad responses.
• Concerns remain about hallucinations and reliability; many see human-in-the-loop chat (like ChatGPT) as the likely “killer app” rather than fully autonomous agents.

Broader Risks and Strategy

• Overreliance on AI without scrutiny is seen as risky for businesses; AI is framed more as augmenting than replacing labor.
• Some advise avoiding OpenAI due to contractual limits on training with user logs.
• Opinions split on OpenAI’s release pace: some view it as a bullish sign of bigger things coming; others think it just means there’s nothing ready to ship.

What “learning Linux” means

• Many equate it to basic Unix CLI skills: navigation, file manipulation, search, and a ubiquitous editor (often vi/vim).
• Others argue that’s really “learning Unix”; true Linux expertise also involves distro specifics, system services, packaging, and networking.
• Some advocate only a thin layer of skills (terminal, Docker, filesystem basics) unless your work demands more.

Arguments for learning Linux

• Most backend/cloud systems run Linux; understanding it helps debug deployments, permissions, container issues, and serverless quirks.
• Unix/Linux knowledge exposes core CS ideas: processes, filesystems, I/O, networking, kernels, and addresses space.
• Valuable for embedded, robotics, SBCs, and infrastructure roles, where targets often run Linux.
• Shell pipelines (bash + sed/awk/jq) are seen as a powerful, general tool for data and automation.
• Software freedom and avoiding vendor lock‑in are cited as both personal and business advantages.
• Some hiring managers treat Linux/mac usage as a signal of curiosity and willingness to “look under the hood.”

Skepticism and limits

• Several feel the article failed to make concrete, persuasive arguments, and might discourage newcomers.
• Many web/frontend roles never touch OS specifics; SaaS and managed services reduce the need for deep Linux knowledge.
• Some prefer to invest learning time in databases or browser internals instead of OS details.
• A few only want Linux as a container/WSL layer and see full-time Linux use as unnecessary.

OS, tooling, and UX comparisons

• Opinions diverge: some rank Linux > macOS > Windows for development; others praise Visual Studio and Windows debuggers.
• Linux debugging tooling is criticized as slower/less interactive than newer Windows debuggers; others are satisfied with gdb/lldb in IDEs.
• WSL2 is debated: some say complaints are wrong; others report portability and tooling mismatches versus “real” Linux.
• Docker: Linux is viewed as the best host; macOS Docker is described as slower and brittle due to VM overhead; WSL2 often preferred.
• Terminal shortcuts (Ctrl-C vs copy/paste) spark a long consistency vs backward-compatibility argument across Linux, Windows, and macOS.

Distros, stability, and learning paths

• Arch/Gentoo praised for learning but criticized for fragility and time sink; Ubuntu/Fedora suggested for stable daily use.
• Immutable or snapshot-friendly systems (e.g., Silverblue, NixOS, btrfs snapshots) are recommended to make experimentation safe.
• Some run Linux bare metal; others happily use Linux VMs/containers on Windows or macOS and still gain most of the benefits.

Debate over an AI bubble

• Many see a “massive AI bubble,” likening it to dot-com, 1990s rail/telecom overbuild, or 80s AI winter: real tech, but overhyped and overfunded, followed by funding collapse and consolidation.
• Others argue this isn’t like pure-speculation bubbles (NFTs/crypto/tulips) because Nvidia and many AI products already generate substantial revenue and real usage.
• Several expect a pop in valuations and capex growth, not in the underlying tech, similar to how the internet thrived after the dot-com crash.

Nvidia, GPU demand, and future glut

• Nvidia’s profits and margins are viewed as “insanely high”; some expect competition or a capex slowdown to compress them.
• Commenters anticipate that any severe shortage will eventually be followed by a glut of used datacenter GPUs (like post-crypto), though quality and reliability of ex-datacenter cards are debated.
• Gamers hope for cheaper GPUs, but many doubt much “trickle down” due to segmentation and higher-margin datacenter priority.

Who are the “mystery whales”?

• Most assume the big four buyers are hyperscalers: Microsoft, Meta, Google, Amazon; some articles cited in-thread say exactly that.
• Other large buyers mentioned: Oracle, CoreWeave, Lambda, Chinese cloud companies, Tesla/xAI.
• Some speculate about indirect government/NSA/DoE demand, but note such purchases would likely be routed through intermediaries.

Custom silicon, CUDA moat, and competition

• Cloud and big-tech efforts: Google TPU, AWS Trainium, Meta MTIA, Microsoft Maia, Tesla D1, plus specialized players (Groq, Cerebras, etc.).
• CUDA is seen as a major moat; AMD/Intel and others have struggled to attract training workloads despite hardware.
• Ideas discussed: CUDA-compatibility layers and open alternatives to gradually weaken Nvidia lock-in; skepticism remains about difficulty and incentives.

Open vs closed and self-hosted AI

• Some companies are commissioning ~$20k in-house AI servers running open-source models, citing flexibility and richer APIs than proprietary services.
• There’s uncertainty whether proprietary “frontier” models or a diverse open-source ecosystem will dominate long term.

Use cases, productivity, and limits

• Reported uses: search/lookup, translation, moderation, coding assistance, writing, design, data analysis, medical and legal document work, recommendation systems.
• Individual experiences vary: some feel AI is a “new power” and increasingly indispensable; others see only modest productivity gains and fading novelty (e.g., image generation).
• Open questions raised:
  • Whether LLM quality is plateauing and hallucinations can be tamed enough for broad deployment.
  • How big non-LLM markets (robotics, autonomy, scientific computing, drug discovery) will be, and whether they sustain current GPU growth.

Concentration and systemic concerns

• Worry that AI progress and infrastructure are consolidating into a handful of tech giants and clouds; some advocate open source and Linux as partial counterweights.
• Others note that even if this is a bubble, like railroads or dark fiber, the overbuilt infrastructure could still provide long-term economic benefit.

Police use of Tesla/Sentry footage & towing

• Many see towing uninvolved Teslas for possible footage as an overreach, especially when owners are “innocent bystanders” and rely on the car for work, emergencies, etc.
• Others argue it’s legally analogous to long‑standing search warrants for premises with evidence, even when the premises’ owner isn’t a suspect.
• Key disputes: whether this is a “fishing expedition,” whether the warrants are truly based on probable cause (e.g., just “a Tesla was nearby”), and whether police should be liable for towing costs, damage, and downstream harms to owners.
• Some point out the article’s examples mostly involve serious violent crimes and not routine crime, and that police say they tow only when they cannot reach the owner.

Broader surveillance and privacy concerns

• Thread repeatedly links Tesla Sentry with Ring, dashcams, CCTV, ALPR, car telemetry and future AV fleets as components of a growing, inescapable surveillance mesh.
• Worries: erosion of anonymity in public, police and government buying data from companies (4A “workarounds,” third‑party doctrine), and mission creep from serious crimes to minor infractions.
• Some defend extensive camera use for solving murders and dangerous crimes; opponents warn of a de facto panopticon and easy repurposing for political or social control.

Effectiveness and incentives of policing

• Several commenters say Bay Area police ignore property crimes even with clear video (e.g., Tesla break‑ins, car theft), yet will tow cars for homicide evidence; this selective attention erodes willingness to cooperate.
• Discussion of resource limits, institutional incentives to maintain fear of crime, and “catch and release” prosecution that discourages enforcement.

Dashcams, Sentry mode, and personal tradeoffs

• Users share Sentry/dashcam success stories (hit‑and‑run, garbage truck damage) and failures (police no‑action).
• Many keep Sentry off due to heavy vampire drain and privacy worries; some prefer simple dashcams not networked to OEMs.
• Legal landscape differs by country: in the US public filming is broadly allowed; in parts of Europe/Scandinavia and Switzerland, fixed or automated public recording can be restricted or technically illegal.

Crowdsourced traffic enforcement & insurance

• A long subthread explores using dashcam + ML to automatically report reckless drivers to insurers or police.
• Supporters think high certainty of being caught would deter maniacs and enable actuarially fairer premiums.
• Critics foresee abuse (targeting disliked people, editing video), over‑policing of minor infractions, “Karenization,” bounty systems like NYC’s idling program, and massive expansion of commercial surveillance.
• Legal constraints in Europe (privacy and ML on public footage) contrasted with looser US environment.

Drones, guns, and “escaping” surveillance

• Heated debate over shooting down drones and other cameras in rural areas:
  • One side fantasizes about moving off‑grid and physically disabling surveillance.
  • Others emphasize it’s a serious federal crime (drones = aircraft), very risky, and practically traceable.
• Meta‑point: even in the countryside, road and consumer cameras make true escape from surveillance difficult.

Law, rights, and future of driving

• 4th Amendment discussions: warrants vs. “unreasonable” seizures; whether towing a bystander’s car strictly for data crosses a new line.
• Some advocate encryption and owner‑only access to camera data; others note police can still compel decryption or seize media under warrant.
• Speculation on autonomous vehicles:
  • Shift from privately owned cars to corporate fleets would concentrate surveillance with operators (e.g., Waymo); blurring may be legally mandated but law‑enforcement backdoors are expected.
  • Future licensing/insurance regimes could make manual driving a costly “luxury” or restricted class, while AVs become the default.

Accuracy and Mechanics of “Buy, Borrow, Die”

• Several commenters challenge the Reddit write‑up’s technical accuracy, especially around U.S. rules on step‑up in basis, estate vs. capital gains taxation, and whether the estate ever owes capital gains.
• There is disagreement on whether moving assets into irrevocable trusts is itself a taxable event and whether the write‑up glosses over estate tax (≈40%) on large estates.
• A key point of contention is whether loans at 0.5–3% interest “maturing at death” are realistic, and whether they can legally avoid minimum interest rules (AFR). Some argue these must still be treated as loans; others say they can be structured as securities with “stock appreciation rights.”

Lender Incentives and Loan Terms

• Skeptics ask why a bank would write very low‑rate, decades‑long, interest‑only loans against volatile assets.
• Supporters say lenders get secured exposure, interest plus a slice of appreciation, and can structure products as hybrids between debt and equity.
• Multiple commenters argue sub‑AFR interest claims are simply false and call the Reddit post a likely “LARP.”

Who Can Use It and When It’s Worthwhile

• The original write‑up (as quoted) claims this only makes sense above roughly $300M net worth, where bespoke credit is available; below that, borrowing costs make it marginal or worse than selling and paying tax.
• Others note similar but more basic versions exist for “normal” people: margin loans, HELOCs, borrowing against whole‑life insurance, but without the same tax leverage.

Risk and Practical Limits

• Several point out leverage risk: market crashes or asset devaluations can trigger margin calls and forced liquidations, with real‑world examples cited.
• Even very rich borrowers typically only draw a small percentage of their portfolio annually; aggressive borrowing is seen as dangerous and has burned some.

Tax Policy Debates

• Many argue the real “bug” is the step‑up in basis at death, not unrealized gains per se. Removing step‑up (possibly with exemptions for modest homes/farms) is widely suggested.
• Others propose treating collateralizing assets as a realization event, or moving to a progressive consumption tax.
• There is extensive argument over taxing unrealized gains: some see it as an obvious fix; others warn it would punish volatile or failed investments and suppress entrepreneurship.

Broader Views on Wealth and Tax Avoidance

• Thread splits between those seeing this as emblematic of unfair elite tax avoidance and those viewing tax minimization as rational, even ethically preferable given perceived government waste.
• Side discussions cover wealth concentration, inheritance fairness, property tax as de facto tax on unrealized gains, and whether rich people’s obsessive tax planning is itself part of why they remain rich.

Meta / Credibility and SEO

• Multiple commenters are suspicious of the brand‑new subreddit and account, seeing it as SEO‑driven or narrative‑shaping ahead of U.S. tax debates.
• Others note the legal citations look plausible but emphasize there is little concrete, public evidence for widespread use of the exact lifetime, ultra‑low‑rate structures described.

Platform migration and comparisons

• Many expect Brazil’s X ban to push users to Bluesky and Threads; some see Mastodon as too complex and fragmented for “normal” users.
• Others counter that Mastodon signups are actually spiking, citing user-count bots.
• Some argue Bluesky has the best chance as a Twitter-like product, but note missing video and small team may limit DAU retention.
• Threads is seen as huge by MAU but tainted by its Instagram tie-in; some refuse to join for that reason.
• Several users report personally quitting X, moving to Bluesky/Mastodon, and not missing it; others still see X as where “the action” is.

Brazil’s X ban: law vs. free speech

• One camp says X refused lawful orders to block accounts tied to warrants and incitement of violence / coup attempts, and to maintain a legal representative, so blocking X follows Brazilian law.
• Another camp calls the key Supreme Court justice an unaccountable “little dictator,” alleging secret, extralegal censorship orders, threats to imprison local reps, VPN fines, and selective enforcement.
• There is sharp disagreement on whether the 2023 Congress invasion was an attempted coup or just a protest that turned violent.
• Polls are cited claiming a large majority of Brazilians oppose the ban, but others question framing and sample.

Musk, hypocrisy, and global double standards

• Many highlight X’s high compliance with censorship requests in Turkey, India, etc., and Musk’s own word bans and suspensions, arguing his “free speech absolutism” is selective and often aligns with right-wing interests.
• Defenders frame the Brazil stance as civil disobedience against unconstitutional orders; critics say you must still obey and appeal.

Bluesky / ATProto tech and moderation

• Bluesky staff describe their architecture (SQLite repos, Go event stream, ScyllaDB views, hybrid on-prem/cloud) and note the Brazil surge stressed systems but didn’t fully break them.
• ATProto emphasizes:
  • Decentralized personal data servers (PDS) and DID-based identity.
  • Pluggable feeds and “stackable” moderation/labeling that users choose.
• Debate over whether this model really helps resist censorship, or just changes which levers governments pull (blocking apps, relays, domains, or local devs).

Decentralized alternatives and fragmentation

• Nostr, Mastodon, and other federated/protocol-first systems are discussed; nostr is seen as more censorship-resistant but overrun by crypto culture and spam by some.
• Several predict increasing regional social-media balkanization, with national bans, VPN workarounds, and multiple overlapping networks rather than a single “town square.”

Normative debates about speech

• Strong divide between:
  • Those who argue free speech should be near-absolute and “disinformation” is an excuse for repression.
  • Those who see limits on incitement, hate speech, and coup advocacy as necessary to protect democracy and minorities.
• Multiple commenters note that every side now accuses the other of weaponizing “censorship” and “disinformation” rhetoric.

Definitions: What Is “Art”?

• Major split: some define art as inherently human expression or communication, so non‑human systems cannot make art; AI is only a tool.
• Others argue that tying “art” to the maker (human vs machine) is arbitrary or gatekeeping; they prefer definitions based on the object or the viewer’s response.
• Several note the lack of consensus and say debates often collapse into semantics or compliments (“art” = “I like this”).

AI as Tool vs Creator

• Many frame generative models as tools like cameras, brushes, or instruments that augment artists rather than autonomous creators.
• Others stress that tools that augment humans are often marketed as replacements, with real labor‑displacement risks.
• Some liken sophisticated prompting, iterative refinement, and workflows (e.g., compositing, inpainting, LoRAs) to directing or photobashing, and thus part of an artistic process.

Process, Effort, and Intention

• A recurring theme is that art arises from countless small, intentional choices; a single text prompt plus one‑shot output is seen as shallow.
• Several artists say the joy is in the process; “prompt engineering” doesn’t satisfy that for them.
• Counterpoint: historically, many “great” artists used assistants or “ghost‑painters,” suggesting that high‑level direction can still count as authorship.

Quality, Novelty, and “Slop”

• Some experience AI output as bland, averaged “slop,” good for stock imagery or spam but not for meaningful work.
• Others cite examples of striking AI‑assisted work, argue Sturgeon’s law (most of everything is bad), and say low skill floors always flood media with low‑effort content.
• There’s disagreement over whether current AI images “all look the same”; critics point to systematic artifacts (e.g., lighting), defenders say serious practitioners know and work around these limits.

Economic and Social Effects

• Concerns about “enshittification”: businesses replacing human content with mediocre AI, plus humans using LLMs to expand bullet points that are then summarized by other LLMs.
• Some think AI lowers barriers for “outsider” creators who lack time or training; others respond that bypassing skill development yields shallow work.

Future Capabilities

• Some insist current systems can’t truly make art and extrapolate that forward.
• Others call it foolish to draw hard ceilings from today’s models; they expect future systems or AGI to achieve genuine artistic sensibility, while noting consciousness/qualia remain unresolved and unclear.

Debian vs bcachefs‑tools: what happened

• Debian’s Rust policy discourages heavy vendoring and expects crates to come from Debian’s own archive, built offline and reproducibly.
• bcachefs‑tools pins exact Rust dependency versions, includes locally modified crates, and expects those versions to be used.
• Debian maintainers tried to “unbundle” and replace vendored crates with distro versions (including an older bindgen), which broke builds and, more seriously, runtime behavior (mount options not passed; degraded mounts failing).
• After technical friction and increasingly hostile interactions with upstream, the Debian maintainer orphaned the package.

Vendoring vs distribution policies

• Distros argue: one version of each library and no vendoring keeps security manageable; they can patch a single copy and avoid chasing embedded forks (citing earlier zlib experience).
• Vendoring advocates argue: for critical tools (filesystems, backups) exact, tested dependency sets are essential; distro‑level rewiring of deps reintroduces bugs and makes debugging and bisection harder.
• Some suggest vendoring is fine for statically linked Rust/Go binaries; others respond that security teams still need to track and patch source, vendored or not.

Rust ecosystem and dependency churn

• Several comments say the problem is ecosystem‑wide (Python, Ruby, ML stacks) but Rust’s culture of many small crates, weak standard library, and Cargo’s model amplifies it (dozens of transitive deps, multiple 0.x “major” versions).
• Others note that in Rust it’s idiomatic to depend on semver ranges, not exact versions, but tooling for multiple major ranges is weak.
• There’s debate on whether distributions misunderstand Rust versioning (e.g., treating 0.2 and 0.4 as “minor” instead of separate majors).

Stability, security, and release models

• Debian is characterized as “waterfall” and long‑term‑support‑oriented, mismatched with fast‑moving software and immature ecosystems.
• Supporters say that’s exactly Debian’s value: audited, reproducible, offline‑buildable, long‑lived systems where the distro—not every upstream—owns security backports.
• Critics argue modern software stacks make that model increasingly untenable; suggest using distros as a stable base plus language/alt package managers (Nix, cargo, containers) for fast‑moving apps.

Perspectives on bcachefs and its maintainer

• Some see a pattern of conflict with conservative processes (Debian, kernel release discipline) and describe upstream as overly aggressive.
• Others emphasize the value of driven “move it forward” individuals for ambitious projects like new filesystems, while acknowledging social/process friction.

Perceived Support and Lobbying

• Some see ChatControl as persistently on the agenda, driven by lobbying and fear of terrorism and CSAM, especially among certain parent groups.
• Others from the same countries say they know no one who supports it and argue it should be put to a referendum, suggesting support may be highly localized or framing-dependent.

Privacy, Security, and Terrorism

• Examples like Lithuania are cited to argue strong privacy protections do not imply higher terrorism risk.
• Opponents stress that terrorism/CSAM are repeatedly used as emotional justifications that short-circuit rational debate.
• Some explicitly prefer a small increase in successful attacks over mass surveillance that reshapes society.

CSAM Scanning and Misclassification

• Commenters mock the idea of AI judging whether someone “looks underage,” noting the absurdity and invasiveness of manual review.
• Others note this is already happening: platforms have scanned private photos and triggered false abuse accusations (e.g., medical photos of children).

Authoritarian Abuse and Political Exemptions

• Many fear such tools will be repurposed to target political opposition, citing Hungary and other states as likely abusers.
• There is widespread suspicion politicians will exempt themselves and government communications, revealing the real intent.
• Some argue if monitoring is “for safety,” it should apply especially to politicians, up to full black-box logging.

Democracy, Hungary, and Electoral Manipulation

• Long tangent on Hungary and similar systems: claims of media capture, gerrymandering, diaspora voting, and patronage networks versus defenders citing electoral majorities.
• Dispute over whether such governments are genuinely democratic or managed democracies with weakened checks and balances.

Surveillance Trajectory and Social Credit

• Several see an “inevitable” drift toward broader surveillance and quasi–social-credit systems, growing from credit scoring, data brokers, and pervasive identity tying.
• Others reject inevitability narratives, arguing that resistance (legal, political, civil-society groups like EDRi/EFF) can still meaningfully slow or stop this.

Technical Workarounds

• Many note criminals can easily add their own encryption on top of monitored platforms or use alternative tools, making controls mainly a dragnet on ordinary users.
• Mesh and alternative networking projects (e.g., Meshtastic, Reticulum) are discussed as imperfect but evolving options for censorship-resistant communication.

Legal and Institutional Responses

• Suggestions include sunset clauses, constitutional challenges in EU courts, mass protest, and sustained political organizing.
• Skepticism remains that once such powers are granted, they will be routinely renewed and difficult to roll back.

Overall concept and claimed advantages

• System reduces iron oxide with hydrogen in summer, then re‑oxidizes iron with steam in winter to regenerate hydrogen and heat.
• Main selling points in the thread: cheap, abundant, non‑toxic materials; solid, inert long‑term storage; usable waste heat (e.g., district heating); potential for campus‑scale demonstration.

---

Efficiency, thermodynamics, and losses

• Reported lab round‑trip efficiency is ~11%, with theoretical values up to ~79% if scaled and well‑insulated.
• Multiple commenters highlight many loss stages: electrolysis (60–80% efficient), iron-oxide reduction, hydrogen release, and final conversion to electricity or motion.
• Some back‑of‑the‑envelope chains suggest overall efficiency could fall well below 50%, especially when including compression, piping, and conversion.
• Others argue efficiency matters less if input electricity is very cheap or surplus; heat co‑use (district heating, industrial processes) could improve system value.

---

Comparison with other storage technologies

• Batteries: Seen as far more efficient; sodium‑ion and sodium‑sulfur mentioned as potentially very cheap for grid storage. Critics say an iron/hydrogen path starting at ~30–60% net efficiency is non‑competitive for anything but seasonal storage.
• Seasonal storage: Some argue this competes with water reservoirs or synfuels, not daily‑cycling batteries.
• Hydrogen storage: Compressed H₂ tanks have near‑perfect storage efficiency but require high pressure and face leakage and embrittlement; supporters of iron note its safety and density.
• Related concepts: iron‑air batteries, direct reduced iron as fuel, peroxide, ammonia, metal hydrides, “baking soda” hydrogen storage, and synthetic hydrocarbons (methane, propane, kerosene) are all discussed as alternatives with their own trade‑offs.

---

Economics, “free” energy, and demand elasticity

• Debate over whether “free” summer solar will really exist: some expect highly elastic demand (smart appliances, flexible industry) to absorb cheap power so prices rarely hit zero.
• Others think overbuild will still create long cheap periods, where even lossy storage or energy‑intensive industries (e.g., smelting, compute) make sense.
• Question raised whether simply adding more PV (especially where winter drop is modest) might be cheaper than complex hydrogen‑based seasonal storage.

---

Safety and practical issues

• Fine iron powder can self‑heat and pose fire risks; direct reduced iron is regulated as a hazardous bulk cargo.
• Hydrogen leakage, material embrittlement, and high‑pressure infrastructure are ongoing concerns.
• Some see handling dense solids at very large scale as a logistical challenge; others note iron oxide can be stored cheaply “in a hole in the ground.”

---

Broader system debates

• Disagreement over the role of hydrogen at all in the energy system; some view it as inherently inefficient, dangerous, and historically unsuccessful at scale.
• One subthread argues nuclear could provide reliable baseload without such storage complexity; others counter with current nuclear cost, deployment time, and safety/approval barriers.
• General consensus that no single storage technology will solve intermittency; a “hierarchy” of options (batteries, thermal, chemical, reservoirs, demand‑shifting) is likely needed, but the precise niche for iron-based hydrogen storage remains unclear.

Calculator bug and immediate reactions

• Screenshots show misaligned digits and controls in macOS Calculator’s programmer mode.
• Some see multiple layout issues: wobbly binary digits, off‑center labels (“Unicode”, 8/10/16 selector), misaligned bit indices, uneven button padding.
• Restarting Calculator reportedly clears the misalignment, suggesting a transient or stateful bug.
• One commenter notes the bug appears fixed in the Sequoia beta.

Possible technical causes

• Several point to font rendering: hinting, antialiasing, and subpixel positioning on low‑DPI or non‑“Retina” displays.
• Others argue it’s not just fonts: some elements look like they’re positioned at half‑pixels, causing inconsistent aliasing.
• There’s debate over kerning (horizontal spacing only) vs other font metrics; some incorrect uses of “kerning” are corrected.
• Some suggest accumulated floating‑point rounding errors in the layout engine; others think that is unlikely.

Hardware, displays, and text clarity

• Multiple reports that modern macOS looks bad on non‑Retina or awkwardly scaled displays (e.g. 27" QHD): blurry text, poor hinting.
• Apple removed subpixel rendering in recent macOS versions; some say this makes anything below ~4K effectively “unusable” for text compared to Windows/Linux.
• Workarounds like fake 2× scaling and third‑party tools (e.g. BetterDisplay) are mentioned.

Random bitflips vs systematic bugs

• A minority speculate about random hardware errors (cosmic rays, marginal timing on GPU/VRAM) causing one‑off visual glitches.
• Others consider that too rare to explain repeatable UI issues and lean toward real software bugs instead.

iOS Calculator and input UX

• Many criticize the iOS Calculator: fast taps often fail to register, apparently due to touch‑handling that cancels a digit if the finger slides off the key.
• Very similar complaints about lock‑screen passcode entry: sluggish, inconsistent feedback, and the same “tap then slide cancels” behavior.
• Some defend this interaction as consistent with traditional “press‑then-drag-to-cancel” buttons; others call it uniquely confusing and buggy on the lock screen.

Broader Apple software quality and feedback process

• Multiple anecdotes of long‑standing bugs in Notes, TextEdit, Screen Time, Maps, etc., and a perceived decline in Apple’s software polish.
• Feedback Assistant and apple.com/feedback are cited; experiences vary from “bugs fixed in next minor release” to “ignored or closed as won’t-fix.”
• Some lament the lack of a public bug tracker and rely on third‑party mirrors (e.g. OpenRadar) or HN itself to surface issues.

Vision, accessibility, and design

• Several tie the bug to a larger theme: subtle misalignments, small/light fonts, and low contrast are hard on aging eyes.
• There’s extended discussion of chromatic aberration and chromostereopsis in glasses, and how high‑contrast colored UI elements can appear misaligned or “3D.”
• Commenters urge designers to anticipate deteriorating vision and favor clearer, higher‑contrast, and less fragile layouts.

Human assumptions about animal intelligence

• Many argue it’s odd we’re still “surprised” by smart animals; they see this as a form of human arrogance, akin to old ideas like Earth being the center of the universe.
• Others say it is surprising because humans still appear qualitatively different, though when you try to define the difference, exceptions (e.g., animal tool use) appear.

Crow cognition and capacities

• Crows and other corvids are cited as having dense, efficient brains, complex social structures, tool use, multi-step planning, and even apparent cooperation with other species (e.g., guiding wolves to carcasses).
• The new result about “mental templates” is seen as an extension of prior work on New Caledonian crows and other birds; some note the real novelty is that such abilities may be more widespread across crow species.
• There is debate over whether behaviors are “true” flexible learning vs. instinctive or mimicry-based.

Anecdotal observations

• Numerous stories: crows recognizing individual humans for years, sharing information about “dangerous” people, holding grudges, leaving gifts (often shiny objects), and “protecting” properties from other animals.
• Several accounts describe crows coordinating attacks, mobbing predators or disliked humans, and seemingly mourning or defending dead conspecifics.
• Some feed crows and receive repeated behavioral responses (communication calls, glass fragments as “payment,” etc.).

Comparisons to AI and definitions of intelligence

• Some compare crow intelligence to large language models: smaller systems can still reason within limited domains; frozen models are likened to “brains” stuck at inference-only.
• Multiple commenters stress that “intelligence” is poorly defined and context-dependent; many species may surpass humans in specific domains (echolocation, distributed control, instinctive skills).
• There’s debate over whether all life is “intelligent” vs. reserving the term for adaptive, general-purpose cognition.

Ethics, status, and human exceptionalism

• Several argue that recognizing animal intelligence should impact how we treat and eat other species, while others emphasize human cognitive uniqueness (language, symbolic thought, meta-cognition, rapid learning in children).
• Some frame human specialness less as raw intelligence and more as hands, fire, social accumulation of knowledge, and ego or motivation.

Methodological and interpretive cautions

• A few question experimental details (number/complexity of shapes, “good enough” fits, imprinting flexibility).
• There is general appreciation for the study but pushback on sensational headlines like “smarter than we thought.”

Neurological and Cognitive Symptoms

• Many posters report post-Covid “brain fog”: word-finding problems, memory lapses, confusion, and reduced mental “resolution.”
• Some describe profound changes: dementia-like symptoms, mistaking voltage for resistance, misrecognizing faces as familiar, or forgetting ongoing tasks (e.g., leaving taps running).
• A few note insomnia or severely disrupted sleep around infection, and wonder if sleep loss contributes to cognitive issues.
• Some individuals say symptoms improved gradually over months; others report intermittent relapses years later.

Cardiovascular, Respiratory, and ME/CFS-like Effects

• Several describe post-Covid exercise intolerance: heart rate spiking with minimal exertion, “air hunger,” and hitting a hard physical “wall.”
• Comparisons are made to ME/CFS and post-viral syndromes; some note ME/CFS diagnoses roughly doubling after Covid and affecting more women.
• Reports include new-onset hypertension, chronic fatigue, and long-term disability even in previously very fit people.

Proposed Mechanisms and Related Research

• Multiple comments link Alzheimer’s and long Covid to chronic inflammation, blood–brain barrier breaches, and infections (viruses, bacteria, candida, mold toxins).
• Herpes-family viruses are repeatedly discussed as possible drivers of dementia, with citations to studies on antivirals and a shingles vaccine–dementia association.
• Some mention hypotheses involving microclots, vascular dysregulation, or T‑cell depletion; others reference histamine/MCAS, which is both promoted and criticized as unscientific.
• A cited UK Biobank study and EEG data are used to argue for real, measurable brain changes after Covid.

Evidence, Diagnosis, and Medical Response

• Skeptics argue long Covid evidence is weak, confounded by lockdown and psychosocial stress; others reply that newer, more rigorous studies exist.
• Several note that many clinicians are unaware of, or dismissive toward, long Covid and ME/CFS, defaulting to depression or “get your act together.”
• There’s frustration that nuanced symptoms are not systematically recorded, hindering pattern discovery and treatment research.

Vaccines, mRNA, and Causality Debates

• One strand blames mRNA vaccines as “gene therapy” causing clots and long-term issues; others strongly reject this, framing Covid infection itself as the primary risk.
• Calls are made to stratify long Covid research by vaccination status, though no clear consensus emerges.

Coping and Rehabilitation

• People mention rest, graded return to activity, cognitive exercises (e.g., category alphabet games), beta blockers, inhalers, and specialized rehab clinics as partial helps, but no consistent remedy is reported.

Feasibility of a Rust Linux-Compatible Kernel Clone

• Many like the idea of a Rust-based, Linux-ABI-compatible kernel to avoid Linux kernel politics and prove Rust’s value.
• Others argue it’s wildly underestimating effort: even with better tools and hindsight, you still need a large fraction of Linux’s total person-hours.
• Prior attempts at Linux-ABI compatibility on other OSes (BSDs, Solaris/Illumos, Windows, Fuchsia’s starnix) are cited as cautionary; all are incomplete or stalled.

Drivers, ABI Complexity, and Hardware Vendors

• Several comments stress that drivers, not the core kernel, are the dominant cost.
• Linux has no stable in-kernel driver ABI; drivers are GPLv2 and tied to fast-changing internals, limiting reuse.
• The Linux userspace ABI is huge, underspecified, and a moving target (syscalls, ioctls, /proc, netlink, driver-specific ABIs), making bug-for-bug compatibility extremely hard.
• Hardware vendors often withhold documentation or require NDAs, effectively “locking in” Linux; this affects any alternative OS.
• Some propose designing a new, stable driver API (with Android’s GKI cited as a partial step), but feasibility is debated.

Rust vs C in the Linux Kernel

• Rust is praised for strong memory safety guarantees that static analysis in C cannot match.
• Demonstrated Rust-in-kernel work includes: Apple M1 GPU driver, Android binder driver, a Rust network PHY driver, QR-code-on-panic work; these are seen as proof-of-concept successes.
• Critics argue maintainers don’t want to learn Rust or maintain Rust breakage; proponents counter that refusing to consider safer tools is poor technical leadership.

Community Politics, Burnout, and Governance

• Some see the Linux kernel process as hostile and “burnout-inducing,” especially for Rust contributors.
• Others insist maintainers are entitled to reject Rust code and owe nothing beyond the right to fork.
• There is disagreement whether forking/clone-building is constructive pragmatism or a grudgebased fragmentation that rewards abusive behavior.
• Burnout of at least one Rust-for-Linux maintainer prompts discussion on whether the project itself is “burned out” or simply under political strain.

Alternatives and Architectural Debates

• Some argue language changes miss the real issue: Linux’s monolithic architecture and lack of isolation for drivers.
• Microkernels and systems like seL4 or other Rust OS projects are proposed as better long-term bets; others dismiss microkernels as historically unworkable at scale.
• Clones and forks are framed by some as healthy (Xorg, GCC, Wine, Ladybird example), by others as condemning Rust devs to forever chase Linux’s decisions without influence.

Rust Ecosystem and Community Perception

• Rust’s community is described both as enthusiastic/productive and as sometimes toxic or evangelistic (“rewrite it in Rust” culture).
• Some question the lack of “killer apps,” while others list multiple widely used Rust tools and note extensive Rust adoption in large companies.

Background and Immediate Trigger

• Brazil’s Supreme Court justice ordered X to block accounts accused of spreading disinformation and supporting a failed coup.
• X refused, closed its Brazilian office, and did not appoint a required local legal representative, leading to an order to suspend X nationwide.
• When fines couldn’t be collected from X locally, the judge ordered Starlink’s Brazilian assets frozen, treating it as part of the same “economic group” due to Musk’s ownership.

Judge’s Powers and Brazilian Legal Context

• Some commenters say Brazilian law requires a local legal representative and allows suspension of services that ignore court orders.
• Others argue the judge is overstepping: opening inquiries improperly, acting as judge/prosecutor/jury, issuing secret censorship orders, and stretching constitutional powers.
• Several cite Brazilian constitutional provisions protecting free expression and argue the “fake news” framework is being invented by judicial “resolutions,” not legislation.

Fines, VPN Ban, and App Store Orders

• The ruling includes:
  • Full suspension of X in Brazil.
  • Mandatory blocking by ISPs, mobile operators, and backbone providers.
  • Daily fines (~R$50k) on individuals and companies using “technological subterfuges” (VPNs) to access X.
  • An initial order for Apple/Google to remove and remotely delete VPN apps, later partially walked back.
• Many see fining users and touching VPNs as a dramatic, authoritarian overreach; a few defend it as lawful enforcement of existing internet law.

Motivations, Politics, and “Authoritarianism” Debate

• One side portrays the judge and current government as de facto authoritarian (or “judge-king”), targeting political opposition and chilling dissent.
• Another side emphasizes that X is defying legitimate court orders related to anti-democratic propaganda and election denial, and Musk is choosing confrontation.
• There is intra-thread dispute about whether the government is “socialist,” “leftist,” or something else, and whether that label matters versus behavior.

Musk/X’s Conduct and Alleged Hypocrisy

• Critics note Musk complied with takedown/censorship demands in India and Turkey but is drawing a red line in Brazil, suggesting political selectivity rather than principle.
• Supporters counter that Brazil’s orders are uniquely secretive, extra-legal and personally threatening to employees, justifying exit and resistance.

Technical Feasibility and Enforcement

• Many doubt large-scale technical enforcement of VPN fines is realistic; selective enforcement against public figures and critics is seen as more likely and more dangerous.
• Discussion covers DPI, SIM/ID linkage, side-loading, Tor/onion services, and alternative tunneling (SSH, SOCKS) as circumvention paths, but with risk.

Impacts on Citizens, Business, and Democracy

• Commenters highlight:
  • Loss of income for creators and small businesses dependent on X.
  • Loss of access to global discourse (politicians, scientists, activists).
  • Chilling effect on VPN use, remote work, and privacy tools generally.
  • Broader signal that Brazil may be becoming a hostile environment for foreign investment and for free speech online.

Debate on the Three Suggested Questions

• IC with an amazing idea:

  • Seen as a strong question for exposing how ideas flow: is there a healthy path to triage and ship IC ideas, or do they die in PM/executive layers?
  • Some say “talk to a PM” as the default path is a red flag (signals “keep your head down”), others argue it’s fine if it means a structured “get buy‑in” process.
  • Discussion notes that in many orgs, real product change must be funneled through managers or exec “sponsors,” revealing status-driven cultures.

• Last major migration & estimates:

  • Critics say this forces interviewers to either admit failure (migrations always slip) or lie, so it yields low-quality or defensive answers.
  • Suggested alternatives: ask about challenges, how scope/technical debt was managed, or how the first 6 months for a recent hire went.

• “Six months later, what’s different?”:

  • Some view this as a bad or self-centered question: impact is primarily the employee’s responsibility.
  • Preferred variants: “What would convince you you’d hired the right person in 3–6 months?” or direct questions about onboarding and expectations.

---

Interviews as a Two-Way Evaluation

• Many endorse using the Q&A portion to seriously vet the team, not just perform.
• Others dislike the pressure to ask “smart” questions, seeing it as selecting for people who rehearse canned questions, especially unfair to desperate or long-term unemployed candidates.
• Some interviewers say candidate questions don’t affect hiring; others admit overall impression and curiosity do matter, even with standardized rubrics.

---

Alternative Questions People Actually Use

• About product and engineering quality:

  • How IC ideas have become real features.
  • What and how they measure (performance, defects, engagement).
  • Build times, test automation, migration frequency, and handling of technical debt.

• About culture and team dynamics:

  • “What sucks / what are the biggest challenges about this job?” or “If you had a magic wand, what would you change?”
  • Average tenure, recent departures, and reasons people leave.
  • “Day/week in the life” and “What would you be doing if you weren’t interviewing me today?”

---

Views on Product Management, Leadership, and Process

• Strong disagreement over PM-heavy orgs: some see them as necessary vision-setters, others as bottlenecks and idea-killers.
• Debate over whether ICs are “leaders”; some emphasize leadership via influence/mentorship even without reports.
• Concerns about standardized hiring reducing bias but also flattening “interesting” candidates and interactions.

---

Critique of Hiring Culture and Advice Articles

• Several comments deride the interview process as performative, hostile to experienced engineers, and dominated by HR gatekeeping.
• Some see this kind of “turn the tables” advice as shallow social engineering or overfitted to one person’s context.
• Others still find the core idea valuable: use questions to uncover real culture and avoid wasting years in a bad environment.

Vegetarianism, Meat, and Relative Risk

• Several commenters say stories like this reinforce their choice to avoid meat or deli products.
• Others counter that vegetables are also frequent sources of Salmonella, E. coli, and Listeria (onions, lettuce, cucumbers, sprouts, melons, peanuts, etc.).
• Point made that raw leafy greens are often the most dangerous items in a kitchen because they’re eaten uncooked and can’t be washed perfectly.
• Some argue being vegetarian doesn’t free anyone from microbial risk; microbes are ubiquitous (water, produce, etc.).

Contamination Sources and Health Outcomes

• Discussion of specific pathogens and vectors: listeria in deli meats, E. coli on salads, Salmonella on produce, rat lungworm from slugs on vegetables.
• Emphasis that cooking meat generally reduces pathogen risk; raw produce is harder to make safe.
• Several people share experiences of food poisoning from salads and highlight severity of listeria for pregnant women and immunocompromised people.

Industry Practices and Scale

• Many see the described plant conditions (blood puddles, mold, flies, heavy meat buildup) as far beyond “normal messiness.”
• Some note that small-scale or pasture-based operations they’ve seen are much cleaner and that large-scale industrialization is what feels “gross,” across both plant and animal agriculture.
• Others say it’s hard to know how atypical this plant is without broader baseline data on violations.

Regulation, Enforcement, and Accountability

• Strong sentiment that food production must be clean and that failure is both managerial and regulatory.
• Debate over whether the U.S. inspection system is “working”:
  • One side argues it failed because issues only surfaced after illnesses and deaths.
  • Another argues rare but serious failures can still coexist with an overall functional system.
• Calls for stricter, more frequent inspections, meaningful shutdowns, and penalties that may bankrupt repeat offenders.
• Some mention regulatory capture, underfunding, and political hostility to regulation as root problems.

Consumer Reactions and Transparency

• Multiple commenters say they will stop buying from the company and shift to local or imported cured meats, or avoid deli meat entirely.
• Support for publishing all inspection reports online so consumers can evaluate suppliers, with the caveat that the results might be disturbing.

GLP-1 History, Missed Opportunity, and Hindsight Bias

• Pfizer/MetaBio abandoned GLP-1 in 1990 despite efficacy and patents; Novo Nordisk restarted work shortly after and eventually created blockbuster drugs.
• Commenters debate whether this was an obvious mistake vs a reasonable decision given then‑limited obesity prevalence, long timelines, and competing pipelines.
• Several note “resulting”/hindsight bias: pharma has many parallel bets (e.g., obesity drugs, Alzheimer’s pathways, CETP inhibitors) and most fail despite smart, persistent teams.
• Another angle: MetaBio’s startup structure was flawed (wholly owned from day 1, founders not fully committed), which left them powerless when the parent shifted priorities.

Injections, Patient Acceptance, and Delivery Tech

• Disagreement over how much injection aversion mattered in 1990: some recall large syringes and clinic visits; others point out disposable syringes and short needles existed and subcutaneous injections are manageable.
• Needle phobia and pain sensitivity are described as real barriers for some patients, influencing route-of-administration decisions.

Obesity, Diet, and Corporate vs Personal Responsibility

• US obesity rose from ~17% (1990) to ~40% today; some estimate many deaths could have been avoided had GLP-1 therapy been available earlier.
• Debate over primary drivers:
  • One side stresses cheap sugar/HFCS, 1970s farm policy, and “low‑fat, high‑sugar” reformulations.
  • Others cite data showing sugar intake has fallen since 2000 while obesity still rose, arguing hyper‑palatable foods (fat + sodium/sugar) and processed food engineering are bigger culprits.
  • Some emphasize parental/individual responsibility; others stress corporate manipulation and structural food environment.

Pricing, Patents, and Compounded Semaglutide

• Strong criticism of US pricing (~$1k/month list, with limited insurance coverage for weight loss vs diabetes indications).
• Compounded semaglutide is much cheaper while the drug is on the FDA shortage list; commenters explain that status temporarily permits compounding without licenses, but this will end.
• Some compounders reportedly crush oral Rybelsus tablets as API; others worry about opaque sourcing and “shady” operators.
• International price gaps (e.g., UK much cheaper) are noted, often blamed on US policy and lobbying.

Regulation, FDA Conservatism, and Access to Risky Drugs

• Tension between protecting patients from harmful drugs vs the “hidden graveyard” of people who die because beneficial drugs are delayed or never developed.
• Proposals: multi‑tier approval (experimental classes, noncommercial/terminal‑only), “accredited ingestor” status, or easier use of drugs approved in other rich countries.
• Counterarguments stress fraud risk, desperation of terminal patients, manufacturing limits, and that even early‑stage drugs often fail on safety/efficacy.
• Historical examples (e.g., thalidomide abroad vs delayed beta‑blocker approval in the US) are used on both sides to argue the FDA is too lax or too strict.