Hacker News, Distilled

Target audience and platforms

• Thread disputes who this editor is “for.” The README says: users unfamiliar with terminals needing an accessible editor, especially on Windows.
• Several argue it’s primarily a Windows 11 terminal editor that happens to run on Linux/macOS, analogous to PowerShell being cross‑platform but not “for Linux.”
• Suggested use cases: Windows Server Core/Nano over SSH; admins who can’t use vi; scientists on clusters needing to edit SLURM scripts; people who just want a Notepad-like tool in a terminal.
• Others say such users will never discover a GitHub repo and that clusters should instead provide higher‑level frontends.

Features, performance, and limitations

• Praised for being a single, tiny, dependency‑free Rust binary (~200–220 KB) with mouse support, menus, fuzzy search, and regex find/replace.
• People marvel at newline‑scanning performance (SIMD up to ~125 GB/s) but debate whether this is a meaningful metric versus “fun optimization.”
• Major limitation: no syntax highlighting or LSP yet. Some see this as disqualifying versus micro/nvim; others note an open issue and focus on keeping binary size small.
• Scripting/plugins are planned via DLLs; some advocate WASM instead, but that’s only under discussion.

Comparison with other terminal editors

• Frequently compared to nano, micro, mcedit, dte, jed, tilde, Turbo Vision‑based editors.
• Some want it as a “saner nano replacement” with CUA keybindings and mouse, especially for beginners.
• Others defend vi/vim as ubiquitous and worth learning; detractors call its UX hostile and overkill for casual users.
• Micro is repeatedly recommended as a richer TUI (Lua plugins, large‑file handling) but heavier; some complain about Go binary size and bloat.

Microsoft motives and ecosystem context

• One explanation: Microsoft needed a small, SSH‑friendly, modeless editor bundled with Windows; this is not just a “for fun” rewrite of EDIT.COM, though it intentionally evokes it.
• Skeptics see it as “nerd‑washing” or groundwork for future Copilot/AI integrations, noting bloat in Notepad and other Microsoft apps.
• Broader tangents discuss PowerShell on Linux, WSL, and speculation about deeper Microsoft moves into Linux userland.

Security and distribution concerns

• winget is criticized as a serious supply‑chain risk (open manifests, many random packages, weak author control); defenders compare it favorably to curl | bash and say manifests plus hashes and human review are sufficient.
• Some prefer alternative installation paths (nix, manual build) and want packages for Flatpak/Snap/BSD.

Conflict-of-interest and indirect ownership

• Some question article disclaimers claiming “no positions” in Berkshire, noting that any broad-market ETF effectively creates indirect exposure.
• Others point out the fine print: such disclosures usually exclude broad-based ETFs and mutual funds, and this is standard practice to avoid conflict-of-interest issues.

“Fair share” and tax loopholes

• “Fair share” is seen as undefined and political; suggestions range from “whatever Congress decides” to “when everyone feels equally unhappy.”
• One side equates fairness with not using aggressive loopholes (e.g., corporations paying zero or negative tax on large profits).
• Others argue that if it’s explicitly legal, it’s not cheating; law is based on text, not intent. Critics counter that lobbying-created loopholes and “dark store” property tactics show how rules are bent beyond the spirit of the law.
• Retirement vehicles (IRAs, Roth, 401(k)s, backdoor Roths, Peter Thiel’s IRA) become a case study: are they legitimate incentives or abusive when used by billionaires?

Corporate vs individual taxation

• Debate over whether corporate tax should exist at all:
  • One camp: corporations are legal persons using public infrastructure, so they should pay.
  • Another: corporations are capital-allocation machines; profits eventually become personal income, so tax individuals and capital gains instead.
• Long thread over “double taxation”: some say taxing both corporate profits and shareholder income is double tax; others say it’s just taxing transfers between distinct legal persons, same as paying a plumber and then a landlord.

Profit, reinvestment, and zero-tax companies

• Berkshire’s 5% share of corporate tax is seen either as proof it doesn’t aggressively avoid tax, or as evidence that many peers underpay.
• Insurance and retained earnings (no dividends, long-term holdings) are cited as reasons its corporate tax bill is large.
• Examples like Amazon and other big firms show how reinvestment and deductions can drive reported profit (and tax) to zero for years; defenders call this an intentional growth incentive, critics say it favors monopoly-building over smaller, profit-taking competitors.

Economic impact and system design

• Linked research claims corporate taxes hurt GDP more than income taxes; skeptics call this corporate-friendly framing and note high-tax eras coincided with stronger labor purchasing power.
• Some argue corporations allocate capital more efficiently than governments or individuals, so taxing them heavily reduces overall output; others highlight inequality, capital concentration, and the limited trickle-down to workers.
• Several comments suggest shifting more burden to payroll or personal income taxes, but others warn individuals already face taxes on earnings, spending, property, and inheritance, while corporations exploit more planning options.

Enshittification, Southwest, and Investor Pressure

• Several commenters frame “basic economy” as part of broader “enshittification”: deliberate degradation of the base product to extract more revenue.
• Southwest is cited as a case study: once differentiated by no bag fees and a distinct culture, it’s now seen as converging toward legacy carriers after its software meltdown and activist investor pressure.
• Activist investors are portrayed by some as short‑term players chasing a single “good quarter,” even at the expense of the practices that made a company durable and profitable.

Is Basic Economy Expanding Access or Just Squeezing?

• One camp argues unbundling and tight seating have made flying affordable to lower‑middle‑income travelers, including in developing countries.
• Others respond that most of the world still can’t fly; quality has collapsed, and basic fares often haven’t truly fallen once fees and “optional” costs are included.
• Some would happily fly less or pay more for a less hostile experience.

Airline Profitability, Margins, and Loyalty Programs

• Multiple comments stress airlines are structurally low‑margin; many go bankrupt despite these tactics.
• Frequent‑flyer programs and co‑branded credit cards are described as major profit centers, sometimes valued higher than the airline operations themselves.
• There’s disagreement whether US carriers are an “abusive oligopoly” with excessive profits or barely scraping by; both PRASM data and survivorship bias are invoked.

Regulation, Bundling, and Public-Good Framing

• One thread debates whether regulators should mandate bundled fares (bags, rebooking, food) versus allowing granular add‑ons.
• Analogies to public health and schools are used to argue for some community‑wide standards; skeptics say cheap restricted fares mostly help leisure/vacation travelers and aren’t a clear public good.
• A compromise view favors “all‑in” price transparency rules over banning unbundling.

Price Discrimination: Efficient Matching or Artificial Misery?

• Some see price discrimination as socially useful: richer or time‑poor travelers subsidize cheaper seats; everyone chooses their own tradeoffs.
• Others object to “artificially worse” products—extra hoops, bad default options, scary UX copy—designed solely to push upgrades, not to save underlying costs.
• Concerns extend to “everything is for sale” norms like paid priority lines, viewed as corrosive to egalitarian norms.

Corporate Behavior, Goodwill, and Exploiting Irrationality

• Commenters connect airline practices to a broader shift away from valuing goodwill toward pure extraction, especially in public and private‑equity‑owned firms.
• Sophisticated pricing algorithms, dark patterns, and intentionally noisy fare fluctuations are seen as exploiting human irrationality rather than improving core products.
• Some blame financialization and weak antitrust enforcement for creating incentives where treating customers well rarely pays.

System-Level Issues: Alternatives and Externalities

• Several note that focusing on fare games distracts from larger questions: lack of US high‑speed rail, airport slot regimes, delays/cancellations, and environmental and noise externalities.
• Examples from Asian low‑cost carriers illustrate how lower labor and operational costs, plus different regulatory and infrastructure contexts, can deliver cheap, tolerable service—highlighting what’s structural versus optional “enshittification.”

Meta: Article Quality and “AI-Generated” Accusations

• A side discussion questions whether the linked article itself is AI‑written or content‑marketing fluff, citing repetitive structure and stylistic tics.
• Others push back, noting inconsistencies that current LLMs usually avoid, and arguing that “AI” is becoming a generic insult for low‑quality writing.

Purpose of social media screening

• Many see the policy as “security theatre”: ineffective at stopping terrorism or crime because bad actors can create burner or purchased accounts.
• Others argue the real goal is to deter or exclude political opposition, especially foreign students engaged in activism.
• A minority supports “vigorous vetting” of non‑citizens, framing entry as a privilege and screening of posts, networks, and associations as common‑sense risk management.

Free speech, human rights, and foreigners

• Strong disagreement over whether foreign visitors should enjoy meaningful free speech and freedom of assembly.
• One camp: universal human rights (speech, assembly) apply regardless of citizenship; suppressing foreign dissent undermines what the US claims to be.
• Opposing camp: foreigners are “guests” who should be “respectful and quiet”; political rights properly belong only to citizens, and it’s acceptable to keep out “anti‑American” voices.

What counts as “anti‑American”?

• Several point out the term is highly fluid and can be weaponized against mainstream positions (e.g., criticism of Israel, views on Jan 6).
• There is concern that whoever controls the executive branch effectively defines “Americanism” and can impose de facto political purity tests on visas.

Effectiveness, fraud, and data permanence

• Debate over whether omitting or deactivating accounts is visa fraud, and how enforceable that is.
• Some note that entry can be denied for claiming “no social media,” making selective disclosure de facto possible.
• Many emphasize that deleted or private posts are unlikely to be truly gone; big platforms probably retain data that may surface years later.
• Anecdotes: visa applicants and students have been asked for social media passwords or to make accounts public well before this news cycle.

Racism, Israel, and broader political agenda

• Several tie the policy to a broader nativist project: ending birthright citizenship, mass deportations, and moving toward a white ethnostate.
• Others argue it is specifically about shielding Israel from criticism and punishing pro‑Palestinian activism.
• Some see knock‑on effects: declining tourism, talented students staying away, and erosion of US credibility on free speech and human rights.

Effectiveness of the strikes and bunker‑buster limits

• Many point out that known centrifuge halls are ~70–80m underground while public specs for the GBU‑57 suggest ~60m ideal penetration, so full destruction was unlikely, especially in hard rock.
• Others argue that real performance is classified and modeling/BDAs are complex; sequential “drilling” with multiple bombs into the same shaft could increase effective reach.
• Counter‑view: even with drilling, you’re still constrained by basic physics, aircraft payload, and geology; ultra‑high‑performance concrete and solid mountain rock give the defender the edge.
• Several note that bomb damage assessment for underground targets is notoriously hard; images show holes in a mountain, not internal damage. Limited bomb inventory is also highlighted.

Status of facilities and uranium stockpile

• Thread cites reports that much enriched uranium was moved ahead of the strike, but others doubt Iran could do that unobserved given Israeli surveillance.
• There’s disagreement over whether stockpiles are under rubble, moved to other sites, or “missing”; both Israeli and Western leaks are seen as politically motivated and unsubstantiated.
• Multiple commenters emphasize that you can’t “destroy” U‑235—only disperse it. Blowing up a stockpile might turn it into a difficult but recoverable “uranium mine,” delaying but not eliminating capability.

Politics, media, and intelligence spin

• Some see the operation as primarily domestic theater: enabling Trump to look reluctantly “tough,” then retroactively selling a success narrative.
• The White House’s rejection of its own leaked assessment and a canceled classified briefing are read as signs the public messaging can’t be squared with intel.
• US mainstream media are criticized as captured and war‑friendly, with anti‑war voices pushed to small outlets and social media.
• Commenters note selective trust in intelligence: embraced when it justifies action, dismissed when it undercuts it.

War, escalation, and incentives for nukes

• Many argue the strikes increase Iran’s motivation to get a bomb; North Korea and post‑Budapest‑Memorandum Ukraine are used as cautionary examples.
• Others say Iran could instead fully disband its nuclear program to remove the casus belli, though skeptics call that unrealistic given regional power politics.
• Personal accounts of Iraq/Afghanistan casualties and PTSD fuel strong anti‑war sentiment and fears of repeating past interventions.

Diplomacy, JCPOA, and compliance

• One camp blames US withdrawal from the JCPOA and targeted killings of negotiators for collapsing a working containment framework and making rearmament rational for Iran.
• Another stresses IAEA findings of past undeclared material and activities, arguing Iran was never fully compliant and used the deal as cover.
• There’s broader concern that attacking NPT‑party facilities—without clear public evidence of weaponization—undermines the entire non‑proliferation regime.

Technical side debates

• Long sub‑threads dissect penetration math, rock fracturing, sequencing of six bombs per shaft, use of air ducts as “blast channels,” and possible blast doors/compartmentalization underground.
• Some argue enriched to ~60% means Iran now needs far fewer centrifuges and could relocate to smaller, covert sites, making future detection harder even if major complexes were badly damaged.

X11 vs Wayland: What’s Actually Possible

• Several commenters say the post “proves” something nobody serious ever denied: X11 has long exposed physical display size/DPI (via RandR) and you can render at whatever resolution you want, including over the network.
• The hard part is not getting DPI but building a full scaling ecosystem: dynamic per‑display scaling, mixed‑DPI multi‑monitor, hairline crispness, protocol semantics, and toolkit support.
• Some argue many of these could have been added to X11 with more extensions or an “X12”; others say the community instead chose to clean up the stack around DRM/Mesa and start fresh with Wayland, similar to Python 2→3.

Fractional Scaling & Mixed DPI

• Major pain point: multi‑monitor setups with different DPIs. Many report that X11 still handles this poorly in 2025; others say it “could” work but toolkits and DEs never did the work.
• Debate over two approaches:
  • “Proper” per‑output scaling with clients rendering at target scale factors (Wayland fractional-scale protocol, Qt/GTK support).
  • The “2x then scale down” Retina-style approach that introduces blur and overhead but simplifies legacy support.
• Some insist fractional scaling is fundamentally flawed and unnecessary; others counter that font/vector rendering and careful snapping to pixels make it good enough in practice.

DPI vs Scale Factor UX

• Long subthread on whether users should think in absolute DPI (or PPI/PPD) vs relative scale factors (100–400%).
• Pro‑DPI side: absolute metric would make matching sizes across displays trivial and predictable.
• Pro‑scale side: people care about “how big it looks” (distance, eyesight, preference), not inches; DPI reporting is often wrong; UI sliders with percentages are more approachable. Some suggest UI could expose DPI‑like semantics while protocols/toolkits keep using scale factors.

Toolkits, Protocols, and Compatibility

• X11 core drawing APIs are pixel‑centric; modern toolkits mostly bypass them using Cairo/Skia/OpenGL and upload buffers to X, which weakens arguments about “X can’t draw circles” but also undercuts the post’s claim of using “X11 scaling” (it’s really GL paths).
• Wayland initially only had integer scale; fractional scaling is a later protocol extension that must be explicitly supported by compositors and toolkits, leading to mixed behavior and blurry XWayland apps.
• Windows and macOS are cited as examples: both rely on app/toolkit cooperation and opt‑in DPI awareness, with varying degrees of legacy blur and redraw jank.

User Experiences and Desktop Preferences

• Strongly divergent anecdotes:
  • Some say Wayland sessions (especially on new hardware) are clearly more polished, with less “jank” and better multi‑monitor behavior.
  • Others stick to X11 because Wayland sessions feel buggy, lack favorite DEs (Unity, ROX, LXDE‑style, CDE clones), or break workflows (window placement persistence, special tiling setups).
• One view: Wayland “solves” problems by dropping old features and forcing new patterns; another: X11’s complexity and bolt‑ons doomed it, and Wayland finally aligns with how toolkits already render.

Remote Rendering, GLX, and Network Transparency

• Several participants defend X11’s SSH‑forwarded UX as still very useful on LANs and in thin‑client or scientific setups, even if laggy on bad links.
• Others note X11’s protocol is extremely chatty; modern remote protocols and things like Waypipe (Wayland‑to‑Wayland) are considered more appropriate going forward, though lacking X11’s universality and ssh integration.
• There’s clarification that the article’s OpenGL example is effectively sending pixmaps, not exercising X11’s original vector drawing model.

Tearing, Performance, and “Real” Issues

• Disagreement over how serious X11 tearing is: some barely notice or fix it with TearFree options; others report chronic tearing that “just went away” under Wayland.
• Critics argue many Wayland wins (no tearing, better scaling) could have been achieved via config changes or modest X11 work; defenders reply that fundamental design (compositor model, DPI virtualization, input semantics) made a new protocol the more sustainable path.

Historical and Conceptual Context

• NeWS/PostScript and Cairo are referenced as examples of truly device‑independent, vector‑oriented models that sidestep many pixel‑grid issues, highlighting that both X11 and Wayland still operate on pixel buffers.
• Some lament that, in 2025, Linux still struggles with what older systems or printers solved via resolution‑independent rendering, and that much debate is about microscopic visual differences invisible to most users.

Enthusiasm for uv’s Speed and UX

• Many commenters report uv feeling “suspiciously fast” compared to pip, pyenv, Poetry, etc.—both for resolving/installing dependencies and starting tools.
• Speed is attributed less to Rust per se and more to design choices: cross-environment installs, smarter caching (uncompressed wheels, hard-links), avoiding bootstrapping pip into every venv, less legacy baggage, and fewer unnecessary imports.
• Several people say uv is the first Python tool that makes dependency management feel “Node-like” or comparable to Maven/Go in convenience.

How uv, uvx, and PEP 723 Work

• uv is a Python package/project manager that also installs isolated Python interpreters; it uses virtual environments, not hardware virtualization.
• uv run --script uses PEP 723 “inline script metadata” in comments to auto-resolve and cache dependencies per script; envs are ephemeral or reused via a content-addressed store, so space usage stays low.
• uvx runs console entry points from PyPI packages; it doesn’t directly run arbitrary .py files, though you can approximate this with --with and python.
• PEP 723 is now part of the official packaging spec and is also supported by tools like pipx.

Virtualenvs, Packaging, and Ecosystem Comparisons

• Long debate on virtualenv design:
  • Critics call venvs a leaky, historical hack that aren’t portable and require awkward activation compared to “just folders + config” (e.g., Java/Maven).
  • Defenders say venvs are simple directories plus pyvenv.cfg; activation is optional; non-relocatability mainly comes from absolute paths in wrapper scripts, which could be changed.
• Broader comparison:
  • Java/Maven and Go praised for single-artifact distribution and no C-extension headaches.
  • Python packaging history (setup.py, C deps, venvs) seen as messy; uv + modern PEPs viewed as finally cleaning this up.
  • Some still prefer Go/Rust for one-off binaries; others now favor Python one-shots thanks to uv + PEP 723 and richer libraries.

Shell Scripts and Cross-Language Tooling

• Several people wish shell had a uv-like dependency layer; others argue that needing dependencies means you’ve outgrown shell.
• Alternatives mentioned: Nix/Guix (nix-shell, nix run, Guix shell), mise, babashka, conda wrappers. Opinions split between “Nix is overkill but universal” and “uv/mise are simpler and more pragmatic.”

Reproducibility, Locking, and Longevity

• Concern: inline scripts without lockfiles may break over time if dependencies update.
• Responses:
  • PEP 723 allows version pins; uv supports script lockfiles and date-based resolution limits.
  • Some argue Python deps are relatively stable; others counter with examples (e.g. NumPy 2.0) and insist strict pinning is necessary.
• Many believe uv is likely to become the de facto standard, though a few remain wary of yet another third-party tool and the “infomercial” vibe of recent hype.

What “stupid” means in the thread

• Several commenters distinguish:
  • Knowledge vs intelligence vs self-control vs values.
  • “Stupid” as: acting against one’s long‑term interests when better options are available, even while knowing this.
• Others argue the word is unhelpful, undefined, and mostly a moral insult; many prefer to say “people do stupid things” rather than “are stupid.”
• Some explicitly assert that large segments of the population are cognitively limited, citing IQ distributions and personal/work experience.

Value of “obvious” research

• Multiple comments defend studies that “prove the obvious,” likening them to foundational theorems in math.
• In psychology, “boring” results tend to replicate while flashy ones often don’t; this supports building a hierarchy of small, verified claims rather than chasing surprising findings.
• Others note that what “everyone knows” is often wrong or never really examined.

AI chatbots, usefulness, and respect for users

• One line of argument: if hundreds of millions use chatbots, it’s unreasonable to say they are always useless; that presumes users are idiots.
• Critics push back:
  • Popularity ≠ value (tobacco, TikTok, etc.).
  • Something can be “useful” in a narrow, immediate sense yet net harmful; disagreement centers on what “useful” should mean.
• Some see the article as a bait‑and‑switch defense of AI that dismisses strong AI criticism by framing it as contempt for ordinary people.

Students, cheating, and rationality

• Several disagree with the article’s claim that AI‑cheating students aren’t stupid:
  • Knowing cheating is bad but doing it anyway is itself a form of stupidity or at least self‑sabotage.
  • School teaches meta‑skills (research, writing, time management); skipping them via AI may be rational only in a very short‑term, grade‑maximizing sense.
• Others emphasize structural incentives: grades, credentials, and hoops make cheating a predictable response, not pure intellectual failure.

Everyday irrationality: food, health, and habits

• Examples cited as counters to “people are smart about their own lives”:
  • Widespread poor diet and obesity despite awareness of health consequences.
  • Addictions (gambling, tobacco, alcohol) where people know the harm but can’t or won’t stop.
  • Belief in pseudoscience and misinformation.
• Some attribute this to evolved shortcuts and high discounting of the future rather than low IQ; others label it straightforward stupidity.

Driving, competence, and practice

• Driving is used as an example: many people do it constantly yet remain bad at it, suggesting limits to learning or lack of deliberate practice.
• Counterpoints:
  • Much driving is on “autopilot”; repetition without feedback doesn’t improve skill.
  • Standards differ by region; “bad” driving is partly norm‑based.
  • By accident statistics, average drivers might actually be “good enough,” indicating that “driving is hard” more than “most are hopeless.”

Politics, ignorance, and stakes

• Some commenters care more about abstract/factual ignorance (civics, minority sizes, etc.) because it affects voting and policy.
• Democracy relies on ignorance being roughly random; commenters worry that systematic misinformation and uncuriosity break this assumption.

Meta‑critiques of the article and test

• Several see the “are you presuming most people are stupid?” test as:
  • A check on arrogance when explaining human behavior.
  • But also as a potential strawman: many critics don’t assume most people are stupid, only that enough are misinformed or short‑sighted to cause real harm.
• Others feel the article projects the author’s own cognitive style onto everyone else, ignoring people who function only by memorizing scripts and avoiding problem‑solving.

Story, evidence, and CBP response

• The incident rests almost entirely on the traveler’s own account; commenters note there are no independent witnesses or documents.
• Some find aspects suspicious or incomplete (why he was flagged initially, other photos like a wooden pipe, the coincidence of his name).
• Later, CBP publicly stated he was refused entry over admitted prior drug use, not a meme; this convinces some and is seen as face‑saving spin by others.
• Several argue media should have sought and published official comment before framing it as “denied for a meme.”

Border powers, phone searches, and rights

• Broad consensus that border law is exceptional: non‑citizens have no right to admission, and search powers are far broader than inland.
• People report being pressured or threatened into unlocking devices; lawyers note that for non‑citizens refusal usually means denial of entry and possible long bans.
• Disagreement over how much constitutional protection (especially the First Amendment) really applies to non‑citizens at the border, and whether it matters in practice.

Authoritarianism, “fascism,” and terminology

• Some see this and similar cases as part of a slide into fascism, evoking East Germany, North Korea, or lèse‑majesté laws in various countries.
• Others push back: East Germany was communist, not fascist; “authoritarian” is more accurate; overusing “fascist” dilutes the term.
• A meta‑debate emerges over whether quibbling about labels misses the lived reality of arbitrary power.

Free speech, cancel culture, and hypocrisy

• Many connect the story to earlier “cancel culture” panics, arguing that some self‑styled free‑speech defenders only oppose censorship when it targets their side.
• Others from that milieu explicitly condemn the incident, insisting consistent principles are possible but rare.
• A long subthread describes “symbolic violence” and arbitrary enforcement: the point isn’t coherent rules, but the ability to punish unpredictably.

Travel behavior and comparative anecdotes

• Numerous stories of rough treatment at US and Canadian borders (secondary screening, strip searches, phone checks) contrast with smoother entries into places like China or some European states.
• Some now avoid US travel entirely or only enter with wiped or burner devices; others emphasize that abusive cases are statistically rare among tens of millions of entries.
• Debate continues over whether such incidents are systemic authoritarian drift or “normal” but troubling border‑bureaucracy overreach.

Scope of the F1 Wallet Ad and Immediate Reactions

• Ad appeared both as a push notification and a banner at the top of Apple Wallet.
• Many saw this as crossing a line: Wallet and notifications are considered “infrastructure” features, not marketing channels.
• Several users disabled Wallet notifications entirely; some are considering canceling Apple Card or other Apple services to avoid future ads.
• A few commenters note the ad “worked” in the sense that it made them consider buying tickets—while still calling it unacceptable.

Expectations of the “Apple Premium” vs Reality

• Strong sentiment that people pay high prices specifically to avoid being treated like an ad target.
• Others argue Apple only needs to be “less bad” than Windows/Android OEMs, not ad‑free, to justify its premium.
• Comparisons to Windows, Xiaomi, Samsung, and Ubuntu show that intrusive ads are increasingly common across platforms; some say only niche OSes (e.g., BSD) remain ad‑free.

Push Notification Abuse and Desired Controls

• Widespread frustration with apps (Uber, Amazon, food delivery, etc.) using push for marketing instead of critical events.
• Users want system‑level separation of “offers” vs important alerts (payments, deliveries, emergencies), and per‑category blocking.
• Android’s notification channels and iOS features like Live Activities/Time Sensitive notifications are cited, but many say apps misuse or ignore them.
• Some adopt a zero‑tolerance policy: any promotional push → all notifications off or app deleted.

Apple’s Own Rules and Enforcement Double Standards

• Commenters quote App Store guideline 4.5.4, which forbids promotional push notifications without explicit opt‑in and an in‑app opt‑out.
• Multiple people assert that big apps routinely violate this and Apple mostly looks the other way.
• Apple pushing its own F1 promo through Wallet is seen as especially hypocritical given those rules.

Privacy, Trust, and Monetization Pressure

• Debate over whether Apple truly offers “industry‑leading privacy” given closed source, push metadata sharing with governments, and remote control over apps.
• Some suggest Apple is compensating for losing high‑margin revenue (search deal, 30% fees) by ramping up ads and services promotion.
• Overall fear: this is one more step in the “enshittification” of even high‑end devices and will erode long‑term customer trust.

Extended Support “Strings” and Business Model

• Extra Windows 10 security updates require enrolling in Windows Backup (Microsoft account, cloud tie‑in) or redeeming 1,000 Microsoft Rewards points (earned via Bing searches, purchases, etc.).
• Many see this as “not free”: you pay with data, time, and attention, and help Microsoft’s ad network—likened to ad‑click “serfdom,” Black Mirror, or Onion satire.
• Others note the points are trivial to earn (a few minutes per day) and even scriptable, but critics argue that normalizing this exchange is corrosive.

Hardware Limits, “Last Windows,” and Trust

• Numerous users have powerful PCs blocked from Windows 11 by TPM 2.0 or CPU lists, calling the requirements arbitrary and effectively paid upgrades via new hardware.
• Workarounds (registry edits, custom ISOs, “server” install trick, TPM enablement in BIOS) are common but fragile and don’t solve app‑level TPM dependencies.
• Strong resentment over Windows 10 being pitched as “the last version of Windows” and “lifetime of the device” support; debate over whether this was official marketing or one evangelist, but many feel the impression was deliberately fostered and later walked back.

Accounts, Telemetry, and UX Regression

• Frustration that consumer Windows 10/11 installs heavily nudge or practically force Microsoft accounts; bypasses exist but are obscure and periodically broken.
• Wider complaints: telemetry, ads in Start, browser/OneDrive pushing, AI “everywhere,” and aggressive control via updates.
• Windows 11 is seen by many as bloated with few real benefits over 10; some praise WSL2 and certain GPU improvements, but UI changes (no vertical taskbar, buggy Explorer, Electron/webview feel) are viewed as a major downgrade, partially mitigated by third‑party tools.

Alternatives, Old Versions, and Adoption

• Several commenters have already moved old machines to Linux (Mint, Fedora, Zorin, SteamOS) or macOS; experiences range from “liberating and stable” to “too much tinkering, went back to Apple/Windows.”
• Gaming is the main reason many still boot Windows; SteamOS/Proton progress makes dropping Windows increasingly tempting.
• Some still prefer Windows 7/10 and question the real‑world risk at home if behind a router, though others flag browser/outdated‑OS exploit concerns.
• The fact that ~53% of Windows PCs still run 10 so close to end‑of‑support is read by many as a failure of Microsoft’s Windows 11 push, prompting this last‑minute extension.

Scope of the ruling: training vs piracy

• Many commenters read the decision as:
  • Training LLMs on copyrighted books can be fair use if the use is transformative and the model isn’t a market substitute for the works.
  • Acquiring books via piracy is not fair use; the judge calls that “inherently, irredeemably infringing,” and leaves damages for a separate trial.
• Several see this as analogous to Google Books: destructive scanning of purchased books and storing full text is allowed if downstream access is constrained.

Transformative use and human analogies

• Supporters argue training is like a person reading books and forming internal representations, then creating new works; copyright protects expression, not ideas or knowledge.
• Critics respond that an LLM is “a tool, not a person,” and that calling its learning “reading” is anthropomorphism; a model is a machine built from copyrighted works.
• Scale is a key counterargument: a human can’t memorize or reproduce millions of works; LLMs can approximate that at industrial scale.

Memorization, outputs, and open weights

• The order assumes, for the sake of argument, substantial memorization, but finds training fair use when outputs are filtered to prevent verbatim reproduction, analogizing to Google’s snippet limits.
• This worries some:
  • Hosted, filtered models may be safe, but open-weight models might be vulnerable if users can extract memorized text.
  • Others point to a separate case holding that model weights themselves are not infringing derivative works; infringement turns on specific outputs and uses.

Contracts, licenses, and “no AI training” clauses

• Commenters debate whether publishers can block training via contract terms or EULAs.
• Physical books generally lack enforceable licenses beyond copyright; ebooks and databases are different.
• Fair use can override the need for a license, but not a signed contract—breach-of-contract remedies would be separate from copyright.

Economic and ethical concerns

• Skeptics see “plagiarism automation at scale”: a small number of firms monetize the distilled product of billions of human hours without compensation, potentially chilling future creation and driving DRM and information silos.
• Others emphasize copyright’s constitutional purpose (promote progress, not guarantee pay for every use) and warn against using copyright to halt a broadly useful technology.
• Some propose intermediary solutions, like an “LLM levy” analogous to cassette-copying royalties, with pooled payments to rights holders.

Copilot Branding & Licensing Confusion

• Many commenters say “Copilot” is unintelligible as a brand: it can mean GitHub Copilot (inline coding), Copilot in VS Code/JetBrains, Windows Copilot, Bing/Edge Copilot, M365 Copilot, domain-specific Copilots (Sales, Service, Fabric, Dynamics, Security), and now even the Office suite (“Microsoft 365 Copilot app”).
• Users can’t easily see what license or model tier they have, especially in M365; “Pro”, “enterprise data protection”, “researcher agent”, etc. are poorly surfaced.
• Several see this as deliberate obfuscation for enterprise sales: marketing can claim “you get all these Copilots” even though quality varies radically.

Product Quality: Copilot vs ChatGPT & Others

• Many report Copilot (especially M365/Web) as the “dumbest” major LLM: short, cautious, often useless answers; bizarre replies like claiming to have run missing Python code instead of just emitting an ffmpeg command.
• Others can’t reproduce those failures and get good ffmpeg commands or solid Outlook/Planner help, underscoring non-determinism and context dependence.
• Repeated theme: the same prompts work well in ChatGPT, Claude, Gemini, Perplexity, or even small local models, but fail or underperform in Copilot.
• Some think Copilot is “lobotomized” via system prompts, shorter responses, or token conservation; others say it’s mostly a Bing‑grounded RAG layer whose ceiling is Bing’s results.

Integration with Microsoft 365 and Enterprise Use

• Expectations were that Copilot 365 would shine on work graph (Teams, mail, files). Many say that’s unreliable: it can’t always see recent emails, markdown docs, or query panes (e.g., in SSMS).
• When it does have graph access, some report good performance for summarizing messy org documentation or generating Planner plans and executive-style summaries.
• Several note Copilot often acts like a generic chat iframe, not a true agent that can actually send emails, create meetings, or consistently act on documents.

UX, Friction, and Adoption

• Heavy criticism of Microsoft UX: confusing portals, modals, auth flows, broken links, inconsistent UIs, and opaque rate limits/quotas (some hit monthly Copilot limits in days).
• Users contrast: “open ChatGPT/Claude/Perplexity and type”, vs. multi-step M365 login plus a “sterile, over-cautious” Copilot tone.
• Some orgs report Copilot hype fading; teams pivot to Claude/ChatGPT or tools like Cursor/Windsurf instead.

Strategy, Control, and “Rivalry”

• Debate on whether Microsoft “wasted” a unique head start (exclusive OpenAI access + Bing) by shipping a messy brand and weak UX, versus still winning via distribution and a large profit share from OpenAI.
• Concern that Microsoft’s main advantage is bundling/sales, not product quality; parallel drawn to Teams vs Slack.
• Some see the tension as about control and positioning: OpenAI wants to be a first-class enterprise platform, not just an Azure feature.

Technical Debates: Models, Prompts, and Non‑Determinism

• Confusion over what models Copilot actually uses (“GPT‑4‑based”, distills, older 4o revisions, smaller internal models) with no transparency and no model picker in many SKUs.
• Extended argument over whether bad outputs are mostly bad prompts vs. bad models; examples show even tiny modern models handle vague prompts that Copilot sometimes fumbles.
• Multiple people stress non‑determinism and hidden context: “worked for me” doesn’t invalidate failures, but reproducible anecdotes also expose real quality gaps.

Security, Governance, and Enterprise Buying

• Some call Copilot a “security nightmare” given broad tenant access, citing at least one published vuln (details not discussed).
• Others say its biggest real advantage is clear corporate legal terms around data use; “safe to buy” often beats “best to use” in large enterprises.

OpenAI’s Position & Alternatives

• Several note that OpenAI’s direct ChatGPT offering feels faster, less constrained, and more capable, which is driving enterprises to test it alongside or instead of Copilot.
• Mention of OpenAI’s rumored “AI super app” (canvas + docs + more) as a potential direct challenge to Office/Workspace, though details are unclear from the thread.
• Some commenters think OpenAI’s enterprise “success” is still mostly marketing and anecdotes; the article’s numbers are viewed as incomplete.

Quality and Validity of XBOW’s Findings

• XBOW claims all reported vulnerabilities were real and accompanied by executable proof-of-vulnerability; some commenters ask directly if that implies a 0% false-positive rate.
• The article mentions automated “validators” (LLM- or script-based) to confirm each finding (e.g., headless browser to verify XSS), but people note it doesn’t quantify how many candidate bugs were discarded before the ~1,060 reports.
• Success rates differ sharply by target (e.g., very high validity for some programs, very low for others), which commenters attribute partly to varying program policies (third-party issues, excluded vuln classes, “never mark invalid,” etc.).

AI Slop, Noise, and Triage Burden

• Multiple maintainers describe AI-generated “slop” reports as demoralizing (e.g., placeholder API keys flagged as leaks) and expect AI to massively industrialize low-quality submissions.
• Others note bug bounty programs already receive an overwhelming volume of terrible human submissions; platforms like HackerOne exist partly to shield companies from this.
• Concern: XBOW’s ~1,060 submissions consume triage capacity; stats from its own breakdown show many duplicates, “informative,” or “not applicable” reports, which still cost reviewer time.

Automation vs. Human Involvement

• Some see XBOW as a strong, pragmatic AI use case because working exploits are hard evidence and reduce hallucination risk.
• Others stress that humans still design the system, prompts, tools, and validators, and review reports before submission; calling it “fully autonomous” is seen as marketing overreach.
• There’s skepticism that such a system could run unattended for months and continue to produce high-value bugs without ongoing human tuning.

Bug Bounty Ecosystem and Ethics

• Several participants describe bug bounties as economically skewed: many low-paying programs, slow payouts, and companies allegedly using them for near-free security work.
• Some argue many companies shouldn’t run bounties at all; they’d be better off hiring security firms.
• Ethical concerns arise over using automated tools where program rules forbid automation; others counter that if a human can reproduce the bug, the discovery method shouldn’t matter.

Broader Impact on Security and Talent

• Many view AI-assisted pentesting as ideal for clearing “low-hanging fruit,” especially in legacy code, and freeing experts for more creative work.
• Others worry about triage scalability, the flood of mediocre AI reports hiding real issues, and long-term effects on training and opportunities for junior security researchers.

Joy of “toy” software vs. production work

• Many relate to the core point: tinkering on “toy” code is fun; depending on it is stressful. Once you try to actually use a toy (e.g., invoicing app, finance tooling), bugs, edge cases, and deadlines quickly erode the joy.
• Several embrace having “two bikes”: a playful, breakable toy and a reliable daily driver. Others say they only enjoy software that’s truly useful, even if small and personal.

DIY vs delegating (self‑hosting, SaaS, and cars/bikes)

• Some stopped self‑hosting critical infrastructure (email) because rare but badly timed failures and maintenance overhead weren’t worth it. Others report decade‑long smooth self‑hosting with minimal babysitting, seeing it as essentially “set and forget.”
• Analogies to bikes vs cars illustrate tradeoffs: tinkerer tools (bikes, self‑hosted stacks) are repairable and customizable but can leave you stranded; SaaS and commercial services are less hackable but offer reliability and support ecosystems.

LLMs: joy, learning, and risks

• Strong split: some say LLMs supercharge toy projects, letting them focus on architecture, UI polish, or weak areas (CSS, infra) while delegating boilerplate. Others argue the joy is in understanding, not in “hand‑setting code,” and LLMs can short‑circuit that.
• Many use LLMs as “search on steroids” or rubber ducks: for overviews, reverse‑searching concepts, navigating bad docs, or scaffolding prototypes, then heavily editing.
• Several warn that over‑reliance degrades deep learning and problem‑solving (“easy chair for the mind”), and that LLM‑generated answers can be subtly wrong or biased. Some recommend using them as teachers, not interns, or even typing out suggested code by hand to internalize it.

Scope, difficulty, and value of toy projects

• Multiple commenters find the author’s time estimates (e.g., GBA game in 2 weeks, physics engine in 1 week) wildly optimistic unless you build extremely stripped‑down versions and already know the domain well.
• Debate over “reinventing the wheel”: one camp calls it pointless; many others defend re‑implementing compilers, shells, git, DBs, etc. as powerful learning exercises that later pay off in real work.
• Stories highlight toy projects directly enabling career advances (e.g., understanding pattern‑matching algorithms for a production language; graphics and game engines leading to dream jobs).

Keeping toys simple: stacks, configuration, deployment

• People praise minimal stacks (single binary, few deps, VPS + systemd, rsync) and warn against over‑generic “configuration engines” that add complexity for hypothetical users.
• A recurring theme: to preserve joy, constrain scope (e.g., one week per project), accept 80–90% solutions, and resist turning every toy into production software.

Scope of the “new abstraction” claim

• Some readers don’t buy that prompting LLMs is a new level of abstraction; it feels more like a different activity entirely, not an abstraction over previous programming work.
• Others argue it’s a “new nature” of abstraction:
  • Up: expressing intent in natural language, specs, and examples instead of code.
  • Sideways: dealing with probabilistic, non-repeatable behavior rather than deterministic compilation.

Reliability vs solving “harder” problems

• Supporters say unreliable LLMs can still be worth it if they address problems that were previously too hard or expensive (e.g., “common sense” judgment, messy edge cases, autonomous behavior in hopeless scenarios).
• Skeptics counter that “90% reasonable, 10% insane” behavior is unacceptable for most production systems; better to fail loudly and fix the root cause.
• Several report LLMs have not solved problems they couldn’t solve themselves, but they dramatically speed up work—mainly turning solvable problems into faster ones, not fundamentally harder ones.

Non-determinism, determinism, and “practical” predictability

• Strong debate on whether non-determinism is really “unprecedented”: fuzzing, mutation testing, and earlier ML already introduced it, though mostly outside the core compiler/toolchain.
• Technically, LLMs can be deterministic (temperature 0, fixed seeds, pinned models/engines), but:
  • Hosted APIs, batching, hardware differences, and implementation quirks often break reproducibility.
  • Even with fixed seeds, tiny prompt changes can lead to drastically different outputs.
• Several distinguish technical determinism from practical determinism: developers can’t reason about prompt changes with the precision they have for code.

Experiences building with LLMs

• Practitioners building LLM-based apps report:
  • Minor prompt tweaks causing major behavioral shifts and downstream effects.
  • Context-window failures that silently degrade quality unless you actively manage tokens.
  • Mainstream business users often give up when behavior feels too fuzzy or inconsistent.
• As coding assistants, LLMs are widely seen as productivity boosters—but they introduce subtle bugs, making tests and strong typing even more important.

Natural language vs formal code

• Some want “English to bytecode” and treat prompts as source, LLM output as compiled target.
• Others invoke classic arguments (e.g., Dijkstra) that natural language is inherently imprecise; precision requires formalism and well-defined machine models.
• A nuanced camp pushes for mixed systems: blend natural language for intent and high-level behavior with traditional code and formal models (e.g., TLA+ + LLM, or languages explicitly designed to interleave NL and symbolic notation).

Skepticism about hype and authorship

• Several commenters think the “unprecedented” framing and talk of fundamental change are overblown or consultant-driven hype.
• Others argue that even observers who “only dabble” can provide useful, contextual perspectives—provided their claims about practice are treated with caution.

Scope of the Problem

• Multiple comments stress that US life expectancy is lower not only overall but at every wealth level compared with Europe; even the richest Americans fare worse than rich Europeans and sometimes only as well as poor Europeans.
• The US also has fewer healthy years, with many living longer but in poor health.

Inequality, Stress, and Healthcare Access

• Inequality, stress, loneliness, and “deaths of despair” (addiction, mental health, etc.) are repeatedly cited as core drivers.
• “Difficulty accessing healthcare” is argued to be as much about cost opacity, insurance networks, and fear of financial ruin as about distance or wait times.
• Several anecdotes describe people avoiding urgent care — and even dying — because of expected bills, despite having insurance.
• There is debate over US poverty: some argue official statistics understate the impact of large welfare spending; others say spending levels are irrelevant if outcomes (e.g., places like Gary, Indiana) remain poor.

Behavior, Environment, and “Social Causes”

• Many point to poor diet, ultra-processed food (major share of calorie intake), car dependence, and low physical activity as central.
• Obesity is highlighted as a key driver of chronic disease and reduced life expectancy, especially among younger adults.
• Traffic fatalities, overdoses (especially synthetic opioids), and homicides are seen as major contributors, particularly for ages 15–49.
• Alcohol is debated: US drinking culture is criticized, but others note per-capita consumption is lower than in much of Europe, where life expectancy is higher.

Regional and Demographic Variation

• Thread repeatedly emphasizes huge state- and county-level gaps (≈10-year differences) and argues national averages hide crucial spatial inequality.
• Some argue multicultural demographics require disaggregation by race/ethnicity and region; others counter that using demographics to “explain away” poor outcomes is morally troubling and risks racist framing.
• Climate and walkability are discussed: some blame southern heat for inactivity; others counter that northern states with harsh winters still manage higher fitness, pointing instead to culture, urban design, and diet.

Obesity, Doctors, and Culture

• Several report doctors downplaying or over-attributing problems to weight, suggesting inconsistent clinical handling.
• There’s disagreement whether doctors avoid discussing obesity due to “body shaming” fears or, conversely, focus on it too bluntly.
• Some propose sugar/fast-food taxes, better urban design, and stronger social safety nets; others emphasize individual lifestyle changes (cooking, walking, everyday activity).

Interpreting the data and “safe” limits

• Several commenters note that even very “contaminated” foods appear far below current federal intake limits, which paradoxically makes them feel reassured.
• Others point to the report section arguing that historical experience with PFOA/PFAS shows regulators often start with limits hundreds–thousands of times too high.
• Many chemicals in the table lack any official intake guideline, raising the question of what “safe” even means for them.

Ubiquity and sources of plastic contamination

• Raw farm milk in glass and grass‑fed ribeye rank surprisingly high, used as examples that even minimally processed or “premium” foods are embedded in plastic-heavy supply chains.
• Discussion highlights livestock feed (baled/wrapped hay, ground-up packaged waste), milking and processing equipment, and conveyor belts as major sources.
• Household sources get attention: plastic pepper grinders, plastic cutting boards, Teflon vs packaging, polyester clothing, dryer vents, and water infrastructure.
• Some note plastics likely enter food long before packaging; processing machinery visibly sheds plastic dust.

Health risk, evidence, and regulation

• One camp argues plastics get outsized attention compared to clearly harmful lifestyle factors like sugar and alcohol, and sees “microplastic-free” marketing as potential hype.
• Others counter with emerging evidence of endocrine disruption, inflammation, and microplastics crossing the blood–brain barrier, and stress that “absence of evidence is not evidence of absence.”
• Historical parallels (asbestos, lead, PFAS) are used to argue for a precautionary approach and skepticism of current regulatory limits.
• Some remain broadly fatalistic: given existing exposures (lead, asbestos, past jobs), reducing microplastics now feels marginal.

Consumer responses and practical advice

• Strong emphasis on prioritizing PFAS in drinking water; distillation and reverse osmosis are frequently recommended, along with PFAS-focused filters.
• Micro-optimizations discussed: metal/ceramic grinders, mortar and pestle, bamboo toothbrushes, wood vs plastic cutting boards, natural fibers, minimizing plastic contact with hot or fatty foods.
• Others warn against trying to “care about everything” and argue for focusing on the largest exposure sources (especially water).

Site design, methodology, and limitations

• The UI receives a lot of praise; commenters identify Next.js, Tailwind, TanStack Table, and specific fonts.
• Some criticize missing context (e.g., whether drinks were tested in plastic-lined cups vs mugs) and inconsistent units.
• Concerns about sample handling in plastic bags are raised, while others note the lab’s controls (isotopically labeled standards, solvent washes) likely keep contamination manageable.
• Several call PlasticList a valuable independent effort as trust in and funding for federal agencies declines.

Expressivity and Theoretical Bottlenecks

• OP claims: with ~15k tokens and 1k-dimensional embeddings, the next-token distribution is limited to rank 1k, constraining which probability distributions are representable.
• Replies note high-dimensional geometry: exponentially many almost-orthogonal vectors can exist, so practical expressivity is much larger than intuition suggests, though not enough to represent arbitrary distributions.
• Some argue nonlinearity and deep networks break the simple linear “1k degrees of freedom” story; others point to work on “unargmaxable” outputs in bottlenecked networks as real but rare edge cases.

Tokenization, Characters, and the “Strawberry r’s” Meme

• Several comments explain that subword tokenization hides character structure: “strawberry” might be a few opaque tokens, so models must effectively memorize letter composition per token to count letters.
• Evidence from in-review work: counting accuracy declines as the target character is buried inside multi-character tokens.
• Others are skeptical, arguing:
  • We lack clear demonstrations that character-level models can reliably “count Rs”.
  • RLHF and training on many counting prompts suggest the limitation is not purely tokenization.
• There’s recognition that models don’t “see” characters; they see embeddings, and any character-level reasoning is an extra learned indirection.

Math, Logic, and Number Tokenization

• Several posts claim logical/mathematical failures are strongly tied to tokenization, especially how numbers are split.
• Cited work shows large gains when numbers are tokenized right-to-left in fixed 3-digit groups (e.g., 1 234 567) and when all small digit-groups are in-vocab.
• Other research: treating numbers as special tokens with attached numeric values so arithmetic is done on real numbers rather than digit strings.
• Some argue LLMs are the wrong tool for exact arithmetic; better is: LLM selects the right formula and delegates computation to a calculator engine.

Bytes, UTF-8, and Raw Representations

• “Bytes is tokenization”: using raw bytes (often via UTF-8) is seen by some as the ultimate generic scheme, avoiding out-of-vocabulary issues with a 256-token alphabet.
• Counterpoint: UTF-8 itself is a biased human-designed tokenizer over Unicode; models are not guaranteed to output valid UTF-8, and rare codepoints can be badly trained.
• New encoding schemes are being explored to better match modeling needs and reduce “glitch tokens”.

Bitter Lesson, Compute vs Clever Tricks

• Debate centers on whether tokenization is the next domain where the Bitter Lesson (general methods + compute beat handcrafted structure) will apply.
• Some say it already did: simple statistically learned subword tokenizers outperformed linguistically sophisticated morphology-based approaches.
• Others highlight counterexamples where architectural tweaks to tokenization (e.g., special indentation tokens in Python, better numeric chunking) give large, practical improvements—evidence that cleverness still matters.
• There’s concern that over-relying on “just scale compute” can obscure simpler, more principled solutions and slow genuine understanding.

Costs, Scaling, and Energy

• A claim that training frontier models costs “around median country GDP” is challenged with data: estimated compute costs for GPT‑4 or Gemini Ultra are in the tens or low hundreds of millions of dollars, far below ~$40–50B median GDP.
• People discuss GDP measures (PPP vs nominal) and note training cost estimates are rough and incomplete (hardware, engineering, data, etc.).
• Another angle compares theoretical human brain energy (a few fast-food meals/day) versus enormous current AI energy use, suggesting large headroom for efficiency improvements.

Determinism, Capability, and AGI Limits

• Clarification: the model function is deterministic; nondeterminism comes from sampling, numerical instability, and changing deployments.
• Some argue DAG-like, immutable-at-runtime transformers can never reach AGI; others counter that with sufficiently long context and high throughput, such models could be effectively general, and that “immutability” is a modeling convenience, not a hard theoretical limit.
• Theory papers showing transformers can simulate universal algorithms are cited; critics note these are existence proofs, not guarantees that gradient-based training will find such solutions.

Future Directions: Learned or Mixed Tokenizations

• Multiple commenters imagine mixtures of tokenizations:
  • A learned module that dynamically chooses token boundaries (e.g., via a small transformer predicting token endpoints) so models can “skim” unimportant text and compress context.
  • Mixture-of-experts where each expert has its own domain-specific tokenization.
• Character-level and byte-level models (e.g., Byte-Latent Transformers) are seen as moves toward end-to-end learned representations, but questions remain about efficiency and performance on math and reasoning.
• Overall sentiment: tokenization is likely suboptimal today; compute scaling will help, but domain-aware or learned tokenization will probably deliver important gains before “just bytes + huge models” fully wins.

Discovery and “Computing Archeology”

• Commenters frame this as “computing archeology”: people deliberately trawling ROMs, binaries, and old systems for hidden content using hex editors, debuggers, and pattern/string searches.
• Some emphasize that the article itself fully explains the “how”; others marvel that anyone spends time on this at all, pointing to communities devoted to uncovering unused game content and hidden assets.

OS Size, Bloat, and AI Models

• Discussion branches into why modern OSes are so large compared to classic Mac OS.
• One view: higher-resolution assets, bundled translations, dual-architecture support (x86 and ARM), and now on-device AI models all inflate size.
• Another counters that on some systems languages are optional downloads and questions how much core OS imagery really weighs.
• A concrete breakdown on macOS shows gigabytes consumed by AI models, fonts (notably emoji), printer drivers, loops, and linguistic data.
• There’s disagreement over whether this is a meaningful problem on 256GB SSDs and whether AI assets are preinstalled or downloaded only with consent.

Easter Eggs: Fun vs Professional Risk

• Many express affection and nostalgia for Easter eggs, seeing them as proof that real humans built these systems.
• Others argue strongly against them in commercial products: they add undocumented code paths and potential bugs, complicate security audits, and can threaten schedules and contracts (e.g., government requirements, “Trustworthy Computing” era).
• Some note that past corporate bans (Apple, Microsoft) were driven by security, reliability, and optics, not jealousy.

Jobs, Apple Eras, and Cultural Shift

• Debate over Steve Jobs banning Easter eggs: some see it as killing whimsy; others cite earlier efforts to credit teams and argue the ban was pragmatic (risk, recruiting/poaching, seriousness).
• Several reminisce fondly about Apple’s “interregnum” years (mid‑80s–mid‑90s): quirky hardware, HyperCard, OpenDoc, strong UI/UX culture, and a “cozy, whimsical” classic Mac feeling later lost under macOS and today’s iPhone‑centric, services-driven Apple.

Humanization, Credit, and Modern Process

• Easter eggs like signed ROM images are seen as a way for “small people” to leave their mark, contrasting with executives taking public credit.
• Others respond that modern products involve thousands of contributors; any selective credits are inherently exclusionary and politically fraught.
• Compliance, audits, secure SDLC, Agile, and constant deadline pressure are cited as making secret features nearly impossible: undocumented artifacts trigger IT controls, SOC findings, and HR issues.

Learning Reverse Engineering

• Reverse engineering is described as hard but approachable. Old console and PC games are recommended as starting points: simple hardware, immediate visual feedback, and rich tooling and documentation.
• Commenters encourage readers that many such old systems still hide “low-hanging fruit” like this ROM Easter egg, especially with modern tools like Ghidra.

Nostalgia for Old Easter Eggs and Small Teams

• People share memories of classic Mac and Windows-era Easter eggs (secret about boxes, mini-games, hidden images, credits screens) and lament their disappearance.
• There’s a recurring wish to “bring them back,” tied to broader nostalgia for smaller, more personal teams and less sterile, more playful software.