The claim submitted for fact-checking was a multi-step causal argument:
Large tech monopolies have laid off many people over the last few years.
The layoffs were caused mainly by Covid-era over-hiring ("bloat"), and partly by over-hiring from before Covid.
These layoffs would have happened anyway, independently of AI.
Those same monopolies almost all sell their own AI services or are otherwise affiliated with the AI business, so they have a stake in AI's perceived success.
After a while, they began framing the layoffs as evidence that AI is so useful it can replace entire organizational layers.
They did this because the framing is good PR and marketing for the product they currently invest in most heavily and on which they struggle to show a return on investment (ROI).
Many other players, large and small, believed the framing, "drank the kool-aid," and joined the bandwagon β laying off staff in the hope of replacing expensive human labor with AI services rented from the monopolies.
As soon as this happened, and as the services became actually somewhat useful (around 2025), the monopolies began tweaking the pricing structure of those services.
They tweaked it enough that the math and viability of replacing humans with AI no longer adds up.
Overall directional verdict: the first half (1, 3, 4-as-fact, 5) holds; the middle (2, 6, 7, 8) is partly true but overstates single-cause certainty; the pricing claim (8-9) was initially judged largely wrong, then substantially upgraded after a justified pushback. Details below.
A. Claim-by-claim fact-check
(1) Large tech firms laid off many people β VERIFIED. Layoff trackers (Layoffs.fyi, Crunchbase, Challenger Gray & Christmas) record roughly 153,000 tech job cuts in 2022 (about ten times the 2021 level), about 226,000 in 2023 (a near-40% rise), roughly 150,000 in 2024, and at least 127,000 at U.S.-based tech firms in 2025. Total U.S. layoffs across all sectors reached about 1.2 million in 2025.
(2) Mainly Covid-era over-hiring β PARTLY VERIFIED, incomplete. Over-hiring is the most-documented driver. Meta nearly doubled headcount between March 2020 and September 2022; a study of laid-off LinkedIn profiles found average tenure of about 2.5 years, implying hiring during the lockdowns. But "mainly Covid bloat" omits co-drivers: the end of zero-interest-rate cheap capital, recession fears, falling venture funding, and a strategic shift from growth toward profit. Not all firms accept the framing β SAP's CEO denied over-hiring. The cause is multi-factor, not solely pandemic-driven.
(3) Would have happened independently of AI β SUPPORTED for the 2022β2024 wave. Yale's Budget Lab examined U.S. labor data from November 2022 to July 2025 and found the labor market little disrupted by AI automation since ChatGPT's release. A strict counterfactual is unprovable, but the evidence aligns with the claim for the early wave.
(4) Same firms sell AI / have a stake β VERIFIED as fact; the conflict-of-interest motive is inference. Microsoft, Amazon, Google, and Meta all sell AI and invested heavily (e.g., Microsoft's roughly $80B AI data-center investment announced alongside 2025 cuts). Analysts (Brookings) note a self-serving framing incentive: investor relations benefit from presenting cuts as forward-looking technological adaptation rather than reactive cost-cutting. The stake is real; the inferred intent is plausible but not directly proven.
(5) Reframed layoffs as AI replacing organizational layers β VERIFIED. Documented and named "AI washing." Challenger, Gray & Christmas reports AI was cited in nearly 55,000 of 2025's layoffs (about 4.5%). An MIT professor characterized AI as "a perfect excuse" that shifts blame to the technology. A CNBC-quoted labor economist expressed skepticism that the cuts reflect true efficiency gains, naming Duolingo and Klarna as likely cover cases given their Covid over-hiring.
(6) PR for AI products they struggle to monetize β PARTLY VERIFIED, with a caveat. The ROI shortfall is well-evidenced: MIT's NANDA report "The GenAI Divide" found that despite $30β40B in enterprise investment, 95% of organizations saw zero measurable return; an IBM survey of 2,000 CEOs found only one in four AI projects delivered the promised return. However, most documented AI-washing is about making the layoff-er look innovative to investors β not specifically a coordinated scheme by platform vendors to market their AI to third parties. That specific motive cannot be verified.
(7) Others believed it and cut staff to rent AI β PARTLY VERIFIED. The bandwagon is real; the canonical case is Klarna, which replaced the work of about 700 customer-service agents with an OpenAI-powered chatbot, then reversed. Orgvue/Forrester research found 55% of companies that rushed to replace workers with AI now regret it. The "kool-aid" causal chain (cut because the monopolies said so) is stronger than the evidence; what is documented is correlated adoption and widespread regret.
(8) Services became "somewhat useful" around 2025 β vague; partly true, partly contradicted. Capability improved, but enterprise value remained the exception, not the rule (MIT's 95% no-return). "Useful enough to replace humans" was contradicted in customer-facing roles, where Klarna's CEO admitted over-prioritizing cost produced lower quality and began rehiring.
(9) Pricing tweaked upward so the math no longer adds up β REVISED VERDICT: PARTIALLY VERIFIED. Initially judged "largely wrong as stated," because headline per-token prices fell sharply rather than rose. After a justified pushback (see Β§B), the verdict was upgraded: the structural mechanism is confirmed; the magnitude conclusion and the causal intent remain partly unproven. The full reasoning is in the next section.
B. The pricing question β the most-developed thread
B.1 The pushback and the concession
The pushback: the original claim was about effective cost (what a company actually pays month to month), never about per-token list price, and effective cost is the only relevant metric when comparing a human worker to an AI service.
The concession: this is correct. For a replace-a-human calculation, effective monthly cost is the right denominator, and leading the first rebuttal with the collapse in per-token list price answered a question the claim did not ask. That was a framing error.
B.2 The corrections that survived
Token price is not meaningless; it is the multiplicand. Effective cost = (price per token) Γ (tokens consumed). The reason effective cost stayed high despite a roughly 60β94% collapse in frontier unit prices since 2023 is that reasoning and agentic workloads multiplied the token volume per task. Remove the unit-price term and you cannot explain why effective cost did not fall as buyers expected.
The causal attribution is entangled with real compute costs. The original wording β "the pricing structure was tweaked by the monopolies" β contains a causal claim about agency. The dominant drivers of rising effective cost are integration/engineering labor and token consumption, not provider list-price hikes; the vendors actually cut the per-unit price they control. There is one genuine provider-driven exception (Anthropic's shift of enterprise billing to usage-based plus a new tokenizer), but it is secondary.
The flagship "math didn't add up" case failed on quality, not price. Klarna reversed because the all-AI approach produced lower service quality, not because prices rose. (Note: the pushback correctly observed that Klarna predates the 2025β2026 pricing restructurings, so Klarna is not evidence against the later pricing claim either way.)
B.3 The concrete pricing-structure changes, by provider (2025β2026)
July 2025: rate limits added to Claude Code with no advance notice, concentrated on heavy users of the $200 Max plan.
28 August 2025: two new weekly rate limits on Pro ($20), Max 5x ($100), Max 20x ($200), layered on the existing 5-hour rolling window β one overall cap, one Opus-specific. Estimated to affect under 5% of subscribers. Max users can buy overage at standard API rates.
Enterprise billing shifted to usage-based, plus a new tokenizer that can raise per-task cost (per Josh Bersin).
6 May 2026 (loosening, opposite direction): doubled Claude Code's 5-hour limits for Pro/Max/Team/Enterprise, removed the peak-hours reduction, and raised Opus API rate limits, on the back of a SpaceX compute deal. Weekly caps unchanged.
API unit prices fell for the flagship: Opus 4.6 at $5/$25 per million tokens (input down from $15), Sonnet 4.6 at $3/$15.
29 August 2025: ChatGPT Team renamed ChatGPT Business.
13 February 2026: GPT-4o, GPT-4.1/4.1-mini, o4-mini, and GPT-5 (Instant and Thinking) retired from ChatGPT (API access unchanged) β forced migration.
Rate-limit fallback: paid users hitting limits on the Thinking model fall back to a "mini" model (a quality downgrade rather than a hard stop).
Plus caps (2026): GPT-5.5 Instant at 160 messages per 3 hours; GPT-5.5 Thinking at 3,000 messages per week. Pro tiers are not literally unlimited; guardrails explicitly target running ChatGPT as a backend for paid services.
9 April 2026: new $100 Pro tier introduced; Codex usage changed for Plus/Pro; analysis of the release notes concludes OpenAI is moving usage beyond a "normal" pattern toward higher tiers, credits, or a lighter fallback model (a hybrid subscription-plus-credits model).
2 April 2026 (opposite direction): ChatGPT Business price dropped $5 per seat.
Google I/O, May 2026: switched from a fixed daily prompt count (previously up to 100 Gemini Pro prompts/day on AI Pro, regardless of complexity) to compute-based metering keyed to prompt complexity, features used, and chat length.
Added a 5-hour rolling window plus weekly quotas.
Tier multipliers vs. a "standard" baseline: AI Plus ($7.99) 2x, AI Pro ($19.99) 4x, AI Ultra 5x and 20x at $100 (new tier) and $200 (cut from $250).
Fallback to a smaller model at the cap; AI Pro/Ultra can buy pay-as-you-go credits for overage.
Included media-generation credits removed (the main user complaint).
B.4 The cross-provider pattern
Within roughly twelve months, all three made the same four moves: (1) replace fixed/predictable allowances with compute- or token-metered limits; (2) impose weekly ceilings on top of shorter windows; (3) auto-downgrade to a smaller model at the cap; (4) sell pay-as-you-go overage past the cap. This convergence raises and de-predictabilizes the effective cost of sustained, high-volume, agentic use β the exact profile of using AI to replace a full-time employee. On this point, the original claim holds.
Two qualifications that constrain the claim without rescuing the providers:
The squeeze is in the limit structure, not the sticker price. In the same window, headline subscription prices and flagship per-token prices mostly fell or held (Gemini Ultra $250β$200; ChatGPT Business β$5/seat; Claude Opus API input $15β$5; frontier output price down ~94.5% since March 2023). "Tweaked upward" is accurate only for effective throughput cost.
The direction is not monotonic. Anthropic loosened limits in May 2026 when capacity arrived. Metering tracks compute, and agentic workloads genuinely consume disproportionate compute, so cost recovery and rent-seeking are entangled in the public record, not separable.
B.5 Correction to the "ideal customer" premise
The pushback also asserted that heavy users are the providers' ideal, ROI-delivering customers. This inverts the economics. On flat-rate subscriptions, heavy users are loss-makers β which is precisely why metering, weekly caps, and PAYG overage were introduced (coverage cited the goal of stopping power users from running tools continuously and degrading capacity for everyone). A heavy user becomes ROI-positive for the provider only once moved onto usage-based or API billing β which is what the restructuring accomplishes. So the replacement economics do live in the heavy-user segment (correct), but heavy flat-rate subscribers are the segment providers are converting from flat to metered, not their ideal customer (the inversion).
B.6 The key reframing: a one-time regime shift, not a trend
The restructurings are mostly a single transition from flat-rate pricing that subsidized heavy users to metered pricing that charges closer to the marginal cost of compute. A subsidy can only be ended once. After the transition, the price a heavy user pays tracks marginal compute cost rather than floating free of it. Therefore "extrapolate the price increases forward" is a category error: what continues forward is the marginal-cost curve the restructuring exposed, not the restructuring itself.
C. Forward prognosis β economic viability of human replacement
This is reasoned extrapolation, not verifiable fact.
The governing race. After the regime shift, effective cost per useful task is set by two opposing trends: unit inference cost (falling at roughly an order of magnitude per year; Epoch AI measures the cost to inference at fixed performance halving about every two months) versus tokens-per-task (rising sharply, because doing a human's job means autonomous, multi-step, long-context, reasoning-heavy work). Effective cost is the product of the two.
Rent-seeking vs. cost-recovery. Durable rent extraction is bounded by the open-weight floor and low-cost entrants (e.g., DeepSeek-style pricing). The more durable component of high effective cost is the compute-scarcity premium, which is physical and temporary, not the oligopoly margin, which is contestable. So the long-run trajectory is set more by compute supply (physics) than by pricing power.
Judgment. The unit-cost decline is the more structural force (hardware, algorithms, competition, open-source floor). For a fixed task, replacement gets cheaper over time even under metered pricing β metered times cheap is still cheap. But tasks do not stay fixed: buyers push AI at harder, longer, more autonomous work, re-inflating consumption at the frontier. The result is a moving frontier, not uniform collapse or triumph.
The reliability threshold may dominate price. Full replacement requires capability high enough to remove the human from the loop. Below that threshold, you pay for AI plus human supervision (more than either alone), and no fall in token price rescues the economics; above it, viability jumps discontinuously, largely independent of token price. Klarna failed on this threshold, not on price.
The dominant swing factor (unresolved). Compute supply. The current scarcity premium on frontier inference is the largest reason effective cost stays high for heavy use. Whether the data-center and power build-out relieves the bottleneck cannot be verified; it sets the height of the cost floor.
Net forecast: bifurcation.
Routine, high-volume, error-tolerant work: effective cost keeps falling; replacement becomes steadily more viable and expands.
Open-ended, autonomous, long-horizon, high-stakes work: stays expensive (token consumption + oversight + error-correction + scarce frontier compute); replacement remains marginal, and augmentation dominates.
The dividing line advances gradually as cost falls and capability rises β the opposite of the "AI replaces organizational layers overnight" marketing.
Metered pricing makes the market sort honestly: it kills the illusory arbitrage (rent a flat plan, fire the worker) while leaving intact the cases where AI is actually cheaper than a human for a bounded task.
Most fragile assumption. That the open-weight floor and competition continue to cap closed-provider rents. If the frontier consolidates and the open ecosystem fails to keep pace, durable rent-seeking on locked-in heavy users becomes possible, and viability would be artificially capped above true compute cost β the pessimistic reading of the original thesis.
D. The technical engine behind the cost curve β LLM performance mechanics
This section answers why effective cost behaves as Β§C describes. Three questions were posed: the shape of the size increase, whether equal increases yield equal gains, and whether scale matters relative to reasoning/context/prompt engineering.
D.1 Shape of the size increase. "Size" splits into three diverged metrics. The cleanest is training compute (FLOP), which is exponential over calendar time: about 4.1x/year from 2010 to mid-2024, roughly 5x/year for frontier language models since 2020 β a doubling every ~5.2 months (Epoch AI). On a log axis this is a straight line. Parameter count decoupled from compute after the Chinchilla result (2022) showed most models were undertrained; compute was reallocated toward data, with tokens-per-parameter rising from about 10 (2022) to roughly 300 (2025) for open models. Mixture-of-experts further separates active from total parameters, and closed frontier counts are undisclosed (cannot verify). Summary: compute exponential and continuing; parameters decoupled, slowed, and partly opaque.
D.2 Does equal increase give equal gain? No. Pretraining loss falls as a power law in compute, data, and parameters: each equal multiplicative step buys a roughly constant loss reduction, while each equal additive step buys steadily less β exponential cost for linear progress on loss. Two further nonlinearities sit on top: the map from loss to benchmark score is uneven (the contested "emergence" effect, possibly a metric artifact; plus saturation near benchmark ceilings), and "perceived" performance β reliability, not hallucinating, agentic coherence β tracks loss even more loosely. This compounding produced the "pretraining wall" narrative of late 2024: diminishing returns and a plateau in the effectiveness of pretraining scaling.
D.3 Scale vs. reasoning vs. context/prompts β a stack, not an either/or. These operate at different layers, and the dominant source of gains shifted over time. From roughly 2018β2023, pretraining scale led. From late 2024, marginal frontier gains came increasingly from post-training (reinforcement learning, including RL with verifiable rewards on math/code) and test-time compute (extended chains of thought; o1/o3, DeepSeek-R1, Claude and Gemini thinking modes). Research shows scaling test-time compute can beat scaling parameters for reasoning, with 2β4x efficiency from compute-optimal strategies. Three constraints:
The base model sets a ceiling you cannot reason past: test-time compute substitutes for pretraining only within the model's capability; genuinely hard out-of-capability problems still require pretraining.
Reasoning/RL has its own diminishing returns; over-long reasoning can degrade calibration.
Context- and prompt-engineering improve the use of a fixed model (high deployment ROI, often where enterprise value lands via retrieval/grounding) but raise no ceiling.
Hierarchy: pretraining sets the ceiling (still growing, diminishing return per dollar); post-training RL and reasoning are currently the highest-leverage gains; test-time compute converts inference dollars into capability; context/prompt engineering convert a fixed model into delivered value. The weight moved down this list since 2024.
The cost coupling (the crux). In the reasoning paradigm, "better" partly means "thinks longer," i.e., more tokens, i.e., higher cost per task. The capability frontier and the inference-cost frontier became coupled in a way they were not in 2022β2023 β the technical mechanism behind the effective-cost squeeze. The countervailing force is distillation: capabilities discovered expensively in large models or long RL runs migrate into small, cheap ones (e.g., the s1 result, where fine-tuning on about 1,000 distilled reasoning traces produced strong reasoning). So for any fixed capability the cost keeps falling, while the frontier stays expensive because it always runs at the longest thinking budgets. This is the mechanism behind the moving frontier in Β§C.
E.1 Manufacturing automation vs. offshoring
The premise tested: cheap (and often exploited) human labor, in regions with weaker labor and environmental protections, has repeatedly beaten automation.
Industry-split, not uniform. The premise holds strongly in apparel, footwear, and electronics assembly, for two compounding reasons. Technically, sewing resists automation β fabric shifts, stretches, and folds unpredictably, and robots lack the tactile feedback and flexibility to match human sewers; flagship "sewbot" efforts (Sewbo, SoftWear Automation) stalled for years and retreated to premium niches. Economically, fast-fashion margins are so thin that firms chase the lowest wages rather than invest capital. The cleanest data point: Crystal Group, the world's largest clothing maker, explicitly bet on humans, stating sewing robots could not compete on cost with developing-country labor. But the premise fails in autos, semiconductors, and appliances, which are heavily roboticized β fewer workers, near-record output. The deciding variable is whether the physical task is amenable to rigid automation and whether margins justify the capital.
The trade-vs-automation attribution is contested. One reading (Public Citizen, citing Acemoglu's acknowledgment) holds China trade displaced about three times as many manufacturing jobs as robots did from 1990β2007; a competing reading (AEI) holds automation/productivity is the long-run driver and the China-shock estimates are overstated; a general-equilibrium model (Caliendo et al.) attributed only about 15% of 2000β2007 manufacturing job losses to the China shock. Cross-country, robots cut manufacturing jobs in the US and France but were offset in Germany and the Netherlands, with labor-market institutions explaining much of the difference. So the macro attribution is disputed, not settled in labor's favor.
Parallels to white-collar AI.
The Moravec divide is relocated: the hard-to-automate residue moves from physical dexterity to judgment, accountability, and tacit context.
A cheap-labor floor exists here too: AI's real comparator for much back-office, support, and junior analytical work is already-cheap offshore labor (Indian IT/BPO, Filipino back-office), not expensive Western staff β which tightens the math as the Bangladesh wage tightened it against the sewbot. Klarna is the white-collar echo of Crystal Group's bet.
The capex/integration barrier repeats (the 95% MIT failure; integration at 35β45% of first-year cost).
The likely employment shape matches the automated-industry pattern: output up, fewer workers per unit, polarization, and hollowing of routine middle-skill cognitive work (Autor's polarization mechanism).
Disanalogies (cutting both ways).
Cost structure: a robot is fixed capital that cheap labor underpriced; AI inference is a variable, scalable, location-independent cost, weakening the capex moat that protected garment workers. But metered pricing makes AI behave more like a wage (pay per task) than like capital β making the AI-vs-offshore-labor contest more direct and ongoing.
Tradability: digital service work is fully tradable, so white-collar work has less geographic protection than local physical services.
Automation fit is partly inverted: AI is native to symbol manipulation (routine cognition, historically expensive) and weak at the physical (historically cheap). So routine white-collar work may be more exposed than routine manufacturing was, and AI can undercut even the cheapest offshore wage on high-volume structured tasks β something the sewbot never managed.
Synthesis. Not "labor saves workers," but a three-way sort β AI vs. cheap offshore labor vs. expensive local labor β running task by task on cost and reliability. The structured core goes to AI or offshore; the judgment-and-accountability residue stays local. Cheap labor is a floor that shapes where automation lands, not a wall that stops it everywhere. Caveat: the manufacturing record spans decades of a relatively stable robot-capability curve, whereas the AI capability curve is fast and unsettled β an informative base rate, not a reliable quantitative prior.
The premise tested: self-driving has been hyped since the 2010s, yet by 2026 it exists only in a handful of US cities with the simplest layouts, and only as taxis; driving is a semi-conscious task any human learns in a few weeks; so if AI cannot be trusted to drive, it cannot be trusted to run businesses, build software, or make financial decisions.
Factual corrections. Deployment is wider, harder, and not taxi-only. Waymo runs roughly 3,000 robotaxis, served over 20 million trips, expanded to 10 US cities by February 2026 (with groundwork for 20-plus, including London and Tokyo), covers on the order of 1,400 square miles, and began freeway service in late 2025. "Simplest layouts" fails: San Francisco β a core market β is among the hardest driving environments anywhere. It fails entirely against China, where Pony.ai holds fully driverless commercial permits in all four tier-one cities (Beijing, Shanghai, Guangzhou, Shenzhen), operating amid tunnels, dense traffic, and irregular intersections. And it is not taxi-only (driverless trucking exists).
The premise that holds. The timeline genuinely blew past a decade of "next year" promises, and the long tail is currently visible: in May 2026 Waymo paused freeway service across markets after a robotaxi drove into a flooded road, struggled in construction zones, and recalled thousands of vehicles. The caution is well-founded.
The inference's flaw (two parts).
Moravec's paradox. Driving is not "easy." The "few weeks" trains the symbolic layer (rules, signs); the hard part β real-time scene perception, physics intuition, predicting other agents, motor control β runs on perceptual machinery evolution pre-installed, which we experience as effortless. That effortlessness is what misleads. Driving sits in the part of the difficulty space that is worst for machines, inverting the ordering the argument assumes. (The human benchmark is not weeks either: newly licensed drivers crash most.)
Failure tolerance. Driving demands near-perfect, unsupervised, real-time reliability β errors are instantaneous, physical, irreversible, potentially fatal, and un-reviewable before they occur; a 95%-reliable driver is lethal. Most white-collar tasks have the opposite profile: recoverable, reviewable before binding, non-fatal, tolerant of a human in the loop. That is precisely why AI is already deployed in software, finance, and operations as supervised augmentation. The property making driving hard to ship is absent in most knowledge work.
The conflation. The inference mixes two deployment modes. Full unsupervised autonomy is what driving requires and what AI mostly cannot do safely anywhere yet β and on that, the self-driving record is good evidence. Supervised augmentation is what knowledge work permits, and AI is trusted with bounded, reviewable pieces of it now. Self-driving is the purest threshold-gated task: near-zero commercial value until near-perfect reliability, then a jump. Most white-collar value releases continuously under supervision β which is why white-collar AI monetizes now while robotaxis took seventeen years to reach limited commercial scale.
The bidirectional caveat. The ROI thesis driving the layoffs depends on removing the human reviewer to capture savings. The moment a firm pushes an AI agent toward unsupervised, irreversible action in a business, it re-enters the driving regime β rare catastrophic edge cases, the demand for near-perfect reliability, the long tail. The robotaxi history predicts that push will be slow, costly, and repeatedly walked back (as Klarna was). So the self-driving record supports skepticism about the full-autonomy end the cost case actually needs β not about augmentation.
The original argument is directionally sound where it matters most and wrong in its strongest single mechanism. The layoff wave preceded and is largely independent of AI; the over-hiring correction is real but not the sole cause; "AI washing" is a documented, named phenomenon; the ROI shortfall is well-evidenced; and the bandwagon and its regret are real. The pricing claim, initially under-credited, is substantially correct as a structural observation: all three dominant providers moved, in the same direction within twelve months, from flat allowances to metered limits, weekly caps, model downgrades, and pay-as-you-go overage β concentrating the cost on the heavy users who define the replacement use case. But that change is a one-time shift from subsidized to cost-reflective pricing, the rise is in metering rather than sticker price, and it is entangled with a real compute-scarcity premium rather than being a pure deliberate squeeze.
The forward picture is bifurcation, supported by both the technical mechanics (a coupled capability/cost frontier offset by distillation) and two historical base rates (manufacturing's industry-split record; self-driving's threshold-gated, long-tail-limited trajectory). Routine, reviewable, error-tolerant work gets steadily cheaper to automate and will be; autonomous, irreversible, high-stakes work stays expensive and resists full replacement, with augmentation dominating. The deciding variables are compute supply (which sets the cost floor and is unresolved) and whether capability crosses the no-human-in-the-loop reliability threshold (which gates full replacement and may matter more than price). The marketing claim of AI replacing whole organizational layers describes the hardest, slowest, most repeatedly-reversed end of the spectrum β not the near-term reality.
stormborns
jacobvanloon
gracieabrams
areudirty
dailybridgerton
snowghoul
amandla
tangentwithtanner