OpenAI Nerfed Your Tokens - CourtneyR.dev's 2nd Brain

OpenAI Nerfed Your Tokens ![rw-book-cover](https://d34adp677peecb.cloudfront.net/static/images/article1.be68295a7e40.png) 🔗 **Source:** [OpenAI Nerfed Your Tokens](https://www.youtube.com/watch?v=P5llBhv8TQ4) > [!note]+ Document Notes > tags: [ai-pricing, agentic-workflows, web-frameworks, frontend-linting, ai-job-displacement] > author: "Robby the Wagner and Adam Thomas (names as the Whisper transcript renders them; likely Robbie Wagner and Adam Thomas, an inference from the audio, not confirmed by the description)" > source: Whiskey Web and Whatnot (YouTube) > reference: https://www.youtube.com/watch?v=P5llBhv8TQ4 > date created: 2026-10-02 09:38 > > > [!info] Generated analysis — not the author's words. Sections below are machine-written > > from the captured source. Only text inside `> [!quote]` blocks is verbatim from the author. > > Do not quote or attribute any other line to the named author. > > # 🏷️ Title: OpenAI Nerfed Your Tokens >  ## 🧭 Table of contents - [📝 Summary](#-summary) - [📖 Main content](#-main-content) - [🗒️ Notes](#️-notes) - [📚 References](#-references) - [🗣️ Paraphrase](#️-paraphrase) - [📰 Newsletter](#-newsletter) - [✅ Action items](#-action-items) - [🧪 Hypotheses](#-hypotheses) - [📊 Methodology](#-methodology) - [❓ Questions](#-questions) - [🔮 Future research](#-future-research) - [💡 Implications](#-implications) - [🧠 Learning styles assessment](#-learning-styles-assessment) - [📊 Accessibility assessment](#-accessibility-assessment) - [🌍 DEIB assessment](#-deib-assessment) - [✅ Logical validity](#-logical-validity) - [🛡️ DARVO analysis](#️-darvo-analysis) - [🎯 Conclusion](#-conclusion) - [🔄 Refactor appendix](#-refactor-appendix--openai-nerfed-your-tokens) ## 📝 Summary - **Source quality warning.** This video had no captions. The transcript is a local Whisper "base" transcription, so it has mishearings, wrong product and person names, missed words, and no reliable punctuation. The opening already shows it: the hosts introduce themselves as "Robby the Wagner" and "Adam Thomas" (likely Robbie Wagner and Adam Thomas), and the bourbon is called "jipped the creed" (a garbled brand name; the real one is not recoverable from the audio text). Other garbles are flagged in Notes and References. Model names in particular ("Opus 5.5," "Astra," "soul 6.1," "Fable") are the transcript's spelling and unconfirmed. Every `> [!quote]` block preserves that spelling. - **What happened.** Whiskey Web and Whatnot, a casual two-host web-dev news-and-whiskey livestream, uploaded a 1:12:56 episode on 2026-10-01. The first 47 minutes cover AI news and opinion. The last 25 minutes are off-topic banter: t-shirt printing, a Zelda Switch, designer toys, college football, Halloween costumes. The description is one generic line and the video has no chapters. - **The claim in the title.** One host says OpenAI "had a conference or something" and announced it would "nerf everything" and "charge you $500 a month" (00:00:00). A second thread about plan changes follows at 00:03:42: "charge you more and you're going to get less, but they're better." Neither host cites a source, a plan name, or a number beyond $500. Treat the nerf and the $500 figure as unverified spoken claims. - **Other news and claims, all spoken and unsourced.** Anthropic made a similar plan change with percentages the host could not recall (00:04:26). A $200 monthly subscription is said to equal about $9,000 of API credits, attributed to "Theo" (00:09:37). OpenAI "dots," a video-call agent called "him," and a Meta handheld the transcript calls "muse charm" are described as personified agents (00:11:54). "Hill climbing" is described as an eval-driven incremental technique built into Claude (00:16:30). - **The opinions with the most practical weight.** Frameworks and abstractions exist for human brains and matter less to agents (00:21:16). Linters work as deterministic guardrails, an "extension of the skill" (00:24:21). Agents using a site in a background tab never see CSS transitions run, so UI that waits on `transitionend` stalls (00:43:08 to 00:44:44). A host who says they work at Shopify reports a linter ban on the CSS `:has` selector as measurably slow (00:42:18). - **Key stakeholders.** The two hosts and their audience of web developers. OpenAI and Anthropic as the named subscription vendors. Shopify as the named employer behind the linter and motion stories. Developers who run agents against live sites and pay for plans themselves. - **Risks and tradeoffs.** Pricing, plan, and product claims are second-hand and unsourced. Names are garbled. Speaker turns are not labeled, so this analysis attributes opinions to "one host" or "the hosts." About a third of the runtime is unrelated to the title. - **Bottom line.** The title claim is not established by this source. The durable material is the engineering content: background-tab animation failure, lint rules as deterministic guardrails, and the debate over whether frameworks still earn their cost in an agent-driven workflow. ## 📖 Main content **Transcript quality, again, because it shapes every sentence below.** Whisper "base" with no caption fallback means proper nouns are guesses. Examples seen in the text: "Robby the Wagner" (Robbie Wagner), "jipped the creed" (a bourbon brand, unrecoverable), "weeded bourbon" and "10% weed" (wheated bourbon, 10% wheat), "Tom MacGachies" (Tamagotchis), "DHA" (almost certainly DHH), "wigum loops" (Ralph Wiggum loops), "Zingad" (Zynga), "Oswary" (ossuary), "AS color" (AS Colour), "Akari no time" (Ocarina of Time), "On Jean de Poir Trin" (a band name, unrecoverable), "cloud" where context means Claude, and "Jeff" in three places where the referent is unclear. Treat any capitalized name here as a lead to verify, not a fact. **News items, as the hosts report them (all unverified in this source).** - *Plan changes.* One host read a thread where OpenAI justified changes to its subscription. The host's summary: higher price, lower usage, with the claim that the models are better. The same host says Anthropic did something similar to its Max plan and "put through a ton of percentages" (00:04:26). A $500 tier is said to be required for a fast mode of the model the transcript calls "Astra" (00:05:08). - *Subsidy economics.* One host repeats a figure attributed to "Theo": a $200 monthly Claude subscription equals about $9,000 of API credits (00:09:37). The host expects the value proposition to erode as lower-cost models from China match capability, and calls the investment situation "propped up." - *Personified agents.* OpenAI "dots" (a personality behind an agent), "him" (an agent that joins video calls with a voice), and a Meta handheld with cloud-hosted agent VMs, rendered "muse charm" (00:11:54 to 00:12:37). A friend who works on Meta glasses says the form factor is good, the field of view is small, and the market is unclear (00:12:37). - *Hill climbing.* An incremental, eval-gated improvement loop, now invocable in Claude according to the host who explains it (00:16:30). The hosts compare it to auto research, Ralph Wiggum loops, and last episode's "quality gauntlet." The explaining host says it requires evals to know a step improved the result (00:17:58). - *Rails.* One host says DHH announced he is "done" with Rails (00:20:25). The other host saw part of the talk and "didn't see him denounce rails" (00:21:16). Unresolved in the episode. - *Shopify admin redesign.* One host says a redesigned admin was announced this week and includes a lot of motion that agents struggle with (00:44:00). **Opinion, grouped by argument.** - *Frameworks are for humans.* Abstractions, component models, and frameworks help "squishy" human brains reason about systems too large to hold at once. Agents memorize IDs and don't conflict (00:21:16 to 00:22:01). A counterweight from the other host: opinionated frameworks like Rails or Ember used to keep agents on a path, and linters, types, and end-to-end tests still give agents fast feedback (00:22:48, 00:24:21). - *Declarative over imperative.* Describing what you want ("what") matters more than specifying how, so one host is dropping TDD for end-to-end journey tests (00:25:56 to 00:26:45). The same host ties this to a long preference for CSS and HTML. - *Lint rules as guardrails.* A rule enforced in the pipeline works where a written skill may not be read or remembered (00:24:21). Examples: a ban on the CSS `:has` selector and a ban on `setTimeout` in favor of `transitionend` (00:42:18 to 00:43:08). - *Background-tab failure.* Browsers don't run animations in inactive tabs, so a click that opens a transitioning modal never finishes. `transitionend` doesn't fire and elements stay at opacity zero (00:43:08 to 00:44:44). - *CLI over website.* Relaying a talk, one host says the guidance was "I don't want your website. I want your CLI" (00:46:26). Both hosts use GitHub, Atlassian, and Shopify CLIs with agents. - *Jobs.* The hosts say white-collar jobs that touch files or equations are at risk (00:37:21, 00:39:01), and discuss shorter work weeks and universal basic income without endorsing either (00:38:12). **Banter and tangents (about 25 minutes of tail plus short detours earlier).** Bourbon tasting and a Kansas City trip (00:00:45 to 00:03:42). Banjo and camping (00:32:28). Composite decking (00:31:41). A t-shirt brand and shirt weights (00:47:58 to 00:55:23). A Zelda-themed Switch purchase, stock-tracker apps, and resale bots (00:55:23 to 01:00:51). Designer vinyl toys and Pokemon cards (01:00:51 to 01:05:34). A bachelor party, Virginia Tech football, and Halloween costumes (01:05:34 to the end). **Contested or unverifiable in this source.** The $500 price, the "nerf," the $9,000 credit figure, the Rails announcement, the speculation that a vendor feeds a better model behind the scenes (00:06:43), the claim that ":has" is "wildly slow and destructive" (00:42:18), the 12-month prediction that streaming music will die (00:29:16), and the Charizard value (01:04:41). Each is a spoken claim with no source on screen that the transcript can show. ## 🗒️ Notes #### ⏱️ [00:00:00](https://www.youtube.com/watch?v=P5llBhv8TQ4&t=0) Cold open: nerf and $500, then the bourbon - The hosts introduce themselves; the transcript spells them "Robby the Wagner" and "Adam Thomas" (garbled; likely Robbie Wagner and Adam Thomas). - One host teases the title story and says they will "get into that more in a bit." - The tasting: a six-year wheated bourbon (transcript: "weeded"), mash bill read as 70% corn, 15% rye, 10% wheat, 5% barley; bloody butcher red corn; a distillery described as 100% woman-owned with a former engineer as master distiller. The brand name is garbled ("jipped the creed"). > [!quote] [00:00:00] > the open AI stuff, like they had a conference or something and announced that they were going to nerf everything they have and then also charge you $500 a month. #### ⏱️ [00:03:42](https://www.youtube.com/watch?v=P5llBhv8TQ4&t=222) Plan changes and model speed tiers - Each host rates the bourbon a six, then moves to OpenAI. - One host read the company's thread justifying subscription changes: pay more, get less, models are better. - The other adds that Anthropic made a comparable change to its Max plan with confusing percentages. - Both say the current frontier models are strong. A fast mode is said to need the $500 plan, and the "Astra" model is described as accurate but slow (00:05:08). > [!quote] [00:04:26] > Everything's better. Everybody and everyone's like, now it kind of sounds like less #### ⏱️ [00:05:56](https://www.youtube.com/watch?v=P5llBhv8TQ4&t=356) Model churn and the "hooked then worse" theory - One host has used a single model for 1.5 to 2 weeks before a new release made them switch. - Theory: vendors ship a stronger version to hook users, then quality slips. The host admits it may be adaptation to the model and says "I don't know." - Modem analogy (28.8k to 56k to DSL): fast releases make last week's model feel dated. > [!quote] [00:06:43] > which makes me suspicious that they're like feeding it into like fable or something behind the scenes. #### ⏱️ [00:08:09](https://www.youtube.com/watch?v=P5llBhv8TQ4&t=489) Who pays, subsidies, and copycats - One host's employer pays for the top tier of every tool, which won't hold for personal use. One host applied to a Codex program for open source and the other plans to. - The "slow down" pitch from labs is mocked, since new models kept arriving. - Subsidy claim: a $200 monthly subscription equals about $9,000 of API credits, attributed to "Theo." Even at that price the host doubts the economics. - A stealth startup (name garbled as "Jeff") is said to be copied within hours by big companies. > [!quote] [00:09:37] > if you did an API cost, like I think Theo is saying, like a, you know, $200 a month, a cloud subscription gets you like $9,000 worth of like API credits #### ⏱️ [00:11:07](https://www.youtube.com/watch?v=P5llBhv8TQ4&t=667) Personified agents and gadgets - One host frames the week as "personification of agents": OpenAI "dots," a video-call agent called "him," and a Meta handheld the transcript calls "muse charm" with cloud VMs. - A friend who works on Meta glasses says the field of view is small despite 4K OLED per eye, and asks who the product is for. - Both hosts used a Quest for company "metaverse meetings" during COVID and call it a gimmick with low-definition screens. - A "Tomagotchi" tangent: one host would like a physical familiar that follows them. #### ⏱️ [00:15:45](https://www.youtube.com/watch?v=P5llBhv8TQ4&t=945) Hill climbing, evals, and loop fatigue - One host explains hill climbing: nudge a result forward in small steps until it succeeds. A Rust rewrite is the example. It is now invocable in Claude. - The other host's reaction: it resembles breaking work into steps, which good engineering already does. - Evals are the gate that tells the loop whether a step improved the result. The explaining host contrasts it with last episode's "quality gauntlet," which loops until the output matches a described ideal. - The skeptic says chasing each new technique has diminishing returns because it may be obsolete in three days. Their team does run evals overnight at work. > [!quote] [00:18:45] > I think it's diminishing returns, like learning all this new stuff #### ⏱️ [00:20:25](https://www.youtube.com/watch?v=P5llBhv8TQ4&t=1225) Frameworks in the way, and what tests are for - One host expects computer science to boil down to iterative development and writing specs, with little need for extra tools. They recall another host installing every skill and ending up with a bloated context. - DHH ("DHA") is said to have called Rails done. The other host did not hear that. - Abstractions exist for human reasoning; agents "don't have trouble with IDs." The counterpoint: opinionated frameworks used to constrain agents usefully. - One host argues for end-to-end journey tests over unit tests, citing the "door doesn't open" meme. The other treats linting and tests as blurring into one deterministic check. > [!quote] [00:24:21] > It's just like that's part of your pipeline, you know? #### ⏱️ [00:25:08](https://www.youtube.com/watch?v=P5llBhv8TQ4&t=1508) Sunk cost, declarative thinking, malleable software - A host predicts people who built or own frameworks will cling to them because of sunk cost. - Declarative tools (CSS, HTML) are unpopular with developers who want control over how things work. The host says the shift to describing "what" is already underway. - Anecdote: a friend with no coding background had an app built in 45 minutes that handled recipe ingredients, store pickup ordering, a database, and authentication. The tool used is unclear ("chat GPT or something"). - Another anecdote: a native banjo-tuner app for a double-C tuning generated in minutes. The host concludes "app stores are cooked" and calls the result "malleable software." > [!quote] [00:25:56] > all we need to be able to do now in an agentic world is to describe what I want #### ⏱️ [00:29:16](https://www.youtube.com/watch?v=P5llBhv8TQ4&t=1756) Generated music and craft - A relayed tweet predicts streaming services die within 12 months because songs get generated on demand. One host half-agrees and grieves coding as "everything I loved is dead." - A band shirt prompts a math-rock tangent. Counter: human craft still impresses, like Lichtenstein as a printer. - Children aged 8 and 10 follow trends regardless of origin. Predicted outcome: live human music becomes a luxury, like hand carpentry. - Decking aside: one host wants composite decks for rot resistance. > [!quote] [00:29:16] > Apple music and Spotify will be dead because music in general will be dead #### ⏱️ [00:32:28](https://www.youtube.com/watch?v=P5llBhv8TQ4&t=1948) Banjo, camping, and a jobless-future riff - Camping trip jamming with a person the transcript names "Chris Koyer" (spelling unconfirmed). A tambourine on one foot kept a snare beat. - Playing an instrument builds "new neural pathways" (spoken, unsourced). Imposter syndrome while learning is normal. - A video of a future where displaced workers power AI plants by bike, with fit workers and fat wealthy owners. One host says only toppling capitalism changes the outlook. A friend predicts most people would drink rather than make art. #### ⏱️ [00:35:44](https://www.youtube.com/watch?v=P5llBhv8TQ4&t=2144) Game dev, white-collar work, and where the hosts land - Game development has many roles: riggers, animators, texture artists, developers. One host says people now make complete games in an evening with the "Astra" model. - Stated view: every white-collar career is at risk. A mechanical or structural engineer friend does formulaic bridge calculations an LLM can do. Accounting is "done." - The hosts reject shorter weeks and universal basic income as the right plan, without offering one. - A bounded take: any job that touches files or equations is under threat. Both hosts say they are trying to create with the tool in its current form. > [!quote] [00:37:21] > I think literally every career that is white collar work is toast. #### ⏱️ [00:39:46](https://www.youtube.com/watch?v=P5llBhv8TQ4&t=2386) Code review flood and lint rules - AI makes it easy to submit a 50-file pull request. One host built a skill to review the way they would, and it closes duplicates of other PRs. Reviewing 10 means 15 more arrive. - Both agree teams will keep caring about code details for another 3 to 5 years, though they say they shouldn't have to. - Code smell is taught to agents once and then enforced by linters. Shopify added a lint rule against the CSS `:has` selector as "measurably impacting our site," with PRs replacing it by JavaScript. - Another rule bans `setTimeout` and points to `transitionend`. > [!quote] [00:42:18] > it's like, measurably impacting our site in so much that we've banned it. #### ⏱️ [00:43:08](https://www.youtube.com/watch?v=P5llBhv8TQ4&t=2588) PSA: agents in background tabs - Browsers don't run animations in inactive tabs; they "don't quite pause them, but they don't run them." - An agent clicking a button that opens a transitioning modal sees nothing, since the transition never plays. Elements remain at opacity zero. - The host says users reported their agents could not use the redesigned admin, which had added a lot of motion. The fix is an elaborate workaround for agent-driven sessions. > [!quote] [00:44:44] > If you have animations or transitions in your application, agents running those in the background, they won't run and they will get stuck #### ⏱️ [00:46:26](https://www.youtube.com/watch?v=P5llBhv8TQ4&t=2786) CLIs as the agent surface - An analogy to separate mobile sites: build a plain surface for agents next to the animated one for humans. - The DHH talk (as relayed) says his agent needs a CLI, not a website. - A host installed the Atlassian CLI to have an agent check which Jira bugs are fixed. The GitHub CLI is described as essential and said to work "half the time" when GitHub is down. - The Shopify CLI is installed for podcast tasks. Prediction: CLI or API will beat the website for agents. #### ⏱️ [00:47:58](https://www.youtube.com/watch?v=P5llBhv8TQ4&t=2878) Off-topic: a t-shirt brand named for an ossuary - A host announces a new shirt business, transcribed "Oswary" (an ossuary: a bone repository). Black garments, white bone art. Candidate domains: skull.style and a ".store" domain. - An SEO question about redirecting multiple domains goes unanswered. - Brand talk: American Apparel (declined), Gildan, then AS Colour (long torso fit, weights). Comfort Colors is criticized as too stiff. #### ⏱️ [00:51:15](https://www.youtube.com/watch?v=P5llBhv8TQ4&t=3075) Off-topic: shirt testing and print methods - The host buys base shirts, prints their art, then washes and wears each before choosing one. - A base tee is mentioned at about $43; target prices are "$10 and $20" for others. Print type is described as direct-to-film, with a back-print "murdered out" series planned. - Solid shirts with small art are called European in style. #### ⏱️ [00:55:23](https://www.youtube.com/watch?v=P5llBhv8TQ4&t=3323) Off-topic: Zelda Switch, bots, and alert agents - A Zelda 40th-anniversary themed Switch and controller (the Ocarina of Time remake is garbled "Akari no time"). A stock-tracker app, name garbled, notifies on restocks. - The host won one from a Nintendo app alert while filling out sleep-doctor paperwork. Bots beat humans to checkout; the host declines to build one, citing terms of service. - The other host runs agents that scan band-show and banjo listings and send a daily digest, using tool names the transcript garbles ("punk ass watching"). - A jokey rant about resellers and a "bag" of items follows; it is hyperbole, not a plan. #### ⏱️ [01:00:51](https://www.youtube.com/watch?v=P5llBhv8TQ4&t=3651) Off-topic: designer vinyl toys and Pokemon cards - Designer vinyl toys ("Dunnies"), a tattoo-artist series, drops in two cities, mystery boxes, and resale at 25 times retail (a $20 toy sold for about $500, per the host). - Pokemon cards are popular again, with Walmart drops. One host's child found a card an app valued at $1,500; condition and lack of plastic sleeve lowered it. - A first-edition Charizard is said to sell for about $350,000. Magic: The Gathering Lord of the Rings packs sit unopened. #### ⏱️ [01:05:34](https://www.youtube.com/watch?v=P5llBhv8TQ4&t=3934) Off-topic: bachelor party and Virginia Tech football - A bachelor party at a lake and a Virginia casino. The host says slot machines are one-button video games. Blackjack was more fun. - Virginia Tech is 4-0 and plays Pittsburgh, also 4-0, the next day. College football is preferred because players seem to try harder, though they are now paid. #### ⏱️ [01:08:03](https://www.youtube.com/watch?v=P5llBhv8TQ4&t=4083) Off-topic: Halloween costumes and sign-off - Family plan: Adventure Time (Ice King with a 3D-printed hat, Marceline, Finn), with Spider-Man variants from a kids' show also in play. - A sibling who wore a Spider-Man suit for about 1.5 to 2 years. - Closing joke: a tortilla chip in a monk robe as a "chip monk." The transcript ends at 01:12:01 with a compliment on the costume; the last 55 seconds of the 1:12:56 runtime have no transcribed text. ## 📚 References The transcript itself contains no links; the following come from the video description or are primary sources for named entities. The description is one generic sentence and has no links. No network was available in this session, so the URLs below are the canonical sites for each entity and were not fetched or re-checked. - [OpenAI Nerfed Your Tokens, Whiskey Web and Whatnot](https://www.youtube.com/watch?v=P5llBhv8TQ4): the source video, 1:12:56, uploaded 2026-10-01. - OpenAI and Anthropic plan and pricing pages: the primary source for the nerf, $500, and Max-plan claims. No URL or plan name is given in the transcript; check the vendors' own pricing and changelog pages before repeating any figure. - [DHH](https://dhh.dk): the person the transcript spells "DHA." The Rails and CLI remarks are second-hand; find the actual talk before citing. - [MDN: transitionend event](https://developer.mozilla.org/en-US/docs/Web/API/Element/transitionend_event): the event the host says does not fire in background tabs. - [MDN: Page Visibility API](https://developer.mozilla.org/en-US/docs/Web/API/Page_Visibility_API): the browser mechanism behind hidden-tab behavior. - [MDN: :has() selector](https://developer.mozilla.org/en-US/docs/Web/CSS/:has): the selector Shopify's linter is said to ban. - [GitHub CLI](https://cli.github.com) and [Shopify CLI](https://shopify.dev/docs/api/shopify-cli): tools both hosts say they use with agents. The Atlassian CLI has no URL here; confirm the product name first. - [AS Colour](https://www.ascolour.com): the shirt brand the transcript spells "AS color." - Ralph Wiggum loop: the technique the transcript spells "wigum loops." No canonical source is named; search for the original write-up. - "Theo": the commentator credited for the $200 versus $9,000 figure (00:09:37). Surname and channel are not given; unverified. - Hill climbing and "automating eval design": the host reads a title aloud (00:18:45). Exact source unknown; the transcript gives no URL. ## 🗣️ Paraphrase The hosts open by saying they missed a week of news and plan to learn it together. One flags that OpenAI announced plan changes: less for more money, with a $500 tier. Neither names a source. They agree the newest models are strong and slow, and that new releases arrive faster than anyone can adopt them. A recurring worry is that vendors subsidize heavy users and that cheaper copies erode the value of a subscription. The center of the episode is a long argument about what stays valuable in software work. One host says frameworks, component models, and abstractions were built for human memory and are now more drag than help, so what matters is describing the outcome and testing user journeys. The other host keeps a place for linters, types, and tests, and reframes lint rules as deterministic guardrails that don't depend on an agent remembering a written skill. They agree that review is the bottleneck: agents produce 50-file pull requests faster than a person can read them. A specific engineering warning stands out. Agents often drive websites in background tabs, browsers don't run animations there, and so UI that waits on transitions can hang the agent. The host who says they work at Shopify describes the problem as new and the fix as laborious. From there the talk turns to CLIs as the better agent interface, then to jobs, music, and game development, with both hosts uneasy and neither offering a plan. The final 25 minutes leave tech behind: a new t-shirt label, a Zelda Switch purchase, designer toys, Pokemon cards, a casino weekend, Virginia Tech football, and Halloween costumes. ## 📰 Newsletter **An agent in a background tab can't see your modal.** This week's Whiskey Web and Whatnot is a loose, funny hour with bourbon and banter, but the 43-minute mark has a PSA worth passing on. One host, who says they work at Shopify, reports that browsers don't run CSS animations in inactive tabs, so an agent clicking a button can wait forever on a modal that never finishes opening. `transitionend` doesn't fire, and elements stay at opacity zero. If your theme or plugin gates UI on a transition event, add a fallback for hidden tabs. MDN's [transitionend](https://developer.mozilla.org/en-US/docs/Web/API/Element/transitionend_event) page is the place to start. Other items, flagged as the hosts' claims and not verified here: OpenAI plan changes with a $500 tier, hill climbing as an eval-driven loop, and a lint rule banning the CSS `:has` selector at Shopify for performance. The transcript is a Whisper "base" pass with garbled names, so check any product or person before you repeat it. Link: [the episode](https://www.youtube.com/watch?v=P5llBhv8TQ4). ## ✅ Action items #### DevRel - Developer advocacy: write a short demo showing a transition-gated modal failing in a hidden tab and the `visibilitychange` fix, with a measured before and after. - Education: add "agents drive background tabs" to a front-end testing session, using Playwright with a hidden page. - Developer experience: put a note in any component docs that rely on `transitionend`, and give the fallback pattern. - Developer marketing: don't repeat the $500, nerf, or $9,000 figures in any material without the vendor's own pricing page as the citation. - Community management: when sharing this episode, label pricing items as "spoken claim, unverified." - Internal/external: ask the Shopify host (via public channels) for the actual write-up of the `:has` finding before citing performance numbers. #### OSPO - Policy and governance: decide whether lint rules that encode agent guardrails belong in contribution policy, since the hosts treat them as stronger than written skills. - License compliance: no direct action from this source; nearest relevance: agent-generated PRs raise provenance questions the hosts mention only as volume. - Security management: no direct action from this source; nearest relevance: agents operating live sites in background tabs deserve a threat-model line. - Software inventory: no direct action from this source. - Developer support: draft guidance on reviewing 50-file agent PRs, including the dedupe-first step one host describes. - Community engagement: no direct action from this source. - Strategic alignment: track review time per agent PR as a KPI, since the hosts report review as the bottleneck. #### Community - Maintainers: pilot a duplicate-PR review skill and measure how many it closes correctly against human decisions. - Contributors: document the lint rules a project enforces, so contributors and their agents hit the rule before review. - Sponsors: no direct action from this source. - Ecosystem: no direct action from this source; nearest relevance: shared lint rule sets across projects. #### WordPress - Core: no direct action from this source; nearest relevance: audit core admin UI for transition-gated flows. - Core Performance: test the claim that `:has` is costly on large DOMs with a published benchmark before anyone bans it. - Core AI: record the background-tab failure as a known constraint for agent-driven admin sessions. - **Core Program Team:** no direct action from this source. - Mobile: no direct action from this source. - Playground: build a reproducible hidden-tab animation test page that loads in Playground. - Community: no direct action from this source. - Training: add a short lesson on writing UI that still completes when animations don't run. - Marketing: no direct action from this source. - Support: no direct action from this source; nearest relevance: user reports of "my agent can't click the button." - Design: review motion budgets with a reduced-motion and hidden-tab pass. - Documentation: add a "works without animation events" note to interactive component docs. - Photos: no direct action from this source. - TV: no direct action from this source. - Polyglots: no direct action from this source. - Accessibility: check that `prefers-reduced-motion` paths don't depend on transition events either; the same failure class applies. - Meta: no direct action from this source. - CLI: WP-CLI as an agent surface matches the "give me your CLI" point; audit command coverage for common admin tasks. - Hosting: no direct action from this source. - Tide: no direct action from this source. - Openverse: no direct action from this source. - Themes: add a theme-check lint idea for `setTimeout` used in place of transition events. - Plugins: same lint idea for plugin review, as a suggestion only. #### The WPOCC - Programs: no direct action from this source; nearest relevance: The WP Open Community Collective could fund shared lint rule sets. - Projects: no direct action from this source. - Funding experiments: no direct action from this source. ### FAIR - Technical independence: no direct action from this source; nearest relevance: dependence on one vendor's plan terms is the risk behind the pricing talk. - Policies and covenants: no direct action from this source. - Productized governance artifacts: no direct action from this source. ## 🧪 Hypotheses 1. Pages that gate UI state on `transitionend` or `animationend` fail to complete a click-through when driven in a hidden tab, and a `visibilitychange`-based fallback restores completion. 2. A lint rule that blocks a pattern reduces agent-introduced violations of that pattern more than the same rule written as skill text. 3. Published usage limits per dollar on at least one major vendor plan fell between the previous plan and the one the hosts discuss (the "nerf" claim). 4. An agent review skill that checks for duplicate PRs first closes at least 20% of incoming agent PRs with at least 90% agreement with a human reviewer. 5. For the same task, an agent using a CLI completes it with fewer steps and fewer failures than the same agent driving the website. 6. Removing the CSS `:has` selector from a large app yields a measurable improvement in style recalculation time. ## 📊 Methodology - **H1.** Build a test page with a modal that appears on `transitionend`. Drive it with Playwright in a foreground tab and a hidden tab, 100 runs each. Record completion rate. Add the `visibilitychange` fallback and repeat. - **H2.** Pick one banned pattern. Run the same agent on 50 tasks with the rule as a skill and 50 with it as a linter hook. Count violations that reach review. - **H3.** Collect the vendor's pricing and usage-limit pages from the Wayback Machine for the dates before and after the announcement. Compare limits per dollar by plan tier. Cite the pages. - **H4.** Run the review skill over 200 historical PRs with known outcomes. Report precision and recall for "duplicate" and "close." - **H5.** Pick 10 tasks available both ways (for example, list Jira bugs and check status). Run each with the CLI and with browser control. Count tool calls, failures, and wall time. - **H6.** Use Chrome DevTools performance traces on a large page with and without `:has` rules. Report style recalculation time at p50 and p95 over 30 loads. - **Controls throughout.** Fix the model, prompt, and network. Pre-register the success threshold before looking at results. ## ❓ Questions - Which OpenAI plan, on what date, with what limit change, is the "nerf and $500" claim about? - Does the $500 tier exist as described, and what does the fast mode of the model the transcript calls "Astra" do beyond a thinking-level setting? - Who is "Theo," and what is the source for the $200 versus $9,000 comparison? - Did DHH actually say he is done with Rails, or was that a misreading of the talk? - How much slower is `:has` in Shopify's app, measured how, and on which browsers? - Do Chrome, Firefox, and Safari treat hidden-tab transitions the same way, and does it depend on throttling settings? - If agents run in background tabs, is the right fix a visibility fallback in each app or a change in how agent frameworks hold the tab? - Which models do "Opus 5.5," "Astra," "soul 6.1," and "Fable" refer to, given the transcript garbles? ## 🔮 Future research - **Hidden-tab behavior across browsers.** A cross-browser matrix of what runs, throttles, or stops in a background tab would turn one anecdote into a reference. - **Agent-ready UI patterns.** There's no shared guidance on building a UI that works for both animated human sessions and agent sessions without a second surface. - **Lint versus skill effectiveness.** Compare the hosts' claim that deterministic rules beat instructions against measured violation rates. - **Plan economics.** Independent tracking of usage per dollar across AI subscription tiers over time, built from archived pricing pages. - **Review load.** Measure how agent PR volume changes maintainer review time in open-source projects. ## 💡 Implications - **Practice.** Test interactive UI in a hidden tab and don't gate state on animation events. Add deterministic lint rules for patterns you don't want agents to write. - **Policy.** Projects need a rule for how agent-written PRs are labeled, deduplicated, and capped, since the hosts report review as the limiting step. - **Security.** Agents operating signed-in sessions in background tabs are a surface that threat models should name. - **Education.** Teach "describe the outcome and write the end-to-end test" next to the fundamentals, since one host argues those are what remains. - **Funding.** The hosts' own claim that subscriptions are subsidized means any project that builds on a plan price carries a pricing risk it can't control. Track the vendor's terms. ## 🧠 Learning styles assessment - **Reading level estimate.** Flesch-Kincaid Grade Level computed with a heuristic syllable counter on the transcript text: about grade 3.3, Flesch Reading Ease about 90. The low score reflects short spoken sentences and the Whisper punctuation, not simple content. Effective comprehension needs working knowledge of CSS, browsers, and agent tooling. #### 👀 Visual learners - Present: the hosts show products on screen at times (a bourbon bottle, a Dunny, a shirt), which the transcript implies but doesn't capture. - Improve: no diagram of the background-tab failure. A two-lane timeline of a foreground versus hidden tab would carry the key idea in 10 seconds. #### 🔊 Aural learners - Present: the format is conversation, so spoken explanation is the whole medium. - Improve: crosstalk and tangents bury the PSA. A 2-minute audio summary of the three engineering takeaways would help. #### 🗣️ Verbal learners - Present: analogies carry the argument, such as the modem speeds and the separate mobile site. - Improve: terms like hill climbing, Ralph Wiggum loop, evals, and `transitionend` are used without definitions. One clause each would fix it. #### 🧩 Kinesthetic learners - Present: the hosts describe doing things: installing a CLI, building a skill, buying shirts to test. - Improve: no repo, snippet, or steps for the fix. A copy-paste `visibilitychange` example would let a viewer try it. #### 🧠 Logical learners - Present: the framework-versus-guardrail debate has two sides argued with reasons. - Improve: the claims have no numbers behind them, such as the `:has` slowdown. A single measured chart would convert opinion to evidence. #### 🧑‍🤝‍🧑 Social learners - Present: two hosts converse and invite chat questions. - Improve: the description has no discussion link or community pointer, and the transcript names none. #### 🧘‍♀️ Solitary learners - Present: a one-hour recording can be watched at the learner's pace. - Improve: no chapters, so a solitary viewer can't jump to the 43-minute PSA. Chapters or timestamps in the description would help. #### 🌿 Naturalistic learners - Present: camping, a lake, and a porch by the water appear as asides. - Improve: none of the technical content connects to the environment; the energy cost of the model churn the hosts discuss goes unmentioned. ## 📊 Accessibility assessment #### Screen reader semantics - Present: YouTube supplies the player controls and landmarks; the transcript is machine-readable text. - Improve: on-screen products and any screen shares carry information the audio never describes. WCAG 2.2 AA 1.1.1 (Non-text Content) and 1.2.5 (Audio Description, Prerecorded) are not addressed for visual content. #### Keyboard navigation - Present: the YouTube player is operable by keyboard with visible focus. - Improve: nothing in this video adds a barrier. 2.1.1 and 2.4.7 pass through the platform. #### Contrast and readability - Present: not assessable from a transcript. - Improve: a livestream layout often uses small overlays; if code or chat appears on screen, 1.4.3 (Contrast Minimum) can't be checked here. #### Text resizing and reflow - Present: the player scales and captions resize independently. - Improve: burned-in on-screen text does not resize, which engages 1.4.4 (Resize Text) and 1.4.10 (Reflow) for any frame text. #### Media and captions - Present: a machine transcript exists, which is how this analysis was made. - Improve: the video had no captions at all, so 1.2.2 (Captions, Prerecorded) and 1.2.4 (Captions, Live) are unmet in the published video. The local Whisper pass garbled host names, product names, and model names, so even auto-captions would fail on every proper noun. The description has no chapters or links, which hurts navigation (2.4.5 Multiple Ways). #### Forms and interactions - Present: no forms or interactions in the video. - Improve: viewers who want to react have only YouTube chat; there is no alternative channel in the description. #### Standards reference - Primary gaps against WCAG 2.2 AA: **1.2.2 / 1.2.4** (no captions), **1.2.5** (no audio description of visuals), **1.4.4 / 1.4.10** (burned-in text). Adding human-corrected captions is the highest-impact and lowest-effort fix. ## 🌍 DEIB assessment #### Cultural breadth - Present: references span music, games, food, sports, and parenting, which gives viewers several entry points. - Improve: the tech content is US-centric (Shopify, Virginia Tech, Walmart, Kansas City). The AI-price discussion has no non-US pricing or regional access view. #### Inclusive language - Present: the hosts use direct, plain language and no slurs appear in the transcript. - Improve: frequent profanity may exclude some workplace viewers, and phrases such as "squishy ass human brains" and "we're all cooked" are casual. The jest about harming resellers (01:00:04) is hyperbole, but it normalizes violent phrasing. The mention of a woman-owned distillery is accurate and neutral, but the aside that a name "doesn't sound like a woman's name" (00:01:30) leans on assumption. #### Neurodiversity support - Present: topic changes are signposted with phrases like "all right, let's talk about." - Improve: frequent tangents and crosstalk, no chapters, and a 1:12 length make it hard to find a topic. Chapter markers would help most. #### Representation - Present: two hosts, and speakers cite friends and family from different roles (an engineer, a game artist, a mortuary worker). - Improve: the two hosts present as a pair and no guest voice appears. Women and other developers are referenced only in anecdote. #### Psychological safety - Present: the hosts disagree (Rails, frameworks, linting) without personal attacks, and admit when they don't know. - Improve: the job-loss talk ("every career ... is toast") is delivered casually, which could land hard on viewers worried about their work. A line naming resources or acknowledging the worry would reduce that. ## ✅ Logical validity #### Strong reasoning - **Mechanistic argument, 00:43:08 to 00:44:44.** Inactive tabs don't run animations, so an agent clicking a button never sees the modal transition, so `transitionend` never fires and elements stay at opacity zero. Each step follows from the last, and the host ties it to a real symptom (user reports of stuck agents). To strengthen it, name the browser and show a trace. - **Analogy, 00:23:34.** The "four unit tests pass but the door doesn't open" meme illustrates why passing unit tests don't prove a journey works. The point holds as a reason to add end-to-end tests. To clarify, state what the unit tests covered and missed. - **Deterministic enforcement, 00:24:21.** A lint rule in the pipeline always runs, while a written skill depends on the agent reading and remembering it. The distinction is sound. To strengthen it, compare violation rates. - **Counterargument, 00:22:48.** One host voices the opposite view, that opinionated frameworks used to constrain agents usefully, before the other answers. Presenting both sides gives the listener material to judge. - **Self-limiting admissions, 00:06:43.** Hosts mark guesses with "I don't know" and "maybe I'm getting used to it," which signals confidence level. #### Fallacies - **Hasty generalization, 00:27:37.** One friend's app "within 45 minutes, it completely all worked" supports "we're all cooked" for developers. One case, one app, unknown complexity, unknown tool. *Fix:* collect several cases with scope and maintenance cost. - **Overgeneralization, 00:37:21.** "every career that is white collar work is toast." Universal claim with no timeline or evidence. *Fix:* name occupations, tasks, and a measured exposure source. - **Unsourced attribution, 00:09:37.** "like I think Theo is saying" a $200 plan equals $9,000 of API credits. The host hedges the source and the number. *Fix:* link the original post and show the calculation. - **Appeal to unnamed authority and possible misreading, 00:20:25 and 00:21:16.** "I think DHA said he's done rails" is contradicted by the other host's "I didn't see him denounce rails." *Fix:* cite the talk with a timestamp. - **Speculation presented as inference, 00:06:43.** "which makes me suspicious that they're like feeding it into like fable or something behind the scenes" reasons from a feeling of quality to a hidden vendor practice. The host hedges, but there is no evidence offered. *Fix:* run a fixed eval over several weeks and compare scores. - **Motive attribution, 00:25:08.** "I think a lot of people are going to really struggle to get rid of these because of some cost fallacy" assigns a psychological reason to framework owners without evidence, and it works as a dismissal of their reasons. *Fix:* address their stated reasons directly. - **Prediction without basis, 00:29:16.** "Apple music and Spotify will be dead" within 12 months is relayed from a tweet with no mechanism for licensing, catalog, or social use. *Fix:* give a measurable indicator, such as share of listening hours from generated tracks. - **False dichotomy, 00:21:16.** "It does not care" about frameworks sets agent indifference against human benefit, but the other host shows at 00:22:48 that frameworks can still constrain agent output. *Fix:* separate what agents need from what teams need. - **Ambiguity from transcript quality, throughout.** Not the hosts' fault: unreadable names ("Jeff," "Astra," "soul 6.1") make some claims uncheckable. *Fix:* human-corrected captions. #### Tone of voice - Casual, profane, joking, and friendly, with hosts finishing each other's thoughts and mocking themselves. Opinions are stated as confident takes ("I'm ditching my TDD" at 00:26:45) with frequent self-doubt ("I don't know," at 00:06:43). The register moves to real worry on jobs and music (00:29:16, 00:38:12), then back to jokes within a minute. The engineering sections (00:42:18 to 00:44:44) are the most precise; the pricing sections are the least. That pattern is the best guide to which parts of this episode to trust. ## 🛡️ DARVO analysis Not applicable to this source. DARVO (Deny, Attack, Reverse Victim and Offender) describes a response to accusation or criticism, and this episode contains none. For completeness across the three prongs: **Deny**: no one denies wrongdoing; the nearest item is one host correcting the other on the Rails remark (00:21:16), which is a factual disagreement. **Attack**: the hosts mock vendors' plan changes ("we're fucking you," 00:04:26) and framework owners (00:25:08), which is criticism of positions and not of an accuser. **Reverse victim/offender**: no such move appears. ## 🎯 Conclusion The title claim, that OpenAI "nerfed" tokens and set a $500 tier, is two hosts relaying a conference announcement they didn't read closely, and this source gives no plan name, date, or limit to check. The transcript also comes from a local Whisper pass with garbled names, so even the model names are uncertain. The useful material is smaller and specific. Browsers don't run animations in inactive tabs, so an agent in a background tab can hang on UI that waits for a transition (00:43:08 to 00:44:44). Deterministic lint rules, such as banning `setTimeout` or the CSS `:has` selector at one company, are treated as stronger guardrails than written instructions (00:24:21). And the hosts disagree usefully on whether frameworks and tests still matter when agents write the code. **Next steps for readers:** test your interactive UI in a hidden tab and add a visibility fallback. Check the vendor's own pricing pages, not this episode, before acting on any plan or price claim. Find the DHH talk and the original "Theo" post before repeating what was said about them. Skip to 00:42:18 through 00:46:26 if you only want the engineering content. ## 🔄 Refactor Appendix — OpenAI Nerfed Your Tokens > Density: 2–5 Literature atoms (H3 pad), 0–3 Permanent candidates (H4 pad). > Emitted: 5 Literature, 2 Permanent. Resonance filter applied. The bourbon, shirt, toy, football, and Halloween segments were dropped as unrelated to the title and unevidenced. > > Konik claim-header rules: present tense, declarative, ≤10 words, searchable, unique. > > Atomicity: one declarative sentence per atom. "and"/"also" → split. > > Tag pool: #bmf · #addendum · #xref · #articleseed · #storystem · #share · #aphorisms ### 📚 Literature Split Pad #### [H1] Hidden-tab agents hang on animated interfaces > [!quote] [00:44:44] > If you have animations or transitions in your application, agents running those in the background, they won't run and they will get stuck **Context & use**: - 00:44:44, the PSA after the host explains the Shopify admin motion problem. - This is the most reusable engineering claim in the episode. It is mechanistic and testable in one afternoon with Playwright, which is why it is the claim worth keeping. **Tags**: #xref #articleseed #share  #### [H2] Lint rules enforce what skills only suggest > [!quote] [00:23:34] > putting in rules that will stop the agent from doing the wrong thing in a deterministic way **Context & use**: - 00:23:34, in reply to the argument that linters are no longer needed. - Keep with the counter-view at 00:22:48: lint also gives agents fast feedback, so the two hosts differ on emphasis, not on whether linting matters. **Tags**: #xref  #### [H3] Describing the outcome replaces specifying the method > [!quote] [00:25:56] > all we need to be able to do now in an agentic world is to describe what I want **Context & use**: - 00:25:56, the declarative-over-imperative argument that motivates dropping TDD for journey tests. - A strong opinion from one host with no measurement. Pair it with the hosts' own "door doesn't open" analogy before using it. **Tags**: #xref #aphorisms  #### [H4] A shipped ban needs measured cost > [!quote] [00:42:18] > it's like, measurably impacting our site in so much that we've banned it. **Context & use**: - 00:42:18, on the CSS `:has` selector linter rule at the host's employer. - The measurement is asserted, not shown. Keep as a prompt to ask for the number before copying the rule. **Tags**: #xref  #### [H5] The title claim is an unsourced relay > [!quote] [00:00:00] > announced that they were going to nerf everything they have and then also charge you $500 a month. **Context & use**: - 00:00:00, the opening line; the later thread at 00:03:42 adds no plan name or date. - Teaching specimen for reading vendor-news commentary: second-hand, no source, one number. **Tags**: #xref #share  ### 🧠 Atomic Split Pad #### [P1] Hidden-tab test for agent-facing UI **Claim** (one sentence, Courtney's voice): If my interface waits on animation events to finish a step, an agent working in a background tab will stall, so I test every interactive flow with the tab hidden. **Evidence pointer**: [[#H1]] **Scope & conditions**: Holds when the page gates state on `transitionend` or `animationend` and the agent drives the page in a non-visible tab. Doesn't hold for agents that run headless with animations disabled, or for flows with a timeout fallback. Browser behavior varies and needs a cross-browser check. **Links**: - [[Page Visibility API]] — because extends - [[Reduced-motion patterns]] — because applies-in the same failure class - [[Agent-ready interfaces]] — because contrasts on building one surface versus two  #### [P2] Deterministic rules beat written instructions for agents **Claim** (one sentence, Courtney's voice): A rule that runs in the pipeline holds every time, while an instruction depends on an agent reading and remembering it, so the rules I care about belong in a linter. **Evidence pointer**: [[#H2]] · [[#H4]] **Scope & conditions**: Holds for patterns a tool can detect statically. Doesn't cover judgment calls, and a rule without measured cost can ban something useful, as the `:has` case may show. **Links**: - [[Agent skills versus hooks]] — because extends - [[Code smell as lint rule]] — because applies-in review automation - [[End-to-end journey tests]] — because contrasts on what each catches  ### ✏️ Fleeting Queue - [F1] Find the actual OpenAI and Anthropic plan announcements that match "nerf" and "$500 a month," with dates and limits; sources: vendor pricing pages and Wayback Machine #fleeting #inbox - [F2] Identify what "Astra," "soul 6.1," "Opus 5.5," and "Fable" mean in the 00:05:08 to 00:06:43 passages given the Whisper garbles #fleeting #inbox - [F3] Open question: do Chrome, Firefox, and Safari all skip `transitionend` in hidden tabs, and does throttling matter? Context at 00:43:08 #fleeting #inbox - [F4] Find the source behind "$200 a month gets you $9,000 of API credits" attributed to "Theo" (00:09:37) #fleeting #inbox - [F5] Article seed: "Test your UI in a hidden tab" for agent-driven sessions #fleeting #inbox #articleseed - [F6] Check the DHH talk for what was said about Rails and CLIs (00:20:25, 00:46:26) #fleeting #inbox ### → Things 3 (max 2 actions) - [ ] Build and run a hidden-tab Playwright test of a transition-gated modal, with a visibilitychange fallback; front-end testing, 1 hr #things - [ ] Verify the OpenAI and Anthropic plan-change claims against the vendors' own pricing pages before citing this episode; research, 20 min #things  📄 **Full text:** [[Full Document Contents/Articles/OpenAI Nerfed Your Tokens.md|OpenAI Nerfed Your Tokens by Robby the Wagner and Adam Thomas (names as the Whisper transcript renders them; likely Robbie Wagner...]] ## Highlights - [00:00:00] at day because who knows today, we got some spicy stuff for you. We are your hosts, as always, Robby the Wagner and Adam Thomas, our guy with the nerd, what is up? Everybody get ready. Sup, sup. Yeah, I missed a lot of news. I feel like this week. So this will be, we will all be learning along with myself from all of the knowledge Adam can give us today. Um, the one thing I did see, which, uh, well, work on opening this whiskey while I mentioned it, but, um, the open AI stuff, like they had a conference or something and announced that they were going to nerf everything they have and then also charge you $500 a month. [00:00:45] So we'll get into that more in a bit. But first we get to start with some whiskey. Okay, I got all the social shared out. Well, yeah, we got ourselves jipped the creed today, six year weeded bourbon. What a name. Um, what a name is a little hard to say, but I loved the fact about the bloody butcher. So we'll have to get into that here in a second. Oh, yeah, I did not read the notes yet, but it's a weeded bourbon, which I tend to prefer. Oh, I was just, so the reason I chose this today, by the way, is because I was just in Kansas City. I was there for a wedding, there for a couple days, tried as many local, uh, bourbon as I could and barbecue, of course. And yeah, anyway, this one. ([View Highlight](https://read.readwise.io/read/01m3ydqy006sbftwcb7f8raqyh)) - Note: [H5] The title claim is an unsourced relay - Tags: atom/literature/h3 - [00:22:48] if you were using Rails or Ember or something that was very opinionated, it was really good for, especially if it was in like TypeScript or like agents to know exactly like, oh, I can't fuck this up because I have to do it this way because that's what the framework tells me to do. But yeah, now it doesn't really need that. Yeah, so a lot of people are arguing with me at work and they're like, I still I'm still going to build linters and I'm still going to build in a typed system because I want my agents to have the fast feedback. And I'm like, well, cool. It doesn't care. I'm going to build end-to-end tests that make sure that their journeys can happen because like you've seen the meme where like all four unit tests pass on this door except the door doesn't open. You know, like there's like things clunking into each other. It's like, oh, [00:23:34] shit, you went all in on TDD, but didn't do any end-to-end tests. And now people can't actually use your app, but all your tests pass. And I'm like, I don't want that. I don't want my linter and test to all pass and my types to be good, but you can't get any shit done. I want to make sure that you can get shit done on the rest of it's noise. Man, the rest of it's for the machines these days. Like I got journeys to fulfill not my new shot to make sure that it runs fast for the stupid agent anyway. So I was just going to say that I think I hear what you're saying, but I think that linting and tests are going to get a little bit blurred. Like the way I'm thinking about linting now is like putting in rules that will stop the agent from doing the wrong thing in a deterministic way because like linting can, you know, sometimes auto-fix itself and like do ([View Highlight](https://read.readwise.io/read/01m3ydqn29fedtbwgks059p7yz)) - Note: [H2] Lint rules enforce what skills only suggest - Tags: atom/literature/h2 - [00:25:56] too much. You're not in control enough. You describe what you want. I want to say how it works because that's what really, really tickles my brain. You know what I'm like, I'm in the same boat, working a bit JavaScript, Java, any imperative language I've written is very challenging, very satisfying, very fun. But guess what? It's gone because all we need to be able to do now in an agentic world is to describe what I want because how is it? How is always going to be done better than whatever I was going to do? And it figures out how to copy cats to not always, but I mean, like, let's imagine Fable 10. Let's imagine Astra 10. You know, we are way better telling it how how to do things. And we'll just be asking for what? I mean, words, we're like close right now. We are over the hump of us defining how things work and to [00:26:45] describing what I want. And that's why I'm ditching my TDD. I'm going straight for end-to-end tests and and journeys that I need to be fulfilled. And because I'm like, I need to, I need to focus my attention on what I want, which is kind of awesome. That's maybe that's why I've always liked CSS is I like describing what I want and letting an engine take care of things. And a lot of other people are control freaks and they don't like that. They're going to have to get over it. Yeah. We're like engineering. But the age of fun engineering is gone. My friend texted me the other day. He was like, Hey, like, he doesn't have any coding experience. And he was like, I wanted an app for like getting ingredients for recipes and like automatically sending them to my like store, like buying them automatically so we can do a pickup order or whatever, like with all ([View Highlight](https://read.readwise.io/read/01m3ydqrb6tdttnk9gz5anzxn7)) - Note: [H3] Describing the outcome replaces specifying the method - Tags: atom/literature/h3 - [00:42:18] So don't do these like, but the thing is is we're adding more and more to them every day. Shopify. We just added to our linter no use of the CSS has selector. Here you go, everybody. You'll see us in this episode. Yeah, no has has is wildly slow and destructive in our application. And JavaScript can be like, huh? But it's so cool. It is cool. It's a super cool feature. And it's really sad to not use it. However, it's like, measurably impacting our site in so much that we've banned it. And now I submitted multiple PRs today that rip it out and replace it with JavaScript. And that's just that's it. And so yeah, and then we've got a linter that goes in there. Now there's no code. LLMs love to do set timeout. And now we've [00:43:08] got a linter that says no set timeout use transition end as much as you can. Oh my god. Actually, there's a bug that showed up today with transition. And can I side rant us for a second? Oh my goodness. Okay. So agents are using websites now and a lot. And they'll use them in the background, meaning the tab will be inactive. However, an agent is still interacting with it. Browser's two things to inactive tabs. One of which is they don't run animations because why would they? There no one there's no one there to see it. You know, so they don't quite pause them, but they don't run them, which means that if an agent has clicked a button that invokes a modal for, for example, that transitions in. ([View Highlight](https://read.readwise.io/read/01m3ydqvbqszb281kscavq2gmn)) - Note: [H4] A shipped ban needs measured cost - Tags: atom/literature/h3 - [00:44:00] I can never test that. You can never tell. And the agent never sees it because it never shows up. It clicked the button, but the transition didn't run because the browser did. So guess what, we're doing a Shopify right now because we added a ton of motion to the new redo, which we announced this week, the new. Anyway, it's what I've been working on for like months. Whole new admin looks beautiful. I think it looks beautiful. Anyway, however we added I added a share load of motion. And the motion is fucking with all of the agents because they don't run. And so all of a sudden people like, Hey, my agent can't interact with the site anymore. And we're like, what? What do you mean? You're like, well, it's in the, I have it as a background tab. We're like a background tab. So basically what I'm about to rant about here is it's like [00:44:44] not only are agents using your application, they're using them in the back. I have to create affordances of our website. When it's in the background, just so that agents can still get their agenda that dev or whatever it was, or it tells you, oh my goodness, dude. Oh my goodness. So yeah, exactly. Like you've got these little linters that go through and link your site to make. Hey, every CSS, this is the PSA everybody right now. If you have animations or transitions in your application, agents running those in the background, they won't run and they will get stuck because the transition end events don't fire. The transitions don't fight. Things are opacity zero if they're animating from that exit stage state. And so I have to go create this elaborate solution. It's so annoying. You're at a similar point where we used to have like a completely ([View Highlight](https://read.readwise.io/read/01m3ydqhjvx8zs1jvpamn5j8w9)) - Note: [H1] Hidden-tab agents hang on animated interfaces - Tags: atom/literature/h1