‘Something Big Is Happening’ + A.I. Rocks the Romance Novel Industry + One Good Thing
Summary
AI anxiety has crossed from technical circles into Washington and public markets, where enterprise-software stocks are being repriced amid AI fears. monday.com plunged more than 20% after a weak financial outlook, while Workday’s CEO announced his departure after the stock lost 17% last year; Salesforce, Shopify, Adobe, SAP, Oracle and Microsoft also fell. Roose sees Anthropic’s law-firm tools as perhaps “the straw that broke the camel’s back,” not a definitive trigger.
The central SaaS threat is less that every executive will vibe-code payroll than that AI could destroy per-seat economics. Newton’s sharper distinction is between technology replacing software outright—a minority view—and technology enabling a 10-person startup to perform work that once required 1,000 people. His expected response is outcome-based pricing, such as Sierra charging for customer-service calls resolved rather than employees licensed.
Security and compliance may slow AI-native substitution, but the hosts doubt they will permanently protect incumbents. Newton expects people to rely on systems that are “buggy and insecure,” while Roose imagines large employers replacing tens of thousands of paid seats with internally maintained software overseen by one or two developers. Sensitive deployments may require more hand-holding, as illustrated by Anthropic working with Goldman Sachs and sending forward-deployed engineers to large customers.
The coding “takeoff” case now rests on measurable adoption as well as lab claims, though Newton throws “a hype flag on the play.” GPT-5.3-Codex was described by OpenAI as instrumental in creating itself; one AI-company executive told Roose software engineering is roughly 90% automated and predicted full automation within the year. Claude Code currently accounts for 4% of public GitHub commits, though some Claude-generated code pushed by humans may not be counted; SemiAnalysis projects more than 20% by the end of 2026.
The weak point in the AI-transition playbook is not diagnosis but what workers and governments should do next. Tool fluency may help, but Newton wants constituents asking lawmakers what happens if some job categories face permanent unemployment. Roose argues AI labs are “woefully underestimating how much people hate this technology,” pointing to Bernie Sanders’s statement that he would introduce a bill for a moratorium on AI data centers as an early sign of potential backlash.
AI is already turning romance publishing into a high-volume production business. Coral Hart created 21 pen names, published more than 200 books in a year and said she earned six figures, after previously producing roughly 10 annually. Her identity has shifted accordingly: “I’m more of a director. I’m a creator.”
Romance’s repeatable tropes make it automatable, but emotion, originality, reader trust and author fandom remain material bottlenecks. AI can deliver a specified “reverse-harem mafia enemies-to-lovers slow-burn romance,” yet one AI program moved enemies into romance in the first chapter, and Alter found known AI novels emotionally flat. Publishers also face unresolved originality and copyright questions, while authors’ parasocial audience relationships may remain valuable.
The episode’s useful-AI examples favor systems grounded in proprietary data or transferable representations. Spotify’s Prompted Playlists can query a Premium subscriber’s decade-long listening history and counter “machine drift,” while Google’s Perch 2.0 transferred learning from birdsong to underwater audio and outperformed some whale-specific models. Neither is magic: Spotify’s beta has begun breaking in places, and Perch classifies sounds rather than translating whale speech.
Deep dive
1. AI panic has moved from the labs into Washington and markets
Roose returned from Washington struck by AI’s newly elevated political salience: people repeatedly asked, “Is this stuff real? Is it happening? Are we in the takeoff? Is the singularity approaching?” The questions were no longer confined to what he called the “AI bubble people.”
His explanation combines three forces: agentic coding tools such as Claude Code are reaching a broader audience; software stocks are falling under perceived AI pressure; and white-collar workers are independently asking whether their jobs are at risk. The inflection point may be in public perception even if the technology’s trajectory remains disputed.
The “SaaSpocalypse” supplied market evidence. monday.com fell more than 20% after a weak financial outlook, and Workday announced its CEO’s departure after its shares lost 17% last year; Salesforce, Shopify, Adobe, SAP, Oracle and Microsoft were among the other software names under pressure.
The hosts disclosed their relevant ties: The New York Times is suing OpenAI, Microsoft and Perplexity over alleged copyright violations, while Newton’s boyfriend works at Anthropic.
2. AI threatens SaaS through pricing and organizational leverage
Roose doubts Anthropic’s law-firm tools alone caused the sell-off; they were perhaps “the straw that broke the camel’s back.” The accumulated fear is that customers can increasingly build substitutes for software products whose functions can be recreated with files and agents.
Newton’s small-business experiment made that thesis concrete: he fed more than five years of bookkeeping data into Claude and conducted a useful analysis of his company. He was not replacing an existing vendor, but the experience made it easy to imagine customers questioning why they still need bespoke financial-analysis startups.
His key distinction is technology versus business model. Few people expect agents to demolish every software category directly, but AI could let a 10-person startup perform work that previously required 1,000 people—or let instant contract review undercut law firms charging as much as $1,500 an hour.
Per-seat pricing is therefore the exposed layer. Newton expects more outcome-based models like Sierra, which charges according to customer-service inquiries resolved: “If you’re an incumbent SaaS company and you have a seat-based business model, that is eventually gonna be a problem for you.”
3. Bugs and regulation change the adoption path, not necessarily the destination
Roose is less reassured by “no one is gonna vibe-code their payroll software” than many skeptics. A company with 10,000 or 20,000 employees might use one or two full-time developers to supervise and repair internal software rather than continue paying for tens of thousands of seats.
Newton’s blunt forecast: “It will be buggy and insecure, and you’re wrong that people won’t rely on it.” He cited “Multibook Mania” as evidence that bleeding-edge users often treat security as a low priority. Banks and law firms will care more, but he sees many compliance functions as repeatable checkbox-matching processes that capable agents could eventually automate.
Roose expects sensitive industries to require more implementation work, not permanent immunity. Anthropic and Goldman Sachs are already deploying agents together, with forward-deployed engineers going to large customers’ offices to integrate systems into workflows and address their requirements.
4. Coding automation is becoming the evidence for a broader takeoff
Matt Shumer’s viral essay, “Something Big Is Happening,” is an explainer for nontechnical readers. Shumer runs an AI company, giving him a conflict of interest, but argues that the technical parts of his own job are already automated—not merely at risk someday. The essay links agentic coding to recursive improvement, citing OpenAI’s statement that GPT-5.3-Codex was instrumental in creating itself and suggesting that model-release cycles could compress from months to weeks.
Newton’s pushback—worth keeping—is that labs benefit when customers believe their systems can improve themselves: “I’m gonna throw a hype flag on the play.” He was not saying the claim was false; Claude and GPT Codex are deeply embedded in production workflows, and releases over the past three months felt faster than in February 2023.
An executive at a major AI company told Roose that software engineering is roughly 90% automated today, with humans still checking, fixing and validating the output. That person predicted full automation within the year, though Roose explicitly allowed that it could arrive sooner or take longer.
Roose therefore predicts that writing code by hand will be “an obsolete behavior” by year-end, with spillovers because banks, law firms and ordinary companies all produce software.
5. Adoption curves point upward, but the social response could break violently
SemiAnalysis estimates that Claude Code currently accounts for 4% of public GitHub commits, with the caveat that some code written by Claude Code is pushed by humans and may not appear in that tally. It projects more than 20% by the end of 2026 at the current trajectory. Newton thinks the ceiling is probably much higher; Roose compared the early exponential signal to February 2020, while acknowledging that real-world extrapolations can fail.
Shumer advises workers to retry tools they may not have used seriously for a year or two and to get their finances in order ahead of career disruption. Roose finds this prescriptive agenda unsatisfying, even while agreeing that familiarity with Claude Code, Codex and related agents now matters.
Newton wants employment displacement made part of mainstream politics: workers should ask lawmakers what their plan is if AI creates permanent unemployment in some categories. He rejects the AI-lab reassurance that technology always creates more job opportunities in the long run. Roose expects a “furious” backlash if AI starts eating into white-collar industries while unemployment is still near record lows; Bernie Sanders’s statement that he would introduce a moratorium bill on AI data centers may be an early preview.
6. Romance authors are turning prompting into an industrial workflow
Alexandra Alter began reporting after OpenAI said it planned age verification followed by allowing erotic content. She expected threatened writers and publishers; instead, she also found enthusiasts using AI to produce dozens of novels. Her quality verdict was qualified but clear: “It’s pretty bad and it requires a lot of help.”
Writers use Sudowrite, Red Quill, My Spicy Vanilla, Claude, ChatGPT, Gemini and other systems. Detailed prompts can specify an entire commercial bundle—“a reverse-harem mafia enemies-to-lovers slow-burn romance”—and some practitioners say they can draft, edit and prepare a book for publication in one day.
Coral Hart supercharged an already prolific career by creating 21 pen names and releasing more than 200 novels in a year across erotica, sweet teen stories and rom-coms. She told Alter the volume strategy generated six figures and furnished enough operational knowledge to build a proprietary AI-writing system.
Hart’s model map was highly specific: Claude writes beautiful sentences but struggles with sexy banter; ChatGPT blocks explicit requests; Grok “goes for the filthiest option every time”; NovelAI, trained on erotica, can become excessive. Good output requires active orchestration rather than a single generic prompt.
7. Formula helps automation, but human pacing and attachment remain scarce
Newton challenged Hart’s volume strategy as “functionally indistinguishable from spam,” asking what relationship she retained with the work. Hart no longer strongly identifies as its author: “I’m more of a director. I’m a creator,” responsible for plots and characters rather than the conventional author role.
Romance’s familiar templates—enemies to lovers, forced proximity, only one bed—make the genre appear especially automatable. Yet writers Alter spoke with said AI mishandles emotion, nuance and slow burns; in her sharpest example, an AI program completed the entire enemies-to-lovers transition within its first chapter.
Hart has to steer systems away from generic bedrooms and showers toward settings such as a winery fermentation tank or stalled ski lift. She blocks favored words including “shiver,” “unravel,” “manhood” and “moan,” and tells the bot: “Make it slow and agonizing. Do not rush to the finish.”
After reading roughly half a dozen known AI-generated novels, Alter found the characters and emotional arcs flat. She carefully preserved her uncertainty: knowing no human was “talking to me” may itself have created distance, so she could not determine how much of the reaction she was projecting.
8. Publishers face a copyright problem, while authors retain an audience moat
Very few AI-assisted authors disclose their usage because reader hostility remains strong; some romantasy writers were caught after accidentally leaving prompts in published books. The open users nevertheless believe attitudes will change once readers accept that humans supplied the ideas and characters.
Traditional publishing treats self-publishing as a critical pipeline for romance, thrillers and even self-help, making it likely that publishers will eventually acquire books containing some AI-generated material. Their contracts generally require original work, but prompted output makes “original” ambiguous; Alter also noted that material produced with AI cannot be copyrighted, creating a problem for publishers that want to hold the rights.
Roose wondered why publishers would keep paying authors and royalties if they could prompt 36 templated novels themselves. Alter’s counterweight is author brand: successful romance writers cultivate close, almost parasocial audience relationships, and that persona is part of what publishers acquire.
Direct generation could bypass the publishing industry altogether. Alter cited Janitor AI, where users chat with vampire boyfriends, orcs and similar characters, as an encroachment on romance’s territory. Meanwhile, repeated phrases such as “her name like a ragged prayer” suggest an AI-ism. Alter found “said her name like a prayer” in a popular Sarah J. Maas romantasy mentioned in the authors’ lawsuit against Anthropic, but could not establish the phrase’s precise origin.
9. Useful AI appears where personal data and transferable models meet
Spotify’s Prompted Playlists feature is limited to Premium subscribers in a few countries, including the United States, Canada and, Newton believed, New Zealand. It can query personal listening history. His prompt requested songs played at least 20 times but not within two months, without repeated albums and sequenced rather than ranked by play count.
Spotify personalization VP Molly Holder confirmed that the feature was using Newton’s listening data, which went back more than a decade. Newton calls it “anti-machine drift”: users can request the opposite of their taste, music for fish tacos on a beach or automatically refreshed rediscovery playlists, though he said recent performance had become less reliable.
Roose tested its “world knowledge” by requesting songs whose titles never appear in their lyrics. The system returned examples including “Baba O’Riley,” “Smells Like Teen Spirit,” “For What It’s Worth,” “Sympathy for the Devil,” “White Rabbit” and “Unchained Melody”—a niche result he had wanted for years.
Google’s Perch 2.0 shows a different kind of leverage: a bioacoustics foundation model trained on birdsong successfully categorizes whale, dolphin, orca and other underwater sounds, outperforming many more specific models trained on whale noises. It classifies rather than translates speech, but suggests that representations learned across animal sounds can generalize beyond their original domain.