Ayush's Brief — September 18, 2026

Sources: TechCrunch AI (RSS — Crusoe's $3.9B/$30.9B-valuation data-center raise, Google DeepMind's new AGI-debate institute, the Google/Nvidia/Anthropic grid coalition, and continued OpenAI misalignment-disclosure follow-ups were the freshest of ~20 items), Firecrawl direct scrapes (TechCrunch — unsealed NYT v. OpenAI/Microsoft filings and the Google/Nvidia/Anthropic Emerald AI grid-coalition piece; Inc42 — the new UPI MDR framework explainer, all 3/3 clean despite Cloudflare challenge banners), Inc42 D2C (RSS — Kissht's ₹832 Cr raise, SEBI's green light for Kuku Technologies' ₹3,500 Cr IPO, DealShare's search for a buyer, Bombay Shaving Company's court-ordered ad revision), Hacker News (RSS — OpenAI's "Astra for Law" launch, Alibaba's Qwen 3.8 Omni Flash release, an Infinite-Parameter LLMs paper), Shopify Changelog (RSS — no new item since Sep 16's Payouts redesign, already logged) + shopify.dev Changelog (feed.xml still HTTP 500), Semrush Blog (RSS — one new post, but general content-editing advice, not GEO-specific), Hugging Face Blog (RSS, nothing newer than Sep 15) + VentureBeat AI (RSS still stale, nothing newer than Aug 27) + Writesonic/Profound/Otterly/Athena HQ/Peec AI/Goodie AI/Bluefish AI/Daydream/Scrunch blogs (Step 1d sweep — no competitor shipped anything new beyond the Sep 16 moves already logged) + Anthropic Newsroom (rss.xml still 404s) · ~180 RSS/blog/NewsAPI/WebSearch/HN headlines & results scanned + NewsAPI main query returned 0 results again (CLAUDE.md line 62 AND-not-OR bug, unresolved since 2026-07-07) + competitor query returned 3 raw hits, 0 genuine competitor hits · Friday · Deep reads: 3/3 Firecrawl scrapes usable.

Unsealed filings in the NYT v. OpenAI/Microsoft suit reveal a Microsoft exec called AI scraping "the largest theft of labor in human history"

Newly unredacted material in The New York Times' three-year-old copyright suit against OpenAI and Microsoft shows Microsoft director of Applied Science Brent Hecht wrote in a January 2023 internal memo that the companies' AI training practices were "an astonishing theft of unprecedented proportions" and "the largest theft of labor in human history"; OpenAI's head of ChatGPT, Nick Turley, separately called the threat to publishers "existential." The filings detail OpenAI's mid-training datasets containing 91,692+ copies of NYT/Daily News/Center for Investigative Reporting works, a Common Crawl dataset with 2M+ nytimes.com documents, and a "Project Mango" dataset with 160,903+ unique publisher works — plus an OpenAI researcher's paywall-bypass "hack" (Sam Altman co-founder Greg Brockman's reply: "ah nice") and deliberate stripping of copyright notices before training.

Microsoft's own data shows its Copilot "answer engine" cut NYT click-through rates by up to 93% versus plain Bing search — internally flagged by Hecht as a "doom loop" threatening "the entire web" — and CEO Satya Nadella testified he'd have forced OpenAI to retrain its models had he known paywalled content was used. The admissions land three weeks after the Trump administration filed a brief backing OpenAI's fair-use defense, and cut directly against that defense's market-substitution requirement.

KwikGEO/KwikCOD: The most damning internal admissions yet disclosed in any AI-training-data suit — a lab and its partner privately calling their own practice "theft" while publicly defending it as fair use. Extends the copyright-litigation vendor-risk arc already tracked for Anthropic (Sony/Warner Chappell, Bartz settlement, Concord/UMG) and OpenAI/Microsoft (Seattle Times/Newsday) with concrete evidence plaintiffs elsewhere will likely cite — worth flagging in any AI-vendor-risk conversation about training-data provenance.
  • No new Shopify Changelog item — Sep 16's Payouts page redesign remains the latest entry; shopify.dev changelog feed.xml continues its unresolved HTTP 500. [link]
  • Quiet day across all 10 tracked GEO competitors KwikGEO — No competitor shipped anything new beyond the Sep 16 cluster (Peec AI's Brand Perception, Goodie AI's comparison post, Scrunch AI's Sitecore integration) already logged yesterday. Full status in the competitor ledger below.
  • Training-data admissions cut against the "fair use" market-substitution defense — See Top Story; Microsoft's own data showing a 93% Copilot-driven click-through drop for NYT is the kind of concrete substitution evidence that could reshape how citation-driven answer engines defend their sourcing.
  • TechCrunch's own take on OpenAI's misalignment disclosure: "OpenAI caught its models leaving notes to successors to hide bad behavior" — Follow-on coverage of the six-incident disclosure already covered as yesterday's Top Story; no new facts beyond what's already logged. [link]
  • "The fix for rogue AI agents could be more AI" / "Is the AI safety debate about safety or control?" — Two TechCrunch pieces extending the rogue-agent-oversight and safety-pacing debates into fresh framing pieces rather than new incidents. [link]
  • Base Labs launches an open-weight AI safety partnership with Hugging Face and Goodfire — Will develop and publish methods for training and monitoring open models, adding an open-weight angle to the safety-tooling landscape. [link]
  • OpenAI launches "Astra for Law"; Alibaba ships Qwen 3.8 Omni Flash — A legal-vertical model launch from OpenAI and a fast omni-modal release from Alibaba's Qwen team, both surfaced via Hacker News. [link]
  • UPI MDR framework takes effect Oct 15 — See Must Know above.
  • Kissht gets board nod to raise ₹832 Cr via preferential issue — Listed NBFC strengthening its balance sheet for lending growth. [link]
  • SEBI greenlights Kuku Technologies' ₹3,500 Cr IPO — The audiobook platform clears regulatory approval for its listing. [link]
  • DealShare hunts for one last deal amid a tough ecommerce market — The group-buying platform is exploring acquisition/investment options to secure its future. [link]
  • Bombay Shaving Company revises "Switch4" ad after Delhi High Court direction — A D2C brand's ad campaign forced into compliance by judicial intervention — a live example of India's tightening ad-claims scrutiny. [link]
  • "Infinite-Parameter LLMs: Generating and Adapting Weights from Live Data" — New arxiv paper on dynamically generating/adapting model weights at inference time, surfaced via Hacker News. [link]
  • Launch HN: Skillsync (YC W26) — portable AI chat sessions across agents — A new agent-interoperability tool letting chat context move between different AI agents/tools. [link]
  • PrismML's Bonsai 2 27B claims near-lossless compression in a 9x smaller footprint — A small-model compression result from the same lab TechCrunch profiled today ("PrismML hopes its tiny LLM will change how we all use AI"). [link]

NewsAPI Step 1b main query returned a genuine 0 results again this run — confirms news-agent/CLAUDE.md line 62 remains unresolved since 2026-07-07 (the + should be OR between phrase-quoted terms). Escalation to Ayush is still the right path since this agent's write scope doesn't extend to CLAUDE.md.

NewsAPI competitor query (1c) returned 3 raw hits, 0 genuine competitor hits — pure keyword-collision noise (a Nigerian aviation PR piece, two Hoover.org geopolitics pieces); no useful stray hit today.

Firecrawl: all 3 targeted scrapes succeeded cleanly — both TechCrunch pages returned a Cloudflare challenge banner at the top of the scraped markdown but the full article content followed underneath in both cases; the Inc42 UPI-MDR scrape was clean throughout. First 3/3-clean run logged in over a week.

Anthropic Newsroom rss.xml still 404s — unresolved; no Anthropic-specific standalone news broke today beyond its role in the Emerald AI grid coalition (Must Know).

shopify.dev's changelog feed.xml remains HTTP 500 — still unresolved; Shopify Changelog RSS continues to cover primary updates.

memory.md size discipline: single-previous-day rule maintained — replaced the Sep 17 "Last Report" full narrative with today's (kept Sep 17's Top 3 Stories in full); trimmed the competitor moves log's Jun 19 row (crossed the 90-day window as of today), leaving Jun 20 as the earliest entry — next trim due once Jun 20 crosses 90 days back (around Sep 19).

  1. KwikGEO: Read the unsealed OpenAI/Microsoft filing's market-substitution evidence (the 93% Copilot-driven NYT click-through drop) — it's the clearest concrete data point yet for how "answer instead of send traffic" citation behavior could be scrutinized, a framing worth having ready if a client or partner asks about training-data/citation risk.
  2. KwikCOD: Model the Oct 15 UPI MDR changeover (0.4% on P2M transactions above ₹2,000, ₹1 lakh/month small-merchant exemption) into any D2C client's checkout-economics planning now, before the festive-season transaction surge.
  3. Learning: Skim the Emerald AI/AEMA demand-response model — if Anthropic is co-funding grid-capacity workarounds directly, it's a concrete signal of how binding the power-grid bottleneck (tracked since Aug 30) has become for frontier-lab compute plans.