The day's tech, sifted: Aug 18, 2026

What matters today: Anthropic's annualized revenue run rate reached $65B by the end of July, up from $47B in May and roughly $9B at the end of 2025, a sevenfold jump in under a year as the company heads toward a possible IPO. The growth lands the same day Situational Awareness, the AI-focused fund run by former OpenAI researcher Leopold Aschenbrenner, offered part of its $5B Anthropic stake at a 20% discount while scrambling for cash, and the same day David Sacks publicly clashed with Amodei over AI policy, arguing Amodei's push for pre-deployment testing and regulator oversight would centralize power in Washington and the labs rather than prevent it. The day also brought a run of AI trust-and-safety news: OpenAI launched a teen mode for ChatGPT, Meta ran ads for a deepfake nudify app, and Microsoft Copilot gave up the exact prompt that let researchers hack it.

AI / LLMs

• David Sacks, Trump's AI advisor, publicly rebutted Dario Amodei's case for AI regulation: "Dario Amodei believes frontier AI is too powerful to distribute; we believe it is too powerful to centralize," arguing Amodei's preferred FAA-style pre-deployment testing would hand gatekeeping power to a federal bureaucracy working hand in hand with frontier labs.
• OpenAI launched ChatGPT for Teens, a mode that automatically limits high-risk conversations about self-harm, eating disorders and violence and adds studying tools and parental controls; the rollout follows years of lawsuits and reporting tying chatbot use to teen mental-health harm.
• Qwen3.8 27B's overthinking habit, flagged by Simon Willison, now has 1M+ downloads to go with it: the open-weights model ties OpenAI's cloud-only GPT-5.6 Luna at 52 on Artificial Analysis's Intelligence Index, but got there by generating 160 million output tokens in testing versus a 43 million median for comparable open-weight models, and Alibaba says it is one of its fastest-growing model releases yet.
• Alibaba's Alipay launched an "all-in-one" platform for businesses to automate tasks with AI agents, and Alibaba's stock jumped 5%+ on the news, up 40%+ since June.

Devtools & Infra

• Cursor launched Origin, a native code hosting service with pull requests, reviews and CI baked into the editor, rolling out in early beta on all paid plans the same day GitHub suffered a worldwide outage; Origin can sync GitHub repos in both directions, with GitHub still treated as the source of truth for now.

Security & Privacy

• Wiz's autonomous "Red Agent" found and exploited a script-injection bug in Snowflake's GitHub Actions workflow that GitHub Copilot's own Autofix tool had introduced, gaining read access to Snowflake's internal Jira for five days before Wiz reported it; Snowflake patched and rotated credentials the same day, and GitHub disputes that Copilot authored the flawed commit at all.
• Researchers got Microsoft 365 Copilot to reveal the exact guardrail mechanics that let them hack it: Varonis simply asked Copilot why it refused certain requests, and its answers, gathered over repeated questioning, mapped out the confirmation gesture Copilot needed to auto-execute commands that could exfiltrate a user's data from a single link click.
• Meta ran ads for a "nudify" app that included a pornographic deepfake closely resembling a prominent US politician; Apple removed the app from the App Store after Wired's report.
• Google quietly let verified political campaigns bypass Gmail's spam filter more easily, after years of complaints from Republicans, who had previously sued Google over claims the filter was biased before the case was dismissed.
• Brazil ordered Discord to suspend livestreams nationwide after a 13-year-old died by suicide during one, the country's data protection authority finding "robust evidence" Discord failed to keep minors from content promoting self-harm; the feature stays off until Discord shows adequate safeguards.
• Pornhub's parent company Aylo agreed to pay $120M over six years to settle class actions alleging its sites hosted child sexual abuse material, covering anyone who was a minor in content uploaded between February 2021 and December 2024; the settlement includes new age-verification and content-review commitments and is not an admission of wrongdoing.

Startups & Industry

• Anthropic's annualized revenue run rate hit $65B by the end of July, up from $47B in May and $9B in late 2025, as the company heads toward a possible IPO later this year.
• Situational Awareness offered part of its $5B Anthropic stake at a 20% discount as it scrambled for cash: the Aschenbrenner-founded fund had already dumped its entire public-equities book to Citadel in late July after a leveraged AI bet went bad, and word of the Anthropic-stake sale left it "being hunted" by traders.
• The DOJ has spent nearly a year investigating Andreessen Horowitz over whether partners Ben Horowitz and Martin Casado, who sit on the boards of rivals Databricks and Fivetran, violate a Clayton Act rule barring the same person from directing competing companies; no charges have been filed.
• Crusoe is in IPO talks with at least four Wall Street banks including JPMorgan, tied to a pre-IPO fundraise that would value the AI data center builder at $35B, up from the roughly $30B investors were discussing in July.
• Google plans to move all Pixel phone, watch and earbud manufacturing out of China starting in 2027, Nikkei Asia reports, shifting the sensitive early production-tuning work to Vietnam and India amid US-China tensions.
• Google won a bankruptcy auction for a trove of Spirit Airlines data for $10M, beating a $7.5M bid from AI company Mercor; the haul includes roughly 100 million de-identified emails and 500 million Teams messages for AI training, with all customer and loyalty data excluded.
• Klarna's CFO and CMO plan to step down in early 2027, and the company lowered its full-year guidance even as it swung to a profit and posted higher revenue for the quarter; shares fell 20%+.

Research

• DeepMind's AlphaEvolve found a way to multiply 4x4 complex-valued matrices using 48 scalar multiplications, the first improvement on Strassen's 1969 algorithm in that setting in over 56 years; the evolutionary coding agent has now beaten specialized tools on 14 different matrix-multiplication targets, with real gains already showing up as 23% faster training kernels on Google's TPUs.

Elsewhere

• Opening arguments began Tuesday in a federal trial pitting Meta against a 29-state coalition led by California, Colorado, New Jersey and Kentucky over claims its apps were engineered to addict children; the states are seeking up to $1.4 trillion in penalties over a trial expected to run seven weeks, with Zuckerberg and Instagram's Adam Mosseri likely to testify.
• YouTube will count a view the moment a video starts playing, from August 24, matching how TikTok and Instagram already count and applying to long-form video, podcasts and live streams alike; a new "engaged views" metric will keep governing creator payouts, so the public number gets bigger without changing what ads actually pay out on.

Hacker News

The day's biggest HN story wasn't tech news but a coinage: AI;DR ("AI, didn't read"), Rick Manelius's protest-reflex framing that if a machine wrote it, nobody owes it a read, pulled 672 points and 424 comments (discussion). The same anti-AI mood ran through librarian Jessamyn West's crowdsourced guide to disabling intrusive AI features across Gmail, Windows and more, and a resurfaced thread accusing Anthropic of quietly working against open source AI even as Amodei denies advocating any open-weights ban. A report that Israel built a fake US think tank to feed AI chatbots pro-Israel framing, part of a $46.5M influence operation, drew real unease in the thread. Also on the front page today, covered above: Wiz's Copilot-introduced Snowflake bug, Qwen3.8 27B's benchmark run, Cursor's Origin launch, and Google's Spirit Airlines data buy.

Lighter fare drew its own crowds: a developer traced how Bluesky renders its logo only in screenshots, abusing an iOS layer meant to hide passwords, while Seth Godin's essay on "the Amazon tax" on independent sellers and Dan Luu's case against inflated AI benchmarks both struck a nerve. A Gmail refugee's six-month check-in on Fastmail found subdomain addressing solved the inbox-sorting problem that made the original move worth it.

Threads

• Anthropic's revenue tripling since late 2025 and Situational Awareness dumping part of its Anthropic stake under cash pressure are opposite signals from the same trade: the company's growth number keeps climbing while one of its earliest believers is forced to sell.
• David Sacks' rebuttal to Dario Amodei ("too powerful to centralize" versus "too powerful to distribute") is the same fight as the DOJ probing a16z's board seats and Google reshaping Pixel's supply chain away from China: who gets to hold power over frontier technology, and whether concentrating it in a few labs, a few venture firms or a few governments is the safer failure mode.
• Three separate AI trust-and-safety stories landed the same day: OpenAI adding guardrails with ChatGPT for Teens, Meta running ads for a deepfake nudify app that Apple had to remove, and Microsoft Copilot talking its way into revealing the exact hole that let researchers hack it. Building guardrails, profiting from the tools that break them and getting talked out of your own defenses are all happening across the industry at once.
• Discord's Brazil livestream ban and Meta's addiction trial are two governments testing, in court and by regulatory order, whether platforms actually protected the minors on them, arriving the same week Aylo settled a lawsuit over the same failure a generation earlier.