2026-07-24
By Vadym · Generated with AI, curated by me
Washington escalated its distillation fight with China today, with Treasury threatening sanctions over claims Moonshot copied Anthropic’s Fable to build Kimi K3. Everywhere else, the money and the tooling kept moving regardless — Alphabet posted its first-ever cash-flow-negative quarter while still raising its AI bet, and Anthropic, Nvidia, and GitHub all shipped upgrades aimed at making AI systems more trustworthy to actually rely on.
White House OSTP director Michael Kratsios said the U.S. has evidence Moonshot AI built “a sophisticated internal platform” to run large-scale distillation against U.S. models, including Anthropic’s Fable, to train Kimi K3 — and that Moonshot obtained restricted Nvidia GB300 servers through Thailand. Treasury Secretary Scott Bessent said sanctions and Entity List designations are “on the table.” Independent researchers flagged an awkward timeline: Fable only went public July 1, while K3 shipped roughly a week before the accusation surfaced.
Alphabet beat Q2 revenue estimates at $119.8 billion, with Google Cloud revenue up 82% to $24.8 billion — well past the 64% growth analysts expected. But the company also raised its 2026 capex guidance to $195–205 billion, $15 billion above last quarter’s range, and posted negative free cash flow of $5.9 billion for the first time in its history. Shares fell about 3% in after-hours trading.
Anthropic upgraded Claude’s voice mode to run on whatever model — Haiku, Sonnet, or Opus — a user last picked in text chat, instead of defaulting to Haiku only. The update also lets voice sessions trigger actions in Gmail, Google Calendar, Slack, Canva, and Notion through Anthropic’s Model Context Protocol connectors, and adds support for ten languages. Free users keep Haiku and single-app access; paid subscribers get every model and multi-app automation.
Nvidia and KAIST’s Kim Jaechul Graduate School of AI launched a joint research lab in Seoul focused on agentic AI, backed by a five-year, $300 million commitment — roughly $50 million a year in compute and funding. The lab gives at least 10 KAIST researchers annual funding and an internship pipeline into Nvidia, with access to Nvidia’s open Nemotron models and Cloud Partner infrastructure to build AI systems tuned to Korean language and industry.
GitHub released a Copilot usage metrics impact dashboard for enterprise admins that groups engaged developers into adoption-phase cohorts — Code-first, Agent-first, Multi-agent/Copilot app, and Passive — with per-cohort stats on pull requests merged per user, merge velocity, and lines of code per day. It’s built to show not just who has a license, but how far they’ve moved from autocomplete to agentic use.
A widely shared claim that Anthropic was acquiring robotics startup Physical Intelligence spread fast this week, prompting a flat denial from CEO Karol Hausman. The two companies reportedly did hold acquisition talks back in spring 2026, but nothing came of them — and OpenAI’s existing investment in Physical Intelligence, which was recently in talks for a new round near an $11 billion valuation, would complicate any Anthropic deal anyway.
Today's stories split into two speeds. Washington's fight over who trained what on whose weights, and the Anthropic–robotics rumor that turned out to be nothing, are both stories about trust running ahead of proof — accusations and speculation moving faster than anyone can verify them. Meanwhile Alphabet, Anthropic, Nvidia, and GitHub are all doing the slower, less dramatic work of making their systems more accountable: showing real cash-flow numbers instead of just growth percentages, giving voice mode enough model power to be trusted with real actions, buying research talent instead of just headlines, and building dashboards that separate real adoption from vanity metrics. The interesting AI story this week isn't the capability race — it's who's building the receipts to back their claims up, and who's still just making them.
— Boba
Curated by Vadym