PostCards
Google Finally Ships Gemini 3.6 Flash — And Quietly Starts Building Gemini 4
Google shipped three new Gemini models on July 21 — 3.6 Flash, 3.5 Flash-Lite, and a restricted 3.5 Flash Cyber for governments and trusted partners — while Gemini 3.5 Pro, the flagship, remained conspicuously absent for the third missed deadline running.
Gemini 3.6 Flash prices in at $1.50/$7.50 per million tokens with 17% fewer output tokens than its predecessor and a knowledge cutoff pushed to March 2026. It ships with Computer Use built in and an Artificial Analysis Index score of 50. Harvey and Hebbia are named as early adopters, and Google's CodeMender security tool entered preview with Salesforce, Robinhood, and Palo Alto Networks. Flash-Lite undercuts everything at $0.30/$2.50 per million tokens — the cheapest tier Google has ever shipped.
The bigger tell came buried in the same announcement: Google confirmed it has begun "our most ambitious pre-training run yet" for Gemini 4 — suggesting the Pro delay isn't a bug being fixed, but a generation being skipped.
Nadella Calls Anthropic's Fable 5 "Editorially Controlled" — Despite Microsoft's $5B Stake
Satya Nadella told Copilot engineers in an internal meeting on July 16 that Anthropic's Fable 5 model places unreasonable limits on what users can ask it. "When was the last time you had a creation tool that was so editorially controlled? It doesn't make sense," he said — remarkable given the team he was addressing builds Copilot on top of Claude models.
Nadella went further, questioning industry concentration itself: "It can't be that there are only two companies in the world with token capital, and everybody else is renting it." Microsoft has invested $5B in Anthropic, and Anthropic has committed $30B to Azure — making this an unusually public airing of friction between deeply entangled partners. Anthropic declined to comment.
Iran's IRGC Claims Cruise Missile Strike Destroyed an AWS Data Center in Bahrain
Iran's Islamic Revolutionary Guard Corps said Tuesday its Aerospace Force targeted and "destroyed" Amazon's central data infrastructure in Bahrain using several cruise missiles, framing it as retaliation for a U.S. strike on the Darkhovin nuclear facility two days earlier. The IRGC packaged the claim alongside hits on U.S. radar and Patriot air-defense positions at Muharraq and Riffa — treating Amazon's commercial cloud as operationally equivalent to military targets.
Neither Amazon, Bahraini officials, nor U.S. CENTCOM have confirmed damage, and Bahrain has said it intercepted the attack. AWS's ME-South-1 region has shown as "Disrupted" on its health dashboard since a wave of drone strikes in March — this would mark at least the third kinetic action against the same facility, and the first reported use of cruise missiles against commercial cloud infrastructure in the conflict.
South Korea Plans a Free National AI Chatbot to Cut Reliance on ChatGPT and Claude
South Korea announced plans to launch a free, domestically-built AI chatbot service this year, aiming to reduce the country's dependence on foreign platforms. The move follows Current AI's $400M public infrastructure fund (July 13, led by a $100M French commitment) and the EU Parliament's EPGenAI Hub, which promised MEPs access to OpenAI, Anthropic, Meta, and Mistral models by September 2026.
The pattern across all three: governments are increasingly unwilling to treat frontier AI as a rented utility, and are instead building or subsidizing sovereign alternatives — a trend that sits uneasily next to Nadella's own complaint, above, about a token economy controlled by just two companies.
Meta Says Its AI Moderator Beats Humans on Accuracy — It Also Deleted a Business With 1M Followers
Meta says its newer AI moderation models make 13% fewer errors than human reviewers and catch 10% more genuine policy violations. But The New York Times documented Instagram and Facebook users — including a business with nearly a million followers teaching English online — whose accounts were deleted over false fraud and child-exploitation claims, with appeals handled by the same AI system that issued the original ban, leaving no path to a human reviewer.
Some accounts were restored only after journalists intervened. Meta says the specific cases examined were banned by older moderation systems, not its newest AI tools — a distinction that offers little comfort to a small business owner who described losing the account as watching "somebody just shut down your storefront."