When I laid out this week's news, it all pointed at one thing: control
Not which model scored a few points higher, but the layer underneath — who holds access to your AI, who gets your data, and whether someone can switch it off with a single directive. Fable 5 cut off from foreign users, Claude asking for your ID, Switzerland shipping an open sovereign model, Europe pushing a digital-sovereignty package — all of it circles the same anxiety: once AI becomes infrastructure, "whose is it" matters as much as "how good is it"
I run a dozen-odd projects and eight Cloud Run services myself, so this week hit the tools I touch every day. That's why this one runs longer than usual
AI Models & Products
The US government ordered Anthropic to suspend foreign access to Fable 5 / Mythos 5 — the biggest story this week. The problem isn't the policy itself, it's how it played out: cloud infrastructure can't tell users apart by nationality, so two flagship models effectively went dark worldwide. For non-US companies running the latest models, that's an overnight cutoff
This turns an abstract risk concrete: your AI vendor is subject to the government of the country it sits in. If you've wired your core flows to a single vendor's newest model, that risk is on your books, not someone else's
Claude starts mandatory identity verification for Free / Pro / Max users on July 8 (government ID + selfie, handled by a third party). The good news: API / Business / Enterprise plans are exempt. If you're mostly on the API like me, this doesn't touch you — but if you run your work off a Pro subscription, think it through now
GPT-5.5 hallucinates 3x more than the MIT-licensed open GLM-5.2 (a top Hacker News thread this week). The contrast is telling: "closed is more accurate" doesn't hold — the highest-scoring model can be the one that makes things up exactly when you need it to be honest. My own rule is that any critical output needs a verifiable source, never the model's own say-so
Switzerland's Apertus — an open foundation model for sovereign AI went live. EPFL / ETH lineage, built around "data, weights, and training pipeline fully open and controllable." Read alongside the three items above, it's no coincidence: when a vendor's control becomes a risk, "can I own the whole stack myself" moves from ideal to a line item in the procurement meeting
Meta had a busy week: Muse Spark (a push toward "personal superintelligence"), four MTIA in-house chips in two years, SAM 3.1 for real-time video segmentation. But The Batch also flags that Meta is pivoting away from open weights — one more once-open player starting to close up. The open-weights camp keeps getting smaller
AI Dev Tools & Agents
Agents are starting to build agents. The Batch's headline this week was literally "Agents Building Agents," and Cursor shipped Composer 2.5. The direction is clear: people who write code are moving up a level — from "write it myself" to "direct a fleet of agents and review the output." That's exactly how I run client work: judgment and review stay with the human, repetitive labor goes to the tools
Google Antigravity 2.0 pitches "100X engineering with AI agents," shifting from a traditional IDE to an agent-first platform, using a "skills" concept to push agent performance higher. If you're still at the "let AI autocomplete line by line" stage, this is the reminder: the tooling already jumped to "hand off the whole task"
Deploy an MCP server to GKE in 30 minutes (a Google Cloud tutorial). MCP (Model Context Protocol — Anthropic's standard for connecting LLMs to external tools and data) is moving from "experiment" to "infrastructure." If you're building agents, learn this protocol now, or you'll be wiring up a one-off integration for every tool later
Ray Serve LLM on GKE, Qwen3.7-Max climbing to third place — the open / self-hosted line accelerated all week. Paired with the sovereignty theme, the signal is consistent: more and more people are seriously weighing "can I use open + self-hosted to keep the core in my own hands"
Expert Takes
"Socrates as a Service" — turning Socratic dialogue into a product. Instead of handing you the answer, the AI asks back and guides you to think it through. I think this is the most underrated direction in education AI: everyone's building "do it for you," but the hard part is "help you think." An AI that asks good questions is far scarcer than one that gives answers
A run of recap pieces this week, on two threads: one, the open-weights camp is shrinking (Meta's pivot, countries evaluating their own models); two, model evaluation has to move beyond SWE-bench. The takeaway: don't just read the benchmark — watch whether it fails at the critical points in your real workflow
a16z published "New Media, One Year In" and opened a "Talent Engineer Fellowship." A VC firm doing its own content and its own talent pipeline is itself a form of control: rather than wait for the market to signal, build the channel and grow the people yourself
VC & Markets
Europe is taking the "open digital sovereignty" route (Google Cloud on the EU's Tech Sovereignty Package). When even the hyperscalers use "sovereignty" as a selling point, it's no longer a fringe topic — it's a core theme of enterprise procurement for the next few years. For anyone doing healthcare, government, or education work in Taiwan, "data stays in-country, architecture stays controllable" will show up in requirements more and more
China's push for green power in AI projects is hitting hurdles (Reuters), and data-center pushback is intensifying (The Batch). AI's compute expansion is colliding with the reality of power grids and local communities — this line will increasingly shape which region you deploy to and how you cost it out
a16z invested in Telepatia and Convey this week. Both bets lean toward "plug AI into real industry workflows," not another chatbot. The money is saying the same thing: general-purpose chat is a red ocean; vertical, on-the-ground use is where the room is
Action Items
-
Don't wire your core to a single model vendor. The Fable 5 export control proved single-vendor risk is real. If your product leans heavily on one vendor's newest model, design a fallback now — drop to the previous generation, or wire in a backup model API. Leave yourself a way out
-
Treat "sovereignty / localization" as a requirement, not a nice-to-have. If you work in healthcare, government, or finance, clients will eventually ask "where's the data, who can see it, can we own it." Have the answer ready early — it's what separates you from a generic consultant
-
Open models for the everyday, flagship models for the hard stuff. GLM, Qwen and friends are good enough now to handle 80% of daily tasks and crush your costs, leaving the flagship models for the architecture-level decisions that actually need them. You can build this tiered routing today
-
Learn MCP now. It's becoming the standard for connecting agents to tools. Standardize your tools on MCP a step early, and swapping models or adding tools later gets much cheaper
Sources
RSS Digest: see research/digests/2026-W26.md (280 articles this week, 8 curated Top Stories)
Main sources: Anthropic, Hacker News, Meta AI, The Batch (Andrew Ng), Google Cloud Blog, Every (Dan Shipper), a16z, Reuters