The reviews are in on Zuckerberg’s superintelligence pitch. They’re not kind. This week’s critics zeroed in on the exact gap you’d expect from Meta. Meanwhile, Claude added a few new tricks: it can now send your emails, manage your Drive files, and coordinate with other Claude sessions on its own.
Edition 026 · 21 August 2026
Zuckerberg wants superintelligence for everyone.
Mark Zuckerberg's case for spreading AI power to everyone met a cold reception this week, while Claude added new agentic features: sending emails, managing files, and coordinating with other Claude sessions on its own.
01
Topic of the Week
Zuckerberg wants superintelligence for everyone.
A few days ago, Mark Zuckerberg published a long personal letter called “The Future is for Everyone”. It lays out Meta’s thesis: give superintelligence to everyone, don’t keep it locked inside a handful of big companies. His argument flips the usual safety story. One tightly-controlled superintelligence can’t represent everyone’s values. So the real danger isn’t too many people having powerful AI. It’s too few. He backs this with real commitments: an independent board has to sign off before Meta releases new models, there’s a “private mode” where even Meta can’t see what your personal agent does, and governments get early access to test versions before the public release.
The reaction since has been cold, and it got colder through this week, not warmer. The criticism isn’t really about the vision. It’s about who’s presenting it. Meta’s own track record with user trust works against the pitch. Critics point out the letter never says what “free or affordable” actually costs, and never says which future models stay open versus closed. It leans on “your personal agent works for you, 24/7,” a line that turns dystopian fast if that agent ends up serving Meta’s ad business instead of you. TechCrunch put it bluntly mid-week: the essay itself is “exactly why people don’t like AI.”
Meta: The Future is for Everyone · TechCrunch: why people aren’t buying it
02
Fresh Papers
Worth naming a pattern first. Agent building went from prompt engineering (what you type) to context engineering (what the model sees) to harness engineering (what environment it runs in). This year the term for the next layer up is loop engineering: designing the cycle that makes an agent act, check its own work, and retry until it hits a goal, instead of you prompting it turn by turn. Addy Osmani and Peter Steinberger get credit for the term, back in June. It’s become the default way people talk about serious agent work now.
This week, the same idea showed up one level down, inside the model itself, not just around it.
Looped Language Models Improve Compositional Tool Calling. Normally a model runs through its layers once per answer. This one loops back through its own layers several times before answering, and it decides on its own how many loops each step actually needs. Tested on API-Bank, BFCL, and NESTful, the gain shows up almost entirely on multi-step tool use: chaining calls, tracking state, handling dependencies between steps. A single, one-shot API call barely improves. The practical read: if your agent struggles with juggling several tool calls in a row, not with making one call correctly, more “thinking loops” at inference time might be a cheaper fix than a bigger model.
03
New Models
A follow-up on something we mentioned before, plus one from OpenAI.
Qwen3.8-27B shipped August 14. Two editions back, we mentioned it as the smaller sibling coming to Qwen3.8-Max. Now it’s here. It’s a dense 27B model, natively multimodal, with a 262K token context window that expands to 1M. Open weights, Apache 2.0, on Hugging Face. Alibaba’s own numbers show real gains over the previous version: SWE-bench Pro went from 53.5 to 61.7, agentic terminal coding from 63.4 to 73.0.
Also worth a mention: ChatGPT’s desktop app can now remember your activity across other apps and websites on your Mac. It’s called Computer History. Opt-in only, with per-app controls and a pause button.
04
Claude Code & Coding AI
As of August 18, Claude can send emails through Gmail and manage files in Google Drive on its own. Ask it to reply to a thread, and it drafts and sends the response itself. It only pulls you in when it decides it needs your approval. Connect Gmail or Drive from the connectors menu. The unsupervised send works on all paid plans.
The same week, Claude Code got better at running itself, not just at running for you. Subagent forking is now on by default: a forked subagent inherits the full conversation and prompt cache instead of starting cold. And typing @ in a prompt now lets one Claude Code session message another one directly, by name. Small feature. But it points at where this is heading: teams of Claude sessions coordinating with each other, not one session doing everything alone.
Claude Academy launched August 20. Free courses at academy.claude.com, from AI-fluency basics to product-specific tracks for Claude Code and the rest. Built for everyone: total beginners, and teams rolling Claude out company-wide.
Smaller items: Anthropic is extending the 50% increase to weekly Claude Code limits through August 31. They say they’re hoping to make it permanent. Claude Code’s own tracked revenue hit $15.12B for the week ending August 10, that’s 21.9% of Anthropic’s total, from one product. That number landed the same week investors started pricing Anthropic’s upcoming IPO at $2 trillion, on a bet that full-year revenue reaches $100-120B, about 10x what the company reported back in May. Also shipped: a new “Concise” output style that skips the preamble and leads with results.
05
Tools of the Week
Slack launched “Slack Code” on August 20: new code channels where a team can direct several coding agents together, in real time, in the same channel. Claude Code and Claude Tag are on the list, alongside Devin, GitHub Copilot, ChatGPT, and Vercel’s agents. Same trend as the subagent-messaging feature in Claude Code above, one level up: coordinating agents is turning into a team sport, not a solo one.
Amazon Bedrock AgentCore payments went generally available on August 18. Two editions back we covered Cloudflare giving agents an actual wallet. This is the bigger, enterprise version of the same idea. Agents pay for APIs and content on their own, inside a “payment session” with a hard spending cap and an expiry time. AWS checks every payment against that cap before it signs anything, and the agent never touches the raw credentials. It supports two payment protocols, x402 and the new Machine Payment Protocol, plus Coinbase and Stripe stablecoin wallets built for microtransactions. Real use cases are live already: Travala runs conversational hotel booking across 2.2 million properties through it, and Anchor Browser uses it to unlock paywalled content for enterprise customers.
06
Fun Break
Researchers built FM-Bench. 15 language models, each one managing a football club for 20 years. They handled budgets, contract talks, youth development, and picking the lineup. Here’s the real finding: how much a model “thought” per move predicted nothing about who won. The model’s size, price, or maker didn’t matter either. What actually mattered was simple stuff: renewing contracts before the deadline, not letting cash sit unused, cutting slow investments as time ran out. And the humans who tried it came in last, behind every single model. Same lesson we keep seeing in enterprise AI surveys, just wearing shin guards this time: how you use the tool beats how big the tool is.
07
AI at Tenvalleys
Zuckerberg’s letter was the one thing everyone on the team had an opinion on this week. Part of it, we agree with. Concentrating superintelligence in two or three companies is scarier than spreading access around. That’s roughly the same instinct behind why we push clients to own their own AI systems, instead of getting locked into one vendor.
But the letter landed the same week as Rafał Brzoska’s very public fight with Meta. Deepfake ads kept using his face, and his wife’s, to push scams. One landed on his own kids as a screenshot. His Instagram account got suspended right after he called Meta’s Polish leadership out by name. Meta’s statement on it never used the word “sorry.” Brzoska’s reply: “No company will voluntarily give up a third of its profit. That’s why appeals don’t work. Only the law does.” That’s the exact tension we kept circling back to. Zuckerberg’s letter asks you to trust that your personal AI agent will work for you, not for Meta’s ad business. The same week, Meta couldn’t get a fake ad off its own platform without a court order and a very public fight from one of Poland’s most recognizable founders.
This week ended up being more of a reflection moment for us than a normal update. Most weeks we’re just moving with the pace of things, building with AI, shipping, not stopping to ask the harder questions. This week we did, at least for a bit. We don’t have neat answers. But it felt worth pausing for, instead of only running with the speed of everything changing.