X is the best way I’ve found to keep up with AI. I like tweets throughout the week, filtering for things I think are actually worth knowing, then use Claude Code to pull those likes automatically and help me turn them into this post (here’s how the pipeline works). This week: 204 tweets liked, filtered down to what’s below.

Check out the previous roundup (June 12) if you missed it. Last week Fable 5 launched and got pulled offline in four days. This week the shutdown turned into a standoff with the White House, and the open-weight world moved in to fill the gap.

AI for Everyone

The Fable 5 shutdown stopped being a model story and became a political one. The best open-weight model ever released showed up the same week to fill the hole, Midjourney announced it’s building a medical scanner, and Tim Ferriss put a hard number on what AI is doing to books.

The Fable 5 Shutdown Becomes a White House Standoff (10+ mentions)

Last week’s export-control order has turned into a slow-motion fight between Anthropic and the Trump administration. The trigger, per Politico’s reconstruction: Amazon CEO Andy Jassy surfaced a claimed jailbreak to the White House, Dario Amodei was at a wellness retreat when officials tried to reach him, the export-control letter landed at 5:21pm on a Friday, and Fable 5 and Mythos 5 went dark for everyone on the planet, including Anthropic’s own foreign-national staff. Anthropic disputes the severity, saying the technique “is used every day by defenders who keep systems safe” and that GPT-5.5 has comparable capabilities. As of this week it’s in direct meetings with Commerce, the CIA, and the White House science advisor. The technical question may resolve fast; the relationship is the longer problem. Frontier model access is now a geopolitical variable, so build fallbacks. (source: @AnthropicAI, @SophiaCai99, @DavidSacks)

GLM-5.2 Takes the Open-Weight Crown, Right on Cue (8+ mentions)

The same week Fable 5 went offline, Z.ai shipped GLM-5.2 and the release post called out “frontier models being abruptly cut off for non-technical reasons.” It’s MIT-licensed (commercial use, no strings), 744B parameters with 40B active, a real 1M context window, priced identically to GLM-5.1 at $1.4/$4.4 per million tokens. It jumped to #1 on the Artificial Analysis Intelligence Index at a score of 51, ahead of every other open model and in the same range as GPT-5.5, and it’s the first open-weight model to clear 30% on ProofBench. It also took #1 on Design Arena, Vals Finance, and Vals Legal. Best-in-class on coding, design, finance, and legal at once isn’t an accident; it’s a signal about where open-source AI is heading. Run it now: ollama run glm-5.2:cloud. (source: @Zai_org, @ArtificialAnlys, @jietang)

Midjourney Is Building a Full-Body Ultrasound Scanner (4 mentions)

The most unexpected announcement of the week came from Midjourney, the AI image company. Its new Midjourney Medical division is building a scanner where you stand in a pool of warm water, 500,000 micro-sensors map you from every angle, and 60 seconds later you have a sub-millimeter 3D model of your insides. No radiation, no MRI tube, and they claim 100x faster and 10x cheaper than an MRI. The framing is the tell: it’s not a hospital machine, it’s a spa where the full-body scan is the attraction, with the first location in San Francisco in 2027. Nine people built the prototype. That detail is the one that sticks, because a team of nine at an image-generation company demoing working medical hardware recalibrates what you think small teams can do. (source: @midjourney, @0xkydo)

Perplexity Brain Gives Its Computer a Memory (3 mentions)

Perplexity launched Brain for its Computer product, a continuously learning memory that builds a context graph across all your sessions, connected files, and tools, then updates it overnight. Every Computer task feeds from and back into that graph, so the next session starts with a warm understanding of your work instead of from zero. That’s the most annoying limitation of agentic tools addressed directly. It’s a research preview for Perplexity Max subscribers right now. Whether the persistent memory turns out genuinely useful or just noise is impossible to call without running it on your own work for a few days. (source: @perplexity_ai, @AravSrinivas)

Tim Ferriss Says His Book Sales Are Down 77% (2 mentions)

Tim Ferriss published his own book-revenue data showing a 77% drop over two years, and he attributes it directly to AI. His team uses Claude daily, so he knows from the inside what replacing “buy a book to answer a question” with “ask an LLM” looks like. His catalog is self-help, productivity, life optimization, exactly the category where an LLM is now a fine substitute. What makes this land is the specificity. Everyone has a vague sense that AI is changing how people get information; Ferriss has the actual sales curve. If you publish anything in the reference or how-to space, this is the data point to sit with. (source: @tferriss)

AI for Developers

Vercel made a serious play for the agent-framework standard, Claude Code and Hermes both shipped meaningful workflow upgrades, two more open coding models landed, and Ramp put out a benchmark worth more than most.

Vercel Eve: an Agent Framework Built Like Next.js (1 mention)

Every team building their second or third agent hits the same wall, rebuilding the same plumbing each time. Eve is Vercel’s answer: an open-source TypeScript framework where an agent is a directory. Tools are files, skills are markdown, channels are adapters, schedules are cron handlers, the same convention-over-configuration bet that made Next.js win. Production features ship with it: durable execution so sessions survive crashes and redeploys, sandboxed compute, human-in-the-loop approvals that pause indefinitely without burning compute, evals, and OpenTelemetry tracing. The credibility marker is internal usage. Vercel runs 100+ agents on Eve, including one that handles 30,000 employee Slack questions a month and an autonomous SDR that returns 32x its $5K/year operating cost. Scaffold one: npx eve@latest init my-agent. (source: @vercel)

Claude Code Ships Artifacts (5 mentions)

Anthropic added Artifacts to Claude Code for Team and Enterprise plans. An Artifact is a shareable web page built from your session, with your codebase, connected tools, plugins, and conversation all in scope, and it refreshes automatically as the session keeps working. An incident investigation becomes a timeline you share before standup that updates itself as you keep digging; a PR becomes a walkthrough a teammate can follow with no session context. Sharing is org-internal only, no public links. Boris Cherny on the Claude team calls it “a game changer for how I work with Claude.” This closes part of the gap between an agent generating work and a human being able to see and trust it. Ask for one in your next session. (source: @claudeai, @bcherny)

Hermes Agent v0.17.0 Adds Async Subagents (8 mentions)

If you use Hermes Agent, run hermes update. The v0.17.0 “Reach Release” adds async subagents, which changes how the tool feels: instead of waiting for a delegated task to finish before your session continues, you kick off work and keep going. That’s the headline, but the list keeps going, with WhatsApp Business Cloud integration that skips the burner phone, Telegram rich formatting, agent distributions to share an agent with your team, live artifacts in the desktop app, and hermesbench for evaluating tool-calling. The shipping pace at Nous Research has been unusual; one developer clocked update notifications “literally every 30 minutes.” This is a project compounding fast. (source: @NousResearch, @Teknium)

Kimi K2.7 Code Lands With Big Jumps and Fewer Tokens (4 mentions)

Moonshot AI open-sourced Kimi-K2.7-Code: +21.8% over K2.6 on its own code benchmark, +31.5% on MLS Bench Lite, while using 30% fewer reasoning tokens. That last part is what matters at scale, more capable and cheaper per task at the same time. It ranks #2 on the ErdosBench hard-math benchmark, behind only the now-offline Fable 5. Unsloth already has a quantized GGUF running at 40 tok/s on 330GB setups, and Moonshot opened a beta at kimi.com/code/beta. Open coding models keep closing the gap fast, and K2.7 is this week’s proof point. (source: @Kimi_Moonshot, @UnslothAI)

OpenRouter Fusion: Frontier Accuracy at Half the Price (3 mentions)

OpenRouter launched Fusion, a compound API that runs a query across a panel of models and synthesizes their outputs. On 100 hard research tasks it claims Fable-level accuracy at half the cost. The interesting finding is the breakdown: 75% of the lift comes from synthesis, how you combine outputs, and only 25% from model diversity. That means a panel of budget models can beat an individual frontier model, and the recombination logic matters more than the model picks. For research workflows where latency is acceptable, this changes the cost math. (source: @OpenRouter)

LiteParse v2.1 Is the Fastest PDF-to-Markdown (2 mentions)

LlamaIndex shipped LiteParse v2.1, adding markdown output to an already-fastest PDF parser: 3.16ms per page with zero LLM calls, versus 141ms for pymupdf4llm, and it beats every model-free competitor on all three standard benchmarks. The license detail matters, since pymupdf4llm (the usual alternative) is AGPL, a problem for commercial codebases, while LiteParse is Apache 2.0. It runs in Python, Node, Rust, and WASM, so client-side parsing with no server round-trip is on the table. If your pipeline has a PDF step, this is a drop-in upgrade: pip install liteparse && lit parse doc.pdf --format markdown. (source: @jerryjliu0, @llama_index)

Ramp’s Private SWE Benchmark Is the One to Trust (2 mentions)

Public coding benchmarks have a structural problem: they become training data. Ramp built a private benchmark from real engineering problems its own team has faced and kept it out of public training sets. The result is more honest than most benchmark posts. Instead of crowning a single winner, Ramp found “the frontier presents as a tradeoff rather than a single winner” once you measure effectiveness against cost. That’s the more useful framing for anyone actually choosing a model: the frontier is a curve, not a point. (source: @RampLabs)

Honorable Mentions

  • Noam Shazeer is joining OpenAI, the co-inventor of the Transformer moving to the company building on it. This doesn’t produce a feature next week; it shows up in capabilities 18 months out and in where foundational researchers choose to work. (source: @NoamShazeer)
  • GPT-5.6 Pro is very likely landing this week (Polymarket at 83%). Testers describe better vision and SVG generation and a December 2025 knowledge cutoff, but say frontend code still trails Claude and it’s back to taking 20-40 minutes on hard tasks. (source: @testingcatalog)
  • SpaceX acquired Cursor in an all-stock deal, and SpaceXAI has reportedly been jointly training a model with the team for months that will ship in Cursor and Grok Build. Cursor’s existing Anthropic and OpenAI partnerships get complicated fast. (source: @SpaceX)
  • SemiAnalysis opened its own chip teardown lab (STEEL, in Oregon, tens of millions in capex), and its first report tears down Huawei’s Kirin 9030: SMIC’s N+3 node hits a 32.5nm metal pitch, tighter than Intel 18A’s 36nm, but only reaches TSMC N6-class density through DUV multipatterning, trailing 18A by about 38% on transistor density. It makes SemiAnalysis a direct rival to TechInsights, and process-node competitiveness is what gates AI-chip supply. (source: @dylan522p, @SemiAnalysis_)
  • Coinbase registered an AI agent with the SEC as an investment advisor. Coinbase Advisor sees your full portfolio, answers in plain English, and does automated tax-loss harvesting, rolling out to Coinbase One members. “SEC-registered” means it meets disclosure rules, not that the advice is guaranteed good. (source: @coinbase)
  • DeepMind published a 60-page AGI-to-ASI paper mapping four paths to superintelligence and six blockers. The least-discussed blocker is the “abstraction barrier”: a model trained only on pre-Newtonian physics may never reason its way to relativity. (source: @kimmonismus)
  • In The Weights is a viral quiz asking whether you’re famous enough to have ended up in a training set. It got traction fast enough that Marc Andreessen was posting about it. Two minutes, and a pointed question under the joke. (source: @wjosephflynn)
  • Google Earth added a flight simulator to the web, free and no install, after years of being desktop-only. The fun part is the sim; the pattern is Google steadily moving pro desktop features to the browser. (source: @googleearth)
  • A Claude Code plugin gates your prompts behind pushups, counted live by webcam, with abandoned sets becoming debt. It’s a joke that works, and a reminder that behavioral gates wired into AI workflows are an underexplored space. (source: @Botchet)

Try This Weekend

For everyone:

  1. Run ollama run glm-5.2:cloud and throw something you’d normally give Claude Opus or GPT-5.5 at GLM-5.2. It’s the best open-weight model available and free to run on Ollama’s cloud.
  2. Take the In The Weights quiz. Two minutes, and more thought-provoking than it has any right to be.
  3. Fly the Google Earth flight sim over somewhere you know. Free, browser-only, no install.
  4. Enable Perplexity Brain (Max subscribers), run Computer on a research task, then come back the next day and see whether it retained context.

For developers:

  1. Scaffold an agent with npx eve@latest init my-agent and see how much production plumbing Eve gives you for free: durable execution, sandboxed compute, evals.
  2. Drop LiteParse into a PDF pipeline: pip install liteparse && lit parse doc.pdf --format markdown. Zero LLM calls, 3ms a page, Apache 2.0.
  3. Ask Claude Code for an Artifact (Team/Enterprise): “Make an artifact walking through this PR, the diff, the reasoning, and what I tested.” Share the link with a teammate.
  4. Run hermes update and use Hermes async subagents to fan out a multi-step task. Notice the difference between waiting and continuing.
  5. Apply to the Kimi Code beta, or if you self-host, grab Unsloth’s GGUF of K2.7 and run it locally.