Subject: Anthropic found a "global workspace" inside Claude — and a free 295B open model dropped

A/B alternates: - Anthropic spotted a "consciousness-like" layer in Claude. Plus: a 295B open model you can try free this week. - Inside Claude's brain, a new open-weight giant, and Tom Riddle's diary on an e-ink tablet.

TL;DR: Anthropic published evidence that Claude has an emergent "global workspace" — a small set of internal patterns that act like a silent scratchpad. Tencent released Hy3, a 295B open-weights model free on OpenRouter until July 21. And Simon Willison shipped sqlite-utils 4.0rc2 with bug fixes mostly written by an agent for $149.25.

Anthropic — "A global workspace in language models" · Anthropic published research on July 6, 2026 showing that Claude has spontaneously developed an internal "J-space" — a small set of neural patterns that light up when the model is "thinking about" something, even when it never says it out loud. The pattern wasn't designed; it emerged during training and behaves a lot like the "global workspace" neuroscientists use to describe conscious access. Why it matters: this is the strongest mechanistic evidence yet that frontier models are doing structured internal reasoning rather than just outputting tokens. Honest take: interesting science, but don't mistake "workspace-like" for "conscious" — Anthropic is careful not to claim that. Source: https://www.anthropic.com/research/global-workspace

Tencent Hy3 · Hy3 is a 295B-parameter open-weights Mixture-of-Experts model from Tencent, Apache 2.0 licensed, with 21B active parameters, 256K context, and a 598GB full / 300GB FP8 release on Hugging Face. Tencent says it rivals flagship open models 2–5× its size. It's free on OpenRouter until July 21, 2026. Why it matters: the open-weight frontier just moved again — and "free for two weeks" is the lowest-friction way to benchmark it against Claude or GPT on your real workload. Honest take: a 598GB download is not a casual afternoon; stick to OpenRouter unless you have the hardware. Source: https://simonwillison.net/2026/Jul/6/hy3/ · https://huggingface.co/tencent/Hy3

sqlite-utils 4.0rc2, written by Claude Fable for ~$149 · Simon Willison ran 37 prompts and 34 commits through Claude Fable on his phone to prep a stable 4.0 release of his sqlite-utils library. The agent caught a serious data-loss bug — delete_where() left the connection in in_transaction=True, poisoning every subsequent write — that Simon had missed himself. Why it matters: a concrete receipt for "AI coding agent, shipped real release, $149 of tokens." Honest take: a great vibe-shift story, but the work was still tightly steered by a human reviewer who caught what the agent didn't. Source: https://simonwillison.net/2026/Jul/5/sqlite-utils-fable/

AMD Ryzen AI Halo — a $3,999 AI dev box with 128GB unified RAM · AMD's first-party AI developer kit pairs a Ryzen AI Max+ 395 (16-core Zen 5) with Radeon 8060S integrated graphics and 128GB of LPDDR5x-8000 unified memory at 256GB/s bandwidth. Why it matters: 128GB unified is the threshold where you can hold serious local models in memory on a desktop-sized machine — competing with NVIDIA's DGX Spark ($3,999-ish, 273GB/s) at a similar price, but on AMD's ROCm stack. Honest take: real local-AI hardware finally exists at a sane size; still $4k, still Linux-driver-shaped, and the Framework Desktop with the same chip is cheaper if you don't need the prebuilt. Source: https://www.lttlabs.com/articles/2026/07/06/amd-ryzen-ai-halo · https://www.amd.com/en/products/processors/desktops/ryzen/ryzen-ai-halo.html

Google quietly opted you into AI training — here's how to opt out · A June 2026 update to Google's "Search Services History" settings expanded Google's AI-training data set to include images, audio, and video you upload across Search, Lens, Maps, Translate, News, Shopping, Flights and Hotels. The toggle defaults to on. Why it matters: this is the third major consumer AI company this year (after Meta) to bury an opt-in behind a buried settings page — worth the 60 seconds to flip it off if you use Google Lens or Translate regularly. Honest take: not a scandal, but a useful nudge to actually read the email Google sent you. Source: https://techcrunch.com/2026/07/06/if-you-use-google-youre-training-its-ai-heres-how-to-opt-out/

One more for fun: Maxime Rivest turned a reMarkable Paper Pro e-ink tablet into Tom Riddle's diary — write with the pen, the words fade, and an answer writes itself back stroke by stroke, powered by an OpenAI-compatible model. 418 GitHub stars in a day. Source: https://github.com/MaximeRivest/Riddle

Subscribe free at buttondown.com/nexttool — one issue, every weekday morning. Forward to a friend who still hasn't tried an AI coding agent.