GPT-5.6 Sol/Terra/Luna land Thursday 🤖, the real AI fear is overwork 💼, the one-token doom-loop fix 🧑‍💻

OpenAI will make the GPT-5.6 family, Sol, Terra and Luna, public on Thursday, ending the government-coordinated preview that kept the models out of ChatGPT since 26 June, with Commerce clearing the broad rollout. Sol runs on Cerebras at up to 750 tokens a second and looks co-designed around the hardware. Sol Ultra adds a subagent mode to Codex, where trained subagents cooperate on a task rather than merging separate runs.

Meta Superintelligence Labs put out Muse Image, its most capable image generator yet, and previewed Muse Video. Muse Image follows instructions closely, edits with precision, composes from multiple reference images, and pulls Instagram for social context, with agentic tool use and Muse Spark. It is live now in the Meta AI app, on meta.ai and in Instagram Stories in the US. Muse Video, built on the same base, carries native audio.

DeepSeek is entering the chip business, Reuters reports, citing three people familiar with the plan. The Chinese lab has spent about a year on it, meeting hardware partners and hiring silicon engineers, focused on data-centre inference chips rather than training. The aim is to cut reliance on Nvidia and Huawei as US export controls tighten. Nvidia shares fell 2.2% on the news.

Chinese authorities have held talks with Alibaba, ByteDance and Z.ai about restricting foreign access to their most advanced models, including unreleased ones, Reuters reports. Options sketched by officials go as far as barring public release or limiting the models to domestic use. It would be a sharp U-turn: open weights are how these labs won global users, and yesterday's data showed US firms run 30% of their tokens on them.

Anthropic is extending Claude Cowork, where you hand Claude a task and it works across your files, calendar, email and connected tools until it is done, from desktop to web and mobile. The beta rolls out over the coming weeks, starting with the Max plan. Separately, the company pushed the date it drops included Fable 5 access from subscriptions back from 7 July to 12 July.

Discord says a bug in its AI moderation wrongly banned more than 8,000 users over two months, after harmless images, spreadsheets, chessboards, game textures and plain white or grey transparent backgrounds, tripped the filter. The system matches uploads against databases of known abusive material, and the human-review step was effectively bypassed. Accounts have been hit since May, with 200 more last weekend before the fix. All are being restored.

A draft report inside the Treasury Department warns that key parts of the AI market resemble the dotcom bubble that broke in the early 2000s, according to NOTUS, which obtained it. Career analysts reached that view privately even as Secretary Scott Bessent publicly praised the $750 billion tech firms are pouring into AI this year. The document breaks sharply from the administration's public message of unrelenting investment.

Small reasoning models often fall into a doom loop, emitting a span like 'Wait, let me reconsider' and repeating it until the context window is gone. Repetition penalties are a band-aid that can hurt quality, and RL fixes need costly rollouts. Liquid AI's Final Token Preference Optimization pinpoints the exact token that starts a loop, trains the model to prefer coherent alternatives at that one position, and leaves the rest of the distribution alone.

OpenRouter tested the common trick of sending images at low detail to cut vision costs. On reasoning models it backfires: low detail scored 13.8 points worse, 65.2% against 79.0%, and cost more per question, because the model compensated for the downsampled image by thinking about 1.6 times harder. Better to send a clearer image and cut reasoning effort than degrade the input.

When OpenHands hit a long-standing race condition in its Kubernetes runtime, the team modelled the Runtime API in Quint, a lightweight formal specification language. The model check caught that a proposed fix was incomplete, and the fixed version reproduced the bug in production exactly as predicted. It is a concrete case for reaching for a spec to validate a distributed-systems fix before shipping, instead of trusting a plausible patch.

Run several agents on one project, a couple in the terminal, a chatbot in the browser, one on AWS, and each keeps its own memory while none knows what you told the others. So you end up copying decisions between them by hand. Neo Kim's guide walks through graph-shaped shared memory as the fix, using the open-source Omnigraph engine as the worked example and showing why a shared folder is not enough.

A year after summing tech workers up as burned out but optimistic, Lenny's 2026 survey reports a workforce splitting in two: 49% feel amplified by AI, more capable than ever, while others feel destabilised and unsure of their place. The sharper fear is not being replaced but overworked, with just 22% worried about losing their job to AI and far more dreading more work for the same pay. Burnout has climbed to 56%.

Ed Zitron makes the case that when the AI bubble breaks, the industry should be left to collapse rather than rescued. His line is blunt: no bailouts, no handouts, no CHIPS-style breaks and no sovereign wealth funds. He reads talk of too big to fail as the industry manufacturing the sense that it cannot die, when each of these firms is as brittle as any other startup.

WSJ's Dan Gallagher pours cold water on hopes that big tech will rein in AI spending. Second-quarter reports this month are expected to show combined capital spending by Google, Microsoft, Amazon and Meta up 74% year over year to about $168 billion, per Visible Alpha consensus. That outlay is squeezing free cash flow and share prices at all four, yet none looks ready to blink first.

After shipping browser agents against real apps, the author concluded the field is solving the wrong problem. Agents stall on checkout buttons with no ARIA role and burn thousands of tokens screenshotting the page to work out what is on it, when the web already exposes a machine-readable structure: the accessibility tree. Driving agents through the DOM instead of pixels cut per-action latency from seconds to under half a second.

Docx-CLI gives coding agents plain commands to comment on and redline .docx files, so a human accepts or rejects the changes in Word afterward. Instead of unzipping the document and hand-writing OOXML, which burns tokens and often yields a file Word will not open, agents address text by stable locators and the CLI edits the XML in place. Custom styles, theme colours and embedded objects all survive.

chromeless is a native macOS browser window built on WKWebView in a single Swift file, about 450KB, with no Electron and no dependencies. It exists for clean product screenshots and kiosk-style dashboards, where you want a browser that is not visibly there rather than hacking Chrome's app flag. Fresh this week, it is a tiny alternative to shipping a whole Electron app for a job that needs one plain window.

herdr is a terminal multiplexer for coding agents: run several at once and see at a glance which are blocked, working or done. Each agent gets its own real terminal, not an app's imitation, so full-screen TUIs render correctly, and you can click, drag and split panes into workspaces and tabs. It runs the agents where they already live, on your machine or any box you can reach over SSH.

otari is Mozilla AI's open-source, OpenAI-compatible gateway: put one endpoint in front of more than 40 providers, then manage API keys, enforce per-key budgets and track usage in one place. Budgets are checked before the call runs, and keys stay in-house rather than scattered across services. It is a self-hosted alternative to LiteLLM for teams that want to own their provider routing and their spend controls.

agentOS from Rivet runs coding agents like Claude Code and OpenCode inside an isolated VM that lives in your own process, rather than a separate container sandbox. It claims near-zero cold starts of about 6 milliseconds and up to 32 times lower cost than sandboxes, and deploys on Vercel, Railway or Kubernetes. It reached v0.2.7 on the day of this edition, a lighter way to isolate agent code execution.