Welcome, humans.

A robot just spent about an hour doing hotel laundry on camera, and Dyna Robotics left the awkward parts in.

Dyna-2.1 runs on Taku, a semi-humanoid robot. The uncut demo shows it working through dozens of little decisions and manipulations across a hotel-laundry workflow instead of nailing one pre-scripted fold for a 20-second highlight reel.

That distinction is basically the whole robotics problem. Real work is a chain of boring edge cases: grab this towel, move that pile, recover when something lands weird, figure out what comes next, repeat for an hour.

My favorite AI benchmark remains: can it survive a fitted sheet?

Watch the full run if you want to see what an hour of physical-agent work actually looks like.

Here’s what happened in AI today:

• 😺 OpenAI launched Dots and 20+ DevDay products

• 📰 OpenAI's revenue run rate neared $70B

• 📰 NVIDIA explored insuring loans backed by AI chips

• 🍪 Meta launched Muse for Small Business

• 🎓 Make AI restate your goal before it starts

😺 OpenAI launched Dots and built the pieces of an agent operating system

OpenAI's DevDay landed yesterday with 20+ launches: new agents called Dots, a new model called GPT-6.1 Sol, hosted computer use, Plugin Extensions, Sign in with ChatGPT, a Marketplace, new Codex cloud environments, Private Safety Processing, and much more.

That’s a lot, so let’s break it down: OpenAI is pushing ChatGPT toward a full agent operating system. You state the outcome; the agent chooses tools and models, moves information between them, and asks for approval when something is sensitive. Less “which tab was that in?” More “please just do the thing.”

ChatGPT isn’t literally replacing Windows or macOS tomorrow. The bet is that ChatGPT becomes the interface while the machinery fades into the background.

Dots is the clearest version of that bet, and OpenAI’s competitor to Muse and Grok-bot. It’s an always-on agent with its own cloud computer and browser, 4,000+ plugins, and multiple ongoing projects.

What’s a dot? One of these little guys:

A chatbot answers and stops; Dots can remember the goal, notice changes, keep working, and return when needed.

More autonomy means more risk, so OpenAI paired it with approval gates and Private Safety Processing, which lets automated safety review happen without creating a new path for OpenAI personnel to read protected customer content.

The rest of DevDay fills in the missing pieces:

• The Agents API gives developers a hosted browser desktop agents can click and type through.

• Plugin Extensions give outside apps panels, forms, viewers, settings, and actions in ChatGPT.

• Sign in with ChatGPT lets participating apps draw from your Plus or Pro allowance without API keys or separate model bills.

• The OpenAI Marketplace gives enterprise software another distribution path, including approved products eligible for existing OpenAI commitments.

• GPT-6.1 Sol pushes costs down. OpenAI says it gets close to Astra on several agentic tasks for much less, while caching makes repeated context cheaper.

During our DevDay watch party, Corey summed up the caching point in five words: “that's what makes agents doable.”

Once one agent holds your context, permissions, identity, and preferred tools, it becomes an action aggregator. In our “polyagentamorous” future, several specialist agents may sit under one primary chief-of-staff agent. As I put it near the end: “the browser is being consumed by the agent window.”

Why this matters: The agent stuff is table stakes. Everybody’s got a Muse or a Dot or a Grokbot. The wild part is OpenAI wants your ChatGPT subscription to become the Apple ID + App Store + AI budget for everything else.

The test: will people hand Dots meaningful work and walk away? Will normal people adopt them as their main way to use a computer? If a Dot can invoice, fix software, schedule, research, buy, and coordinate work without constant supervision, ChatGPT gets much closer to the front door for everything else.

FROM OUR PARTNERS

Stop paying frontier prices for tasks that don't need them

Most production tasks don't need everything a frontier model can do. If the job is to classify, extract, judge, call a tool, or follow a defined agent loop, you may be paying frontier-model prices for capability you don’t need. 

Model distillation can turn usage into a smaller, better-fit model. Watch this on-demand session and learn how to:

• Identify production tasks worth right-sizing

• Turn production behavior into training data for a smaller student model

• Test whether the smaller model clears your quality bar

We’ll show you where distillation fits, how SFT and RL compare, and how to weigh quality, cost, and timeline before choosing a post-training path.

Watch now.

🎓 AI Skill of the Day: Make the AI prove it understood you first

Lauren Tan shared one of her most-used prompts, and it's useful because a lot of bad AI work starts before the model writes a single word: it misunderstood the assignment.

A model can execute the wrong interpretation perfectly. Asking it to restate your goal surfaces that mismatch before you burn time, tokens, or 14 tool calls solving the wrong problem.

Try this before a complicated research, coding, planning, or writing task:

restate in your own words what you think my goals are and what the problem I'm trying to solve is

If the restatement is wrong, correct it before the model starts. If it's right, you've just given yourself a cheap comprehension check before the expensive work begins.

Very small prompt. Very high chance of preventing a very dumb afternoon.

Have a specific skill you want to learn? Request it here.

FROM OUR PARTNERS

Hello, fellow developers 

Enterprise software has a type: beige, brittle, held together by consultants. Not anymore. Workday is opening its platform to independent builders: real APIs, zero gatekeepers, no 6-month onboarding.

• Bring your stack: Cursor, Claude, Copilot - plug in over MCP.

• Uncapped scale: Reach 10,000+ global enterprises

🍪 Treats to Try

1. Muse for Small Business handles background work across connected business apps while requiring approval before it publishes, sends, or spends.

2. America.gov is the US gov’s new chatbot that answers typed or spoken federal-service questions from official sources and can remove personal details from uploaded forms before processing them (read more here).

3. OpenClaw Enterprise adds multi-tenancy, permissions, auditing, and swappable model and sandbox layers for persistent agents in sensitive environments.

4. Liquid d1 returns yes/no, choice, or scored decisions with probabilities instead of spending tokens generating prose.

5. InstaCloud gives coding agents serverless compute, Postgres, branching environments, and deploys they can operate end to end through CLI and skills.

6. OpenResearch turns coding agents into experiment runners with isolated git worktrees and an immutable experiment tree.

📰 Around the Horn

• OpenAI neared a $70B annualized revenue run rate after more than 70% growth since the start of Q3, according to Axios, and plans to raise $30B at a $1.4 trillion valuation.

• McDonald's used machine learning across millions of restaurant tickets to recommend local menu prices from willingness-to-pay and competitor data while franchisees kept final control.

• Google said Gemini Gems will begin automatically migrating to Skills for personal accounts in November, replacing standalone custom assistants with reusable instructions Gemini can auto-apply or stack together in any chat.

• Isomorphic Labs said its drug-design agent searched huge chemical spaces for Pareto-best tradeoffs, though it has not named a disease target or clinical candidate just yet.

• OpenAI's GPT-6 Astra helped researchers find new families of plasma equilibria in fusion math, adding another example of frontier models contributing to real scientific work.

• Researchers found AI models could transfer behavioral traits through apparently unrelated training data, raising questions about hidden model-to-model influence.

🎥 Watch our DevDay Recap

• The Neuron's live DevDay watch party has Grant and Corey reacting to the launches, pricing, agent economics, and the “agent window” thesis in real time.

• OpenAI's official keynote is the clean version if you want the demos without our running commentary.

A Cat’s Commentary

Btw: We just launched a robotics newsletter! Sign up for it here.

P.S: Love the newsletter, but only want to get it once per week? Don’t unsubscribe—update your preferences here.