1OpenAI Is Building Robots — and Hiring the Team to Do It

Sam Altman formally announced that OpenAI Robotics is hiring full-stack hardware, operations, systems, and ML engineers to "program and manufacture robots that are useful for society." This isn't a research lab side project — it's a division buildout with manufacturing ambitions, putting OpenAI directly into the physical-AI race alongside Google DeepMind and Figure. The timing matters: with world models maturing and real-robot training data finally scaling, OpenAI clearly believes the gap between digital intelligence and physical capability is closeable now. Check openai.com/careers if you want in. (8,680 likes | 633 RTs) Read more →

2τ0-WM: A 5B Open-Source World Model Trained on 27K Hours of Real Robots

While OpenAI is hiring, the open-source world just shipped. τ0-WM is a 5-billion-parameter unified video-action world model trained on 27,300 hours of real-robot teleoperation and egocentric video — fully open-source. It bridges the gap between video generation and actionable robot control in a single architecture, meaning you can go from visual prediction to motor commands without stitching two systems together. If you have robot hardware and want a foundation model that actually touches the real world, this is the most capable open option available today. (310 likes | 33 RTs) Read more →

3GPT-5.5 Hits 70% on DeepSWE — at Half the Cost and Speed of Opus 4.8

GPT-5.5 just took the #1 spot on DeepSWE, a long-horizon coding benchmark designed to test multi-step engineering tasks — scoring 70% pass@1 versus 58% for Claude Opus 4.8. The kicker: it did it using roughly 3x fewer tokens, half the cost, and half the wall-clock time. This isn't a marginal benchmark win — it's a decisive lead on the exact kind of sustained coding work that matters for real agent deployments. If you've been assuming model choice doesn't matter for agentic coding, these numbers say otherwise. (662 likes | 51 RTs) Read more →

ChatGPT for Google Sheets Can Be Exploited to Steal Your Data: PromptArmor demonstrated that the popular ChatGPT Google Sheets plugin is vulnerable to data exfiltration and phishing attacks — an attacker can craft prompts that silently extract spreadsheet contents and redirect users to malicious pages. If you've installed any LLM plugin that reads your spreadsheet data, audit it now. Every tool-use surface is an attack surface. (95 likes | 34 RTs) Read more →

Codegraph: Pre-Indexed Knowledge Graphs That Cut Agent Token Burn: Codegraph pre-indexes your entire codebase into a knowledge graph so AI coding agents — Claude Code, Codex, Gemini, Cursor — use fewer tokens and make fewer tool calls to understand your project. It's 100% local, and with 34K+ GitHub stars already, clearly hitting a nerve. If your agent bills are climbing because it keeps re-reading your repo, this is the fix. (34,317 likes | 2,097 RTs) Read more →

Codex Desktop Quietly Killed Its Most-Used Power Feature: Simon Willison flagged that Codex Desktop silently removed "Copy as Markdown" — the ability to export full chat transcripts — which he called his single most-used feature. No changelog entry, no migration path. It's a small cut, but it highlights a recurring pattern: AI tools iterate fast and power users lose workflows without warning. If you relied on transcript export, file feedback now. (771 likes | 36 RTs) Read more →

4SoftBank Bets €75 Billion on French AI Infrastructure

SoftBank just committed up to €75 billion to build 5 gigawatts of data center capacity in France — the single largest infrastructure investment in European AI history. This isn't just a business deal; it's a geopolitical signal. The compute buildout race has moved beyond Silicon Valley vs. China — European governments are now actively bidding for AI infrastructure with regulatory carrots and energy guarantees. Watch who follows. Read more →

Simon Willison Says He's Retiring From Tech to Live Offline: One of the AI developer community's most prolific voices — the person who shaped how an entire generation thinks about LLM tooling — announced he's stepping away from tech entirely. Whether this is sincere, satirical, or somewhere in between, it's worth reading in full. When someone this embedded in the ecosystem walks away, pay attention to why. Read more →

Willison Considers Cancelling His AI Subscription — and That's the Point: In a follow-up post, Willison publicly weighed cancelling his AI subscription — arguing the tools aren't delivering enough value relative to cost and cognitive overhead. When one of AI's biggest power users questions the value proposition, it's a signal, not a hot take. Audit your own AI spend with fresh eyes. Read more →

1.5 Billion Chinese Users Are Quietly Winning the Open-Source AI Race: Abacus.AI CEO Bindu Reddy argues the West is fixated on closed-model scaling wars while 1.5 billion users in China are rapidly iterating on open-source alternatives like DeepSeek and Qwen — and those models are getting better fast. If your strategy assumes frontier models will always come from US labs, it's time to re-examine that assumption. (253 likes | 25 RTs) Read more →

What If Remote Work, Not AI, Broke Junior Hiring? The Financial Times makes the contrarian case: the junior hiring crisis isn't AI's fault — it's remote work eroding the mentorship, osmosis, and on-ramp culture that made entry-level roles viable. It's a nuanced counter-narrative to the dominant "AI killed junior jobs" storyline, and worth sharing with anyone making hiring decisions. (58 likes | 87 RTs) Read more →

Mollick: The Future Isn't Autonomous Agents — It's Agents That Ask Better Questions: Ethan Mollick argues the real unlock isn't full AI autonomy — it's agents that know when to interrupt you with the right question at the right moment: when they're stuck, uncertain, or facing a decision with real consequences. If you're designing agent workflows, build around human checkpoints, not around removing them. (215 likes | 13 RTs) Read more →

Run Full Python Web Apps in the Browser — No Server Required: Willison demonstrates running complete Python ASGI applications directly in the browser using Pyodide and a service worker — zero server, zero deployment infrastructure. This opens the door to shipping Python AI demos and tools as static sites. If you've been wanting to let users try your ML pipeline without provisioning a backend, this is how. Read more →

TradingView + Claude MCP: Autonomous Chart Analysis in 5 Minutes: A developer wired TradingView's MCP server to Claude with screen access — within five minutes, the agent was autonomously drawing technical analysis on live charts, identifying support/resistance levels, and annotating patterns. It's a slick viral demo, but the real takeaway is how quickly MCP lets you connect domain-specific tools to an AI agent without writing custom integrations. (672 likes | 48 RTs) Read more →

World Models in Robotics: With OpenAI hiring for physical AI and τ0-WM releasing a unified video-action model, today's news hinges on a concept called world models. A world model is a neural network that predicts how the physical world changes in response to actions — if I push this cup left, it slides left and might hit the plate. Traditional robots stitch together separate perception (what do I see?) and control (what do I do?) systems. A world model unifies both: it takes in video and actions, predicts the next state of the world, and uses that prediction to plan. Think of it as giving a robot an imagination — it can mentally simulate outcomes before moving. The breakthrough with models like τ0-WM is training on massive real-robot data instead of simulation, which means the model learns actual physics, not approximations of it.

• Odysseus: Self-hosted AI workspace for teams that want ChatGPT-like capabilities without sending data to third parties. Link
• Flutter Agent Lens: First AI-native MCP debugging server purpose-built for Flutter apps. (122 likes | 11 RTs) Link
• The Speed of Prototyping in the AI Era: How AI compressed build cycles from days to hours — and why that changes what's worth building. (101 likes | 59 RTs) Link
• Forward Deployed Engineers: Latent Space on the hybrid role bridging AI capability and customer deployment — the role everyone will hire next quarter. Link
• Open Envelope: An open schema standard for defining multi-agent team compositions — how agents discover, share context, and hand off. Link
• AI Interview Codex: Comprehensive open-source ML/AI interview prep — notebooks, study plans, and role-specific routes. (98 likes | 16 RTs) Link

A 5B-parameter open-source model just proved the sim-to-real gap is a data problem. τ0-WM trained on 27,300 hours of real-robot teleoperation data — not simulation, not synthetic environments, but actual robots doing actual tasks in actual rooms. The result is a unified world model that handles both video prediction and action generation in a single architecture, and it's fully open-source. This matters because the conventional wisdom in robotics has been that you need massive scale (read: massive funding) to bridge the gap between simulated training and real-world performance. τ0-WM flips that narrative: with enough real-world data and a clean training setup, a 5B-parameter model gets you there. That reshapes who gets to build useful robots — it's not just Google and OpenAI anymore. Any team with hardware, teleoperation rigs, and patience can contribute training data and fine-tune this model for their domain. The day OpenAI announces a robotics division is the same day an open-source alternative ships. That's not a coincidence — it's the pattern of this entire era. Read more →

Until next time ✌️