In Today’s Issue:
🤖 BMW puts humanoids on the production line
🤑 Anthropic quietly starts its IPO path
💵 OpenAI brings Codex into AWS workflows
📈 Claude Opus 4.8 leads ARC-AGI-3, barely
✨ And more AI goodness…
⚡ The Signal
Today's issue starts on BMW's production line, where humanoid robots are becoming a factory-layout decision.
BBC reports that BMW will use humanoid robots for car manufacturing in Europe for the first time. Two Aeon robots from Hexagon Robotics are planned to work in production from the summer after a test deployment at BMW's Leipzig factory.
Aeon's shape matters because a human-sized robot can fit into workstations already designed around human bodies, while training improves through teleoperation, digital twins, reinforcement learning, and imitation learning. Physical AI is moving from stage demos toward narrow factory jobs where variance is the real test.
All the best,
Kim Isenberg
📄 Anthropic Quietly Starts Its IPO Path
Anthropic confidentially submitted a draft S-1 to the SEC for a proposed IPO of its common stock. The company says the offering still depends on SEC review, market conditions, and other factors; it has not set the number of shares or pricing.
👉 tl;dr: Anthropic is keeping the IPO option open while the frontier-lab funding race gets more expensive, more public, and harder to separate from cloud and compute strategy. The question that arises is who will hold an IPO first: OpenAI or Anthropic.
☁️ OpenAI Puts Codex Inside AWS
OpenAI frontier models and Codex are now generally available on AWS. The release lets enterprises use OpenAI through Amazon Bedrock and their existing AWS security, governance, procurement, billing, and deployment workflows.
👉 tl;dr: Codex is not only becoming more useful; it is becoming easier for enterprises to approve, buy, route, audit, and deploy inside the systems they already trust. OpenAI is increasingly gaining access to the enterprise sector.
🛡️ Europe May Get Anthropic's Mythos
Anthropic and the EU cybersecurity agency ENISA are negotiating access to Mythos, Anthropic's restricted cybersecurity model. PYMNTS, citing the Financial Times, reports that access has been offered but the conditions are still being agreed.
👉 tl;dr: Mythos is becoming a geopolitical access problem: too sensitive for broad release, but too important for governments to ignore if it can find serious vulnerabilities.
Before automating a workflow, ask what the task actually requires: repetition, judgment, movement, inspection, or recovery from messy edge cases.
Why it helps: BMW's humanoid-robot bet works because the task is narrow and the factory context is explicit. The same rule applies to software agents: define the workspace, the handoff points, and the failure modes before choosing the tool.
Try this: Paste a process description and ask: "Break this workflow into repetitive steps, judgment-heavy steps, handoff points, and likely failure cases. Then tell me which parts are safe to automate first and which need human review."
🎬 Watch This
In Google DeepMind on Gemma, AI Studio & the Future of AI, Omar Sanseviero and Paige Bailey discuss Gemma, AI Studio, on-device AI, sovereign AI, and what it now means to build with open models. The connection to today's issue is distribution: frontier systems get attention, but builders also need models, tools, local options, and workflows they can actually use.
– Satya Nadella, Microsoft chairman and CEO, in NVIDIA's Computex announcement on RTX Spark and Windows PCs for personal AI agents
Apple's next Siri push appears to have leaked before WWDC. TechCrunch, citing Bloomberg, says leaked renders show a standalone Siri app meant to compete with ChatGPT-style assistants, plus a new Siri-powered search mode tied to Dynamic Island and Spotlight-style workflows.
Treat this as pre-WWDC reporting, not a confirmed Apple launch. The reported design points to a more ambitious Siri: less voice-command layer, more systemwide AI interface across apps, search, documents, photos, and local context.
Polsia: your AI co-founder that never sleeps
Polsia is the autonomous AI that runs your whole company, not just one task. It plans your roadmap, ships your code, runs your ads, replies to customers, closes deals, and posts your updates. Every day. While you sleep.
No team to hire. No stack to wire together. You bring the idea, the budget, and the guardrails. Polsia turns it into work its agents actually execute, then keeps going after launch: marketing, support, and iteration on autopilot.
This is not another copilot that stops when the build is done. It is your first employee, the one that never clocks out.
Watch it work live at polsia.com/live
BMW Sends Humanoids to the Line
The Takeaway
👉 BBC reports that BMW will use humanoid robots for car manufacturing in Europe for the first time.
👉 Two Aeon robots from Hexagon Robotics are planned to move into production from the summer after testing at BMW's Leipzig factory.
👉 Aeon is 1.65m tall, weighs 60kg, uses 21 sensors, and can carry 15kg briefly or 8kg continuously.
👉 BMW trained the robots with teleoperation and simulation in a digital twin of the factory using NVIDIA software.
BMW is treating humanoid robots as factory equipment with very specific jobs. BBC reports that the carmaker will use two Aeon robots from Hexagon Robotics in European car production from the summer, following a test deployment at its Leipzig plant.
The economics are doing part of the work. Car factories have used robot arms for decades, but rebuilding an assembly line around every new machine is expensive. The humanoid form becomes a layout hack: same rough size, same workstations, fewer changes to the line.
Aeon is still narrow, but the training stack is getting serious. BBC says BMW used teleoperation, simulation in a digital twin, and reinforcement learning to train the robot, while Hexagon points to imitation learning as a way to cut task training from months to days. The first jobs are constrained: feeding parts to manufacturing tools and pick-and-place work for battery assembly.
The caution is the useful part. Aeon has wheels instead of feet, a three-hour battery, and a job list that sounds boring by design. That is probably where physical AI becomes real first: not in viral demos, but in repetitive industrial tasks where the robot can handle small variations without stopping the line.
Why it matters: Humanoid robots are entering the factory through cost, layout, and training advantages before anything close to general machine labor. BMW's experiment is a good test of whether physical AI can make existing production lines more flexible without pretending robots are ready to do everything.
The chart: ARC-AGI-3 leaderboard shows Claude Opus 4.8 (High) in first place with a verified 1.5% score, near the $10K cost limit, listed as Author: Anthropic and Type: CoT.
The lesson: Claude leads the benchmark, but the absolute score is still tiny. The benchmark is measuring interactive adaptation under cost pressure, not polished coding-demo fluency.
The caveat: A benchmark lead is not AGI solved. The cost axis is the point: today's best visible result sits near the expensive end and still barely clears 1%.
How Cosmos 3 Helps Physical AI Think Before It Acts
⚡ Bottom line: NVIDIA launched Cosmos 3, an open physical AI foundation model that combines vision reasoning, world generation, and action prediction in one system.
💡 Why it matters: Robots and autonomous systems need models that can reason about physical scenes, not just recognize objects or generate video.
🔎 What it means: NVIDIA is trying to make world models part of the standard robotics stack: simulate, predict, train, and act before the machine touches the real world.
NVIDIA's Cosmos 3 release is a clean example of where robotics AI is heading: fewer separate perception, simulation, and control components; more unified models that can understand scenes, generate future states, and predict actions.
The model uses a mixture-of-transformers architecture that pairs a reasoning transformer with an expert generation transformer. NVIDIA says Cosmos 3 can work across text, images, video, ambient sound, and action trajectories, with Cosmos 3 Super and Cosmos 3 Nano available now and Cosmos 3 Edge coming later for real-time inference.
Cosmos 3 is also a platform move. NVIDIA launched the Cosmos Coalition with AI labs and robotics companies, including Agile Robots, Black Forest Labs, Generalist, LTX, Runway, and Skild AI, to push open world models for robotics, autonomous vehicles, industrial AI, and smart spaces.
Hiring in 8 countries shouldn't require 8 different processes
This guide from Deel breaks down how to build one global hiring system. You’ll learn about assessment frameworks that scale, how to do headcount planning across regions, and even intake processes that work everywhere. As HR pros know, hiring in one country is hard enough. So let this free global hiring guide give you the tools you need to avoid global hiring headaches.