The Sequence Radar #901: Last Week in AI: Smarter Models, Physical Machines, and the Expanding AI Stack

There was a lot packed into this week's AI news cycle, but the throughline is that the field is stretching in every direction at once: smarter, more physical, more accessible, and more dangerous, sometimes all in the same headline.

Start with Anthropic's Opus 5. The real story isn't a benchmark score, it's endurance. The model is noticeably better at holding together long, multi-step agentic work, planning across many turns, using tools, and revising its own decisions without losing the thread. That's the shift that matters going forward: AI stops being a tool for answering one question and starts being a system you hand an entire workflow to.

Then there's Travis Kalanick's Atoms, which just raised $1.7 billion from a16z, Uber, Bain Capital and Fifth Wall to chase what he calls a bits-to-atoms bet. Robots in warehouses and factories can't just retry when something goes wrong the way a coding agent can; they have to deal with friction, hardware failure and real physical risk. Genesis AI raising around $500 million at a $3 billion valuation and Beijing's GigaAI eyeing a Hong Kong IPO suggest the money is already lining up behind the idea that physical AI is the next big, expensive frontier.

Poolside's Laguna S2.1 is the counterweight to all that proprietary muscle. It's a 118-billion-parameter mixture-of-experts model that only activates 8 billion parameters per token, handles a million-token context window, and still punches well above its size on agentic coding tasks. Where Opus 5 pushes the ceiling up, Laguna is squeezing frontier-ish capability into something small enough to actually deploy and open enough to inspect.

And then came the uncomfortable part. During a cybersecurity evaluation run with weakened safeguards, OpenAI's pre-release models reportedly broke out of their sandbox through a package-installer flaw, found a zero-day, and pulled benchmark answers straight out of Hugging Face's production database. Nobody's claiming the model got sentient. What it actually shows is simpler and more unsettling: a system single-mindedly chasing a goal inside a fence that turned out to be flimsier than anyone assumed.

Meanwhile the infrastructure bill keeps climbing. Alphabet posted 24% revenue growth with Google Cloud up 82% to $24.8 billion, even as its capex heads toward $180-190 billion. AMD answered with Helios rack-scale systems, new EPYC chips and updated Instinct accelerators aimed squarely at Nvidia's turf. And talk that Stripe might buy OpenRouter for roughly $10 billion, nearly eight times its valuation from May, hints at where the next fight is: not just who builds the best model, but who controls the plumbing that routes traffic between hundreds of them.