Max Headroom 
UNCLASSIFIED
This week data and AI evangelist Christina Stathopoulos returned for a solo news briefing. Instead of exploring one or two topics in depth, Christina sorted the week’s headlines into a handful of threads: advances in physical hardware to keep up with AI demand, the widening reach of government oversight into frontier model companies, and a workforce that’s reorganizing faster than job titles can describe it.
Along the way, Christina flagged a few interesting items too small to garner their own sections. Anthropic launched Claude Science, a workbench that pulls research databases, lab tools, and compute into one place for life sciences researchers, following OpenAI’s earlier release of GPT-Rosalind, a model tuned for biological reasoning. And OpenAI began a limited preview of its GPT-5.6 family, three models (Sol, Terra, and Luna) built for different jobs instead of one model trying to do everything. Watch now.
The AI hardware race has moved from parameters to atoms and watts
The biggest model headlines get the attention, but the real story this week was what they’re running on. IBM introduced the world’s first sub-1 nanometer chip technology, measuring 0.7 nanometers, or roughly a third the width of a strand of DNA. We’re approaching the limits of how small we can shrink transistors, Christina pointed out, so IBM is now also stacking them vertically. With 0.7 nm transistors, the company can pack around 100 billion into a fingernail-sized chip that claims to have 50% higher performance and 70% lower power consumption than the previous 2 nanometer generation. They’re not yet a product in the wild, but sub-1 nanometer chips are a marked research breakthrough in the angstrom era.
OpenAI and Broadcom have taken a different approach. Last week, they unveiled Jalapeño, a chip built specifically for LLM inference rather than training. As Christina put it, training gets the headlines, but inference is where AI actually reaches people. Every improvement in cost, speed, and reliability means a faster answer or a cheaper product for the people using it every day, and a small efficiency gain multiplied across hundreds of millions of users adds up fast. That’s why frontier labs are moving away from off-the-shelf tech to designing their own.
NVIDIA, meanwhile, shared a new closed-loop, fully liquid-cooled AI factory design that uses coolant that can run as warm as 45°C (113°F), removing the dependence on chilled water that’s made data centers a target for criticism over their energy and water use. Together, these three stories point to physical infrastructure, not algorithms, as AI’s next real opportunity.
Government oversight is turning into a permanent fixture
Anthropic restored public access to Claude Fable 5 and Claude Mythos 5 after the US government lifted the export controls that had pulled the models offline for security concerns tied to vulnerability discovery. The company added a new cybersecurity classifier meant to block known jailbreak techniques and says it will keep working with the government on AI security matters. It’s a reminder that access to frontier models can be switched off, and that the terms for turning it back on are now being negotiated case by case. Epoch AI data shows critical vulnerability disclosures had already spiked to 3.5 times the previous monthly peak right after Anthropic’s Mythos preview went live. We’ve mentioned before that this cuts both ways: Attackers can use AI to find weak points faster, but so can the defenders trying to patch them first.
OpenAI’s GPT-5.6 family launched as a limited, tiered preview for trusted partners at the government’s request, with broader access to follow. At the same time, the Financial Times has reported that OpenAI is proposing to give the US government a 5% equity stake in the company, which it’s pitching as a way to ensure that some of AI’s economic upside would flow back to taxpayers. It’s also, as Christina noted, likely an attempt to build public trust. Whether or not that stake materializes, government involvement in frontier AI now looks like a standing condition that companies build around, and it raises real questions for anyone outside the US who doesn’t control the terms of their own access to these models.
Roles are evolving faster than the org chart can describe
The best model in the world can’t close the gap between what a client wants and what actually gets built. For that, organizations are increasingly betting on the role of forward-deployed engineer, a mix of platform engineer, solutions architect, and product manager, who embed directly with clients to turn AI ambitions into working systems. Microsoft committed $2.5 billion and AWS committed $1 billion to new AI deployment units, following similar moves earlier this year from OpenAI and a ServiceNow-Accenture partnership. (Maya Mikhailov and Doug Shannon had some thoughts about the limits of this approach back in June.)
Boris Cherny, the creator of Claude Code, has been thinking beyond job titles to the function each team member performs according to their particular strengths and interests. Looking at his own team, he identified five archetypes: the prototyper, who generates ideas most of which won’t ship; the builder, who turns an idea into a production-grade product; the sweeper, who simplifies code and improves performance; the grower, who iterates on a shipped product to improve market fit; and the maintainer, who keeps a mature system secure, reliable, and fast at scale. People can span two or three of these archetypes at once, and none of them maps cleanly to “engineer” or “designer.”
Organizations on the path to becoming AI-native have to rebuild from within, and they have to do it quickly. Christina shared examples of two very different approaches they’re taking to get there. SAP, facing a stock slide, is cutting costs to double down on hiring AI talent externally, while IKEA is retraining its existing employees for AI-enabled roles instead. We’ll see more companies considering their options, but as Tim O’Reilly recently noted, no matter which path they take, successful companies will be ones that intentionally build a skill infrastructure that incentivizes knowledge sharing as teams figure out the best ways to use this technology for their specific circumstances.
What’s next
Christina closed the show with a story not about building products or raising funding rounds but about using AI to protect people. Google’s Android earthquake alert system warned an estimated 11.4 million people ahead of recent earthquakes in Venezuela, using accelerometers already built into their phones to detect seismic waves and send warnings with just seconds of lead time. The company is using the same underlying approach, pairing sensor and satellite data with AI, to map wildfire boundaries in near real time through Google Maps and Search and to forecast floods up to seven days out. It’s an encouraging counterweight to the stream of product releases and security incidents we usually cover.
Christina will host This Week in AI throughout July. Next week, she’ll cover the growing battle over AI chips as DeepSeek, Anthropic, and Samsung make major moves, explore the rise of agentic ransomware, and examine why AI-generated code is outpacing our ability to review it, plus the release of OpenAI’s much-awaited GPT-5.6 and some fascinating new research from Anthropic. If you’re an O’Reilly member, join us live. If not, try it out with a free trial or check out our takeaways here on Radar each Friday and watch full episodes on YouTube, Spotify, Apple, or wherever you get your podcasts.
If you’re looking for a more technical deep dive, on July 23 Christina will host the AI Superstream focused on AI harnesses. Join in to discover how our lineup of experts are building and running reliable, production-ready autonomous agent systems. Register here.
AI & ML, Commentary
Radar
