Christina Stathopoulos, the info and AI evangelist behind Dare to Information, continued her run sorting the week’s most impactful tales right into a handful of themes we’ve been watching play out over the previous month: extra companies investing within the compute AI runs on, extra considerations about who controls a mannequin’s borders, and extra AI-generated code posing challenges to scaling AI enterprise-wide.
Christina additionally rapidly shared two updates from the frontier labs that we gained’t get into beneath. First, OpenAI completed rolling out GPT-5.6, its household of fashions tuned for various workloads with an choice to dial reasoning up or down, and launched ChatGPT Work, an agent workspace that connects the mannequin to Slack, calendars, paperwork, and different enterprise instruments. Anthropic, in the meantime, revealed analysis describing a hidden inner workspace it’s calling the “J-space” that implies that Claude organizes and manipulates concepts earlier than producing a response. It isn’t proof of something like consciousness, as Christina was fast to notice, however it’s one of many clearer steps but towards inspecting what a mannequin is definitely doing between enter and output. That sort of visibility is important for catching issues like deception or unsafe conduct earlier than they present up in a solution.
Extra AI labs are turning into chip corporations
Final week, Christina coated the opening strikes in an AI {hardware} race, with analysis from IBM and NVIDIA and a joint OpenAI and Broadcom challenge. Now there’s information that Chinese language firm DeepSeek is creating its personal inference chips to chop its dependence on NVIDIA and Huawei, and Anthropic is in early talks with Samsung to construct a customized AI chip. And as we noticed with IBM’s sub-1 nanometer tech, chips are getting denser. Researchers in South Korea have developed a producing method that stacks greater than 10 ultrathin reminiscence chips, packing about 4 instances the density of as we speak’s business high-bandwidth reminiscence into the identical footprint. The layers align inside about six micrometers, roughly a tenth the width of a human hair. The quick distances between layers imply the sign doesn’t should journey as far, making the entire stack run quicker and extra effectively.
For AI corporations, proudly owning extra of the stack is a approach to management the associated fee and efficiency of operating fashions as soon as they’re constructed. As chip entry turns into a lever in commerce and safety coverage, it’s additionally a approach to circumvent obstructions associated to a provider’s roadmap or a rival’s export coverage.
A brand new safety risk underscores the broader geopolitical stakes
JADEPUFFER is the primary documented ransomware assault through which an AI agent carried out your entire operation finish to finish. A human selected the goal, then the agent took over, exploiting a identified vulnerability, looking for passwords and API keys, shifting into the manufacturing database, encrypting it, and even writing its personal ransom be aware, all with out a individual directing every step. Safety groups have been bracing for this type of subtle AI-driven assaults. JADEPUFFER is probably going the primary of many.
That rising risk floor was one purpose why AI safety took up a lot of the dialog on the current NATO summit in Ankara, the place leaders mentioned how AI is reshaping cyberattacks, drone warfare, disinformation, provide chain danger, and the pace at which leaders are anticipated to make high-stakes choices. Paralleling US restrictions on who can entry home fashions, China may be shifting to restrict abroad entry to its personal frontier techniques, and Alibaba is banning US-made fashions for its personal staff. We’ve been monitoring this story since Might, when the US authorities’s on-again, off-again restrictions on Anthropic’s Fable and Mythos fashions provided an early signal that frontier mannequin entry was turning into of nationwide curiosity. Christina shared findings from Our World in Information that present simply how a lot the market share of Chinese language fashions has grown from a yr in the past: Per knowledge from OpenRouter, Chinese language mannequin utilization at US-based corporations, measured in tokens, is approaching parity with US mannequin utilization. For technical leaders, that’s a reminder that mannequin selection is now as a lot a provide chain determination as a technical one, and it’s more and more one with geopolitical repercussions.
Two challenges to observe for as enterprises scale AI
Now that code is effortlessly easy to generate, the actual engineering work is ensuring that AI-created code is right, safe, and secure to run in manufacturing. As many within the area at the moment are realizing, that’s simpler mentioned than executed. A current research of practically 200,000 pull requests throughout greater than 800 builders discovered that AI practically doubled coding productiveness, and reviewers couldn’t hold tempo. Every reviewer is now chargeable for roughly twice as many pull requests as they have been within the years earlier than widespread AI adoption, and the share of pull requests getting human evaluation fell from 89% to 68%, with automated evaluations filling the hole. It’s a part of the identical story Matt Palmer informed on the present a couple of weeks in the past when he in contrast operating a staff of brokers to managing a mid-size staff of human builders: “You’re simply sending messages on a regular basis, and also you’re checking in to verify issues are being executed,” he defined. The rise in velocity units up an actual danger of cognitive fatigue and burnout.
Right here’s one other problem enterprises are going through as they scale AI: They’re connecting an increasing number of of their knowledge, workflows, content material, and enterprise processes to a single AI supplier. As we already discovered within the knowledge house, the extra connected you grow to be to that supplier, the tougher it’s to change down the road. The answer to this vendor lock-in is to construct an AI stack and the workflows round it that hold you in command of your knowledge and guarantee you’ll be able to swap fashions because the expertise evolves. Enterprises that deal with mannequin selection as a one-time determination are organising the identical dependency downside that OpenAI’s GPT-5.6 and Anthropic’s chip talks are attempting to keep away from, only one layer up the stack.
What’s subsequent
Christina will return subsequent week with one other sweep of AI information, together with a primary have a look at Apple’s lawsuit in opposition to OpenAI, New York’s pause on new hyperscale knowledge facilities, and a landmark ruling in Germany holding Google accountable for misinformation generated by AI Overviews, plus updates on DeepSeek’s IPO plans, OpenAI’s first AI {hardware} system, and Anthropic’s new enterprise deployment unit. Be part of her dwell on the O’Reilly studying platform or catch up after the actual fact on YouTube, Spotify, Apple, or wherever you get your podcasts.
And if you wish to continue learning between episodes, take a look at our new weekly present Zero to Agent in 30 Minutes, our AI Codecon dwell occasion on August 31, and The Agentic Enterprise now in early launch on O’Reilly. Christina’s additionally internet hosting the AI Superstream on AI harnesses subsequent week on July 23. Hope to see you there for this four-hour deep dive on turning fashions into brokers and operating them securely at scale.
