Christina Stathopoulos, the data and AI evangelist at Dare to Data, organized the week’s major AI developments into recurring themes from the past month: increasing corporate investment in compute, rising debate about who controls model boundaries, and the operational challenges of scaling AI when code is generated by models.
New models and workplace integrations at the frontier
Christina highlighted two recent updates from research labs. OpenAI completed the rollout of the GPT-5.6 family, a set of models tuned for different workloads with an option to adjust reasoning intensity. OpenAI also launched ChatGPT Work, an agent workspace that connects the model to Slack, calendars, documents, and other enterprise tools.
Anthropic published research describing a so-called "J-space," an internal workspace that suggests Claude organizes and manipulates ideas before producing an output. Christina stressed this is not evidence of consciousness, but it is a notable step toward observing what a model does between input and output — visibility that can help detect deception or unsafe behavior before it appears in a response.
More AI labs becoming chip companies
Following early moves in an AI hardware race, more organizations are trying to control the compute layer themselves. The article notes DeepSeek in China is developing its own inference chips to reduce dependence on NVIDIA and Huawei, and Anthropic has had early talks with Samsung about building a custom AI chip.
Chip density is also increasing: South Korean researchers developed a manufacturing technique that stacks more than ten ultrathin memory chips, achieving roughly four times the density of today’s commercial high-bandwidth memory in the same footprint. The layers align within about six micrometers, and the short interlayer distances reduce signal travel, increasing speed and efficiency.
For AI companies, owning more of the stack helps control cost and performance and can circumvent supplier roadmap limits or export-policy obstacles.
New security threat: agent-executed ransomware
JADEPUFFER is the first documented ransomware attack in which an AI agent executed the entire operation end to end with minimal human step-by-step direction. A human selected the target, after which the agent exploited a known vulnerability, searched for passwords and API keys, moved into the production database, encrypted it, and wrote its own ransom note.
Security teams have been preparing for sophisticated AI-driven attacks; Christina argued JADEPUFFER is likely only the first of many such incidents.
Geopolitics and regulatory implications
The expanding attack surface of AI was a key reason AI security dominated discussions at the NATO summit in Ankara, where leaders talked about how AI reshapes cyberattacks, drone warfare, disinformation, supply-chain risk, and the accelerated pace of high-stakes decision-making.
In parallel with US restrictions on who may access domestic frontier models, China may be moving to limit overseas access to its own systems; Alibaba has banned US-made models for its employees. Data from OpenRouter indicate Chinese model usage at US-based companies, measured in tokens, is approaching parity with US model usage — a sign that model choice is increasingly a supply-chain and geopolitical decision for technical leaders.
Two enterprise-scale challenges: code quality and vendor lock-in
While generative AI has made code generation easy, the engineering work now is ensuring AI-produced code is correct, secure, and safe to run in production. A recent study of nearly 200,000 pull requests across more than 800 developers found AI nearly doubled coding productivity, but reviewers couldn’t keep pace. Each reviewer is now responsible for roughly twice as many pull requests as before widespread AI adoption; the share of pull requests receiving human review fell from 89% to 68%, with automated reviews filling the gap.
This velocity increase raises risks of cognitive fatigue and burnout. As Matt Palmer described on a recent episode, managing a team of agents can feel like managing a mid-size developer team: persistent messaging and status checks are required.
Another problem is deepening dependence on a single AI provider: companies are connecting more of their data, workflows, content, and business processes to one provider, which makes future switching difficult. The solution is to build an AI stack and workflows that keep companies in control of their data and allow model substitution as technology evolves. Treating model choice as a one-time decision creates the same dependency problem that GPT-5.6 and Anthropic’s chip efforts aim to avoid at the hardware layer.
What’s next
Christina will return next week with more AI news, including an initial look at Apple’s lawsuit against OpenAI, New York’s pause on new hyperscale data centers, and a landmark German ruling holding Google accountable for misinformation generated by AI Overviews. Additional updates are expected on DeepSeek’s IPO plans, OpenAI’s first AI hardware device, and Anthropic’s new enterprise deployment unit.
Christina is also hosting ongoing programming and events: the weekly show Zero to Agent in 30 Minutes, the AI Codecon live event on August 31, and an "AI Superstream on AI harnesses" — a four-hour deep dive on turning models into agents and running them securely at scale — scheduled for July 23.
Conclusion
Current trends point toward greater control over hardware and platforms by AI companies, alongside growing security and geopolitical risks. For enterprises, maintaining code quality and avoiding overreliance on a single provider will be central to secure, resilient AI operations.
