OpenAI has released a Codex paper showing that Codex has become the dominant surface for work-related AI output inside OpenAI.

Government restriction on ChatGPT 5.6 release creates pressure on both Anthropic and OpenAI to improve the utility of existing models through context integration features (like Claude Tag) rather than waiting for frontier capability improvements, because they must provide customer value despite constrained access to new models.

Apple, Anthropic, and OpenAI should be understood as being in the same competitive boat—all battling for access to and control of context (Apple for personal context, Anthropic and OpenAI for work context)—rather than as separate market segments.

Slack has had bots for a long time, but Claude Tag in Slack is not merely a Slack bot because Anthropic is positioning it as an embedded team assistant with the ability to build context over time through access to channels, tools, data, and codebases.

OpenAI has released ChatGPT 5.6, but access is restricted to a small group of government-approved partners while Washington reviews cybersecurity risk, creating a significant slowdown in frontier availability.

The most recent Apple announcements about Siri relaunch include promises of more natural conversations, richer answers, and a dedicated Siri app.

Anthropic's Claude Tag in Slack represents a shift from formal context (fed through prompts and code) to informal, messy work context (engineering decisions, customer tickets, pricing debates, people information), requiring careful permission scoping and boundary management to prevent context leaks and corporate liability.

The Codex study at OpenAI shows that even at an AI-native company, AI tools must earn trust gradually—Codex first earned trust with engineers for code, then earned trust to access legal, recruiting, and sales context only after version 5.5 demonstrated sufficient capability.

Codex and Claude Tag represent opposite product philosophies for context integration: Codex asks users to identify and point to local files they care about (file-shaped), while Claude Tag embeds in Slack and asks teams to tag Claude into conversations (conversation/chat-shaped).

A model does not need to maximize benchmarks or be extremely intelligent to have tremendous utility if it is seamlessly integrated with rich contextual data sources, as demonstrated by Apple's approach to Siri.

Open-source models like GLM 5.2 can now compete visibly with frontier models because government slowdown on frontier releases gives open-source time to catch up publicly, even if Anthropic and OpenAI maintain private leads of 6-8 months.

The shift from intelligence wars to context wars means future AI competition will focus on 'how quickly and easily can an AI model apply intelligence to context' rather than on capability benchmarks or raw model improvement.

Users should build personal 'harnesses' (tools and integration layers) to retain control over where they route their context, rather than accepting default options from platform providers, ensuring they are not locked into any single model provider.

Companies are inherently bad at governing access to messy organizational context like unwritten engineering decisions, customer interactions, pricing debates, and personnel information, which makes AI access to this information a governance and liability risk.

Devin (an AI coding assistant) has been successful in accessing and working within team development context, showing that AI agents can handle messy team contexts in specific domains, creating a precedent for tools like Claude Tag.

The speaker has been building pieces of a context integration harness in public through projects like Open Brain and Open Engine to give users choices about where to route their context rather than accepting vendor lock-in.

Z.ai's GLM 5.2 has made cheap, open-source, frontier-quality intelligence feel much closer to a near-term reality than it did just weeks prior.

The real competition in AI is not about model capability but about context—understanding which message matters, which file is current, what the customer meant, what the team decided, and what can be shared or done. The next useful AI product will know where the work is, what it's allowed to see, what it's allowed to do, and do this seamlessly.

Companies and individuals should think carefully about what context they are comfortable giving to Apple, Anthropic, OpenAI, and other AI companies, and whether they want to retain context versus share it.

Apple's Siri redesign is not fundamentally about improving conversational capability, but about connecting Siri to contextual data sources on the user's phone (messages, photos, email, notes, screen, apps) through on-device processing and private cloud where necessary, so Siri doesn't need to be extremely intelligent to be useful.