Anthropic says it had to suspend access to Fable 5 and Mythos 5 for all customers after a US export-control directive tied to concerns about a jailbreak method. Other Anthropic models remain available, but this already shows that access to frontier models is no longer just an internal product decision.

The deeper signal is the precedent. Once a government can force a frontier lab to shut off powerful models at scale, the control layer around AI starts to look more like governed infrastructure than a roadmap choice. From here, the market will have to track three things at once: how capable the model is, who gets access to it, and under what conditions that access can be constrained.

Anthropic says 64% of Americans fear AI-driven job loss, more than 70% support a government role in regulating AI, and only 15% trust AI companies to decide how the technology should be built and used. More notably, cognitive dependency is now surfacing alongside jobs and safety as a mainstream concern.

That suggests the next bottleneck for AI will not be capability alone. Once users start worrying that heavy AI use could weaken their own judgment, labs are pushed into a harder problem: not just shipping faster, but shipping in a way that preserves trust and a sense of user agency.

According to the WSJ, a coalition of state attorneys general has opened an investigation into OpenAI and subpoenaed documents tied to advertising, user retention, data handling, the treatment of minors and seniors, model sycophancy, and internal company policies. If the investigation is moving along the lines described, this marks a real shift from abstract AI-safety debate into scrutiny at the product-operations layer.

What stands out is the scope. Retention, consumer data, vulnerable-user treatment, and model behavior are increasingly being pulled into familiar state-level consumer-protection language. In practice, that means AI governance may be shaped less by policy papers alone and more by how products behave in the market.

TCS and Anthropic push Claude into regulated industries: Anthropic says TCS will bring Claude to 50,000 employees across 56 countries and package it into workflows such as claims processing and lending advisory. Enterprise AI is increasingly won at the deployment layer, where service rails and compliance packaging matter as much as model quality.

Copilot code review adds runner controls and content exclusion: GitHub has added organization-level runner defaults and controls that keep Copilot review away from specified paths. As AI moves into real code-review workflows, governance rails are becoming part of the default product rather than an enterprise afterthought.

Kimi K2.7-Code is aimed at cheaper, tighter coding loops: Moonshot AI says Kimi K2.7-Code improves long-horizon software-engineering performance while cutting thinking-token usage by about 30% versus K2.6. The open coding race is starting to shift toward which loop solves work at lower cost, not which model produces the longest reasoning trace.

Meta's new Applied AI unit is near internal revolt: TechCrunch, drawing on Wired and earlier reporting, says Meta's 6,500-person Applied AI unit is facing backlash over forced reassignment, data-generation work, and morale pressure. At lab scale, AI ambition is now colliding directly with org design and workforce consent, not just capex and hiring optics.

Demis Hassabis - CEO Google Deepmind is pushing a “become superpowered with AI tools” message: Demis Hassabis is quoted telling young people to become “superpowered” with AI tools and lean into where the technology is heading. It is still a clip-led attribution, but the framing is revealing: frontier-lab leaders are increasingly treating AI fluency as a baseline career expectation.

If today's issue is about AI moving into the control layer and production rails, the matching practical lesson is simple: do not let transcripts stay as passive records. In about 30 minutes, you can turn almost any call-based workflow into a tighter loop by capturing the transcript, extracting role-specific feedback, and pushing the result into concrete next steps that can actually be followed up.

The value in the Preply pattern is not “AI summarizes well.” The value is that a transcript becomes a structured follow-up system with clear goals and enough personalization that people can act on it. That makes it portable across sales calls, customer success, coaching, recruiting, and team reviews.

STEPS:

Your finance close runs in #finance. Stripe and QuickBooks reconciled, runway updated, posted Sunday night without you asking.

Engineering review lands in #eng. Viktor pulled the open PRs, left comments on auth-refactor, flagged a dependency blocking api-pagination.

Campaign brief lands in #growth: Meta CPA up 18%, recommendation to pause broad match, a draft landing page already deployed for the variant test.

You hired him on day zero. He lives in Slack and Microsoft Teams alongside your contractors and investors, connects to 3,000+ tools, pushes back when you ship something dumb.

"Viktor is now an integral team member, and after weeks of use we still feel we haven't uncovered the full potential." Patrick, Director, Yarra Web.

Get Started for Free

This paper lays out four paths from AGI to ASI: scaling, paradigm shifts, recursive improvement, and large-scale multi-agent collectives, while also mapping the bottlenecks that could slow or redirect the path. Its strength is that it keeps a long-range frame without slipping into deterministic futurism.

At a time when the market is constantly pulled into launches and product tweaks, this kind of framework helps operators hold a longer-term mental model. Not every capability jump points straight to ASI, and the bottlenecks may end up being the more important thing to watch.

This piece pulls attention down to a less visible layer of AI: the labor chains behind data labeling, moderation, and operational support. While labs talk about scale and efficiency, part of the cost and risk is being pushed onto workers far away from the top of the value stack.

Placed next to today's rollout stories, it is a reminder that AI is not only a software stack. It is also a labor stack, and what gets squeezed at the bottom often comes back as ethical and operational risk at the top.

The argument here is that once AI is described as if it “wants,” “fears,” or “decides,” accountability becomes easier to blur. That language may sound harmless, but it makes it harder to ask which company designed the system, deployed it, and should own the consequences.

That lens is especially useful in an issue centered on the control layer. The more AI gets personified, the easier it becomes to forget that policy, access, and product behavior still reflect human and institutional choices.