A Guardian exclusive reveals that the £20B Cobalt data-centre site, trumpeted by UK ministers and OpenAI in 2025 as part of the Stargate initiative, was essentially hypothetical: OpenAI and partner Nscale never visited the site or filed planning applications. This exposes a pattern of governments and AI vendors co-announcing aspirational infrastructure deals that lack operational substance
For AI PMs in regulated industries, this is a cautionary signal to pressure-test vendor infrastructure commitments — especially when those commitments underpin cloud capacity or data-residency SLAs
Developer Armin found that Claude Opus 4.8 — not a small model — was hallucinating extra, undefined keys inside the <code>edits[]</code> array of a structured tool call, causing Pi to reject the call and retry, degrading reliability despite the model getting the underlying edit right
PMs building agentic pipelines should not assume that upgrading to a more capable model improves tool-use reliability; schema validation and graceful retry logic remain essential guard-rails, and this warrants explicit regression testing when swapping model versions
Simon Willison used Claude Fable via Claude Code on an iPhone to conduct a final pre-release review of sqlite-utils 4.0, with the model doing the bulk of the code and review work across a multi-hour session; rc2 is now published
The cost-per-outcome metric ($149 for a significant chunk of a major open-source release) is the kind of concrete data point AI PMs can use to build internal business cases for agentic coding subscriptions, while also prompting questions about IP ownership and audit trails in regulated contexts
Tesla expanded its Robotaxi service to Miami without a safety monitor in the vehicle, now operating in five cities with plans to reach a dozen US states by end of 2026
The rapid, monitor-free geographic expansion of autonomous vehicle services is a bellwether for how quickly regulators may need to respond to unsupervised AI deployments — regulated-sector AI PMs should track this as a leading indicator of the risk appetite regulators will tolerate in other domains
Armin's hands-on investigation into Opus 4.8's schema hallucination behaviour is a short but important engineering post that challenges the implicit assumption that capability improvements uniformly reduce failure modes
Any PM maintaining an agentic product should read this as a reminder to maintain per-model regression suites for tool-call conformance, not just benchmark scores
Willison's detailed write-up of his iPhone-based Claude Code session provides a rare, honest account of what AI-assisted major-version development actually looks like end-to-end, including cost, prompting strategy, and failure modes encountered
For PMs considering AI-assisted development programmes, this is one of the more transparent field reports available and worth sharing with engineering leadership to set realistic expectations