The dominant thread running through this week's curated links is accountability: who pays for AI, who controls it, who loses their job because of it, and who gets to decide when it is safe to release. From Oracle's 21,000 layoffs to Accenture's token-rationing crisis, from Apple's memory-chip price hikes to the US government's partial lifting of export controls on Anthropic's Mythos model, the week's stories collectively mark a shift from AI as aspiration to AI as a system with real costs and real consequences. We have grouped them into four themes below.
AI Economics: Cost, Spend, and the Reckoning
Companies are scrambling to stop employees from maxing out AI budgets with small tasks: Accenture has been rationing tokens after employees depleted reserves on trivial tasks like PDF-to-slide conversions, a sign that the tokenmaxxing era is over and CFOs are now demanding that AI spend be tied to measurable output.
OpenAI and Anthropic face new AI reality as users shift from 'tokenmaxxing' to efficiency: Enterprise buyers are pulling back on uncapped AI spend as open-source alternatives gain ground, putting pressure on both frontier labs to demonstrate unit economics ahead of their respective IPO timelines.
Microsoft considers DeepSeek as OpenAI costs mount: Microsoft is reportedly evaluating DeepSeek's open-weight models for a lower-cost version of Copilot Cowork as it shifts to usage-based pricing, a concrete signal that even OpenAI's closest partner is stress-testing its dependency on frontier inference costs.
Apple, Microsoft hike prices as AI drives memory chip shortage: Apple raised Mac and iPad prices by $100 to $300, attributing the increases to memory chip costs driven by AI data-centre demand; Tim Cook described the shortage as a "hundred-year flood", and Apple stock fell 6% on the news.
Oracle sheds 21,000 roles over the past year amid wave of AI layoffs: Oracle's 13% workforce reduction, attributed in part to AI-driven operational changes, adds to the growing body of evidence that enterprise AI adoption is producing measurable headcount reductions, with the company spending $1.8bn on restructuring costs in fiscal 2026 alone.
Governance, Regulation, and Geopolitics
Trump admin allows Anthropic to release Mythos AI model to some companies and government agencies: Commerce Secretary Howard Lutnick confirmed that Anthropic can restore access to Claude Mythos 5 for more than 100 US institutions, while the consumer-facing Claude Fable 5 remains blocked, establishing a two-tier access model that may become the template for future frontier model releases.
The AI world is getting 'loopy': Claude Code creator Boris Cherny's appearance at Meta's @Scale conference made the case that agentic loops, where agents continuously prompt other agents in the background without a fixed stopping condition, are as significant a shift as the move from manual coding to single-agent AI.
AI was supposed to kill engineering jobs, but new data suggests they're the most resilient: SignalFire's analysis of 80 million career records shows engineering roles fell only 11% in large tech companies versus a 25% overall hiring decline, and engineers now make up 55% of new hires at Tech Majors, up from 46% in 2019, complicating the narrative that AI is eliminating technical roles.
Infrastructure, Compute, and Hardware
Databricks' former AI chief thinks he can cut AI's power bill by 1,000x: Naveen Rao's startup Unconventional AI has released Un-0, an image-generation model built on a software simulation of an oscillator-based chip architecture that the company claims will eventually reduce inference power consumption by a factor of 1,000, with actual chip schematics due soon.
Hyundai to fully acquire Boston Dynamics in $325m deal: Hyundai is buying SoftBank's remaining 9.65% stake in Boston Dynamics for $325m, making the US robotics firm a wholly owned subsidiary and consolidating Hyundai's position as the dominant force in commercial humanoid and quadruped robotics.
Introducing computer use in Gemini 3.5 Flash: Google has made computer use a native built-in tool in Gemini 3.5 Flash, enabling developers to build agents that click, type, and navigate desktop, mobile, and browser interfaces, pushing agentic device control into a lightweight, widely available model.
Tools, Products, and AI in Practice
NotebookLM's latest upgrades make it a stronger research tool: Google upgraded NotebookLM with Gemini 3.5 and Antigravity, a secure cloud computer for code execution, and the ability to generate PDFs, spreadsheets, and slide decks directly from source material, with a 78% win rate on advanced web research tasks against its prior baseline.
Notion Mail shuts down amid agent takeover: Notion is discontinuing its email inbox on 22 September because more than half of users are already handing email workflows to Notion's AI agents without opening the inbox, a concrete example of agents collapsing standalone productivity tools into background automation.
Patronus AI lands $50m to build 'digital worlds' that stress-test AI agents: The startup, founded by former Meta AI researchers, uses simulated digital environments and reinforcement learning to evaluate agent behaviour before deployment, with revenue growing 15-fold in the past year and virtually every frontier AI lab now a customer.
How GPT-5 helped immunologist Derya Unutmaz solve a 3-year-old mystery: GPT-5 Pro identified disrupted N-linked glycosylation as the mechanism behind a three-year T-cell puzzle at The Jackson Laboratory, correctly predicted the outcome of an unpublished experiment, and is now being used to compile large-scale cancer mutation datasets, offering a concrete proof point for AI-assisted scientific research.
Meta launches its own $299 smart glasses: Meta's new smart glasses, built with EssilorLuxottica but without Ray-Ban branding, start at $299 and come in three designs, positioning them as a more accessible entry point into AI-enabled wearables as competition from Google and Snap intensifies.
Hope you enjoyed this week’s collection… We'll be back with more curated links in the next edition.
All content reflects our personal views and is not intended as professional advice or to represent any organisation.
