🏥 Dr GPT Will See You Now, 🏳️ Apple Surrenders to Google, ☢️ Zuck Goes Nuclear
Plus: AI cracks unproven math, the end of shopping websites, and China blocks Nvidia.
🎵 Podcast
Don’t feel like reading? Listen to it instead.
📰 Latest News
This week’s image aesthetic (Flux 2 Pro): Renaissance / Baroque
Dr GPT Will See You Now
The race to digitise biology has officially begun. OpenAI has launched a two-pronged assault on the sector with ChatGPT Health, a consumer portal that connects directly to medical records via partners like b.well and MyFitnessPal, and OpenAI for Healthcare, an enterprise suite for hospitals powered by the new GPT-5.2 model. This launch follows the company’s quiet $100 million acquisition of Torch, a four-person health data startup, signalling a desperate hunt for specialised talent. Simultaneously, Anthropic has opened a second front with Claude for Healthcare, a HIPAA-compliant platform designed to plug directly into electronic health records (EHRs).
Why it Matters
This is no longer about chatbots; it is a battle for the most valuable, private dataset on earth: the human body. By launching consumer portals that ingest medical records, OpenAI is attempting to build the “ultimate longitudinal dataset” that no public crawl can provide. The simultaneous push by Anthropic and OpenAI into HIPAA-compliant infrastructure reveals that the frontier labs have hit a revenue ceiling with generic chat and are pivoting to high-trust, regulated industries where open-source models cannot easily compete due to compliance barriers. For OpenAI specifically, the deployment of GPT-5.2 (validated against the new “HealthBench” standard) suggests that their next-generation “reasoning” models are being tuned primarily for high-stakes diagnostics rather than just creative writing.
This follows on from last weeks story in which we covered that 40 million people globally are using AI for health-related queries every single day.
GPT-5.2 Cracks a 30-Year Mathematical Mystery
The boundary between human and machine mathematics has been breached. OpenAI’s GPT-5.2 Pro has successfully solved multiple “Erdos problems”—mathematical conjectures posed by the legendary Paul Erdos that have remained unsolved for decades. In a breakthrough verified by Fields Medalist Terence Tao, the AI used a “harmonic approach” via Harmonic’s Aristotle system to auto-formalise its proofs in the Lean 4 programming language, ensuring they were logically flawless before human review. The most notable success was Erdos Problem #728, which the model solved “more or less autonomously” by identifying a “creative and elegant generalization” that had eluded researchers, effectively turning a weeks-long manual proof process into a 15-minute generation task.
Why it Matters
This proves that AI has graduated from “calculator” to “creator.” By pairing the creative reasoning of GPT-5.2 with the rigid logical verification of the Harmonic Aristotle system, OpenAI has solved the hallucination problem in high-stakes math: the AI can now check its own work against formal logic before showing it to a human. For the field of mathematics, this signals the end of the “lonely genius” era; the future of discovery is a hybrid workflow where AI generates the “candidate proofs” and humans simply guide the strategy. It also gives OpenAI a massive competitive moat in the “AI-for-Science” sector, validating that their “Thinking” models (like GPT-5.2) possess genuine novel reasoning capabilities that competitors’ purely pattern-matching models lack.
Why You Will Never Visit a Shopping Website Again
The era of “click and buy” is officially ending. In a massive bid to standardise the chaos of AI shopping, Google has unveiled the Universal Commerce Protocol (UCP), an open-source framework designed to let AI agents browse, negotiate, and purchase products across the web without a human ever visiting a website. Co-developed with retail giants Walmart, Target, and Shopify, the protocol creates a shared “language” that allows synthetic agents to connect directly with inventory systems, bypassing the need for custom integrations for every single store. Major payment networks including Visa and Mastercard have already endorsed the standard, ensuring that these autonomous transactions are backed by cryptographic proof of user consent.
Why it Matters
This is the “TCP/IP moment” for agentic commerce. By establishing a universal standard, Google is preventing the fragmentation of the market where every AI assistant (like Gemini or ChatGPT) would otherwise need a unique partnership with every retailer to function. The protocol effectively turns the entire internet into a headless storefront; for Shopify merchants, this means their products are instantly purchasable by millions of AI agents without lifting a finger. For consumers, it signals a shift to “intent-based” shopping—where you simply tell your AI “buy me a winter coat under $200” and it handles the browsing, cart management, and payment in the background, only surfacing for final approval.
Zuck Goes Nuclear: Meta Kills VR to Build a $600B Energy Empire
Meta has officially pivoted from the Metaverse to the power grid with the launch of Meta Compute, a massive initiative to add hundreds of gigawatts of energy capacity over the coming decades. Co-led by infrastructure chief Santosh Janardhan and former SSI co-founder Daniel Gross, the project is backed by a staggering $600B commitment in US infrastructure spending by 2028. To clear the runway for this capital-intensive sprint, the company is reportedly cutting 10% of its Reality Labs staff, effectively sacrificing its virtual reality dreams to fund its nuclear reality. Newly appointed President Dina Powell McCormick will lead the government negotiations required to secure the 20-year nuclear power agreements needed to keep these clusters running.
Why it Matters
This signals that the constraint on AGI is no longer silicon but electricity. By effectively defunding the Metaverse to finance nuclear power deals Meta is admitting that “scale” is now a physical engineering challenge rather than a software one. The appointment of a former national security official to broker government energy deals highlights that building AI clusters has moved beyond corporate strategy into the realm of national policy. With Daniel Gross at the helm this is also a clear message that Meta intends to build the largest training runs in history regardless of the cost.
The $1 Billion Surrender: Apple Admits It Can’t Build AI
The walled garden has outsourced its brain. Apple and Google confirmed a multi-year, ~$1B/year partnership integrating Gemini into iOS to power the long-delayed Siri overhaul, expected Spring 2026. After evaluating OpenAI and Anthropic, Apple selected Google, tacitly admitting its internal “Ajax” efforts hit a compute wall. While Apple retains its “Private Cloud Compute” privacy protocols, the iPhone’s core intelligence layer is now effectively rented from Mountain View.
Why it Matters
This confirms the “Model Layer” is now a commodity utility. Apple’s concession proves that building frontier models is a CapEx game, not a design problem—and Apple lost. Google secures a monopoly on mobile intelligence (powering both Android and iOS), relegating OpenAI to an optional plugin. Strategic winners are Google Cloud and TSMC; the loser is Apple’s internal AI team. Regulators will view this consolidation as a massive single point of failure for Western computing.
China Bans Nvidia to Force a Domestic Tech War
China has fired a warning shot in the global chip war by ordering its domestic tech giants to pause orders for Nvidia’s H200 AI chips. The directive, revealed by The Information, instructs companies to halt purchases while Beijing reviews “special circumstances” under which imports might be allowed—primarily restricting them to university R&D labs rather than commercial data centres. This comes despite a recent US policy shift that conditionally allowed H200 exports; however, Nvidia has reportedly responded to the uncertainty by demanding full upfront payment from Chinese clients to hedge against sudden regulatory blocks.
Why it Matters
This is a strategic blockade designed to force domestic reliance on “homegrown” silicon. By choking off the supply of Nvidia’s hardware, Beijing is artificially creating a market vacuum that Huawei’s Ascend 910C is positioned to fill. While the 910C still lags behind the H200 in raw training performance, the pause signals that China is willing to sacrifice short-term AI speed for long-term supply chain sovereignty. For Nvidia, this endangers a massive revenue stream—Chinese firms had placed orders for over 2 million H200 chips—and shifts the narrative from “global dominance” to a fragmented market where Western and Asian AI stacks are completely decoupled.
The last newsletter: