Top Stories
UnitedHealth Bets $3 Billion That AI Can Fix Health Insurance
UnitedHealth Group is making the largest single AI investment in healthcare history: $1.6 billion this year alone, part of a broader $3 billion commitment to automate the insurance bureaucracy that patients and doctors have been fighting for decades.
The company’s 22,000 software engineers — more than 80% of whom now use AI daily to write code and build agents — are rebuilding the systems that process billions of medical claims, detect fraud, generate clinical documentation, and select the billing codes that determine what care costs and who pays for it.
The flagship consumer product is already live. Avery, a generative AI companion launched in late March, is currently available to 6.5 million members with employer-sponsored plans and 160,000 Medicare Advantage members, with expansion to 20.5 million members across commercial, Medicare, and Medicaid by year’s end. Avery handles benefits questions, finds in-network providers, and coordinates care — the kind of routine tasks that currently eat up call center hours and patient patience.
The privacy implications are significant. STAT News’ investigation raised the central question: whose interests does the AI serve? When the same company that decides whether to approve a medical procedure also builds the AI that processes the claim, patients often have no way of knowing when or how AI influenced decisions about their care. UnitedHealth controls both the insurance side (UnitedHealthcare) and the technology side (Optum), creating a closed loop where efficiency gains and cost-cutting incentives can look identical from the outside.
Source: STAT News, UnitedHealth Group
OpenAI, Anthropic, and Google Team Up Against Chinese Model Distillation
Three of the biggest rivals in AI are cooperating on something they apparently agree on: Chinese competitors are systematically extracting knowledge from their models.
Bloomberg reported Sunday that OpenAI, Anthropic, and Google are sharing intelligence through the Frontier Model Forum — the industry nonprofit they co-founded with Microsoft in 2023 — to detect and prevent adversarial distillation. The technique involves running massive numbers of carefully crafted queries against a frontier model and using the outputs to train a cheaper competing model, effectively copying the expensive model’s capabilities without paying for the research that produced them.
The collaboration marks a notable shift. These companies fight aggressively for market share, talent, and enterprise contracts, but the threat of state-backed model copying apparently outweighs competitive concerns. The specifics of what evidence they’ve found and what countermeasures they’re deploying remain undisclosed, though the joint effort suggests the problem is serious enough to override the usual secrecy.
The timing coincides with growing U.S. government pressure on AI export controls and increasing evidence that several Chinese labs have closed the capability gap with Western models faster than their own R&D spending would suggest.
Source: Bloomberg
Meta Plans to Open Source New AI Models — With Caveats
Meta is preparing to release the first AI models developed under Alexandr Wang, the former Scale AI CEO who joined Meta last year in a $15 billion deal to lead a new team focused on artificial general intelligence, Axios reported Sunday.
The plan is to eventually release versions of the models under open source licenses — but not all of them. The largest, most capable models will remain proprietary, a departure from the more open approach Meta took with earlier Llama releases. It’s a hybrid strategy: open enough to attract the developer ecosystem that made Llama popular, closed enough to maintain competitive advantages on the frontier.
The pivot comes at a difficult moment. Previous Llama models saw limited real-world adoption despite the open licensing, and Llama 4 “wildly underperformed expectations” according to Gizmodo’s reporting, with a planned release delayed last month over continuing benchmark issues. Meta has committed over $600 billion to AI development and is projecting capital expenditure of $115–135 billion for 2026 alone, but has yet to produce a model that seriously challenges Claude or GPT on the benchmarks that enterprise buyers care about.
Wang’s consumer-focused strategy — targeting everyday users rather than the enterprise market that OpenAI and Anthropic dominate — represents a significant bet that Meta’s distribution advantages (3 billion users across its apps) matter more than raw model performance.
Quick Hits
-
Oracle cuts 30,000 jobs via 6am email to fund AI buildout: Oracle launched what may be the largest layoff in its history, terminating up to 30,000 employees — roughly 18% of its workforce — with automated emails at 6am local time. No advance notice from HR, no manager conversations. The cuts are expected to free up $8–10 billion in cash flow to fund an aggressive AI data center expansion, despite Oracle posting $6 billion in quarterly income. CNBC, Inc
-
Utah approves AI for psychiatric drug refills: Utah expanded its first-in-the-nation AI prescription program with a new pilot letting Legion Health’s AI chatbot renew 15 specific lower-risk psychiatric medications. The system cannot write new prescriptions, change doses, or handle controlled substances, antipsychotics, or lithium. Human escalation is required for safety flags, and physicians must review the first 1,250 requests before wider rollout. WinBuzzer
-
ChatGPT arrives on Apple CarPlay: OpenAI’s voice mode now works through CarPlay, letting drivers have hands-free conversations with ChatGPT while driving. The catch: you have to manually open the app (no wake word like Siri), and it can’t access maps, vehicle data, or control car functions. It’s audio-only — useful for brainstorming and questions, but walled off from everything else in the car. Engadget
-
Google cuts Veo 3.1 video generation prices: Google dropped the price of its Veo 3.1 Fast video generation model today and launched Veo 3.1 Lite, which generates video at less than half the cost of Fast with no speed penalty. Lite starts at $0.05 per second for 720p video. The move comes as OpenAI shut down Sora, effectively conceding the AI video market to Google. The Decoder
-
Foxconn posts record Q1 on AI server demand: Revenue jumped 29.7% year-over-year to T$2.13 trillion ($66.6 billion), with March alone surging 45.6% — a record for the month. The cloud and networking division drove the growth as AI infrastructure demand shows no signs of slowing. NewsBytes
Worth Watching
The Mercor breach continues to ripple through the AI industry. The $10 billion training data startup — whose clients include Anthropic, OpenAI, and Meta — confirmed a supply-chain attack through the widely used LiteLLM open-source library. Lapsus$ claims to have exfiltrated 4 terabytes of data, potentially including training datasets and confidential AI project information. Meta has already pulled its contract. Meanwhile, Microsoft published a report warning that AI itself is becoming a cyberattack surface, with threat actors achieving 450% higher phishing click-through rates using AI-enhanced attacks. The security story in AI isn’t just about protecting models from distillation — it’s about the entire ecosystem of tools, libraries, and data pipelines that the industry depends on.