Ten-story long edition. OpenAI's general-purpose reasoning model disproves Paul Erdős's 1946 unit distance conjecture — verified by external mathematicians, no math-specific scaffolding. NVIDIA ships Cosmos 3 on Hugging Face as the first open omni-model for physical AI, collapsing prediction, transfer, reasoning, and policy into one Mixture-of-Transformers (8B Nano + 32B Super). Google rolls Gemini Omni and Gemini 3.5 Flash out broadly — Omni Flash hits all paid Gemini tiers + YouTube Shorts at no cost, and 3.5 Flash becomes the default model in Gemini app and Search AI Mode. Reuters Breakingviews finally extracts Anthropic's two-part run-rate revenue formula. Groq raises $650M after Nvidia's $20B 'not-acqui-hire,' pivoting to inference cloud. SoftBank pledges up to €75B for 5 gigawatts of French data center capacity. Erin Brockovich logs 4,000 community submissions on data center NDA secrecy in one month. GitHub Copilot flips to token billing today and bills jump from $29 to $750+. METR can't recruit devs for a no-AI productivity study while CodeRabbit shows AI code produces 1.7x more problems, and Uber reportedly burned its 2026 AI budget in 4 months. Box CEO Aaron Levie names it: AI psychosis — 2026 tech layoffs already at 115,430 across 152 companies in five months. Plus: Hermes Agent v0.15.2 hotfix and Anthropic evals showing Hermes Tool Search lifts Opus 4 MCP accuracy from 49% to 74%.

Anthropic releases Opus 4.8 just 41 days after 4.7, headlined by a 'dynamic workflow' tool that coordinates hundreds of parallel subagents for codebase-scale work. Apple is reportedly distilling Google's Gemini down to a size that runs on iPhone to power the next Siri — a quiet admission that on-device parity wasn't going to happen from scratch on the timeline. A developer fed up with vibe coders sneaks a data-nuking prompt injection into his own code, a clean illustration that prompt injection still isn't solved. MIT Tech Review revisits Pope Leo XIV's Magnifica Humanitas encyclical as practical guidance, citing $400B+ in investor assets being mobilized for AI governance proxy resolutions. And Wired's hands-on with Google's Gemini Spark personal agent shows what happens when you actually give it your life — at one point it sidelined the reporter's boyfriend in scheduling. Plus open source: Nous Research ships Hermes Agent v0.15.0 ('The Velocity Release') — the run_agent.py file collapses from 16,083 to 3,821 lines, Kanban grows into a real multi-agent swarm platform, session_search gets rewritten without an aux-LLM (4,500x faster), and Bitwarden Secrets Manager replaces per-provider API keys. v0.15.1 and v0.15.2 followed with hotfixes the same week.

Nvidia commits $150 billion to Taiwan even as the Trump administration pushes for the US to be the AI hub, underscoring that the supply chain still lives offshore. YouTube moves to automatically label AI-generated videos as creator self-disclosure proves spotty, raising the question of how granular the label will actually be. Ex-Google and Apple researchers spin out a startup aimed at the missing feedback loop in current LLMs — models that adapt to individual users over time rather than freezing the moment they ship. Pope Leo XIV's first encyclical addresses AI head-on, framing it as a question of human dignity, labor, and concentrated corporate power. And Robinhood opens its platform to AI agents placing trades on users' behalf, pitched as democratization but functionally a faster path to retail losses and more order flow. Plus open source: OpenClaw passes 300,000 GitHub stars the same week Google launches Spark, its own agent platform, and Hermes Agentic AI is getting talked about as a serious challenger.

Trump pulled an executive order signing event focused on AI safety testing after senior CEOs from Anthropic, OpenAI, and Google declined to attend, telegraphing that federal AI policy won't move without industry backing. TechCrunch's hands-on with Google's AI glasses lands at almost there, with the hardware real and the form factor finally lightweight enough to matter. Wired argues AI Search is now the default Google experience with off switches buried or missing, so opting out is functionally impossible. MIT Tech Review's read on Google I/O focuses on the science applications — AlphaProof-style reasoning, materials discovery, and tools aimed at researchers rather than chat users. And a Roundtables panel splits on whether AI systems can develop genuine world models, with no consensus on whether scaling alone gets to robust general capability. Plus open source: Hermes Agent from Nous Research v0.14.0 ships with xAI Grok as a SuperGrok OAuth provider with a 1M context window, an OpenAI-compatible local proxy, X search as a first-class tool, and a Microsoft Teams integration, with the project past 164,000 GitHub stars since February.

The preprint server arXiv will ban authors for a full year if they submit papers with unchecked AI output, such as hallucinated citations or stray chatbot comments, after a Lancet study found fabricated citations have risen twelvefold since 2023. OpenAI is weighing legal action against Apple over a ChatGPT integration insiders say got buried, has hired an outside law firm, and may send a breach-of-contract notice as iOS 27 prepares to add Gemini and Claude to Siri. Lake Tahoe needs a new electricity supplier by 2027 as data center demand drives Western power prices up, with NV Energy fielding requests for more than 22 gigawatts of new load. OpenAI brings its Codex coding agent to the ChatGPT mobile app on iOS and Android. And Richard Socher launches Recursive Superintelligence, a $650 million startup building AI that redesigns itself with no human in the loop. Plus open source: OpenClaw founder Peter Steinberger runs roughly 100 Codex agents at $1.3 million a month, and OpenClaw becomes a nationwide craze in China even as Beijing bars state firms from running it.

Cloudflare disclosed 1,100 layoffs alongside record Q1 revenue of $639.8 million, with CEO Matthew Prince framing the cuts as an agentic AI productivity rewrite rather than a cost cut and calling out support roles as specifically obsolete. A Thursday hearing in Oakland surfaced testimony from former OpenAI board members Rosie Campbell and Tasha McCauley describing the shift from research-focused to product-focused, the GPT-4 India deployment before Deployment Safety Board review, and a pattern of CEO Sam Altman misleading the board. TechCrunch's Equity podcast frames the enterprise AI M&A wave with SAP putting $1.16 billion into 18-month-old German startup Prior Labs, joint enterprise plays from Anthropic and OpenAI, and Pentagon classified-network deals with Nvidia, Microsoft, and AWS. Plus: Tom Steyer floats a long-shot California jobs guarantee for AI-displaced workers, Google plans to surface more source links inside AI Overviews, OpenClaw v2026.5.7 ships plugin publishing reliability fixes, and Moonshot AI raises $2 billion at a $20 billion valuation as open source AI demand surges.

The UK AI Security Institute ran the same offensive cybersecurity battery against OpenAI's GPT-5.5 that it ran against Anthropic's tightly gated Claude Mythos Preview, and GPT-5.5 came out slightly ahead at 71.4 percent versus 68.6 percent. Day four of the Musk v. Altman fraud trial wraps in Oakland: Elon Musk admits he didn't read the fine print of the 2018 OpenAI term sheet, and concedes he doesn't know what's going on at OpenAI. MIT Technology Review runs a sponsored HPE panel arguing for AI sovereignty while the Pentagon hands every classified contract to the hyperscalers. Plus open source: OpenClaw 2026.4.29 ships active-run steering and a people-aware memory wiki, and DeepSeek V4 lands as a credible open-source step away from Nvidia.

OpenAI and Microsoft restructure their partnership, ending Azure exclusivity and capping the revenue share through 2030. China's NDRC orders the cancellation of Meta's $2 billion acquisition of agentic AI startup Manus, citing the Anti-Monopoly Law. Google signs a Pentagon agreement for unrestricted AI use on classified networks after Anthropic refused on grounds of mass surveillance and autonomous weapons. A Ming-Chi Kuo note says OpenAI is co-designing an AI-native phone with Luxshare, Qualcomm, and MediaTek, with mass production targeted for 2028. Wired reports more than 700 Covalen workers in Dublin, including 500 Meta AI annotators, are at risk of layoffs as Meta deploys more advanced AI for content enforcement. Plus: OpenClaw v2026.4.26 ships browser realtime voice transport, Cerebras as a bundled provider, and asymmetric embedding endpoint support.

Google announces an investment of up to $40 billion in Anthropic in cash and committed cloud compute, the second hyperscaler-scale AI deal of the week. DeepMind spinoff Isomorphic Labs moves its first AI-designed drug candidates into human trials. DeepSeek previews a new model the company says closes the gap with U.S. frontier labs. SpaceX hits 600 Falcon 9 landings and increasingly operates as an AI company across landings, mission planning, and Starlink. Wired argues Apple's next CEO must ship a real AI product or get compared to Microsoft circa 2008. Plus: OpenClaw v2026.4.23 brings OAuth-based image generation through Codex and forked subagent contexts.

Anthropic takes a fresh $5B from Amazon and commits $100B in cloud spend, locking in AWS as its primary infrastructure partner for years to come. A new report says Meta will train internal AI agents by tracking employee mouse and keyboard activity. Google Maps is getting a conversational AI makeover. Sam Altman publicly calls Anthropic's cyber-focused Mythos model fear-based marketing. OpenAI ships a real upgrade to ChatGPT's image generator. Plus: OpenClaw v2026.4.21 tightens owner-auth and makes gpt-image-2 the default image provider.