Key Highlights

  • Google I/O 2026 was an all-agents show. Gemini 3.5 Flash launched at ~4× the speed of competing frontier models, Gemini Omni generates video from any combination of text/image/audio, Gemini Spark becomes a 24/7 cloud-resident personal agent across Gmail/Docs/Workspace, and Antigravity 2.0 + a new Managed Agents API let developers spin up sandboxed Linux environments through a single Gemini API call.
  • AI-content provenance got a real industry handshake. OpenAI is adopting Google’s SynthID invisible watermark and C2PA Content Credentials; Google has now watermarked 100B+ images/videos and 60,000 years of audio, with verification rolling out to Search, Chrome, and Pixel cameras.
  • Andrej Karpathy joined Anthropic’s pre-training team to use Claude to accelerate Claude’s own training — a clear bet that AI-assisted research, not raw compute, is the next moat.
  • A California jury dismissed Musk’s $100B+ suit against OpenAI on statute-of-limitations grounds, not the merits; Musk is appealing and warning the ruling sets a “nonprofit-to-for-profit looting” precedent. Across two interviews this week he also predicted digital intelligence will exceed all human intelligence within ~5 years.
  • Theo (t3.gg) ran a self-exposing 50-session attack on GitHub Copilot’s $40 plan, burning Microsoft an estimated $15–46K of inference using cryptography puzzles that kept GPT-5.4 running 16 hours per “message” — and used the demo to argue this is why every agentic coding tool has had to abandon message-based billing.

Analysis & Opinion

Musk’s OpenAI case runs out of time — Rundown

Musk’s lawsuit alleging Altman and Brockman converted OpenAI from charity to ~$800B for-profit was thrown out as time-barred, not adjudicated on its core claim of unjust enrichment. OpenAI’s defense: Musk himself backed the for-profit pivot and only sued after launching xAI. Musk’s appeal argument, which he repeated at a Forbes dinner the next day, is that the precedent enables founders to “start a nonprofit, take charity money, then flip to for-profit once it’s successful” — a structural risk to American charitable giving, in his framing.

Google Search as you know it is over — TechCrunch

Google reframed Search around an “intelligent search box” that takes long conversational queries, suggests follow-ups, and routes into AI Mode by default — the biggest change to the box in 25 years. Background “information agents” will run 24/7 to track stocks, flights, sports, housing, and notify users when something changes. The ten-blue-links era is being explicitly retired.

An update on recent Claude Code quality reports — Anthropic Engineering

Anthropic published a multi-bug postmortem for Claude Code’s March–April quality regressions. A March 4 default-reasoning change from “high” to “medium” introduced latency and made the tool feel less capable (reverted April 7), and a March 26 caching bug silently cleared thinking history every turn, producing the “forgetful Claude” complaints that took nearly two weeks to root-cause. Worth reading for the reminder that small infrastructure changes in agentic stacks can mimic model-quality regressions and be very hard to attribute.

New Products & Tools

Google I/O 2026 — the agentic Gemini stack

A single I/O released a stack rather than a flagship. Grouping the major shipping pieces:

  • Gemini 3.5 Flash (TechCrunch coverage) — frontier-class intelligence at ~4× competitor speed, beats Gemini 3.1 Pro on most benchmarks, and was used internally to bootstrap an entire OS from scratch via Antigravity agents. 3.5 Pro arrives next month.
  • Gemini Omni Flash — multimodal video generation from any input combo, with consistent characters across edits and SynthID embedded in every output. Anti-deepfake gating requires users to record themselves saying numbers before generating their own avatar.
  • Gemini Spark — cloud-resident personal agent with deep Gmail/Docs/Workspace hooks and its own email address; Android Halo gives an at-a-glance status pill at the top of the phone.
  • Antigravity 2.0 — desktop app, CLI, and SDK for running parallel coding subagents, replacing the Gemini CLI.
  • Managed Agents in the Gemini API — one API call provisions an ephemeral Linux sandbox with persistent state, tool use, and AGENTS.md/SKILL.md customization. Ramp, ResembleAI, and Stitch were launch partners.
  • Universal Cart + Agent Payments Protocol — cross-merchant cart spanning Search, Gemini, YouTube, Gmail; AP2 lets agents transact within user-set caps, brand allowlists, and audited spending limits.
  • Smart glasses with Warby Parker + Gentle Monster — audio glasses this fall, display variant later. Direct shot at Meta’s Ray-Ban lead.
  • AI Studio → native Android apps — Kotlin + Jetpack Compose generation in-browser with an embedded Android emulator and one-click Play Console upload. Plus voice-first authoring in Docs, Keep, Gmail (“Gmail Live”).
  • New $100/mo AI Ultra tier with 5× Antigravity quota; the existing top tier drops $250 → $200.

Cursor Composer 2.5 — Cursor

Cursor’s in-house model now trains on 25× more synthetic tasks than Composer 2 and uses a targeted-feedback RL setup that injects feedback at the exact trajectory point where the model could have done better. Pricing lands at $0.50/$2.50 per million input/output tokens — substantially cheaper than the major frontier models for coding.

OpenAI adopts C2PA + SynthID for image provenance — TechCrunch / OpenAI

OpenAI is stamping its image outputs with both the C2PA Content Credential (visible metadata) and Google’s SynthID (invisible, harder to strip) and shipping a public verifier. The pairing matters because C2PA metadata is trivially removable while SynthID survives screenshots and resizes. The cross-lab adoption of Google’s watermark is arguably the most consequential AI-safety win of the week — provenance only works if it’s an industry standard, not a single-vendor signal.

OpenAI co-founder Andrej Karpathy joins Anthropic’s pre-training team — TechCrunch

Karpathy will report to Nick Joseph and lead work on using Claude to accelerate large-scale pre-training research. The hire is a signal: Anthropic is betting that AI-assisted research, not just GPU spend, decides the next model generation — and Karpathy’s combination of theoretical depth and large-scale training experience is rare enough to move the needle for a frontier lab. It also continues a pattern of high-profile OpenAI alumni landing at competitors.

Research

Simulating real places with Project Genie + Street View — Google DeepMind (TechCrunch)

Genie 3 can now ground generated worlds in any U.S. Street View location, with Waymo already using it to train autonomous vehicles on rare events like tornadoes. Still flagged as experimental on physics accuracy.

Fast-tracking genetic leads to reverse cellular aging — Google DeepMind

DeepMind’s Co-Scientist surfaced 20+ novel genetic factors for cellular rejuvenation; wet-lab validation confirmed several reverse senescence markers. The researchers say analysis that previously took six months now finishes in a day.

Interviews & Conversations

Nvidia’s Huang & Michael Dell on Agentic AI, Memory Demand, and China — Bloomberg Television (20:29)

Huang and Dell’s joint Bloomberg appearance was the clearest articulation yet of why enterprise AI is moving on-prem: “intelligence has to be produced at the point of context” — proprietary data, secure data, and domain skills all sit inside the enterprise. Dell sees 1,000 new AI-server customers in the quarter and frames Vera CPU + Grace Blackwell as the “brain + harness” for agentic workflows, with billions of agents using tools 24/7 driving CPU demand alongside GPU demand. On memory: both confirmed the textbook “boom and bust” framing no longer applies, with Huang noting he laid out the exact roadmap to Micron and SK Hynix three years ago and they’re investing accordingly. Supply is “quadrupling annually” and still can’t keep up — Huang projects at least a decade of buildout before physical-AI infrastructure even begins. On China: H200 export licenses exist; Huang says Beijing signaled openness, but no Taiwan-supply-chain reassurance was discussed.

Elon Musk’s “Jesus-level technology” interview — Farzad (11:57) + Forbes 250 dinner — Forbes (26:00)

Two Musk appearances this week largely converged on the same outlook. Tesla FSD running with no safety monitors in three Texas cities, expected to cover most of the U.S. by year-end. Neuralink moves to “blindsight” implants for fully blind patients this year and is targeting motor-cortex bypass to restore walking for spinal-injury patients — what Musk repeatedly framed as “Jesus-level miracles.” His 5-year predictions: digital intelligence exceeds the sum of human intelligence, 100M–1B humanoid robots deployed, economy ~2× current size. On the OpenAI verdict (Forbes), he echoed the Rundown’s framing — the ruling was a statute-of-limitations dismissal, not a merits decision, and he sees the appeal as necessary to prevent “nonprofit-to-for-profit looting” as a replicable playbook. He also flagged data centers in space, tunnels, electric aircraft, and programmable RNA medicine as underrated frontiers.

I exploited Copilot and burned $46,000 (it cost $40) — Theo / t3.gg (39:05)

Theo’s 39-minute breakdown of GitHub Copilot’s billing economics is the best public explainer of why agentic coding tools cannot survive message-based pricing. He ran 50 parallel sessions of GPT-5.4-extra-high against deliberately unsolvable cryptography puzzles on his $40 Copilot Plus plan, getting individual “messages” to consume 111M input + 1.6M output tokens over 16+ hours of inference — roughly $62 per request after caching. Multiplied across his 1,500-message monthly quota: $93K of inference per user for $40 of revenue. His core technical argument is sharp: when models could only do one step per message, message-billing worked; agentic loops with unlimited tool calls broke the model years ago and every other vendor — Claude Code, Codex, Cursor — already moved to rate-limit or token-credit models. Copilot’s June 1 switch to token-based credits isn’t a “rug pull,” it’s the inevitable correction, and Theo argues users who feel scammed never had a sustainable deal in the first place. The behind-the-scenes commentary on T3 Chat’s own pivot away from message billing (after one user did $2K of inference in five days) is a useful case study for anyone shipping AI products.

Tom Lee — markets, Fed transition, and crypto — Savvy Finance (16:14)

Brief macro entry, not AI-focused. Lee sees turbulence later in 2026 driven by a new Fed chair (11 of 13 first-year transitions have triggered ≥10% drawdowns), tight oil markets, and pricing-in of private credit stress, but stays constructive on software and structurally bullish on tokenization. The provocative data point: Jane Street is on pace for ~$40B profit in 2026 with 3,000 employees, vs JPMorgan’s $60B with 300,000 — and Tether expects $15B with 300 employees. His prediction: five of the world’s ten most profitable banks will be digitally native in a decade.

References

  1. Musk’s OpenAI case runs out of time — Rundown, 2026-05-19 [blog]
  2. Google Search as you know it is over — TechCrunch, 2026-05-19 [blog]
  3. An update on recent Claude Code quality reports — Anthropic, 2026-05-19 [blog]
  4. Gemini 3.5: frontier intelligence with action — Google, 2026-05-19 [blog]
  5. With Gemini 3.5 Flash, Google bets its next AI wave on agents — TechCrunch, 2026-05-19 [blog]
  6. Introducing Gemini Omni — Google, 2026-05-19 [blog]
  7. Google introduces Gemini Spark — TechCrunch, 2026-05-19 [blog]
  8. Google launches Antigravity 2.0 — TechCrunch, 2026-05-19 [blog]
  9. Introducing Managed Agents in the Gemini API — Google, 2026-05-19 [blog]
  10. Universal Cart + Agent Payments Protocol — TechCrunch, 2026-05-19 [blog]
  11. Google audio-powered smart glasses — TechCrunch, 2026-05-19 [blog]
  12. Google AI Studio → native Android apps — TechCrunch, 2026-05-19 [blog]
  13. Google’s AI now lets you talk to your Gmail inbox — TechCrunch, 2026-05-19 [blog]
  14. Everything new in our Google AI subscriptions — Google, 2026-05-19 [blog]
  15. Introducing Composer 2.5 — Cursor, 2026-05-18 [blog]
  16. OpenAI adopts C2PA + SynthID for image provenance — TechCrunch, 2026-05-19 [blog]
  17. Andrej Karpathy joins Anthropic’s pre-training team — TechCrunch, 2026-05-19 [blog]
  18. Project Genie + Street View — TechCrunch, 2026-05-19 [blog]
  19. Fast-tracking genetic leads to reverse cellular aging — Google DeepMind, 2026-05-18 [blog]
  20. Nvidia’s Huang, Michael Dell on Agentic AI, Memory Demand and China — Bloomberg Television, 2026-05-18 [video]
  21. Elon Musk’s Inventing ‘Jesus-level’ Technology (Miracles) — Farzad, 2026-05-18 [video]
  22. Elon Musk’s First Interview Since Jury Rejected Claim Against Altman’s OpenAI — Forbes, 2026-05-19 [video]
  23. I exploited Copilot and burned $46,000 (it cost $40) — Theo - t3.gg, 2026-05-19 [video]
  24. Tom Lee - You’re Not READY for What’s Coming for BTC & Crypto — Savvy Finance, 2026-05-18 [video]