Key Highlights

  • GitHub’s internal repos were exfiltrated through a poisoned VS Code extension. Microsoft confirmed a compromised employee device — via the malicious NX Console extension — pulled an estimated 3,800 internal repos. Theo (t3.gg) and security firm Aikido lay out the kill chain: a contributor’s GitHub token stolen in the earlier Shai-Hulud worm wave was used to publish a malicious version that auto-updated to ~2.2M installs in 18 minutes. The marketplace has no staging window, no audit gate, and no takedown push. This is now the second VS Code extension–driven supply chain breach in six months (after Async API in Nov 2025), and credentials harvested by that earlier worm are still being weaponized.
  • The Trump White House paused its AI security executive order over the requirement that frontier labs share models with federal evaluators 14–90 days before launch. The order was triggered by Anthropic’s Mythos and OpenAI’s GPT-5.5 Cyber — both of which can autonomously find and exploit security flaws — but Trump said the pre-release review language “could have been a blocker” and that he doesn’t want anything in the way of “leading China.” Expect a softer redraft and a continued split between national-competitiveness framing and pre-deployment evaluation regimes.
  • Two Day-After-I/O takes converged on the same thesis: Google has the platform but not the execution. The Rundown’s Pichai interview pitches agents as flip-phone-obvious within three years; Theo’s “I’m scared to make this video” walks through Gemini 3.5 Flash’s tripled per-token price, its 2× cost-to-completion vs. 3.1 Pro on real agentic tasks, the closure of the open-source Gemini CLI in favor of a closed Antigravity CLI, and Google Cloud abruptly suspending Railway’s $2M/month account. The pattern shared across both: capability gains are real, but distribution/trust is fraying.
  • OpenAI says a general-purpose reasoning model disproved an 80-year-old Erdős conjecture on unit distances — a novel discrete-geometry result reviewed by Tim Gowers and Noga Alon. Notable because the work came from an upcoming general model, not a math-specialist system like AlphaProof.
  • Sam Altman (with Patrick Collison) and Jensen Huang (with Michael Dell) gave essentially the same diagnosis in different words this week: coding models inflected hard in late 2025, demand has gone “parabolic,” and compute per task is up 100–1000× because agents now plan-act-iterate instead of one-shot replying. Altman thinks OpenAI can stay near-flat on headcount (2× over five years) while scaling output; Huang says Vera CPU + Rubin/Blackwell racks are the substrate for “unmetered intelligence” on-prem.

Analysis & Opinion

Trump delays AI security executive order, saying language “could have been a blocker” — TechCrunch

Trump postponed signing an executive order that would have required AI labs to share advanced models with the Office of the National Cyber Director and partner agencies between 14 and 90 days before public launch. The order was prompted by recent releases from Anthropic (Mythos) and OpenAI (GPT-5.5 Cyber), both capable of rapidly identifying and exploiting security vulnerabilities — exactly the dual-use frontier the EO was meant to address. Trump’s stated reason: he doesn’t want anything “to get in the way” of US leadership over China. CNN reported the unofficial reason was scheduling — too few tech CEOs could fly in for the signing ceremony — but the substantive disagreement over pre-release review remains unresolved. The outcome will likely set the template for how the second Trump administration handles frontier-AI oversight: voluntary commitments and red-team reporting rather than statutory pre-deployment review. Watch for a redraft with the disclosure window stripped or relaxed to post-launch reporting.

Six search engines worth trying now that Google isn’t really Google anymore — TechCrunch

Reid’s I/O 2026 announcement reframing the search box as the “biggest upgrade in 25 years” — AI Mode by default, chat-style AI Overviews — drew a sharper backlash than Google’s previous AI rollouts, with users explicitly citing it as the trigger to switch. The piece runs through Kagi, DuckDuckGo, Startpage, the &udm=14 web-only override, Brave, and Ecosia as practical alternatives, framed against the 2024 District Court monopoly ruling that’s already eroding default-engine inertia.

Coding agents are giving everyone decision fatigue — Stack Overflow Blog

Stack Overflow argues the engineer’s job has quietly shifted from writing code to “structuring prompts and reviewing code” — and that the cognitive load of constant micro-decisions about agent output is the under-discussed cost of the current generation of tools. (Pairs naturally with Cursor’s cloud-agent retrospective below.)

Google is pitching an AI agent ecosystem to consumers who may not buy it — TechCrunch

A skeptical read of I/O 2026’s consumer-facing announcements — Information Agents, Gemini Spark, Android Halo, Daily Brief — most of which are gated behind the $100/month Ultra tier. The author’s read: Google demonstrated capability without a clear use case the median user is asking for, and the price tag suggests Google itself is unsure of the demand curve.

Building Token-Metered AI Services on Telco AI Factories — NVIDIA Developer

A blueprint for telcos to shift from selling GPU-hours to selling tokens, with three-layer architecture (compute / metering & billing / application services). The headline economics: a single H100 generating 30M billable tokens/hour at $1/M tokens yields ~$157,680/year versus ~$18,400 on hourly rental — almost a 9× revenue uplift. The piece is interesting less as NVIDIA pitch than as evidence the inference-as-utility business model is hardening into a reference architecture.

New Products & Tools

What we’ve learned building cloud agents — Cursor

A year-in retrospective on Cursor’s cloud agents. The core takeaway: cloud agents aren’t a remote port of local agents — they need full developer environments (dependencies, network access, parallel execution) and the team ended up building “an operating layer” around the agent rather than just hosting it. The single biggest quality lever they identified is environment parity with a real developer’s machine.

Here’s what developers can do with the latest Google Play updates — Google

Play Shorts (app previews), Ask Play (conversational app discovery), Gemini-app surfacing of Android apps, broader Engage SDK reach, and Play Games Sidekick (in-game overlay with tips/rewards/social). Mostly distribution plumbing for the agentic surface Google is building above it.

We’re launching the Google DeepMind Accelerator program in Asia Pacific — Google DeepMind

DeepMind’s first APAC Accelerator — a three-month program for startups, research orgs, and nonprofits working on AI for nature conservation, climate adaptation, agriculture, and renewable energy. Mentorship plus technical access.

NVIDIA GTC Taipei at COMPUTEX: Live Updates — NVIDIA

Rolling coverage from Computex. COMPUTEX 2026 Best Choice Awards went to Vera Rubin NVL72 (Golden + Sustainability), Jetson Thor (edge AI/robotics), and the Alpamayo AV platform. Huang’s keynote is June 1.

Get Real-Time Visibility into GPU Usage Across Kubernetes Clusters — NVIDIA Developer

Open-source GPU Usage Monitor — a Helm chart that bundles DCGM Exporter, kube-state-metrics, Prometheus, and Grafana with pre-built GPU dashboards. Targets the two failure modes of GPU-on-K8s shops: idle over-provisioning and silently queued pod starvation.

Spotify and Universal Music strike deal allowing fan-made AI covers and remixes — TechCrunch

Premium add-on for fan-generated AI covers and remixes, with a revenue share to participating artists. Spotify is explicitly framing this against Suno and Udio’s litigation history (Suno settled with Warner for $500M last November) — consent, credit, and compensation as the differentiator.

Spotify adds AI-powered Q&A and briefing generation features to podcasts — TechCrunch

Personalized prompt-driven podcast briefs (daily/weekly), an “Ask YouTube”-style Q&A on episodes for US/Sweden/Ireland Premium mobile, and creator-side monetization (sponsorships, paid subscriptions).

Spotify launches an ElevenLabs-powered audiobook creation tool — TechCrunch

ElevenLabs-generated audiobooks, beta in June, English-only at launch, expanding to ten more languages. Authors aren’t bound to exclusive contracts.

Hark raises $700M Series A for its “universal” AI interface — TechCrunch

Brett Adcock’s third major company (after Figure.AI and Archer) raised $700M at a $6B post-money — NVIDIA, AMD Ventures, Intel Capital, Qualcomm Ventures, ARK, Parkway leading. Multimodal models this summer, dedicated hardware after. Notable for the consumer focus where most peers are chasing enterprise.

The Path: safer AI therapy — TechCrunch

$14.3M seed (Prime Movers Lab) for an AI therapy app from Calm and Mental alumni, with Tony Robbins as co-founder. Framed around safety guardrails for a category that has been chronically underprepared for its userbase.

Aluminum recycling startups bet on AI to cash in — TechCrunch

With aluminum prices up ~20% on Gulf supply disruption and US designation as a critical mineral, sorting startups like Sortera are doubling capacity using laser/camera/XRF sensor fusion feeding classification models. Current US recovery rate is only ~20%.

Research

OpenAI Cracks an 80-Year Math Belief — The Rundown

OpenAI claims one of its reasoning models independently produced a counterproof to a 1946 Erdős conjecture on unit distances between grid points — a foundational discrete-geometry result. The proof draws on algebraic number theory and was validated by Tim Gowers and Noga Alon. The flag this plants: the discovery came from an upcoming general-purpose model, not a math-specialist system like DeepMind’s AlphaProof.

Dissecting ThunderKittens: Anatomy of a Compact DSL for High-Performance AI Kernels — Lobsters

Deep dive into Stanford Hazy Research’s ThunderKittens, an embedded CUDA DSL that finds a middle ground between full hardware exposure and Triton-style abstraction — abstracting tile layouts and tensor-core descriptors while leaving data movement and scheduling visible.

Antigravity 2.0 Tops the OpenSCAD Architectural 3D LLM Benchmark — ModelRift

Practical bench: six tools (Codex 5.5 High, Claude Sonnet, Claude Opus, Cursor Composer, Google Antigravity, ModelRift) generating parametric OpenSCAD of the Pantheon from reference photos. Worth noting alongside Theo’s much harsher take on the Antigravity stack below — benchmark wins and real-world agentic-coding usability are diverging.

Automating and Optimizing Financial Signal Discovery with Multi-Agent Systems — NVIDIA Developer

NeMo Agent Toolkit coordinating signal / code / evaluation agents to discover quant trading signals. YAML-driven workflows let researchers swap model sizes per role; demo discovers momentum-based signals on S&P 500 data using Information Coefficient as the eval metric.

Unlock Exascale Performance on NVIDIA GB200 NVL72 with Slurm Topology-Aware Job Scheduling — NVIDIA Developer

NVIDIA + SchedMD’s new topology/block plugin in Slurm 23.11 — schedules around NVLink rack topology rather than fragmenting jobs across racks. Validated by 7-day simulation across 5,000 nodes / 20,000 GPUs / 15,000 jobs.

Interviews & Conversations

“This is bad…” — GitHub’s compromise via a poisoned VS Code extension — Theo - t3.gg (27 min)

Theo walks through GitHub’s confirmation that an employee device was compromised by a malicious NX Console VS Code extension (since confirmed by Narwhal CEO), leading to the exfiltration of ~3,800 internal repos. The mechanics are worth understanding: a contributor’s GitHub token was stolen during the earlier Shai-Hulud worm wave, used to publish a poisoned 2.2M-install extension that auto-updated everywhere in an 18-minute window before takedown. Microsoft’s marketplace has no staging buffer between publish and auto-install, no audit gate, no email notification under 6 minutes, and — most critically — no rollback push to clients who already pulled the bad version. (Async API’s November 2025 compromise infected machines for a month after takedown because clients kept the malicious version installed.) Theo’s three asks of npm/Microsoft are concrete and reasonable: automatic agent-based audit on update for any package above a usage threshold; a mandatory staging delay before clients auto-install; and an actual takedown-push mechanism. The deeper point: Shai-Hulud harvested an unknown quantity of valid tokens, an agent can be churning through them in the background indefinitely, and “auto-update” — the security industry’s mantra for a decade — is now the primary attack surface.

“I’m scared to make this video” — Google’s I/O misses — Theo - t3.gg (24 min)

A pointed takedown of three I/O 2026 decisions: Gemini 3.5 Flash, the new Antigravity CLI, and Google Cloud’s suspension of Railway. On 3.5 Flash: the headline benchmark wins (Terminal Bench, SWE Pro, MCP Atlas) hide that per-token prices roughly tripled vs. 3 Flash ($1.50 in / $9 out per M), token efficiency is poor (72M tokens vs GPT-5.5 Medium’s 22M on the same artificial-analysis suite), and the actual end-to-end cost on real agentic tasks lands at ~2× Gemini 3.1 Pro. He couldn’t get it to complete a working version of a Fish Slop game rebuild that GPT-5.5 finished and then extended to 3D on follow-up. The Antigravity CLI replaces the open-source Gemini CLI (100K stars, 6K merged PRs) with a closed Go rewrite; subscription users will lose access to Gemini CLI on June 18. Railway’s $2M/month Google Cloud account was suspended without warning, taking the service offline — Theo connects it to the precedent of UniSuper’s full account deletion in 2024. The throughline: capability gains are real, but the operating model around them at Google is corroding the trust that makes them usable.

Sam Altman in conversation with Patrick Collison — Stripe (57 min)

Wide-ranging conversation at Stripe Sessions. Altman places the late-2025 coding-model inflection at a specific training mix change — verifiable reasoning rewards pulled forward — and says the metrics curves at OpenAI “went parabolic” starting January. On headcount, he wants OpenAI at 2× current size in five years, not 10×, on the bet that the same tools they’re shipping will let them stay near-flat per output dollar. On power concentration: defends iterative deployment as the reason OpenAI exists in its current form (“an ivory tower discovering the gift and sharing the fruits” was the consensus alternative they explicitly rejected). On what excites him outside of model/product: data-center infrastructure, energy, robotics, brain-machine interfaces (slow but real progress), defensive biotech, and materials science — which he calls “an AI-shaped problem” the discourse undercovers. Predicts profitable fusion within five years and Mach-4 commercial flight within a decade or so. The non-obvious through-line: he’s noticeably more interested in physical-layer compounding than in model-layer benchmarks.

Jensen Huang on The Future of Computing | AI, Infrastructure & What’s Next — Dell Technologies (13 min)

Huang’s Dell Tech World pitch alongside Michael Dell. The argument is structural: generative → reasoning → planning → agentic, and agentic workloads are 100–1000× more compute-intensive than chat because they iterate plan-act-tool-evaluate-replan instead of replying once. Cue the “parabolic” demand framing — same word Altman used at Stripe Sessions days earlier. Architecture pitch: NVL72-scale brains running 1T-parameter models, with agent harnesses (NeMo Cloud) running locally on Vera CPUs inside NVIDIA’s open-sourced sandbox, plus confidential computing so the operator never sees your data. Vera is explicitly designed against the hyperscaler CPU-cores-as-product model — they’re optimizing single-thread + 3× memory bandwidth because agents pound databases and the GPU sits idle if the CPU can’t keep up. Token generation is now the unit of business.

References

  1. The Rundown, “Exclusive Interview: Sundar Pichai on AI’s Flip Phone Moment,” The Rundown AI, 2026-05-22 [blog]
  2. The Rundown, “OpenAI Cracks an 80-Year Math Belief,” The Rundown AI, 2026-05-21 [blog]
  3. Google, “Here’s what developers can do with the latest Google Play updates,” The Keyword, 2026-05-21 [blog]
  4. Google DeepMind, “We’re launching the Google DeepMind Accelerator program in Asia Pacific to tackle environmental risks,” DeepMind Blog, 2026-05-21 [blog]
  5. Cursor, “What we’ve learned building cloud agents,” Cursor Blog, 2026-05-21 [blog]
  6. NVIDIA, “NVIDIA GTC Taipei at COMPUTEX: Live Updates on What’s Next in AI,” NVIDIA Blog, 2026-05-21 [blog]
  7. NVIDIA, “Automating and Optimizing Financial Signal Discovery with Multi-Agent Systems,” NVIDIA Developer Blog, 2026-05-21 [blog]
  8. NVIDIA, “Get Real-Time Visibility into GPU Usage Across Kubernetes Clusters,” NVIDIA Developer Blog, 2026-05-21 [blog]
  9. NVIDIA, “Unlock Exascale Performance on NVIDIA GB200 NVL72 with Slurm Topology-Aware Job Scheduling,” NVIDIA Developer Blog, 2026-05-21 [blog]
  10. NVIDIA, “Building Token-Metered AI Services on Telco AI Factories,” NVIDIA Developer Blog, 2026-05-21 [blog]
  11. Ryan Donovan, “Coding agents are giving everyone decision fatigue,” Stack Overflow Blog, 2026-05-21 [blog]
  12. ModelRift, “Antigravity 2.0 Tops the OpenSCAD Architectural 3D LLM Benchmark,” ModelRift Blog, 2026-05-21 [blog]
  13. Hamza El Shafie, “Dissecting ThunderKittens: Anatomy of a Compact DSL for High-Performance AI Kernels,” via Lobsters, 2026-05-22 [blog]
  14. Sarah Perez, “Spotify and Universal Music strike deal allowing fan-made AI covers and remixes,” TechCrunch, 2026-05-21 [blog]
  15. TechCrunch, “Six search engines worth trying now that Google isn’t really Google anymore,” TechCrunch, 2026-05-21 [blog]
  16. TechCrunch, “Trump delays AI security executive order, saying language ‘could have been a blocker’,” TechCrunch, 2026-05-21 [blog]
  17. TechCrunch, “Spotify adds AI-powered Q&A and briefing generation features to podcasts,” TechCrunch, 2026-05-21 [blog]
  18. TechCrunch, “Spotify launches an ElevenLabs-powered audiobook creation tool,” TechCrunch, 2026-05-21 [blog]
  19. TechCrunch, “The Path, founded by Tony Robbins and Calm alums, hopes to offer safer AI therapy,” TechCrunch, 2026-05-21 [blog]
  20. TechCrunch, “Hark raises $700M Series A for its secretive ‘universal’ AI interface,” TechCrunch, 2026-05-21 [blog]
  21. TechCrunch, “Google is pitching an AI agent ecosystem to consumers who may not buy it,” TechCrunch, 2026-05-21 [blog]
  22. TechCrunch, “With aluminum prices up 20%, recycling startups bet on AI to cash in,” TechCrunch, 2026-05-21 [blog]
  23. Theo Browne, “This is bad…,” Theo - t3.gg, 2026-05-21 [video]
  24. Theo Browne, “I’m scared to make this video,” Theo - t3.gg, 2026-05-20 [video]
  25. Sam Altman with Patrick Collison, “Sam Altman in conversation with Patrick Collison,” Stripe, 2026-05-19 [video]
  26. Jensen Huang with Michael Dell, “Jensen Huang on The Future of Computing | AI, Infrastructure & What’s Next,” Dell Technologies, 2026-05-19 [video]