Key Highlights

  • The Anthropic export-control saga grinds into its 11th day with no model back online. Theo’s latest video walks through the escalation — a customer lawsuit against the U.S. government, a bipartisan Congressional demand for transparency (response due June 26), leaks of stalled negotiations, and the uncomfortable fact that open-weight GLM-5.2 sits just below the capability line that got Fable 5 and Mythos 5 banned.
  • Karpathy declared a “third paradigm” of LLM UX — and it’s a Slack bot. Anthropic’s Claude Tag turns Claude into a persistent, multiplayer, channel-scoped teammate; Anthropic says 65% of its product team’s code now comes from its internal version. Both a TechCrunch report and a Theo video dig into why channel-level context might be the right abstraction nobody had found yet.
  • The “are the unit economics real?” question is getting loud. A widely-shared analysis pegs AI subsidies at up to 70× for OpenAI enterprise customers, TechCrunch reports companies are now rationing employee AI budgets, and Cerebras stock plunged on margin worries — even as OpenAI unveiled its first custom inference chip (Jalapeño, built by Broadcom) and Amazon committed $13B more to India AI infrastructure.
  • The “AI kills engineering jobs” narrative took a data-driven hit: SignalFire figures suggest engineers are actually a growing share of new hires, while Coinbase reports agents now write three-quarters of its pull requests and cut idea-to-production time by 90%.

Analysis & Opinion

AI’s Affordability Crisis — David Rosenthal (HN)

A blunt accounting of how heavily AI platforms subsidize usage to manufacture demand. The numbers are stark: on a $200/month plan, a user could burn through roughly $8,000 in Anthropic tokens or $14,000 in OpenAI tokens, and SemiAnalysis estimates Anthropic subsidizes enterprise customers up to 40× and OpenAI up to 70×. OpenAI’s 2025 financials reportedly showed $13.07B in revenue against $34B in costs — a $20.92B operational loss, with 44% of revenue going to sales and marketing. The piece argues this is not a path to durable profitability but a land-grab that someone eventually has to pay for, and frames the looming price corrections as a systemic risk to everyone who has built on top of subsidized inference.

Companies are scrambling to stop employees from maxing out AI budgets — TechCrunch

The flip side of subsidized tokens: now that real costs are surfacing, “the tokenmaxxing era was brief — we now appear to be entering the era of token rationing.” Companies are putting cost controls in place as usage patterns reveal employees burning premium inference on trivial tasks. Read alongside the affordability piece, it’s a concrete sign that the cheap-AI honeymoon is ending and that org-level governance of AI spend is becoming a real line item.

A Pragmatic Approach to AI Governance in America — Google

Google stakes out a deliberate middle ground in a new policy white paper, rejecting the binary between heavy oversight and no rules. For frontier models it proposes “an independent, federally overseen, industry-backed organization” to set safety standards and verify voluntary audits — explicitly modeled on how other critical sectors self-regulate under government supervision. For widely-deployed AI applications, it argues for updating existing laws (child protection, IP, employment) rather than writing sweeping new ones. The framing lands pointedly this week, against the backdrop of the Anthropic export ban that has both parties questioning the government’s evidentiary standards for restricting frontier models.

For most of the world, open-source AI is the only way forward — Yann LeCun (via HN)

At UN Open Source Week, Meta’s former chief AI scientist argued open-source AI is the only viable route to “AI sovereignty” for most countries. His thesis: AI is becoming an essential infrastructure layer mediating humanity’s digital interactions, and letting a handful of proprietary Western and Chinese systems dominate that layer is a civilizational risk. Representatives from Morocco, Sierra Leone, and Jamaica backed him, casting open weights as the only way Global South nations move from AI consumers to participants. The argument resonates uncomfortably with the export-ban story — where the most-restricted models are precisely the most capable, leaving open weights as the default for everyone the U.S. won’t sell to.

Anthropic says Alibaba illicitly extracted Claude AI model capabilities — Reuters (via HN)

Anthropic has publicly accused Alibaba of unauthorized extraction (distillation) of Claude’s capabilities — a charge that drew heavy discussion (509 points, 866 comments on HN). It crystallizes a core model-security tension: capable closed models are valuable precisely because their behavior can be cheaply copied into cheaper students, and the line between legitimate API use and IP exfiltration is hard to police. The accusation also feeds directly into the national-security framing driving the current export controls.

Coinbase reduces time from idea to production by 90% with Cursor — Cursor

Coinbase reports a wholesale shift to an “agent-first” model: 2,400+ developers on Cursor, agents authoring three-quarters of all PRs, deployment timelines cut from 20 days to under 2, and engineers merging 55% more PRs annually. The notable claim isn’t the tooling but the org redesign — leadership argues the real bottleneck was broken process, not missing AI, so they rebuilt sprint planning (claim-and-delegate instead of assignment cycles) and shifted engineers toward architecture and outcome validation over line-by-line review. A real-world instance of the “the constraint moved off code-writing” thesis.

Shorter takes

New Products & Tools

Anthropic’s Claude Tag is learning your company, one Slack message at a time — TechCrunch

Anthropic’s new Claude Tag joins Slack as a multiplayer team member: one Claude per channel, shared across everyone in it, building channel-scoped memory and able to plan, work asynchronously over hours or days, and schedule its own tasks. It’s gated to team and enterprise plans. (See the Interviews section for Theo’s deeper take on why channel-level context may be the right primitive.)

Introducing computer use in Gemini 3.5 Flash — Google

Computer use is now a native capability in Gemini 3.5 Flash, letting developers build agents that “see, reason and take action” across browser, mobile, and desktop. Google pairs it with adversarial training against prompt injection and optional enterprise safeguards, and targets long-horizon automation like continuous software testing.

OpenAI unveils its first custom chip, built by Broadcom — TechCrunch

Named Jalapeño, the processor is purpose-built for OpenAI’s inference systems — a vertical-integration move that, alongside Amazon’s and NVIDIA’s announcements, underscores how the AI compute race is increasingly about owning the silicon.

How Notion used the Cursor SDK to embed coding agents — Cursor

Notion embedded Cursor directly into its product — tag Cursor in a doc, thread, or issue and it plans, builds, tests, and opens a PR — built “from nothing to a full integration in a couple of weeks” on the Cursor SDK, which provides the same models and runtime as Cursor’s native app.

Also shipping

Research

Towards a world where no one is surprised by a natural disaster — Google

A decade-in-review of Google’s crisis-resilience work, charting the move from surfacing helpful information to actively forecasting and detecting wildfires, floods, earthquakes, and extreme weather with ML at global scale.

Shorter findings

Interviews & Conversations

Is it ever coming back? — Theo - t3.gg (20:00)

(Transcript-based summary.) A status report on the Anthropic export ban, now 11 days in with no model and no timeline. Theo traces the chain of events: a White House request (June, pre-ban) to revoke SK Telecom’s access over alleged reselling to China, which Anthropic immediately complied with; then a reported “jailbreak” — which Anthropic argues is just intended behavior (asking a model to read a codebase and fix flaws) that’s “widely available from other models, including GPT-5.5.” The restriction isn’t a sales ban but a bar on all foreign nationals using the models, sweeping in allies and even Anthropic’s own employees (Karpathy included). He covers the 43-page lawsuit from customer Legion against the Department of Commerce (arguing the IEEPA can’t bar export of “informational materials”), the bipartisan Congressional letter demanding transparency by June 26, and leaks that negotiations have stalled and that Anthropic may ship a cheaper, dumber Sonnet-5-era model as a stopgap. The deeper worry he names: the most-capable models are now export-controlled while open-weight GLM-5.2 sits just below the line — pushing the world toward open weights that the U.S. can’t restrict but that may carry other governments’ censorship. He partly blames Anthropic’s own safety-doom messaging for inviting this response. Connects directly to this week’s regulation (Google’s white paper) and open-source-sovereignty (LeCun) threads.

The next paradigm shift (according to Karpathy) — Theo - t3.gg (20:47)

(Transcript-based summary.) A reluctant defense of Anthropic’s Claude Tag Slack bot, which Karpathy called the “third major redesign of LLM UI/UX” — from website, to downloaded app, to “a self-contained persistent asynchronous entity with org-wide tools and context working alongside teams.” Theo argues the genuinely novel parts are multiplayer (one shared Claude per channel that anyone can pick up) and channel-scoped memory — which he thinks is the closest anyone has come to the right boundary for agent context, since “global vs. project-specific” is too coarse and channels map naturally to how teams actually split work. He draws on his own team’s experience hand-building isolated Discord/Hermes agents per task (e.g., a daily cron that scrapes top meme posts into a saveable HTML page) and notes Claude Tag delivers most of that value without the bespoke setup. His main critique: it’s locked to Anthropic’s models, whereas his own agents let him swap between GLM-5.2, GPT-5.5, and Claude per task — and he wants this org-level-harness pattern cloned by others so teams aren’t reliant on a single lab.

References

  1. AI’s Affordability Crisis — David Rosenthal (via Hacker News), 2026-06-23 [blog]
  2. Companies are scrambling to stop employees from maxing out AI budgets — TechCrunch, 2026-06-24 [blog]
  3. A Pragmatic Approach to AI Governance in America — Google, 2026-06-25 [blog]
  4. For most of the world, open-source AI is the only way forward — Techstrong.ai (via Hacker News), 2026-06-23 [blog]
  5. Anthropic says Alibaba illicitly extracted Claude AI model capabilities — Reuters (via Hacker News), 2026-06-24 [blog]
  6. Coinbase reduces time from idea to production by 90% with Cursor — Cursor, 2026-06-23 [blog]
  7. AI was supposed to kill engineering jobs, but new data suggests they’re the most resilient — TechCrunch, 2026-06-24 [blog]
  8. AI researchers continue to leave Google for its rivals — TechCrunch, 2026-06-24 [blog]
  9. Europe is pushing back on Washington’s chip war — TechCrunch, 2026-06-24 [blog]
  10. Cerebras stock plunges after earnings — TechCrunch, 2026-06-24 [blog]
  11. Why big AI labs are hiring so many philosophers — The Economist (via Hacker News), 2026-06-24 [blog]
  12. Anthropic’s Claude Tag is learning your company, one Slack message at a time — TechCrunch, 2026-06-23 [blog]
  13. Introducing computer use in Gemini 3.5 Flash — Google, 2026-06-24 [blog]
  14. OpenAI unveils its first custom chip, built by Broadcom — TechCrunch, 2026-06-24 [blog]
  15. How Notion used the Cursor SDK to embed coding agents — Cursor, 2026-06-25 [blog]
  16. Figma adds code layers, support for animations, more AI features — TechCrunch, 2026-06-24 [blog]
  17. Chrome is bringing advanced autofill capabilities to your phone with Google Wallet — Google, 2026-06-23 [blog]
  18. Google Play updates let U.K. developers do more and pay less — Google, 2026-06-25 [blog]
  19. Facebook rolls out an AI companion app for creators — TechCrunch, 2026-06-24 [blog]
  20. Amazon ups India bet with fresh $13B AI infrastructure investment — TechCrunch, 2026-06-25 [blog]
  21. India’s MoEngage bets marketing’s future on millions of AI agents — TechCrunch, 2026-06-23 [blog]
  22. Fika Jobs raises $4M to build a video-first hiring platform where AI agents interview candidates — TechCrunch, 2026-06-23 [blog]
  23. RubyLLM: A Ruby framework for all major AI providers — RubyLLM (via Hacker News), 2026-06-24 [blog]
  24. NVIDIA launches BioNeMo Agent Toolkit — NVIDIA News, 2026-06-23 [blog]
  25. NVIDIA powers over 400 of the world’s 500 fastest supercomputers — NVIDIA Blog, 2026-06-23 [blog]
  26. Towards a world where no one is surprised by a natural disaster — Google, 2026-06-23 [blog]
  27. VibeThinker-3B: Exploring the Frontier of Verifiable Reasoning in Small Language Models — arXiv (via Lobsters), 2026-06-23 [blog]
  28. Unlimited-OCR: One-shot Long-horizon OCR — Baidu (via Lobsters), 2026-06-24 [blog]
  29. Using the Gini Coefficient to Plan Edge Capacity — Fastly (via Lobsters), 2026-06-24 [blog]
  30. Boost Inference Performance up to 15x on NVIDIA Blackwell Using DFlash Speculative Decoding — NVIDIA Developer Blog, 2026-06-23 [blog]
  31. Maximize AI Factory Energy Efficiency Through Full-Stack Inference and Training Optimizations — NVIDIA Developer Blog, 2026-06-23 [blog]
  32. Accelerating BEV Pooling on NVIDIA GPUs for Physical AI Applications — NVIDIA Developer Blog, 2026-06-24 [blog]
  33. Is it ever coming back? — Theo - t3.gg, 2026-06-24 [video]
  34. The next paradigm shift (according to Karpathy) — Theo - t3.gg, 2026-06-25 [video]