Key Highlights

  • Access to frontier AI is closing. Anton Leicht argues compute scarcity, security concerns, and U.S. government oversight will lock most users out of the most capable models — Anthropic’s Mythos and OpenAI’s Daybreak-gated gpt-5.5-cyber are early signals of a structural shift rather than one-offs.
  • Anthropic’s June 15 monetization change reframes “programmatic” usage. Paid Claude plans will get a separate, smaller dedicated credit for Agent SDK / Claude-P usage, cutting effective rate limits by up to 40x for tools like T3 Code, Zed, OpenClaw, and any CI integration — a hard line that many developers are calling an attack on open-source harnesses built around Claude Code.
  • Mobile coding agents go mainstream. OpenAI launched Codex Mobile in preview in the ChatGPT iOS app with a secure relay layer for live thread management, while Ramp data shows Anthropic has overtaken OpenAI in enterprise paid-user adoption (34.4% vs 32.3%) for the first time.
  • Local-AI infrastructure keeps maturing. Osaurus (open-source macOS app) and whichLLM (hardware-aware local model selector) ship the same day Anthropic open-sources Claude for Legal — 14 practice-area plugins, scheduled agents, and 20+ MCP connectors — signaling a pivot from chat UIs toward harnesses and embedded workflows.
  • Bun is being rewritten in Rust by agents in under a month. Jared Sumner reports 99.8% of Bun’s pre-existing test suite passes on Linux x64 in a Rust port that is already ~960k LOC — but with ~13,000 unsafe calls (vs ~73 in UV), raising concerns about whether AI-driven line-by-line ports trade known bugs for an unknown long tail.

Analysis & Opinion

Access to frontier AI will soon be limited by economic and security constraints — Hacker News AI

Anton Leicht argues that broad API access to frontier models is structurally unsustainable. He identifies three compounding forces: compute economics (high marginal cost per query makes wide distribution loss-making), security and misuse concerns (model theft, distillation, and weaponization push labs toward gated access and stronger identity verification), and U.S. government involvement (export controls and national-security review of frontier deployments). He points to Anthropic’s restriction of Mythos to vetted cybersecurity firms and OpenAI’s selective Daybreak distribution of gpt-5.5-cyber as early structural moves rather than one-offs. The implication is that the assumption of “powerful AI for everyone” underlying much current policy discourse may be wrong within 12–24 months.

No Dumb Questions: What is cloud computing and why is everyone doing it? — Stack Overflow Blog

A primer interview with Stack Overflow’s infrastructure lead Josh Zhang covering containers, Docker, Kubernetes, pods, and how AWS productized excess data-center capacity. The piece’s sharper note is the pushback on the “cloud is cheaper” assumption — Zhang notes “the one thing in the cloud that definitely scales is your bill,” echoing the Basecamp/37signals repatriation discourse without explicitly endorsing it.

New Products & Tools

OpenAI Takes Codex Mobile — Rundown

OpenAI launched Codex inside the ChatGPT iOS app in preview, letting developers approve code changes, manage plugins, and start new tasks while agents continue running on desktop or remote hosts. A “secure relay layer” syncs across ChatGPT instances without exposing the host machine. The launch directly answers Anthropic’s earlier Remote Control (February) and Dispatch (March) mobile flows, and lands as multi-hour autonomous runs become common — reducing the need to keep laptops open in public to babysit agents.

Osaurus brings both local and cloud AI models to your Mac — TechCrunch

Open-source macOS app from Dinoki founders (ex-Tesla/Netflix) that acts as a “harness” — a single consumer-friendly interface that switches between local and cloud models while keeping user data on-device. Uses hardware-isolated virtual sandboxing for safety. Positioning targets users who balk at paying subscription on top of token costs.

whichLLM — Find the best local LLM for your hardware — Hacker News LLM

CLI that ranks local LLM candidates not just by VRAM fit but by weighted benchmark quality (LiveBench, Artificial Analysis, Aider, Chatbot Arena), quantization level, runtime compatibility, and hardware-aware inference-speed estimates. Auto-detects NVIDIA, AMD, Apple Silicon, and CPU-only systems and supports GPU simulation for planning.

Claude for Legal — Lobsters AI

Anthropic-published suite of 14 practice-area plugins (commercial, employment, privacy, AI governance, etc.), scheduled monitoring agents (renewals, docket watching, regulatory change), 20+ MCP connectors (Slack, Google Drive, Ironclad, CourtListener), and 100+ named workflow agents (Vendor Agreement Reviewer, DSAR Responder). The repo is explicit that every output is “a draft for attorney review — not legal advice.” Notable as a reference implementation showing how Anthropic envisions vertical agent stacks built on Claude.

From policy to practice: supporting the future of AI in education — Google (The Keyword)

Google for Education’s AI Policy & Guidance Labs convened educators across Brazil, India, Malaysia, Mexico, Spain, and Sweden to produce vendor-neutral 12-month AI implementation roadmaps. The framing rejects the “teacher as AI usage monitor” model in favor of “teachers must lead” — determining when and how AI is deployed by classroom and student need. The piece reads as Google staking out a position in the regulatory-and-norms layer that will eventually shape procurement.

Interviews & Conversations

I’m done. — Theo - t3.gg (46 min)

Theo’s most direct break with Anthropic to date. Starting June 15, paid Claude plans will receive a separate, must-be-claimed monthly credit for “programmatic” usage (Agent SDK, Claude-P, GitHub Actions, third-party tools built on the Agent SDK) — $20/$100/$200 matching the plan tier, with no rollover. Interactive Claude Code keeps its existing subscription limits; everything outside Anthropic’s own UI drops to API rates or the new bucket. Theo argues this effectively cuts allowed usage by 25–40x for open-source harnesses like T3 Code, Zed, gene, and CI integrations — tools that were explicitly told this path was supported. He frames it as Anthropic drawing the “subsidization line” too far left and punishing open-source UIs because Anthropic can’t reliably distinguish them from token-farming wrappers like OpenClaw and Hermes Agent. Notable position: he canceled his Claude plan on air and is shipping a CLI-only fallback mode in T3 Code.

Stop letting your agents write Markdown — Theo - t3.gg (36 min)

Discussion of Lurk Thoric’s argument (and Karpathy’s endorsement) that agents should output HTML rather than Markdown for plans, specs, code reviews, and reports. HTML carries richer info (tables, SVG diagrams, code with syntax highlighting, interactive sliders for design iteration, mobile-responsive layouts), and people actually read it instead of skimming a 100-line markdown file. Theo agrees in cases like PR explainers and design exploration but pushes back on the novelty-vs-substance question and on Karpathy’s longer prediction that we’ll move toward fully neural-generated interactive UIs. Useful concrete tip: ask for several distinct UI directions in one HTML grid rather than iterating sequentially — variety is higher in a single-shot multi-option prompt.

We all fell for it… — Theo - t3.gg (57 min)

Walkthrough of Lars Fey’s “Agentic coding is a trap” essay, focused on cognitive debt rather than technical debt. Cites Simon Willison and Martin Fowler reporting reduced familiarity with their own codebases, and Reddit/Twitter posts from devs with 9+ years’ experience describing their skills atrophying after 6–12 months of heavy agent use. Anthropic’s own research acknowledges the “paradox of supervision” — supervising Claude effectively requires the same coding skills overuse erodes. Theo’s nuance: AI separates code that will run thousands of times a day (where quality matters and AI should help you understand it better) from code that runs once (where AI lets you write throwaway tools that weren’t worth writing before). Both ends improve; the failure mode is using AI as a slot machine because feeling-dumb is painful and learning the underlying pieces is no longer forced. Strong recommendation: senior engineers and OSS maintainers benefit most; juniors who learned to code with AI struggle hardest to debug code they didn’t write.

I wish this was clickbait — Bun rewritten in Rust — Theo - t3.gg (27 min)

Jared Sumner is rewriting Bun (Zig) into Rust using AI agents in under a month. 99.8% of the pre-existing test suite passes on Linux x64 in the Rust port. The codebase is ~960k LOC and “basically the same codebase” — a line-by-line translation rather than an idiomatic Rust rewrite. Theo’s concern: the port has ~13,044 unsafe calls vs ~73 in UV (Rust’s idiomatic Python package manager) across a similar code surface. It’s Zig with Rust syntax, not Rust. Compared to the TypeScript-to-Go port that succeeded by deliberate suboptimal line-by-line translation, Bun’s situation is harder because Bun’s primary economic incentive is now Claude Code (shipped as a Bun binary) — concerns about non-Claude-Code-affecting bugs being deprioritized after the rewrite. Open Code’s Dax has separately moved off Bun to Node.js citing Windows stability, Electron compatibility, and uncertainty about Bun’s future under Anthropic.

Anthropic just…wait what — Theo - t3.gg (35 min)

Anthropic and XAI/SpaceX announced a compute deal: Anthropic is leasing ~300 MW / 220k+ H100s of Colossus 1 from SpaceX (XAI moved training to Colossus 2). Notable because Anthropic banned XAI from using Claude in January over policy concerns. Theo’s framing: each major lab is short one of {research, data, compute}. Anthropic has research+data but is short compute (revenue annualizing at 80x growth, not the 10x they planned for); XAI has compute but is short data — explaining their separate $60B option on Cursor for the chat-history corpus. OpenAI is the only player with all three, plus imminent AWS hosting, which is what’s forcing Anthropic to swallow the optics of the SpaceX deal. Direct user impact: peak-hour rate limit reductions are reversed, 5-hour Claude Code limits doubled, API rate limits for enterprise tiers raised 4–5x.

What’s next? — GitHub alternatives — Theo - t3.gg (51 min)

Survey of GitHub alternatives prompted by GitHub’s recent multi-day outages and reverted merges. GitLab dismissed as “worse GitHub with better uptime” — built on aging Ruby with deep UX problems. Bitbucket positioned as the Jira-integrated discount option. The strong recommendation is Forgejo (and its hosted instance Codeberg), the fork of Gitea that took the project after the original turned commercial — open-source, MIT-licensed, Go-based, governed by a democratic nonprofit (Codeberg e.V.). Theo donated $1,200 and committed to $400/mo on-stream. Separately calls out Pierre’s code.storage as the first Gen-3 primitive (ultra-low-latency Git cloud built for agentic throughput — 9M repos stored in 30 days, peaks of 15k repos/minute), Graphite (now acquired by Cursor), and Entire (founded by GitHub’s last CEO, $60M seed, building agent context tracking on top of/alongside Git). Theo’s bittersweet closing: leaving GitHub fractures the unified developer profile and community history that has lived there for 20 years — no tool can rebuild that.

How We Grew Koch Industries to $150 Billion Without Going Public — All-In Podcast (95 min)

Charles and Chase Koch on principle-based management, capability-bounded vs industry-bounded growth, and why staying private and Kansas-based was a competitive advantage. Key operational principles: hire on values first / skills second / credentials last; experimental discovery (treat every business as a laboratory); reward employees for overall contribution to the future rather than short-term outcomes; comparative advantage at the individual level (Chase fired himself as Koch Fertilizer president when he realized he was a builder, not an operator). The Koch Disruptive Technologies discussion ties to AI: their app “Principal Companion” delivers the book’s 41 principles via a Socratic agent. Their broader political-economy frame — that today’s malaise comes from “the means to live and no meaning to live for” — frames their Stand Together philanthropy work, including funding 5,000+ micro-schools via the Veila Fund post-COVID.

Nvidia’s Jensen Huang Joins Trump’s China Trip — Bloomberg Television (92 min)

Jensen Huang accompanying the Trump administration on a China trip — relevant context for ongoing export-control debates affecting frontier compute. Full discussion covers Nvidia’s positioning between U.S. policy constraints and Chinese demand for AI infrastructure.

Anthropic’s Dario Amodei and JPMorgan’s Jamie Dimon on AI — CNBC Television (5 min)

Amodei warns Chinese AI models will close the gap in 6–12 months and predicts some SaaS incumbents will go bankrupt as moats erode — though prepared incumbents will adapt. Dimon argues the ~$1T in AI infrastructure spend will pay off in aggregate even if individual bets fail. On regulation, Amodei pushes back on an FDA-style pre-release model for frontier models (calls FDA a cautionary tale for slowing progress) and prefers an NHTSA-style continuous-monitoring approach. Both speakers treat the “20–30% unemployment in 2–5 years” thesis (Dan Schulman’s recent claim) as overstated but acknowledge legitimate concern about transition pain. Notable for safety/social-impact framing from two of the most influential voices in the AI economy.

Alex Karp on Palantir’s earnings call — David Carbutt (19 min)

Palantir reported 100% US growth, Rule of 40 at 145, $791M free cash flow on $1.4B revenue (56% FCF margin). Karp’s framing is a sustained attack on “AI slop” — the idea that AI products that look like working software in demos collapse in enterprise production because they lack an ontology grounding agents in real objects with real permissions. Argues only the Palantir foundry/Apollo/ontology stack delivers real enterprise AI today and that demand exceeds their capacity. Talent angle: Palantir is recruiting against the AI labs by selling mission (“build something that runs in war zones, hospitals, and factories”) over chatbot work. Worth listening to as a counter-narrative to the “AI as commodity infrastructure” thesis dominating the rest of this week’s coverage — and a useful data point on how the AI-deployment-in-real-enterprises problem is being framed by a company that actually deploys it.

1 Hour Ago: Elon Musk Interview — The Money Investing (45 min)

Long-form interview with Elon Musk at the Cannes Lions advertising festival. Musk endorses Geoffrey Hinton’s ~10–20% probability of catastrophic AI outcome but says the modal outcome is universal high income and “work optional” — with a likely “crisis of meaning.” Frames OpenAI’s creation as offsetting Google’s AI monopoly and laments that the org went closed-source. Predicts ~20 billion humanoid Optimus robots eventually deployed, including consumer use. On news: AI aggregation of millions of real-time posts will increasingly replace conventional journalism. On Neuralink: first telepathy implant recipient now controls phones and plays video games by thought.

References

  1. Access to frontier AI will soon be limited by economic and security constraints — Hacker News AI, 2026-05-15 [blog]
  2. No Dumb Questions: What is cloud computing and why is everyone doing it? — Stack Overflow Blog, 2026-05-15 [blog]
  3. OpenAI Takes Codex Mobile — Rundown, 2026-05-15 [blog]
  4. Osaurus brings both local and cloud AI models to your Mac — TechCrunch, 2026-05-15 [blog]
  5. whichLLM — Find the best local LLM for your hardware — Hacker News LLM, 2026-05-15 [blog]
  6. Claude for Legal — Lobsters AI, 2026-05-15 [blog]
  7. From policy to practice: supporting the future of AI in education — Google (The Keyword), 2026-05-15 [blog]
  8. I’m done. — Theo - t3.gg, 2026-05-14 [video]
  9. Stop letting your agents write Markdown. — Theo - t3.gg, 2026-05-13 [video]
  10. Nvidia’s Jensen Huang Joins Trump’s China Trip | The China Show 5/13/2026 — Bloomberg Television, 2026-05-13 [video]
  11. How We Grew Koch Industries to $150 Billion Without Going Public: Charles & Chase Koch — All-In Podcast, 2026-05-12 [video]
  12. I wish this was clickbait — Theo - t3.gg, 2026-05-12 [video]
  13. We all fell for it… — Theo - t3.gg, 2026-05-11 [video]
  14. Anthropic just…wait what — Theo - t3.gg, 2026-05-07 [video]
  15. What’s next? — Theo - t3.gg, 2026-05-06 [video]
  16. Anthropic’s Dario Amodei and JPMorgan’s Jamie Dimon on AI boom, AI regulation & impact on jobs — CNBC Television, 2026-05-06 [video]
  17. 1 Hour Ago: Elon Musk Gets Seriously Confronted In Interview But FIRES Back! — The Money Investing, 2026-05-06 [video]
  18. Alex Karp Leaves Audience Speechless — David Carbutt, 2026-05-05 [video]