GitHub Copilot Tokenpocalypse, Bun's Rust Port Hint & Karpathy's Three New Horizons

Theo's GitHub Copilot $221-on-a-$40-plan exposé

The day's biggest story in agent economics: Theo sent a single Copilot message that ran for ~7 hours, did 60M+ tokens, and consumed an estimated $221 of inference. After 15 messages he had logged $221 of token cost on a $40/month plan — 1.6% of his quota used.

"I sent a single message on Copilot and it did over 60m tokens. It's still going. $30 of inference so far. In their current billing model, you get 1,500 messages, regardless of how expensive each is. I'm pretty sure I can do $45,000 of messaging on this plan." — https://x.com/theo/status/2051218167780041147

His subsequent breakdown:

This is the cleanest live demonstration to date that per-message billing for agent products is structurally broken — on a long-horizon agent, message-count is decoupled from cost by 3+ orders of magnitude. (Compare the OpenAI Codex/Claude Pro/Max pricing-postmortem cycles from the past 3 weeks; the wrapper-vs-substrate margin debate is converging on the same answer at every layer.)

Bun's Rust port (?) — coding agents discover the smoking gun

Late Monday, Luke Parker spotted a commit in oven-sh/bun and fired the alarm:

"what!? is Bun moving from Zig to Rust??? github.com/oven-sh/bun/commit/9f3917e979fecd2dae1327e159c6b2fd258bb67f" — https://x.com/LukeParkerDev/status/2051431276205461635

Simon Willison confirmed by finding the actual coding-agent guide in the repo:

"Sure looks like Bun is at least exploring a port from Zig to Rust given this docs/PORTING.md guide for coding agents github.com/oven-sh/bun/blob/…/docs/PORTING.md" — https://x.com/simonw/status/2051476878712840407

The fact that the migration is shipped as instructions for agents to execute, not for humans is the actual story. Robobun (Bun's autonomous-PR agent) has been the top contributor to the project for months, per Jared Sumner:

"Robobun is the top contributor to Bun for months now. Even if we only merge a fraction of the PRs, it's worth it. We didn't have to write the code ourself or debug it ourselves. As long as the tests are good, the tests fail on main, and the fix is correct." — https://x.com/jarredsumner/status/2051450571060867148

Mitsuhiko captured the surreal flavor of the resulting maintainer experience:

Context for the Zig angle: Zig project leader Andrew Kelley recently issued a blanket ban on AI-assisted contributions (covered in our Apr 30 roundup). If Bun does end up porting to Rust because the agent ecosystem has consolidated there, that's a much bigger story about language ecosystems where coding agents will / won't be welcome.

Mitsuhiko's "Content for Content's Sake" — the slop math

Armin's blog post the same day is the empirical companion piece to mattpocock's "AI negligence" thread from yesterday. He took 90 days of his local coding sessions, looked for medium-frequency words that were inflated relative to wordfreq baselines, then cross-referenced US Google Trends. Words that pop in agent output (like "substrate" and "capability") also spike on Google Trends.

The personal anxiety inside the post is the more interesting part:

"I am increasingly worried that I'm starting to write like an LLM because I just read so much more LLM text. The first time I became aware of this was that I used the word 'substrate' in a talk… Since then, however, I am reading this word everywhere."

His thesis, in three pieces:

  1. Engagement farming: "There are entire companies now that just exist to automate sending LLM-generated shit and people evidently pay money for it… real genuine human signal is losing out quickly."
  2. Speed Should Kill (but doesn't): if quality won, slop would lose; instead, slop gets posted faster than thoughtful replies and wins on the algorithm anyway. Open Source has the same dynamic — projects get "remixed" with marketing sites within hours.
  3. Friction & Rate Limiting is the missing primitive: "We have known for a long time that certain things should not be easy, because of the misuse that happens. We know it in engineering; we know it when it comes to governmental overreach. Now we are probably going to learn the same lesson in many more situations because LLMs make almost anything that involves human text much easier."

Adjacent: he also analyzed his pi (pi-mono) issue tracker and found that the bulk of slop tickets are agents acting on people's accounts without the human even being aware. "One person, 4 tickets in 15 minutes, all useless slop. How did we end up here." https://x.com/mitsuhiko/status/2051204508844196097

Karpathy's Sequoia Ascent fireside — three new horizons, jaggedness, agent-native economy

Karpathy posted his recap of the Sequoia Ascent 2026 fireside chat. His core argument: LLMs are not just for speeding up existing things; they enable categories of products that classical code couldn't even attempt.

Three examples of new horizons he pushed:

  1. menugen — "an app that can be fully engulfed by LLMs, with no classical code needed: input an image, output an image and an LLM can natively do the thing."
  2. Install .md skills instead of .sh scripts — "Why create a complex Software 1.0 bash script for installing a piece of software if you can write the installation out in words and say 'just show this to your LLM'." LLM = English interpreter that can target your specific setup and debug inline.
  3. LLM knowledge bases — "an example of something that was impossible with classical code because it's computation over unstructured data (knowledge) from arbitrary sources and in arbitrary formats."

The jaggedness theme:

"How it can be true that a single artifact will simultaneously 1) coherently refactor a 100,000-line code base and 2) tell you to walk to the car wash to wash your car… revenue/TAM dictates what the frontier labs choose to package into training data distributions during RL. You're either in the data distribution (on the rails of the RL circuits) and flying or you're off-roading in the jungle with a machete, in relative terms."

Closing with the agent-native economy: products & services as sensors / actuators / logic, "making information maximally legible to LLMs," and "hints/dreams of fully neural computing handling the vast majority of computation with some help from (classical) CPU coprocessors."

Post: https://x.com/karpathy/status/2049903821095354523 (tweeted Apr 30 but actively recirculating today)

Codex 5/5 token party: rate limits 10x'd

LLMJunky's Sunday call ("you should use absolutely all of the Codex usage you can before 5/5 because there's absolutely no way we don't get a reset. finna be a token party") paid off today. OpenAI 10x'd Codex rate limits as a 5/5 gift to subscribers:

"Codex rate limits 10x'd as a gift to the 5th of next month for everyone who signed up for the 5/5 party. This is madness. @sama you crazy bro ✌️" — https://x.com/LLMJunky/status/2051465480620089456

This connects back to the Theo Copilot story above: token-based plans are the dominant pricing primitive now, and OpenAI is using subscriber-friendly token resets as a hard moat against Copilot's collapsing per-message economics.

LLMJunky's macOS-vs-Linux Codex Computer Use challenge

A Linux maximalist (LinuxBTW) claimed that Codex Computer Use on macOS isn't special — Omarchy can do the same thing. LLMJunky's response:

"Challenge accepted. So I had Codex draw him a painting using only its mouse, no other tools. Holler at me when you can do this." — https://x.com/LLMJunky/status/2051454763347181676

The video shows Codex driving the macOS pointer through a paint app to produce an actual image, no DOM/accessibility shortcuts, just mouse moves. The wider thesis from his earlier post in the window:

"I'm increasingly becoming macOS pilled. And it's really because of two apps: Codex and Clicky. Since the dawn of computing, machines have been designed around people using them directly… But the next shift is already starting: humans won't always be the ones at the keyboard." https://x.com/LLMJunky/status/2051342251477946521

The frame: macOS's accessibility surface + apps like Clicky (local computer-control bridge) make Mac a more capable agent host than Linux right now, regardless of the kernel-level argument.

Steipete: Crabbox 0.5.0 ships with VNC + AWS Windows

Crabbox went from "ephemeral machines for agents" (0.4 yesterday) to a real-time shared-desktop primitive overnight:

"Crabbox 0.5.0 is live 🦀 🖥️ Desktop/browser leases 🧑‍💻 VNC + authenticated WebVNC 🪟 AWS Windows + WSL2 📸 Screenshots + app launch Remote CI boxes, now suspiciously usable." https://x.com/steipete/status/2051485798613111116 — github.com/openclaw/crabbox

The 0.4 → 0.5 jump in one day is interesting on its own — desktop leases and authenticated WebVNC bring this a lot closer to Anthropic's Claude Computer Use developer experience but as an OSS layer. Combined with WSL2 + Windows AWS support, it's the most platform-agnostic agent-controlled-machine primitive in OSS today.

Also noted in steipete's RTs: Veritas Kanban 4.0.1 (OpenClaw task manager) shipped a security pass on its n8n webhook — required shared secret + timing-safe checks, addressing the kind of issues the OpenClaw security postmortems have been flagging. https://x.com/BradGroux/status/2051151592754180128

Theo on the OpenAI ↔ Microsoft breakup

Theo dropped a long video he says he's been working on:

"OpenAI and Microsoft broke up. The impact of this is massive and I don't think enough people understand why. Put a lot of time into tracking the history of the deal to help you all understand 🫡" https://x.com/theo/status/2051535315287060790 (video)

He pinned the post. Combined with the Copilot tokenpocalypse story above, today's Theo throughline is "Microsoft's AI stack is structurally fragile, and that fragility is now visible at both the partnership-strategy level and the product-billing level."

swyx — current AI lab valuations & the Plan-and-Review pivot

swyx revisited his August 2025 ARR-multiples chart with new numbers:

"OAI 850B valuation, ~30B ARR now Ant ~900B valuation, ~44B* ARR now *revenue recognized differently, per Denise Dresser its probably closer to 8-10B lower if using OAI methodology" https://x.com/swyx/status/2051440392722391180

For comparison, his Aug 2025 numbers were OAI 300B/13B ARR and Ant 170B/5B ARR — both labs roughly 3× their valuation and ~3-9× their ARR in nine months.

Also from swyx today, recirculating from AIE Europe: Steve Ruiz's "What if chat is the wrong interface for managing agents?" talk is now public on YouTube — the "agents on a shared canvas" thesis from TLDraw. https://x.com/swyx/status/2051329419860758932

Mattpocock — 200K YouTube subs, vocab thread, AI-as-compiler-output meme RT

A milestone day for Matt's content arm:

Also confirmed: he's giving an AI-focused talk at ETN this Tuesday at 1:00 PM BST (live + filmed). https://x.com/etnshow/status/2051259837447536784

Theo's vesting-cliff vocabulary fight (off-topic, but unavoidable)

Theo spent a chunk of the day correcting people on what "vesting cliff" actually means (the 1-year initial wait, not the 4-year fully-vested point). The thread is mostly off-topic, but a useful glossary if you've ever been confused: https://x.com/theo/status/2051546385946947823.

Simon's experiment running the canonical "Generate an SVG of a pelican riding a bicycle" prompt against 21 different quantized variants of IBM's new Granite 4.1 3B model:

"the results weren't as interesting as I had hoped" — there's no clean quality-vs-size pattern, "they're all pretty terrible." Will retry with a model that draws pelicans better. https://x.com/simonw/status/2051449431686525102

Separately, Redis just dropped an array data type (PR by Salvatore Sanfilippo). Simon used Claude Code for web to build an interactive WebAssembly playground for the new ARGREP/ARGET/ARSCAN commands. The ARGREP command runs server-side grep against a range using a newly vendored TRE regex library. Salvatore's writeup covers the AI-assisted dev process for the Redis side. https://x.com/simonw/status/2051310232022986772

Misc / shorter

Videos worth watching

News / longer reads

Non-AI / off-topic