Anthropic ↔ SpaceX on Colossus 2, Theo Builds a Cloud, and the Tokenmaxxing Pod
Claude Code & Anthropic Updates
Anthropic expanding the SpaceX partnership — GB200s on Colossus 2 throughout June. Tom Brown (@nottombrown) confirmed it from the Anthropic side: "We're expanding our partnership with @SpaceX, and will be scaling up on GB200 capacity in Colossus 2 throughout June. Appreciate @elonmusk and the team helping us find good homes for the Claudes." Quoting a SpaceX post promising "we are going to need to move a lot of atoms in order to keep up with AI demand." 6.1k likes, 1.3M views, 306 replies in under 12 hours. The most quoted reply, from @aabyzov, reframes the year: "2024: AI labs at war. 2026: AI labs sharing GPUs because the actual scarce thing is fabs." Latency concerns surfaced too — Colossus is centralized in Memphis, which adds 50-150ms for users outside the US Midwest.
trq212 joining the Claude Code team. Aditya Agarwal (adityaag) confirmed Thariq's landing spot publicly: ".@trq212 is a builder's builder. After research at MIT, exiting his company, raising millions for another, and exploring research threads as an SPC member, he's now on the team building Claude Code." Thariq is back at South Park Commons next week. Pairs with yesterday's news of Karpathy joining Anthropic to use Claude to accelerate Claude — back-to-back senior hires onto the model and harness teams.
Computer use, in production. ClaudeDevs posted (retweeted by bcherny) a new blog post on making Claude's computer-use mode reliable in production: "getting click accuracy right, choosing thinking effort levels, kicking off async work." Pitched as the "agent that can operate real UIs" — which fits the Antigravity / Spark / Codex agentic browser race.
Agentic Coding & Agent Harnesses
Composer 2.5 lands third on Artificial Analysis — at 10-60× lower cost. Artificial Analysis posted (retweeted by leerob) that Cursor's Composer 2.5 takes third on the Coding Agent Index "and is ~10-60× lower cost than the higher-effort Opus 4.7 and GPT-5.5 variants above it." Coupled with Lee Robinson's own public feedback request: "Where could we improve Composer 2.5? We're working on the next model and would love your feedback." 572 replies, 2.3k likes. Common asks: UI/frontend design (leerob asked back: "are you using any skills for this?"); "5.5 medium still feels better"; "review the code and it sometimes looks like AI slop"; "see how Claude Code does orchestration — context:fork, hooks etc." (leerob noted Cursor now ships hooks too).
"$60B for an IDE LOL" — am.will defends Cursor. LLMJunky's morning broadside: "People hate on Cursor, or even go as far to laugh at it. '$60B for an IDE LOL', 'It's just a VSCode fork'. Yeah, and Tesla is just a car. Youtube is just a video player. And the iPhone is just a phone. It's honestly hysterical how wrong they are. GPT 5.5 High Fast and Cursor 2.5 Fast feel unbelievably good in this harness. I have been building non-stop for the last week." Lobster-emoji defenders of "Cursor + Codex sub, I want for nothing" out in force in the replies.
"If Codex maintains a high quality bar and refuses to ship slop, they will win." Theo, quoting an Andrew Mason line about resisting "the temptation to ship garbage." 724 likes, 44k views. The strongest reply from TravelerOfCode: "every product hits the same inflection: ship to win the news cycle or ship to keep the user. codex picking the second has a real shot at sticky users + a smaller but real revenue arc." Reads as a thinly-veiled jab at Google after this week's Gemini 3.5 Flash debacle.
Simon Willison untangles "the Antigravity harness." simonw: "Anyone understand what Google mean by 'Gemini Spark runs on Gemini 3.5 and uses the Antigravity harness' — is 'Antigravity' a generic term they're using for their agent harnesses now or is their Claw-competitor running the same closed-source Go binary we can download ourselves?" Andrew Garrett (ex-Googler) confirms in the replies: "I use the 'antigravity SDK' internally, it's being used for ~everything and I believe it's more or less the same thing as the Go binary packaged with agy." Simon then observed that the Python package is "mainly a wrapper around a 95MB Go binary." His follow-up blog post is his usual "only write about products that have shipped" filter applied to this year's I/O — the most-liked reply applauds the policy: "'Only write about things I can try' is the most underrated policy in AI coverage right now."
LlamaIndex's free due-diligence template + LiteParse. Jerry Liu announced a fully free, open-source SEC-filings agent with audit trails: "~600 lines of Next.js. No vector DB. Just LiteParse." The pitch is that LiteParse — their model-free parser — returns text plus exact bounding boxes, so the agent can cite the precise PDF page for every number it surfaces. llamaindex.ai/blog/building-…. cdiamond's one-line summary in the replies: "audit trail back to source is the enterprise gate. most AI finance demos die at 'show me where that number came from'." The team also got an official shoutout at the Google I/O developer keynote for their Gemini API integration.
Builders & Tools
Theo teases a new cloud — built from the metal up. Theo: "So I have a working prototype and I think this is the coolest thing I've ever built." In a follow-up post he listed the surface area he had to ship to get there: "a sync engine, a JS runtime, a database platform, a CLI, a full stack JS framework, a JSX interpreter." No name yet — but he did confirm someone is already squatting t3.cloud for $2.5k. 46k views in 4 hours, 393 likes. The replies are roughly: "AWS wrapper?", "Vercel/Railway alternative?", "is this notgoogle.cloud?" — Theo declined to clarify.
The tokenmaxxing pod: $1.3M/month at Steipete-scale. Theo posted the video drop with Steipete: "We saw that @steipete is spending over $1.3M in tokens every month. We're not tokenmaxxing hard enough." Sections: Tokenmaxxing → Secure Agents → Anthropic Rules → Security → Token Taxxing → AI Psychosis → macOS hell → AI Video is dangerous. The thread's best critique came from ByteCrafter: "$1.3m/month is loud but I doubt raw tokens are the real cost. our triage agent's spend stayed flat until context pressure forced us to haiku-summarize every tool output before re-entry. once you stop re-reading prior calls each turn, the bill compresses fast." Abenz_mato echoed: "Output per token is what matters. We run Mato's audio pipeline through Claude Code. 9 of 11 production files are AI-written. The flex isn't the spend. It's shipped product per dollar."
Steipete: "Can't recommend cotypist enough. Autocomplete everywhere." Tweet — 1.4k likes, 125k views. cotypist.app — macOS-only autocomplete across every text field, ML-driven, runs locally. Steipete had previously been all-in on dictation; he clarified in the replies: "typing is useful when you are on a call." Plenty of Linux users asking for parity in the thread.
swyx on Exa + the Agent Labs thesis. swyx on Exa's $250M raise at $2.2B (a16z-led): "we did a bake off of Exa vs competitors and it took all of 1.5 hrs for the team to unanimously converge on exa lol." In a separate thread he tied this to Sam Altman's "build a business that gets better when models get better" — "basically what I called Agent Labs" — and pointed at a Q4 2025 discontinuity correlating directly with model performance. The thread thinks Exa is the canonical case: the product is the model roadmap.
OpenClaw on Meta Ray-Bans. This Week in Startups (retweeted by steipete): "Someone connected Open Claw to their Meta Ray-Bans. Shawn looks at a box of lens wipes and tells his agent to add them to his Amazon cart. No phone. No keyboard. No app. Just glasses." 240k views. Half of the replies are "I am building this exact thing" — but also a sober "Looking forward to billboard prompt injection" from @WesAField.
Security & Supply Chain
Mitsuhiko's new minimum bar for Shai-Hulud post-mortems. Armin Ronacher: "If you get Shai Hulud, the minimum I want to see is a comment about what you're going to do to prevent it going forward. Just linking to a security advisory is not going to cut it for me to feel confident." He then made the more interesting structural point: "The 'problem' is that we need to work with stuff that just has irresponsible dependency chains (looking at you Google genai and Bedrock). Those chains look the same with or without npm. Pinning and reviewing on upgrade is what matters more. Wish though npm would default to minver resolution :(" — i.e. the package manager is downstream of the real issue. Earlier that day he had asked, "Where is the polymarket for which npm package/org/account gets hacked next?"
Pi supply-chain hardening, from the same team. @badlogicgames (retweeted by Mitsuhiko) shipped "supply-chain hardening release" for Pi after last week's mistralai package was Shai-Huluded. "We were not affected, due to pinning." Munawar Shah's reply captures the broader posture shift: "I have not been hit but it was enough of a close call with the tanstack attack that I scrubbed my AWS long lived tokens in favor of hourlong expiry, and all my .env files seeded through 1Password CLI at runtime. For ALL my projects!!!"
Industry & Career
Matt Pocock — Tactical vs Strategic Programming. The essay tweet — 944 likes, 61k views, 71 replies. The argument:
Agents have eaten the tactical part of programming. When you can pay below minimum wage for code, there's no point going into the trenches yourself. But AI cannot code strategically… So, a developer's day-to-day job has become 100% strategy. Long-term thinking, all the time.
What makes me nervous is that we've pulled down the only bridge that brought juniors into the industry. We used to train juniors like this: 1. Give them only tactical tasks, 2. Let them build up their strategic experience slowly. But what happens when all tactical code is written by AI? What is the point of a junior?
The best reply, from @pythonics00, has Pocock convinced: "We have a very talented intern who started writing code only 18 months ago. He's religious about using grill-me (now with docs) for everything… Deploying a junior on lower-risk, but valuable, projects that they can deliver slices of end-to-end and giving them macro strategic feedback seems to be a winning formula." Pocock's response: "So give them big chunks of responsibility on projects with a small blast radius." He also conceded a second-order point: working with agents may accelerate strategic learning because the feedback loop is tighter — "your mistakes hit you faster."
Gemini CLI sunset. Mitsuhiko notes, via a Google announcement: the open-source gemini-cli is being deprecated in favor of the Antigravity harness. His framing: "A good reminder that AI years are like dog years. Multiply by 7. So gemini-cli made it to impressive 7 years which is more than most other Google projects." Will Chen (ex-Googler) chimed in: "it's a bit of a shame they're deprecating the open-source gemini CLI in favor of a closed-source harness."
Off-Topic / Lighter Notes
Bezos returns the favor on no-tax-bottom-half. LLMJunky retweeted Bezos: "The important part is zeroing out taxes on the bottom half. Best way to put money in someone's pocket is to not take it out in the first place. Bottom half is only 3% of total revenue." Outside our coding lane but the engagement loop is the same Twitter — Bezos and Musk both in this week's roundups on AI infra topics.
Astronomy moment. LLMJunky: "Its too hazy here, but tonight we got both Jupiter and Venus in a line with the moon. You will know them when you see them; you cannot miss them; they are the brightest 'stars' in the sky." Genuinely worth looking up tonight if your sky is clearer than his.