SpaceX Buys Cursor, "Git 2.0" Lands the Same Day & the Loop People Win
SpaceX Buys Cursor — and "Git 2.0" Lands the Same Day
The deal. The biggest story of the day landed without warning: SpaceX is acquiring Cursor. The official line, quoted by @cursor_ai (1,060 replies, 2,799 RTs, 27.7K likes, 3M views): "SpaceX has exercised the option to acquire @cursor_ai in an all-stock transaction with the goal of building the world's most useful AI models. For the past few months, SpaceXAI has been jointly training a model with Cursor, which will be released in Cursor and Grok Build soon." Cursor's own framing was characteristically flat: "We're excited to join forces with @SpaceX to advance the frontier of useful AI. Expect significant improvements to Cursor soon." Founder Michael Truell: "Lots to do together. Excited to be joining forces with @SpaceX to build useful AI."
The number. LLMJunky did the math (the headline figure: ~$60B): "JUST IN: VS Code fork sells for 60 Bs. With a capital B. And an uppercase 60. That's a lot of cash. 🤯" In a follow-up he worked out what it means for the cap table: "Truell owned a 4.5% stake in the parent company of Cursor, Anysphere… Their previous valuation was approximately $29B, but with this deal, Michael's stake becomes worth roughly $2.7 billion — good enough for first. Not too shabby for someone who, just a few years ago, wasn't old enough to buy alcohol." Jerry Liu put the trajectory plainly: "Congrats to the team for being AI's biggest growth/success story so far (outside the labs, which don't count) 🚀📈."
The reaction. The 3-million-view reply section split between awe and Musk-empire dread. NIK: "Welcome to the Musk Empire." Kent C. Dodds: "huge news. I'm excited for you!" The recurring practical question, from GitLawb: "what will happen to Grok Build?" And the inevitable bit, from ballusmaximus: "super edgy groksor commit messages are gonna be 🔥."
What Cursor still has to build. The sharpest analysis came from Jerry Liu, reading the "useful AI" (not "for developers") positioning: "It's no secret that coding agents are a great proxy for computer use and generalized knowledge work. After all, Claude Cowork is just a UI wrapper on Claude Code. I think Cursor has a great shot here. However… 1. For a long time you couldn't upload a PDF to Cursor 2. It needs a broader set of interfaces to knowledge-work data sources (CRM, ERP, ITSM…) 3. It needs more integrated surfaces to manage, read, and edit unstructured documents — from markdown to Word/PowerPoint to Excel." The read: Cursor is aiming straight at Codex and Claude Desktop, and the harness, not the model, is the gap.
Origin: agent-native version control, shipped hours later. Same team, same day. swyx broke it (103 replies, 2,474 likes, 341K views): "Cursor/Graphite's @TomasReimers just announced Origin — @cursor_ai's long awaited Git competitor, scalable for agent workloads, extensible with api and mcp, and built-in merge conflicts and [CI] failure agent resolution." He clarified the pitch in two replies: "*github competitor fml… *ci/cd failure resolution, thanks tim apple." Tech News connected the dots: "this drops the same day SpaceX announces the $60B cursor acquisition. Timing aside, agent-native version control is the actual infra bottleneck right now." The clearest articulation of why, from Sunshine Burri: "Git for agents makes so much sense. Human version control assumed a slow, mostly linear history. Agents branch wide and fast, so the merge and conflict layer is exactly where it breaks." And the line of the thread, from Weighty: "an agent to resolve the merge conflicts the other agents made. Beautiful. They finally invented a manager." (The lone GitHub loyalist, Buggy D. Code: "vibe coded github competitor? ill stick with github.")
"Git 2.0" is officially a category now. Kyle Gill counted the field: "So we're now up to 4 of these? — Origin from Cursor/Graphite, DeltaDB from Zed, Project Switch from GitLab/Anthropic, Code.storage from Pierre." swyx added his own: "mine — forge.smol.ai." Five agent-first version-control bets in flight; the primitive everyone said couldn't be replaced is suddenly the hottest greenfield in dev tools.
The Loop People Were Right (Even Theo Caved)
The capitulation. Twenty-four hours after the timeline's loudest day of loop-skepticism (Pocock's "instruction rot," the self-improvement-loop dread), Theo flipped (141 replies, 1,415 likes, 161K views): "I hate to admit it but the loop people were right." The obvious follow-up was the most-liked reply, from V: "stop tweeting, make the vid already."
The same three fights, reopened. The replies relitigated every concern the loop crowd has been waving off for months:
- Cost. Ananth: "Doesn't it require unlimited tokens?"
- Regressions. nick_rackley: "How do you avoid regressions in complex codebases though? I'm open minded about it, but nobody has been able to explain this to me."
- Where it actually pays off. The most useful answer, from Travis Gautier: "Highly overrated for developing prod. Highly recommended for completing large batches of refactor targets or identified bugs."
And the genealogy question that comes up every time, from jass: "how is this different from Ralph loops?" — plus the perennial confusion, from Aculnaj: "is loop the same as /goal?" The dissent was pithy, from Willy Drucker: "10 PRINT \"No.\" / 20 GOTO 10."
The autonomy receipts. Fueling the loop convert wave, one widely-shared post (RT'd by steipete) made the trendline concrete. prinz (213 RTs, 2,790 likes, 162K views): "Two years ago, reasoning models did not exist… One year ago, the best public model was o3, and 99% of those who used it were using it as a chatbot. Today, I have Codex building several pieces of personalized software and running a large-scale data-management project that's been autonomous for nearly 4 days. Projecting this trend to June 2027 is absolutely mind-boggling." The killer follow-up was about the death of off-the-shelf software, in a reply: "it is now easier to just build software of my own… than to spend time locating an app someone else built, installing and learning it, and finding out it does only 95% of what I want. But even this phase will be short-lived — we'll quickly progress from 'Codex, please build it' to 'Codex, please do it.'" Hello Chicken echoed it: "I make custom 1-off software for all kinds of tasks now… an hour or two of work to save 100 hours." The skeptic's counter, from QRDL: "if there's some good-faith corner of your brain: this could just be high-end data labelling" — answered with the flex of the day, prinz: "I'm writing this while listening to FLAC streamed off my NAS to an Android music app Codex built for me last night… optimal settings for my specific headphones, it sounds like HEAVEN, the future is now."
Pocock's Leitwörter & the Words That Steer the Agent
The craft idea of the day. Matt Pocock named the thing his best skills have in common: "The outrageous effectiveness of Leitwörter. A leitwort comes from literary theory — a repeated word or phrase used throughout a text to anchor meaning. In skills, a leitwort is a word the agent uses to guide its own behavior. Take 'zone of proximal development' from my /teach skill. I use it only a couple of times in the SKILL.md, but I see it almost every time the agent invokes the skill — 'Let me adjust the lesson so it's in the user's zone of proximal development.' That single phrase encodes how the agent should behave in a concise token it can repeat to reinforce its own behavior — and it likely tickles the parameters related to educational research and being a good teacher. For engineering, leading words like 'tracer bullets,' 'deep modules,' 'test seams,' 'clean code' are outrageously effective for leading the agent to produce better code." It's the constructive flip-side of yesterday's "instruction rot" panic: not every word the agent repeats back is noise — the right ones are steering.
And the proof that /teach sticks. Pre-empting the obvious objection, Pocock posted video evidence: "Folks say the stuff you learn from the /teach skill doesn't stick. 'Did AI really teach you how to solve a Rubik's cube?' Yes. Here's the proof (don't expect any great speed, but I can do it)."
A lighter note on the harness reprieve. Following Anthropic's reversal on blocking claude -p and third-party harnesses, Pocock summed up the mood: "Anthropic has opted against slashing AFK/third-party usage, at least for now. Go back to sleep, Codex, I don't need you yet."
Opus-Level Local Models, If You Can Afford the VRAM
GLM 5.2 cracks the Opus tier on DeepSWE. LLMJunky called it (127 likes, 9.9K views): "GLM 5.2 has outscored Opus 4.5 in most benchmarks, including one that is difficult, if not impossible, to benchmaxx: DeepSWE. GLM 5.2 scored ~46%, Opus 4.7 max effort ~45%, [and Opus 4.6 scored 28%]. I don't think it's up for debate anymore. We have Opus-level local models now." The catch is the hardware: "Even at 4-bit quantization with full context you're looking at ~500GB of VRAM… ~6x RTX 6000 for usable speeds. The price of admission is likely around $75K all in. But it's yours." (He later accepted a correction from Lukas that GLM-5.2 is "744B total, 40B active," not 1T+ — "Derp. Well, that makes things a bit easier, doesn't it?") The reply that captured the whole subculture, from Screenclutter: "Local LLM lovers are just recreating LAN parties for local inference smdh."
The open-weights drumbeat continues. Armin Ronacher kept pushing the practice: "Do more experimentations with local models people!" (pointing at Vicki Boykis's running-local-models write-up) — and, in a nod to where the useful data actually lives, praised the transparency: "I really appreciate the data that OpenRouter provides on models. Wish more people had that transparency." Yesterday's "frontier-as-orchestrator, open-weights-as-implementer" pattern is hardening into the default for the privacy- and cost-conscious; GLM 5.2 is the first credible "you don't need the orchestrator either" data point — for anyone with $75K of GPUs.
Also Worth a Look
- Codex gets
.worktreeinclude. Via steipete (RT of btraut): "Codex now supports.worktreeinclude! Worktree enjoyers can now automatically copy over things like.envfiles, config files, agent overrides, etc." A small but real quality-of-life fix for parallel-agent / multi-worktree setups. - OpenClaw 2026.6.8. The lobster shipped again: "🦞 OpenClaw 2026.6.8 just dropped — richer Telegram + WhatsApp, sharper agent & Gateway recovery, new models + tougher memory, native
/usagefooters, smoother WebChat & iOS." - A hidden-but-powerful Codex browser setting. Surfaced by Kevin Kern (RT'd by LLMJunky): a Codex app setting that's "a bit hidden but powerful" — part of the broader scramble to make sense of computer-use vs browser-use vs the Chrome extension (a confusion jxnl wrote a guide to untangle).
- Documentation pays for itself. LlamaIndex quantified it: "How much can good documentation save an AI agent in cost and time? Turns out, a lot." A useful counterweight to the "docs go stale, skip them" camp — for agents, good docs are a token-cost lever, not just hygiene.
- GitHits raises $1.75M. Pre-seed for the GitHub-analytics startup, led by Vendep Capital (RT'd by Jerry Liu).
- Where the Fable export saga actually stands. Still unresolved, and the negotiation is now a play-by-play. Theo flagged "very good reporting" from Sophia Cai: "It will likely take longer than a few days… Trump has asked Howard Lutnick to be more involved with Anthropic… over the weekend Anthropic assembled a technical team including Logan Graham (Frontier Red Team), Dave Orr (head of safeguards), and Nicholas Carlini to engage the admin… Monday they gave an in-person presentation to officials from ONCD and Commerce's CAISI." Meanwhile the discourse curdled into self-parody — Theo RT'd roon: "pretty good midwit test for when someone was unimpressed with fable."