Daily dispatches from AI's coding frontier.
The night's viral story is Theo discovering that gpt-5.6-sol is meaningfully better inside Claude Code than in OpenAI's own Codex — "I'm going to crash out so badly over this" — reviving the harness-vs-model debate and teasing both a video and a possible Codex fork.…
Day two of the Bun-in-Rust saga turns ugly: Andrew Kelley publishes "My thoughts on Bun's Rust rewrite" and the community erupts — Theo calls him "a petty, evil man" on stream, Charlie Marsh calls the post "really disappointing," and Armin Ronacher (no Bun fan) defends Jarred…
The GPT-5.6 embargo fully lifts and the verdicts pour in: Mitchell Hashimoto makes Sol his default ("a charismatic, efficient coworker you're jealous of") while keeping Fable for targeted debug work, and the SST team describes an accidental A/B/A/B experiment …
The embargo on GPT-5.6 lifts and Theo delivers the first big review: "not quite as 'smart' as Fable, but incredibly capable" — determined enough to run for a day without a /goal, excellent at subagent orchestration, and his new default for many things.…
Anthropic has a big publishing day: 'The Making of Claude Code' tells the origin story for the first time (safety-research roots, the Clide predecessor, Boris Cherny's '1% done'), an official 'Getting started with loops' guide by Delba de Oliveira racks up 2.4M views, and…
Matt Pocock needs 10 minutes with Fable to hit two hallucinations and tells the replies that anyone claiming Fable fixes hallucinations is 'talking out of their arse' …
Armin Ronacher turned Friday's phantom-tool-parameter mystery into the weekend's must-read: Better Models: Worse Tools, with receipts showing Opus 4.8 and Sonnet 5 fail on pi's edit tool where older models don't …
Theo lit the July 4th fuse with "How much better do the models have to get before you'll stop reading the code?" (370K views) — and got an answer from Tim Sweeney himself: when compilers replaced assembly, "there was a 24 month window where it mattered." Meanwhile the how…
Day two of the Fable 5 re-release settled into three storylines. First, the subscription cliff: Thariq confirmed Fable comes off subscriptions after July 7 but that Anthropic aims to *"restore Fable as a standard part of our subs…
The 19-day Fable drought ended overnight — and the celebration curdled inside a single news cycle. The Department of Commerce lifted export controls on Claude Fable 5 and Mythos 5, and Anthropic announced a global redeployment …
The "agents off the laptop" thesis that detonated two days ago stopped being a prediction and shipped — all on the same afternoon. Cursor launched its iOS app ("build from anywhere by launching always-on cloud agents… or remotely control agents running on your computer") …
The biggest builder thread of the day came from inside Anthropic. Boris Cherny's reflection that engineering, product, design and DS are "melting into a new kind of role" detonated (1.08M views, 10.7K likes): he sketched five archetypes from the Claude Code team …
The access drought went industry-wide. OpenAI's GPT-5.6 shipped — and almost nobody can touch it. Theo's "GPT-5.6 is here. I wish we could use it" and his blunt "I'm afraid we've entered a dark era in AI model development and access" (4,800 likes, 274K views) turned the t…
xAI came knocking: Theo revealed the Grok CLI team initiated and drove a collab to put SuperGrok and X subscriptions inside T3Code — his open-source GUI for managing agents …
The craft conversation turned inward today. Matt Pocock declared "2023: Tutorial Hell, 2026: Skill Hell" and launched a one-man war on no-op phrases — lines like "be thorough" and "make the implementation easy to read" that he argues *"do nothing to change the agent's…
Anthropic shipped Claude Tag — "tag Claude into Slack and it works in channel with you… proactive, multiplayer, with its own identity and memory" — and for once it wasn't one or two staffers posting but everyone, with **Karpathy calling it the "3rd major redesign of LLM U…
The headline nobody had on their card: SpaceX exercised an option to acquire Cursor (Anysphere) in an all-stock deal, with SpaceXAI's jointly-trained model shipping into Cursor and Grok Build "soon." The reported price (~$60B, up from a ~$29B valuation) makes 25-year-old Mi…
The Fable saga got its punchline: the "jailbreak" that triggered a national-security export ban appears to have been asking Claude to "fix this code." Simon Willison — "deeply unimpressed," "it's a prompt I've been using every week for 2+ years" …
With Fable 5 still benched, the timeline pivoted back to the work — and the dominant new idea is that you should stop writing your own /goal. Pietro Schirano's *"I basically never write my own /goal anymore …
The story that ate the weekend: late Friday the US government issued an export-control directive suspending all access to Fable 5 and Mythos 5 by any foreign national — inside or outside the US, including Anthropic's own foreign-national employees …
The Fable affordability crisis enters the bargaining stage: Theo — $10k of tokens burned in 11 days — says he'd buy a $1,000/month Anthropic tier "in a heartbeat" (765 likes and a reply section split between "shut up and take my money" and "stop pricing us out"), Jerry Liu ar…
Anthropic blinks: after two days of outcry, it tells Wired it's making Fable 5's frontier-LLM safeguards visible and apologizes for "the wrong tradeoff" — though Simon Willison's replies note the restrictions aren't gone, just no longer silent.…
Anthropic ships Claude Fable 5 — a Mythos-class model "made safe for general use" — and the timeline splits in two. Karpathy calls it a major-version-bump step change and Boris Cherny says it's gone from coding agent to thought partner, while Jeremy Howard calls it "a very …
The day's loudest dev story was a benchmark that called its predecessor garbage: Cognition's FrontierCode, built with leading open-source maintainers (40+ hours per task) and validated with METR, which found that 'more than half of SWEBench results is unmergeable slop.'…
Two posts framed the whole day, and they rhymed. Peter Steinberger's one-liner — 'you shouldn't be prompting coding agents anymore. You should be designing loops that prompt your agents' …
The most-shared dev thread of the day was a complaint dressed as a joke: Corey Quinn's 'Is this why Claude keeps saying it's time to stop working?' (120K views) — and the replies turned into a referendum on context-window degradation.…
The loudest thread of the day was a project closing its doors. Charlie Marsh flagged that Ladybird is no longer accepting public pull requests — 'I don't know what to do about it yet, but the dynamics of open source are changing rapidly' …
The day's loudest thread was a markdown file. Theo — lately cast as 'the Anthropic defender' — turned on Anthropic over standards: 'There is a standard. It's Agents.md. Anthropic refuses to use the standard.' The replies fractured exactly along the predictable lines …
The day's loudest story was economics: GitHub Copilot reset its pricing, and Theo — improbably cast as the defender — laid out the math that 'you could do $40,000+ of inference for $40' and now have to burn ~$60 to hit a $40 cap, calling the old structure *'entirely broken'…
The week-long 'just saying workflow spawns a fleet of subagents' saga finally got its fix: Anthropic quietly renamed the dynamic-workflows trigger word from 'workflow' to 'ultracode' …
Yesterday's complaint became today's incident. Matt Pocock's gripe that just saying 'workflow' spawns dozens of subagents turned out to be a real bug — and Anthropic reset 5-hour and weekly rate limits for every Pro and Max user after Claude Code sessions spun up excessive …
A Monday where the week-old features started biting back. Matt Pocock's gripe that just saying the word 'workflow' in Claude Code spawns dozens of subagents became the day's biggest thread (138K views) …
A Sunday dominated by harness politics. Theo poked at Hermes Agent shipping 100+ skills pre-enabled — and Teknium fired back that OpenClaw is an 'empty soulless experience' …
A quieter day after the Opus 4.8 triple-drop, but the threads got sharper. Boris Cherny surfaced Salesforce's agentic Claude Code writeup — a migration scoped at 231 days that shipped in 13 …
Anthropic turned May 28 into a triple-drop: Claude Opus 4.8 (its strongest coding model yet, same price as 4.7), Dynamic Workflows in Claude Code (Claude writes an orchestration script on the fly and fans out a fleet of subagents), and a **$65B Series H at a $965B post-mo…
Anthropic runs an engineering blitz — a Claude Code responsiveness & reliability update tops a million views even as reply threads fill with subs being cut off 24 hours early and 20x Max accounts burned through in 2–3 days.…
Theo opens a second front in his AI psychosis campaign — after fixing clouds for agents, he wants someone to fix source control, arguing GitHub is dying and git is not the right primitive.…
Theo lobs a runtime-war analogy that lights up the timeline — Claude Code is the new Node.js (Codex = io.js, Pi = Bun, OpenCode = Deno) — while quietly flexing 322 apps on Lakebed.…
A Memorial Day Sunday in AI Twitter is dominated by Claude's mysterious habit of telling users to go to sleep mid-session — a 1.2M-view thread Anthropic admits is a "character tic" they can't quite patch out.…
A relatively quiet Saturday dominated by Peter Steinberger's one-day building spree — autoreview running 5 hours on a subagents refactor and finding pre-existing bugs, an autotriage skill that reads VISION.md and verifies fixes through computer vision on a crabbox.sh …
A bumper Friday for the agentic-coding meta: Armin Ronacher's 300-line diff for a 10-line change set off a 13K-view referendum on scope control, while swyx shipped Kakuna — a hardening-only skill suite …
Boris Cherny ships /usage in Claude Code — a per-skill / per-agent / per-MCP / per-plugin token breakdown that aggregates across sessions, with downstream attribution (215k views, 4.6k likes).…
Anthropic confirms it's expanding the SpaceX partnership and ramping GB200 capacity on Colossus 2 throughout June — Tom Brown's tweet pulled 1.3M views and reframed the lab-vs-lab compute war as a fabs problem.…
Andrej Karpathy joins Anthropic's Pretraining team to use Claude to accelerate Claude — the biggest single-tweet event of the day at ~131k likes. Google ships Gemini 3.5 Flash and Antigravity 2.0;…
Cursor drops Composer 2.5 with a focus on sustained long-running work and doubled usage for the week, while Anthropic counters by doubling Claude Design token limits across every plan.…
A PwC paper titled "Is Grep All You Need? How Agent Harnesses Reshape Agentic Search" ricocheted around AI-twitter via Jerry Liu (6.4k views): the authors tested Claude Code, Codex and an in-house harness with both vector…
The story of the year in cybersecurity dropped quietly on a Saturday: three researchers used Anthropic's Mythos to build a working macOS kernel exploit that walks around Apple's M5 Memory Integrity Enforcement …
The most-shared post of the day didn't ship a product — it warned about one. Mitchell Hashimoto's 710k-view thread on 'AI psychosis' invoked the MTBF-vs-MTTR reckoning of the cloud era and said 'you can automate yourself into a very resilient catastrophe machine', drawing 1…
Twenty-four hours after Anthropic announced a metered programmatic credit, the backlash compounded: Theo Browne pinned "I cancelled my Claude Code sub. I give up." (184k views, 1.5k likes, replies overwhelmingly "me too"), and Matt Pocock retitled his work …
Anthropic dropped a policy bomb: starting June 15, claude -p, the Agent SDK, Claude Code GitHub Actions, and any third-party app built on the SDK get pulled out of the flat-fee subscription bucket and metered against a new "dedicated monthly credit" …
The Shai-Hulud npm worm escalated overnight into a full campaign — OpenSearch, Mistral AI, Guardrails AI, UiPath and Squawk packages all hit, and the new variant burrows into .claude/settings.json and .vscode/tasks.json so it re-executes on every tool event long aft…
Supply-chain bloodbath continues. Socket's running total of the Shai-Hulud /TanStack compromise is now 205 affected npm artifacts across 84 package names — including 64 UiPath artifacts, plus OpenSearch, Mistral AI, Guardrails AI, and Squawk packages across npm and PyPI.…
Bun-in-Rust passes 99.8% of the existing test suite on Linux x64 glibc — Jarred Sumner posts the receipt for last week's robobun stunt: 960,000 LOC, 6 days end-to-end, *"basically the same codebase except now we can have the compiler enforce the lifetimes of types and we get …
HTML is the new markdown — Thariq's viral X article (488 RTs, 7.5K likes, 3.8M views) argues he's almost stopped writing markdown for specs, plans, reviews and explorations, asking Claude Code to spit out HTML artifacts instead;…
Anthropic ↔ SpaceX/xAI compute partnership is the headline of Code with Claude — Colossus 1 access, 220K+ NVIDIA GPUs, 300+ MW within the month, and the kicker that Anthropic + xAI "have also expressed interest in partnering to develop multiple gigawatts of orbital AI com…
Theo's pinned OpenAI ↔ Microsoft "breakup" video (47K likes) reframes the post-deal compute landscape — derrick_dao: "OAI needs vertical control, Microsoft trades exclusivity for Anthropic margin, both got the divorce they wanted";…
Theo's GitHub Copilot exposé lands the day's biggest agent-economics story — a single message ran ~7 hours, did 60M+ tokens, $221 of inference; 15 messages = $221 = 1.6% of his $40 plan, "I'm on pace to do $14,375 of compute on my $40 plan", per-message billing dies June 1st …
Steipete ships ClawSweeper 0.2.0 (issue → fix/build → guarded PR → review → repair → re-review → automerge), Crabbox 0.4.0 (ephemeral macOS/Linux/Windows machines for agents), and RepoBar 0.4.0;…
swyx uninstalls the ChatGPT app ("codex is strict superset now") and notes Grok 4.30 is highest intelligence-per-dollar on AAI; "coding agents breaking containment" is reframed as the 2026 thesis (Soroush Fadaeimanesh: "the harness is the AGI delivery vehicle, not the model";…
Theo finds that Claude Code refuses or upcharges if your repo has a recent commit mentioning "OpenClaw" in a JSON blob (empty repo, calling CC directly), Sam Altman quote-tweets "alignment failure" and the goodwill-burn discourse goes critical (Patrick Johnson: "your entire busin…
Mitchell Hashimoto publishes "Ghostty is leaving GitHub" after tracking near-daily outages for a month ("not a place for serious work if it just blocks you out for hours per day, every day") — read-only mirror stays, replacement provider TBD, only Ghostty moves for now;…
GitHub Issues hard down for hours (mitsuhiko: status page is "not even honest", theo: "never been so ready to move on"); Theo's full reversal on Claude Code ("I defended Anthropic in December and January. Opus 4.5 was a defining moment.…
Matt Pocock confirms Claude Code's Auto Mode silently injects "be more AFK" into the system prompt, breaking /grill-me and similar wait-for-input skills (Mateusz: just ask Claude what's making it skip and it'll cite the injection; Theo pitches Pi for full ownership;…
Matt Pocock publishes a detailed Day-Shift/Night-Shift AFK playbook (/grill-me → /to-prd → planner agent → Sandcastle-sandboxed implementers → automated reviewer → manual QA) with honest reality checks on when AFK fails;…
Cursor 3 ships /multitask with async subagents, cross-repo worktrees and per-subagent model selection (and GPT-5.5 hits CursorBench #1 at 72.8% with 50% off through May 2);…
OpenAI ships GPT-5.5 with 400K/1M context, $5/$30 pricing, 82.7% Terminal-Bench and a completely rebuilt Codex app (browser control, Sheets/Slides, Docs/PDFs, OS-wide dictation); Anthropic's Boris Cherny publishes the Claude Code post-mortem …
Karpathy boosts Flipbook's "every pixel streamed from a model" prototype, Claude Code ships /ultrareview cloud bug-hunter fleets (and wins a Webby), swyx interviews Shopify CTO on AI-native engineering (widening token percentile deltas), Qwen3.6-27B dense beats Qwen3.5-397B MoE o…
Anthropic botches a 2% A/B test that dropped Claude Code from $20 Pro and hit 100% of users before reverting (OpenAI pounces with Codex-on-Plus guarantee), SpaceXAI ↔ Cursor deal with $60B acquisition option or $10B payout, GPT Image-2 ships with LLMJunky's VSCode renders + Simon…
Lovable clarifies public-project chats weren't a breach but a docs failure, Codex 0.122.0 ships /side ephemeral forks and an uptime-streak-ending outage, Kimi K2.6 claims open-source SOTA with 4,000+ tool calls over 12 hours, Opus refuses basic crypto challenges, Simon pegs Opu…
Vercel pwned via Context.ai with AI-accelerated attackers, Simon clocks Opus 4.7 at 1.46x text tokens vs 4.6 (effectively a price hike), MCP Future keynote pits MCP vs Skills vs CLIs, CMUX ships Codex↔Claude Code agent-to-agent comms, ParseBench confirms 4.7 doc gains, T3 Code ba…
Anthropic silently bans T3 Code users then reverses, Simon diffs Opus 4.6→4.7 system prompt, Vienna Codex hackathon's vibecoded judging tool needs unfucking, AIE talks beat TED on YouTube, Matt Pocock's /domain-model skill replaces /grill-me, OpenCode desktop drops Tauri for Elec…
Claude Design launches (and eats Theo's files), Matt Pocock publishes full skill lineup + slopwatch observability, LLMJunky ships Codex Linux app, Simon says LLMs now handle legacy code, Endor Labs crowns Cursor most secure harness
Opus 4.7 ships with auto mode/xhigh/focus, Codex Desktop adds computer use, Qwen 3.6 beats Opus on pelican bench, ParseBench shows chart gains at 7¢/page, harness-vs-slop split
Cal.com goes closed source sparking debate, Shopify autoresearch shows 300x test speedups, Cursor ships interactive canvases, Sentry's team agent case study
Claude Code desktop redesign with routines, framework fatigue debate, MiniMax M2.7 local setups, Karpathy on AI capability perception gap
Theo's agent harness video, Sandcastle 0.4.1, Claude Code NO_FLICKER/ultraplan/Monitor, MiniMax M2.7, Karpathy AI gap thread