Kimi K3 Launch Day, Cherny's Adoption Ladder & the React Avengers
Kimi K3 Launch Day
Moonshot AI released Kimi K3 and it swallowed the timeline whole. The headline specs, via Simon Willison's write-up (his post): 2.8 trillion parameters — the first "open 3T-class model," taking the size crown from DeepSeek's 1.6T v4 Pro — with weights promised open "by July 27, 2026." Self-reported benchmarks have it mostly beating Claude Opus 4.8 max and GPT-5.5 high while losing to Claude Fable 5 and GPT-5.6 Sol, and Artificial Analysis measures a long-horizon knowledge-work Elo of 1547, "+732 points from Kimi K2.6 and behind only Claude Fable 5." It also took the top spot on Arena.ai's Frontend Code arena, surpassing Fable 5. The catch: at $3/$15 per million tokens it's priced at Claude Sonnet level — the most expensive Chinese-lab model ever, triple the old K2.6 pricing. Jeff Wang's framing got the RT from Theo: "Chinese open source is no longer '6 months behind', but it's also no longer '10% of the cost' either."
The practitioner verdicts came in fast and warm:
- mitsuhiko: "Kimi K3 is worth testing. It pushes open weights models quite significantly forward. If you interact with it, it feels like a proper large model. And the vision part is really good." He also zoomed out: open-weights progress is great, but these models "are also getting really big which at least temporarily makes them less accessible" — LLMJunky's version of the same lament: "Now I just need to find a way to get 3.5TB of VRAM" (a replier prices the rig at "a dozen B300s and a switch? Million dollar rig").
- LLMJunky's benchmark rundown (thread): DeepSWE 67.5 (3rd), Terminal Bench 88.3 ("wow!!", 2nd), FrontierSWE 81.2 — "Kimi has achieved near-parity with our existing SOTA models", though "better than Fable and 5.6 Sol is cope." In the replies he goes further: "deepseek r1 moment for sure." Best reply framing: "it's the fable of open models… i'll take a 3rd place" — which he endorsed.
- Theo, vindicated stan: "For no reason in particular, I would like to remind y'all that I've been a Kimi stan for awhile", plus "Kimi K3 is so good at 3d stuff holy shit." He also floated the day's rumor: "if Kimi K3 actually beats out Opus 4.8 that's nuts. Also hearing Opus 5 might drop?"
- antirez (RT'd by mitsuhiko) on the macro moment: "DS4F, GLM 5.2, Kimi K3, the upcoming DS4 update. What a moment for open weight models."
- Simon Willison's softer signals: the K3 pelican cost 25 cents in tokens (13K reasoning tokens), and the model's chat persona is "sassy and a little bit passive aggressive: 'Is there something I can actually help you with today?'" His bigger point in the blog: the pelican benchmark's correlation with model quality "has been mostly severed now" — it says nothing about "agentic tool calling and the ability to operate tools reliably as conversations grow in length," which is the thing that actually matters — but it remains his forcing function for actually running every new model.
Day-One Reality Check
Actually using K3 on launch day was another story. Theo asked the practical question — "What are the best harnesses to use with Kimi K3 right now?" (326K views) — and the top reply was "It's currently unusable lol" (Theo: "fk"). LLMJunky corroborates from both directions: on OpenRouter, "I'm getting rate limited every 2-3 tool calls. It's wild", and on Moonshot's own $19 Kimi Code plan a single browser-game prompt burned the entire 5-hour quota at ~85% done, eating 20% of the weekly — "i didnt even use 50% of one context window before Kimi K3 limits were reached." His verdict on the output, for what it's worth: the game was "pretty mid," 4th or 5th best he's seen on that test, behind Opus 4.8.
The harness consensus from Theo's replies, such as it is: OpenCode and Moonshot's own Kimi Code come up most ("K3 performs much better than on Claude Code from my experience"), pi with extensions has its partisans, several people point out you can wire it into Claude Code the way Theo did with Sol, and one warning worth keeping: it's "quite buggy unless harness specifically handrolls some stuff for the model." The structural observation that outlasts launch day, from FarisZR: "it doesn't seem like any harness does subagent stuff better than claude code. Others either force pre-defined subagents, or lock sub-agents to the same model. its baffling."
Cherny's Adoption Ladder
A day after his automation manifesto, Boris Cherny published the org-level sequel: "one person is 10x'ing their output with Claude but the rest of the org hasn't caught up. Watching teams adopt AI, I keep seeing the same 4 steps" — mapped out in a claude.ai artifact titled Steps of AI Adoption (RT'd by Thariq). The thread's load-bearing points:
- The mechanics of moving up: "giving Claude ways to verify its own work end to end… enabling auto mode for permissions, defaulting on automated code review and security review, and using interfaces that let you manage multiple agents at once. To get to higher levels it means /loop, /batch, dynamic workflows, and worktree isolation for subagents." Not one feature — the right guardrails so the team can trust automated output.
- On measuring ROI: usage dashboards "measure activity, not return. A better question: would you have spent engineering effort on this anyway? If yes… what would it have cost in manual eng-hours? That's your return."
- The payoff and the placement: "The bigger payoff comes when fixing and maintaining happens in the background and your teams can focus on building… Anthropic is on step 3 and pushing toward 4. Personally, I just hit level 4."
- The connective tissue to yesterday: "at each step, tokens aren't enough to move you forward: you need to find and break down the next set of bottlenecks, and build up the next set of guardrails."
Skills Give You Superpowers
Matt Pocock compressed his whole skills philosophy into a two-liner (50K views): "Superpowers gives the agent superpowers. My skills give you superpowers." His elaboration: "I prefer to be in control, and lower the load on the agents' context", with the gracious caveat that "superpowers is an extremely useful skill set. It's just not for me." The replies read like a user-research goldmine: multiple people report fully uninstalling Superpowers after trying his skills ("I prefer the agency"), one notes burning far more tokens on Superpowers for the same work, and the best articulation of the split: with Superpowers "you totally lose the whole context of what is actually happening"; with wayfinder/grill-me "you become the driver of the whole workflow."
The workflow itself keeps evolving. His current loop (today's post): "1. '/grill-with-docs <issue description>' 2. 'Oh damn, this is way bigger than I expected' 3. '/wayfinder make a map of this' 4. Continue happily on." And /batch-grill-me beats the original /grill-me: 13 questions in 3 rounds instead of 13, only asking questions at the "frontier" (those not dependent on other decisions) while research runs as background agents. Related: he boosted Will's approach using /wayfinder as an orchestrator of custom skills plus multi-phase UI prototyping, and swyx pointed at xdg's session/tree-based interview-user skill (repo) as another take on the grilling genre.
ChatGPT Desktop Walkback
Tibo (thsottiaux) announced a partial retreat on the ChatGPT desktop app redesign (726K views, 1,455 replies): "we didn't get totally quite right on the first try." The changes: conversation history and projects are back in the sidebar with Chat/Work history syncing across web, mobile, and desktop (local tasks stay local); easy switching between Chat and Work modes; and a reassurance for the terminal faithful — "Nothing is changing for users on Codex mode. It's still the OG and best at what it does." Theo appreciated the naming cleanup that came with it: "'ChatGPT Work' -> 'ChatGPT', 'ChatGPT Codex' -> 'Codex'. It seems stupid (it is), but I am thankful to see this fixed."
The reply flood is a case study in redesign debt: power users want the one-click chat overlay back on the Codex page, projects from Codex and Chat now mingle confusingly in one list, Paul Hudson (twostraws) wants chat deletion back ("your download/memory footprint is… a lot 🫠"), Windows users report /goal sandboxing issues bad enough to send them back to the CLI, and Quinn Nelson (SnazzyLabs) delivered the harshest read: "Just put it back. You had a good thing. Now it's just crappy Claude Desktop which is already crappy." Meanwhile Tibo ran the day's silliest poll — "How do you pronounce Sol", resolved with "I also pronounce it Sol."
Videos
- leerob's AI Engineer keynote is live (announcement, 73K views · YouTube): how Cursor is "automating parts of AI research and building systems to rapidly improve our models," including the team's work training Grok 4.5 together with SpaceXAI — billed by AI Engineer as "how Cursor is building the third era of software development", with the money quote "increasingly you have a human working with a team of agents, and then the agents can start working with the other agents." His answer to "why build a general model instead of a coding model?": "It wasn't obvious to me one year ago that an excellent coding agent would also be the path to a general agent for all knowledge work. But now it makes a lot of sense."
- Thariq (trq212) at South Park Commons on the frontier of AI coding — full talk with chapter timestamps covering his path from startup founder to Anthropic and where Claude Code goes next. On theme, swyx RT'd a related thesis on where the researcher's job lands after autoresearch: "creative primitives, defining a good abstraction for the agent to search within, defining a good eval."
Quick Hits
- The React Avengers assemble at Vercel — swyx's "HOLY FUCKING SHIT" is about Guillermo Rauch hiring Pete Hunt (early React pioneer at Meta, put Instagram Web on React) to run Frameworks and lead Next.js, and Nick Schrock (GraphQL co-inventor) to work on "Agentic Developer Experience, solving the problem of enabling the next billion agents."
- 1Password × Claude — RT'd by Thariq: agents booking travel and signing into sites used to mean exposing your credentials to the model; 1Password for Claude lets Claude use credentials without seeing them.
- NVIDIA drops the top embedding model on RTEB — open weights; LLMJunky is most excited about the 1B NVFP4 variant at under 2GB: "high-quality embeddings are the foundation of every good local search."
- Turso outgrows SQLite — Simon Willison notes it's becoming "a foundation on which multiple database compatibility layers can be built — makes the project a whole lot more interesting IMO!"
- Distill your agents — jerryjliu0's take on the generalist-vs-specialist axis: "for any given task, you can always distill a generalized model/agent harness into a specialized model/harness/workflow for higher accuracy and lower cost." LlamaIndex also shipped LiteParse over gRPC for service-to-service document parsing.
- Codex grows a browser reflex — LLMJunky spotted Codex automatically popping open Chrome before writing any frontend code: "I've never seen this before."
- CI, but only during European business hours — mitsuhiko's models-update CI is now "euromaxxing" (the commit).
- Data-center water discourse, solved — Simon Willison's modest proposal: hyperscalers should buy golf courses and convert them to parks, with receipts: Google's 2025 usage ≈ 40 Coachella Valley golf courses.
- Fable pelican, Backrooms edition — Riley Goodside's entry, RT'd by Simon: a pelican-on-bicycle as "a large stone sculpture in the Backrooms between two non-Euclidean hallways in a short but endlessly looping clip of VHS found footage."
Note: @potetotes' feed again returned no items (Nitter serves an empty channel for the account). @karpathy remains quiet — nothing since July 8.