AI Digest — July 27, 2026, 8 PM
Summary
The day splits cleanly into two threads that reinforce each other: the arrival of Kimi K3 and the maturing craft of agent-driven development. Three sources cover K3 from different angles — Willison on its more restrictive “open weight” license and pricing, vLLM on the deep serving engineering behind its hybrid KDA cache, and Nate B Jones using it to argue that lumping all Chinese models together obscures a 17x price gap between K3 and DeepSeek. Meanwhile the GitHub and Tessl pieces converge on a shared “harness/factory” philosophy — autonomy over hand-approval, cheap verifiers and loops over hero prompts — though they disagree on tooling emphasis, with GitHub’s post insisting the bare harness is nearly all you need while Tessl builds an entire platform of registries, maintenance agents, and eval suites on top of it.
The harness is all you need (mostly)
The author argues that success with AI coding comes not from stacking up skills, MCPs, custom agents, or clever prompts — most of which is "gimmicks" and "slop" — but from deeply understanding and using the harness itself, treating GitHub Copilot as the agent harness. The recommended starting point is the Copilot CLI, because its bare text interface keeps you close to the raw interaction rather than hiding it behind UI. A key step is turning on "YOLO mode" (Allow All / /allow-all), since agents need autonomy to deliver real productivity — approving every command trains you to rubber-stamp without reading. The safety caveat: never run YOLO mode on your local machine, especially at work, because private org data and costly mistakes are at stake.
Read the source →GitHub Copilot app for Beginners: Getting started
The GitHub Copilot app reframes AI coding away from a single chat window toward a workspace that manages multiple concurrent agent sessions, each tied to a project so it carries the repository context needed for a task. You can keep several threads moving at once — using Quick Chat to ask questions or investigate the codebase while another session continues a project update — without losing your place. The app also adds an interactive browser canvas via the /create-canvas slash command, letting you preview a running app inline, and with Canvas Dev Mode's "Pick & Polish" you can select page elements directly as context for the next request. Agent Merge extends the workflow through the standard pull-request, CI, and review process.
Read the source →moonshotai/Kimi-K3
Moonshot released the weights for Kimi K3, a 2.8-trillion-parameter model, weighing a hefty 1.56TB on Hugging Face. The licensing is the notable story: where K2 used a "modified MIT" license merely requiring large commercial users (100M+ MAU or $20M+ monthly revenue) to display a "Kimi K2" attribution, K3 drops the MIT framing entirely and now requires any Model-as-a-Service business exceeding $20M revenue over any 12 months to enter a separate agreement with Moonshot before commercial use. To Moonshot's credit, they never call it "open source," consistently using "open weight" instead. OpenRouter is already serving K3 from seven providers, most matching Moonshot's own pricing of $3/million input and $15/million output tokens.
Read the source →An opinionated guide to which AI to use to do stuff
Willison tracks the evolution of Ethan Mollick's recommendation guide, which a year ago centered on chat (ChatGPT, Claude, Gemini with o3, Claude 4 Opus, Gemini 2.5 Pro) but is now dominated by agentic systems capable of "many hours of real human work in one go." Gemini has fallen off the list because Google lacks an established entry in the Codex/ChatGPT Work/Cowork category. The guide untangles the confusingly named agent modes: ChatGPT offers Work and Codex, Claude offers Cowork and Code, with no meaningful mapping between the names. A notable practical detail: flipping ChatGPT mobile from "Chat" to "Work" mode removes the internet restriction from its Code Interpreter container, and desktop ChatGPT Work is effectively a friendlier skin over Codex.
Read the source →Harness Engineering: Building an AI Software Factory
Dru from Tessl presents "harness engineering" (aka loop/meta engineering, software factories) — a shift from writing code to building and monitoring loops of agents that build and review the product, with humans shaping tickets and improving system autonomy, automation, and quality. The technique rests on three loops: an inner loop (unit tests, linters, TDD) that drives per-attempt correctness, an outer loop (agentic code review, agentic QA clicking through the built product) that builds confidence to reduce human review, and a meta loop of maintenance agents that watch CI, PR comments, and logs to propose their own fixes. A central discipline is that when an agent produces a broken PR you don't correct it — you throw it away and add a test or skill so the mistake can't recur, a "slot machine" workflow that's organizationally painful but compounds. Tessl cites OpenAI's reported 5x productivity gain from moving engineers from interactive sessions to ticket/PR flows, and shares that its own 24 engineers now ship ~700 PRs/week while still human-reviewing everything touching production. Tessl also pitches cheap LLM "verifiers" (100+ per PR at ~$0.30/day) as a shift-left alternative to $25-per-PR agentic reviews.
Read the source →US AI Dominance Is Over: Here's Why
Nate B Jones pushes back on the lazy shorthand of "Chinese models," arguing that DeepSeek, Kimi, and Qwen share only a country of origin — not price, deployment path, hardware burden, or best use — even though people conflate "Chinese" with cheap, open, or local (sometimes all three, sometimes none). He illustrates with Kimi K3, a 2.8-trillion-parameter model with a million-token context window that charges $15 per million output tokens, versus DeepSeek V4 Pro at just 87 cents for the same volume — a roughly 17x price gap between two "Chinese frontier models." He notes Moonshot calls K3 open weight but the downloadable weights weren't yet available at recording time (expected imminently). The video's stated aim is a practical framework for deciding whether and when to include a Chinese model in your stack rather than treating them as a monolith.
Read the source →Kimi K3 Is Here: Efficient Day-0 Support on vLLM
vLLM announces day-0 serving support for Kimi K3, a 2.8-trillion-parameter Mixture-of-Experts model (16 of 896 experts active per token) with a 1M-token context and native vision, built on Kimi Delta Attention (KDA) and Attention Residuals. Most layers use KDA — a linear-attention mechanism keeping a fixed-size recurrent state instead of a growing KV cache — interleaved with periodic full-attention layers for exact global recall, which is what makes the 1M context affordable. vLLM introduces a single hybrid KV-cache manager holding paged KV blocks and compact recurrent-state blocks under one scheduler, plus a dedicated KDA backend running FlashKDA for prefill and fused CUDA kernels for decode. The hardest engineering problem was prefix caching across the hybrid cache, solved by decoupling large physical KDA state blocks from fine-grained prefix matching so long shared prompts can reuse both state and KV. Recommended hardware is 8 NVIDIA B300 or 8 AMD MI355X GPUs, and for now only Docker images work due to pre-release dependencies like FlashInfer.
Read the source →TIRx
The transcript for this video was unavailable. Per the video description, Bohan Hou introduces TIRx, an open-source programming DSL for GPUs and AI accelerators that combines native-level control with tile APIs to reduce boilerplate, built on a simple layout system, and being used to build frontier ML kernels and mega-kernel systems while exploring agentic kernel programming.
Read the source →