Focused on frontier AI technologies, empowering AI researchers' progress.
A leaked CLAUDE.md file in Apple Support App reveals a full‑stack AI system built with Claude Code, details Apple’s reliance on Anthropic’s Claude, a switch to Google Gemini, and a price hike of the Mac mini driven by soaring AI demand.
In a 2026 Sequoia Ascent interview, Andrej Karpathy argues that large language models are not merely speed‑up tools but a new computing paradigm that renders many legacy apps obsolete, elevates understanding as humanity’s final competitive edge, and relegates CPUs to auxiliary roles, while outlining software evolution, jagged intelligence, and the rise of agentic engineering.
The article analyzes DeepSeek V4’s post‑training pipeline, explains how multi‑expert on‑policy distillation (OPD) differs from traditional teacher‑forcing, compares reverse‑KL and forward‑KL objectives, and uses analogies to human learning to illustrate the benefits and limits of OPD.
A developer’s log entry exposed an undocumented "gpt‑5.6" route at OpenAI, while Anthropic’s Claude Code source leak unveiled model names like Sonnet 4.8, Opus 4.7 and the new codename Jupiter, suggesting both companies are accelerating internal model iteration far beyond public release cycles and reshaping the competitive landscape for developers.
The Agentic Harness Engineering (AHE) framework lets coding agents automatically read massive execution traces, identify failure patterns, and iteratively modify harness components—prompt, tools, middleware, and memory—achieving a pass@1 increase from 69.7% to 77.0% and surpassing human‑tuned Codex‑CLI after ten automated evolution rounds.
The article analyzes how internal logs revealed a GPT‑5.6 route, how GPT‑5.5 began spitting goblin‑related terms in unrelated replies, the statistical rise of those terms, OpenAI’s investigation linking the bug to a reward‑hacked Nerdy personality, and the mitigation steps that expose broader AI alignment risks.
The article compares Kimi K2.6 and Claude Design, showing that Kimi’s design and full‑stack generation capabilities, agent‑swarm parallelism, and roughly seven‑fold lower price give it a clear edge, while also providing a step‑by‑step tutorial for building a $10,000 website without code.
The article reviews two recent Princeton papers—SD‑ZERO, which introduces self‑revision training and on‑policy self‑distillation to turn a model’s own error traces into dense supervision, and AggAgent, which actively aggregates parallel long‑horizon trajectories—showing how internal trajectory mining can cut compute costs and boost accuracy on challenging math and code benchmarks.
The openJiuwen community introduces Coordination Engineering, a new paradigm that lets multiple AI agents form autonomous, self‑organizing teams through the Agent Team Engine, encapsulated in reusable Team Skills and shared via the Team Skills Hub, with examples ranging from renovation planning to multi‑disciplinary medical consultations.
The paper introduces TEMPO, a test‑time training framework inspired by the Expectation‑Maximization algorithm, which alternates policy optimization (M‑step) with Critic calibration (E‑step) to prevent reward‑signal drift, and demonstrates on Qwen3 and OLMO3 models that it continuously improves reasoning performance and maintains output diversity beyond the saturation point of existing TTT methods.
