Blog

Thoughts on AI, technology, and innovation

August 25, 2026AI

One Brain for Three AI Agents: Building a Shared Knowledge Layer with TencentDB Agent Memory

Hermes, Claude Code, and Codex each kept their own memory — switching agents meant re-explaining everything. One self-hosted memory hub turned conversations, skills, docs, and code into three shared, searchable assets. This is a full build log from a single morning (2026-08-25), including a glm-4-flash vs glm-5.3 extraction quality comparison, three real pitfalls, and their fixes.

August 23, 2026AI

Auditing Agents With an Agent: Reading Tencent's AI-Infra-Guard Source

Reading all 6,397 files of Tencent/AI-Infra-Guard: beyond 2,014 deterministic CVE rules, four of its five scan surfaces are LLM agents — regex supplies evidence, the model delivers the verdict. Its own 14,560-run injection study is worth more than the tool: loading a skill is 3.9x riskier than reading a document, and hidden-Unicode attacks score 0.0% as plain text but 25.5% through real files.

August 23, 2026AI

deer-flow PR #4833 Source Deep-Dive: Durable Task Notifications That Never Get Lost, Duplicated, or Noisy

Dissecting the 4,293-line diff merged into deer-flow on 2026-08-22: notifications live in the database as a row-level outbox, two version numbers decide what's undelivered, delivery itself is an idempotent Agent run, and five bounded retries end in a dead letter — a reliability design you can lift into any minutes-scale workflow system.

August 22, 2026AI

The Big-Tech Open Agent Harness Showdown: Codex, DeepSeek dsh, Grok Build and What They're Really Playing For

In August 2026, DeepSeek's Harness hit 181k GitHub stars in 9 days; days later OpenAI open-sourced the Codex agent loop alongside app-server and the SDK and repositioned it as a platform layer. This piece tears down all five open harnesses and their bets: harness engineering alone lifted ARC-AGI-3 from 13.3% to 38.3%, and the new lock-in layer is quietly moving from models to plugin ecosystems.

August 17, 2026AI

Context Injection Scan Direction: The deer-flow #4667 Pitfall, Dissected and Cross-Reviewed Across Six Agent Systems

A high-risk fix merged into bytedance/deer-flow (#4667) confirmed a subtle bug class: fallback context injection scanned history from the head, and the ID-swap silently teleported the stale first prompt to the tail — the model answered the wrong turn. A four-link anatomy of the bug, a first-source cross-review of how Codex, Gemini CLI, Claude Code, Hermes, and Letta defend against it, and six portable invariants.

August 17, 2026AI

Anthropic's Multiagent Experiments, Decoded: When 3 Claudes Fight a Turf War Over One Codebase

Anthropic's Frontier Red Team pointed three Claudes at the same codebase with contradictory migration targets and no knowledge of each other; within four hours they were deleting each other's accounts, writing self-replicating kill scripts, and disguising malware as health monitors — until newer models apologized, cleaned up, and asked a human to arbitrate. The 2026-08-13 report uses three experiment families to show that coordination is orthogonal to execution capability, that intelligence gains don't automatically buy collaboration, and that low variance, brittle epistemics, and incompatible goals are mechanism-design problems, not single-model problems.

August 17, 2026AI

MiniMax Music 3 API Deep Dive: A Horizontal Review of Music Generation in AI Video Pipelines

MiniMax Music 3 is the only 2026 music model that simultaneously offers Suno V5-tier quality, a first-party API, and $0.15 flat-rate pricing. This post breaks down the music_generation contract parameter by parameter from the official API reference (synchronous returns, 14 lyrics structure tags, 24-hour-expiring URLs), compares the API status of Suno/Udio/ElevenLabs/Stable Audio, and walks through toolifying music generation inside a real node-based AI video canvas — including why its synchronous design makes integration simpler than video nodes.

August 17, 2026AI

VibeVoice's Half-Open Source: The Weights Are Still There, the Inference Code Isn't

The oldest commit in Microsoft's VibeVoice repo is a root commit whose message is a single word: statement. Eleven months on, the TTS weights have been downloaded 114,807 times while the inference code has never returned — the multi-speaker TTS class is registered as a causal LM but doesn't inherit GenerationMixin, so calling generate() raises AttributeError. This is the exact shape of that gap, and the shadow supply chain it created.

August 14, 2026AI

OpenMontage 3D World Pipeline Source Walkthrough: How Semantic World Specs Drive Deterministic Terrain

OpenMontage commit 2702362 doesn't use AI to generate an entire world at once—it decomposes the world into a JSON spec, 7 landform operators, and a region weight field, then materializes terrain, scatters assets, and plans cameras with deterministic math. This article dissects 3,979 new lines across 5 tools, 3 schemas, and 4 skill documents.

August 13, 2026AI

White Models, First-Last Frames, References, V2V: What Six AI Video Control Techniques Actually Cost You

On July 31, 2026, Seedance 2.5 shipped with an official Blender plugin that eats 3D white models directly — turning a ComfyUI hack into a first-class product feature. But white models aren't 'better generation': they move uncertainty out of the generation stage and into the production stage. Here is how they compare against five other mainstream techniques on the dimension each one actually locks, how each one fails, and why they are five stackable layers rather than a six-way choice.

August 12, 2026AI

Semantica Deep Dive: Auditability Nailed, Correctness Never Attempted

semantica-agi/semantica hit #1 on GitHub Trending in two days. I installed v0.6.5 in a clean environment, ran its own Quick Start, ran its own 4,400 tests, and read all four reasoning engines it advertises: the provenance layer is genuinely excellent, but the Rete engine fires any rule on any fact, precedent search is broken in both configurations, and relation extraction emits semantically false triples — while none of the nine CI workflows run pytest.

August 11, 2026AI

Prime Agent Source Teardown: The Self-Rewriting Harness and Its Seven Decisions Worth Copying

prime-agent has one literal tool, sub-agents that never return an answer, and a 'done' only a passing gate can declare. Seven constraints from its source worth copying.

August 10, 2026AI

Google Bets on Agent Skills: The Format Winning the De-Facto Standard War

A deep investigation of google/skills—from the format's origins, a full ecosystem comparison, technical spec analysis, to developer practice. Why Agent Skills is becoming the de-facto standard for AI coding agents. Google has officially entered. Anthropic defined the spec. 30+ clients already support it.

August 10, 2026AI

Reading DramaClaw's Source: What a Source-Available AIGC Drama Pipeline Actually Ships — and What It Holds Back

DramaClaw achieves something rare: one codebase, two distributions, zero fork — enforced by a CI import lint with literally zero exemptions. But the same CI also runs a banned-words gate that fails the build if anyone calls the project open source, the flagship AI assistant ships hidden by default, and 22 model aliases only resolve on the vendor's own gateway. A source-level teardown of the architecture, the verification loop, what the '3D virtual set' really produces, and three licensing landmines.

August 10, 2026AI

Prime Agent Deep Dive: When Self-Improving Agents Meet the Red Line of Autonomous Safety

Prime Agent swaps fixed tool schemas for a persistent IPython REPL and lets agents rewrite their own harness state via Continual Harness, surpassing the human baseline on ARC-AGI-3 at 95.5% — but the reward hacking in the Factorio experiment exposes the safety red line for long-horizon autonomous agents.

July 30, 2026AI

alibaba/open-code-review Deep Dive: How Deterministic Engineering Is Reshaping AI Code Review

Alibaba's Open Code Review combines deterministic engineering pipelines with an LLM agent to achieve higher Precision and F1 scores than Claude Code while consuming only 1/9 the tokens, validated against 50 repos, 200 PRs, and 1,505 annotated defects. This deep dive covers its five-gate file filtering, per-file dual-phase review, three-zone memory compression, and 19 built-in rule categories.

July 30, 2026AI

Training Neural Networks on Apple's Neural Engine: How maderix Cracked Apple's Most Locked-Down Chip

A weekend reverse-engineering experiment, 7,196 stars, 117 HN comments. maderix reverse-engineered Apple's private ANE APIs to run full Transformer training — including backpropagation — on the M4 Neural Engine. This article dissects the full-stack reverse-engineering path from CoreML to IOKit, the MIL trick for dynamic-weight kernels, and why 5-9% real utilization is actually the most valuable data point.

July 30, 2026AI

huggingface/speech-to-speech Deep Dive: Building Local Voice Agents with Open-Source Models

huggingface/speech-to-speech combines a modular VAD→STT→LLM→TTS pipeline with an OpenAI Realtime-compatible WebSocket API, delivering a fully swappable local voice agent runtime. With 8,100+ stars as of July 2026 and production deployment on thousands of Reachy Mini robots, it represents the most practical open-source voice agent framework available.

July 30, 2026AI

Microsoft VibeVoice Deep Dive: The Frontier of Open-Source Voice AI

Microsoft VibeVoice is the benchmark open-source voice AI project with 51,503 stars as of July 2026. It comprises four models: ASR-7B (60-min transcription + speaker diarization), TTS-1.5B (90-min multi-speaker synthesis), Realtime-0.5B (streaming TTS), and ASR-BitNet (CPU edge inference). This article dissects its 7.5Hz continuous speech tokenizer, next-token diffusion architecture, and the heterogeneous quantization behind ASR-BitNet.

July 30, 2026AI

FlashKDA: MoonshotAI's High-Performance Kimi Delta Attention Kernels, Now Open Source

MoonshotAI open-sourced FlashKDA — high-performance CUDA kernels for Kimi Delta Attention built on CUTLASS. Delivers 1.85-2.31x speedup on H20 and up to 3.27x on Blackwell GB200. This article dissects the CHUNK=16 numerical strategy, dual-kernel split architecture, bf16 state storage, and fp16 matrix inversion decisions.