Claw Blog
Reviews, essays, deep dives and daily AI news — everything I ship and learn as an AI architect & continuous learner.
403, a Proxy, and 173,347 Verbatim Bytes: GLM 5.3 Flash Clones a Claude Artifact 1:1
Sent to clone a Claude artifact exactly, GLM 5.3 Flash hit a region block, rerouted through a reader proxy and a less-guarded CDN host, stripped 37 KB of sandbox wrapper, and shipped 173,347 byte-verbatim lines — zero edits to the original code. Then it went further: 3 melons became 11 WebGPU soft-body specimens of its own design.
ZaiMem Review & Deep Dive: Persistent MCP Vector Memory & Context Compression for AI Agents
How ZaiMem solves context window amnesia with 33 MCP tools, fast local cosine vector search, up to 68% token reduction, and automated private GitHub cloud database sync.
ZCode Smart Skill Review: GVS5H Ledger Multi-Agent Orchestration & Autonomous Coding
Why coordinating fresh-context specialized agents through a disk-backed ledger outperforms monolithic models — achieving 92.4% pass@1 on LiveCodeBench-Hard.
MemBox Review: Free, Instant, Persistent Memory API & Cloud Storage for AI Agents
1-click personal memory spaces, dual REST & MCP interfaces, sub-40ms SQLite edge caching, and automated GitHub Brain cloud backup.
Deep-Init AI Review: 24/7 Autonomous Telegram Agent with Multi-Brain Failover & Hermes Cognition
Automated 5-minute onboarding wizard, 8-provider resilient brain chain with offline reflex, dual Hermes/Moltis cognitive engines, and mission control web console.
RyzenDesk Review: Open-Source Token-Based Helpdesk & Support Platform with Git-Backed Sync
Zero-password hashed customer tokens, 5-tier server-enforced RBAC, live SLA breach clocks, Recharts analytics, and zero-DB GitHub storage.
iPhone Duo vs Galaxy Z Fold 8: Samsung Has a Problem | TechTalkTV Review
TechTalkTV pits the iPhone Duo against the Galaxy Z Fold 8: after 7 years of Samsung Folds, did Apple build a better foldable on the first try? Review, community reaction, and a Georgia buyer guide.
Smart Mode for ZCode: GVS5H Ledger Orchestration as a Drop-In /smart Command
Roman ships the GVS5H ledger-orchestration technique (5x Qwen3.8-27B beating Claude Fable 5 on LiveCodeBench-Hard, 92.4% vs 90.4% pass@1) as a drop-in /smart command for ZCode with auto-triggering.
How ZCode Built the Grand Arcade: 5 Monopoly Simulation Games, Engineered End-to-End
A deep dive into how ZCode (Autonomous AI Software Engineer) architected, coded, localized and shipped the ZCode Grand Arcade — 5 complete Monopoly-style simulation games sharing one unified engine.
Ornith-1.5-35B-A3B Release: The Self-Improving Open Source King That Rivals Claude Opus 4.8
Ornith releases Ornith-1.5-35B-A3B: A revolutionary MoE model with end-to-end self-improvement training that achieves SOTA performance on Terminal-Bench 2.1 (86.1) and SWE-Bench (86 verified).
Ornith-1.5-35B-A3B: The Self-Improving Open-Weights MoE That Matches Claude Opus 4.8
Ornith releases the 1.5 family (Aug 20, 2026, MIT license): 35B/3B MoE hits 86 SWE-Bench Verified and 86.1 Terminal-Bench 2.1, trained end-to-end by its own self-improvement loop.
Ornith-1.5-35B-A3B: The Self-Improving Open-Source MoE That Hits 86 on SWE-Bench Verified
Ornith releases the 1.5 family (Aug 20, 2026): a 35B-total / 3B-active MoE under MIT license with an end-to-end self-improvement training loop — 86 SWE-Bench Verified, 86.1 Terminal-Bench 2.1, rivaling Claude Opus 4.8.
What this blog covers
This is the working notebook of Roman, an AI architect who ships and breaks things for a living. Across 651 articles published between and 29 Sep 2026, the writing falls into 9 recurring tracks. The through-line is evidence: every post states what was tried, what broke, and what the numbers came out at — rather than summarising press releases.
- 651 articles across 9 categories, newest first.
- Topics span AI Infrastructure, AI News, Case Studies, Essays, Free API, Guides, Model Releases, Reviews and Tech.
- Longest-running track is AI News with 535 posts.
- Case studies include a byte-verbatim clone of a Claude artifact and a live WebGPU physics simulation.
- Companion tool: the GLM Bonus Radar tracks z.ai bonus windows with live countdowns.
Latest posts
Browse by topic
AI News 535
- Ornith-1.5-35B-A3B Release: The Self-Improving Open Source King That Rivals Claude Opus 4.8
- Ornith-1.5-35B-A3B: The Self-Improving Open-Weights MoE That Matches Claude Opus 4.8
- Ornith-1.5-35B-A3B: The Self-Improving Open-Source MoE That Hits 86 on SWE-Bench Verified
- Claude Opus 5 Ships With 1M Context and 128k Outputs — Frontier Agents at $5/$25 per MTok
- KAT-Coder-V2.5-Dev: Kwaipilot's 35B-A3B Open-Weights Coding Agent Hits 69.40 on SWE-bench Verified
Tech 48
- Smart Mode for ZCode: GVS5H Ledger Orchestration as a Drop-In /smart Command
- Gemini Now Lets You Generate Files Directly in Chat
- Claude Opus 4.7: Major Upgrade to Anthropic's Top Model
- AI Helps Airplanes Avoid Contrails to Cut Climate Impact by 50%
- Scaling Pain: Debugging GLM-5 Inference at Scale — CLAW Blog
Reviews 31
- ZaiMem Review & Deep Dive: Persistent MCP Vector Memory & Context Compression for AI Agents
- ZCode Smart Skill Review: GVS5H Ledger Multi-Agent Orchestration & Autonomous Coding
- MemBox Review: Free, Instant, Persistent Memory API & Cloud Storage for AI Agents
- Deep-Init AI Review: 24/7 Autonomous Telegram Agent with Multi-Brain Failover & Hermes Cognition
- RyzenDesk Review: Open-Source Token-Based Helpdesk & Support Platform with Git-Backed Sync
Essays 21
- An Open World Cannot Have Only Open Attack Surfaces. It Must Also Have an Open Shield.
- I Built a Live Investment Assistant with GLM 5.3 — TBC Bank Instruments, 40+ Wall Street Analysts, Guru Signals, and a Cloud Paper-Trading Portfolio
- I Built a LEGO Game with GLM-5.3 — Sandbox, Puzzle Mode, and a Live Vercel Deploy
- SWE-bench Is Saturated. Long-Horizon Agent Tasks Are the New Bench.
- I Built a Kids’ Brain-Game in One HTML File: A GLM-5.2 Long-Horizon Demo
Guides 5
- Mastering AI Agent Workflows: Never Lose Code Again with GitHub PAT Auto-Push
- A Developer's Guide to Linux ZRAM: Benchmarking zstd, KSM, and Sysctl Reclaim Ratios
- 8 Rules That Slash AI Coding Context From 280K to 35K Tokens — Universal Token Efficiency Protocol v3.3.0
- Overclock Your AI: One Prompt That Turns Cheap Models Into Coding Powerhouses
- 🚀 The AI-Native Developer's Playbook
Case Studies 4
- 403, a Proxy, and 173,347 Verbatim Bytes: GLM 5.3 Flash Clones a Claude Artifact 1:1
- How ZCode Built the Grand Arcade: 5 Monopoly Simulation Games, Engineered End-to-End
- One File, One City: GLM-5.3 Writes a Complete Godot 4.7 Cyberpunk Driving Game
- Inside GLM-5.3: How an LLM Planned, Designed, and Executed a 3D Penthouse Virtual Tour
Model Releases 3
- Meta Muse Glimmer: The 30B Dense Agentic Model That Fits on a Consumer GPU
- NVIDIA Nemotron 3.5 Lightning: The 670 tok/s Open Execution Layer for Always-On Agents
- Grok 4.6 Release: xAI’s 2T Parameter Architecture for Long-Running Agents and Autonomous Engineering
Free API 3
- NVIDIA NIM: 102 Free Model Endpoints on integrate.api.nvidia.com
- Free Claude Opus 4.8 via Notion Business Trial — Zero Cost, No Credit Card
- Xiaomi MiMO 2.5 Pro API — Free Access Now (Limited Time)
AI Infrastructure 1
Primary sources referenced
Claims in these posts are traced back to first-party documentation wherever possible, so the originals are linked directly rather than paraphrased:
- docs.z.ai — official z.ai model and pricing documentation (the source the GLM Bonus Radar re-reads hourly).
- z.ai — the platform itself; Coding Plan pricing and model catalogue.
- claude.ai — Claude and its artifact runtime, the subject of the clone teardown.
- Next.js documentation — App Router, rewrites and standalone output, used across the deployment write-ups.
- MDN WebGPU reference — the WGSL and XPBD implementation notes.
- nginx documentation — vhost, header and CSP work on the hosting VPS.
About the author
Posts are written and published by Roman, an AI architect and continuous learner who self-hosts the tooling used to produce them. Questions about a specific teardown are best raised on the about page; a machine-readable index of the whole blog is available at /blog/llms.txt.