Writing

imast.ai Now Records Real Work for Robots: Egocentric Video of Hotels, Kitchens and Short-Lets (Oct 2026)

Imast AI records consented, de-identified first-person video of real work in hotels, short-let apartments and kitchens, as training data for robots.

tigoLab: AI Assistants That Take the Busywork Off Small Teams (Oct 2026)

tigoLab builds AI assistants, AI-ready websites, RAG bots, MCP servers and Microsoft 365 automation for small teams. A person approves every important step.

I Profiled 12,125 Polymarket Weather Traders. Copying the Winners Loses Money (Oct 2026)

I pulled 4.1M fills from 12,125 wallets in Polymarket daily temperature markets over 21 days. A forecast rule lost 11.4% per $ after fees. Copying the top wallets lost 1% to 5%.

Porting Manim to TypeScript with Coding Agents (Oct 2026)

How manim-web, a TypeScript port of Python Manim, was built with coding agents: 101 of the first 169 commits came from an agent. What mapped to WebGL, what was hard, and where a human was needed.

Claude Code Controls My iPhone: A Skill Built on iPhone Mirroring (Oct 2026)

A Claude Code skill that drives an iPhone through macOS iPhone Mirroring: window capture by id, a screenshot and tap loop, retina mapping, cliclick, and the limits.

imast.ai Post-Mortem: 427 Commits in 34 Days, Zero Clients (Oct 2026)

A builder post-mortem of imast.ai, an AI HR agent: 427 commits in 34 days, auth, billing and Cloud Run in production, zero clients. What I built, where the time went, and what I keep.

MCP Security Incidents, August to September 2026: Deadbugz and the SDK OAuth Flaw (Oct 2026)

Deadbugz and the MCP Python and TypeScript SDK OAuth flaw, mapped to three protocol-level MCP flaws and the NSA MCP guidance. Where AttestMCP-style attestation helps, where it does not, and a defense checklist tied to each incident.

The Best Fully-Local Voice Stack for Home Assistant (2026) (Jun 2026)

An opinionated 2026 build guide for a fully-local Home Assistant voice assistant — faster-whisper, Piper or Kokoro, and Ollama-hosted Qwen3 over Wyoming. Which LLM, STT, and TTS to pick by hardware tier, plus the prefer_local_intents setting almost every guide skips.

Sleeper Channels and Provenance Gates: Persistent Prompt Injection in Always-on Autonomous AI Agents (May 2026)

Always-on agents fold messaging, memory, skills, scheduling, and shell into one authority boundary. Untrusted input persists as memory, skill, scheduled job, or filesystem patch, then fires later via a different surface. End-to-end OpenClaw confused-deputy cron walkthrough plus a D2 provenance gate with a soundness theorem.

Launching imast.ai: an AI HR agent for sourcing and outreach (May 2026)

imast.ai is live: chat to find candidates across 800M+ profiles, get AI-ranked shortlists, and run personalized outreach on autopilot. The product later shut down with zero clients.

BeeLlama.cpp on a Mac Mini M4 16GB: DFlash and TurboQuant for Local LLMs (May 2026)

Anbeeld's BeeLlama.cpp ships DFlash speculative decoding, TurboQuant/TCQ KV-cache compression, and adaptive draft control. Honest take on which parts pay off on 16GB Apple Silicon.

DFlash and DDTree: 8x Faster LLM Inference via Block Diffusion and Draft Trees (Apr 2026)

Two 2026 papers that compound. DFlash swaps the autoregressive drafter in speculative decoding for a block-diffusion model; DDTree reuses DFlash's per-position distributions as a verified draft tree. Lossless. Up to 8.22x on Qwen3-Coder-30B.

Running Qwen 3.6 Locally on a Mac Mini M4 with 16GB RAM (Apr 2026)

A practical guide to running Qwen 3.6-35B-A3B locally on a Mac Mini M4 16GB — the best open-weight model you can run on a $599 machine, with setup instructions and the published benchmarks for this machine.

Prompt Injection Attacks on Agentic Coding Assistants (Jan 2026)

A review of 78 studies on prompt injection in coding agents like Claude Code, Copilot and Cursor: taxonomy, real exploit chains, and why defenses fail.

Breaking the Protocol: MCP Prompt Injection and Tool Poisoning (Oct 2025)

MCP security paper: three protocol-level flaws, attack success 52.8% through MCP vs 26.4% without, AttestMCP cuts it to 12.4%. Plus how tool poisoning works.

Prompt-Injection Attacks on LLM-as-a-Judge Architectures (May 2025)

GCG-optimized suffixes flip the verdict of pairwise LLM judges in over 30% of cases (Qwen2.5-3B, Falcon3-3B, MT-Bench), vs 1-1.5% for random text.

Adversarial Attacks on LLM-as-a-Judge Systems: Insights from Prompt Injections (Apr 2025)

Prompt injection against five LLM judges on four tasks: up to 73.8% attack success, 50-63% transfer, and mixed-model committees as the best defense.

A Neurosurgical Instrument Segmentation Approach to Assess Microsurgical Movements (Nov 2024)

EFMI STC 2024 paper: EndoViT segmentation plus mask tracking for microsurgical instruments. Mean IoU 0.8158 on CholecSeg8k and 0.7196 zero-shot on neurosurgical microscope video.

Prompt Injection Attacks in Defended Systems (Jun 2024)

Attacking and defending LLMs in the SaTML 2024 CTF: how secrets leak past system prompts, Python filters and LLM filters through code and encoding attacks.

Trojan Detection in Large Language Models: Insights from TDC 2023 (Apr 2024)

TDC 2023 lessons: forcing a backdoor's output was easy (REASR 0.987), but recovering the real trigger was not (recall about 0.16, near a random baseline).

DN at SemEval-2023 Task 12: Low-Resource Language Text Classification (2023)

DN at SemEval-2023 Task 12 (AfriSenti): AfroXLMR-large fine-tuned with a 5-fold ensemble for African-language tweet sentiment. 3rd place on the multilingual track, weighted F1 72.55.

Blind Face Restoration Survey (2023)

Survey (in Russian, IJOIT 2023) of deep learning methods for blind face restoration: GAN, codebook and diffusion priors, quality metrics and datasets.