RESEARCH · BUILD · EXPLAIN
AI Frontier Research & Engineering
Tracking emerging technology, dissecting system architecture, and publishing what I actually build and evaluate.
- Public notes
- 89
- Reading paths
- 3
- Technical topics
- 46
FULL ARCHIVE
Find a specific article
Search first, then narrow by content type or technical topic.
- AI Software Development Environments: Choosing Between Vibe Coding and Verified Agent Workflows
InfoWorld surveys GitHub Copilot, Google Antigravity, JetBrains Air, Kiro, Zed, and Zenflow; this article turns that tour into a selection framework based on autonomy, context, isolation, and verification.
- GitHub Copilot MCP Governance: From Allowlist Semantics to Agent Adoption Telemetry
GitHub has added enterprise MCP allow and deny lists alongside third-party agent activity in usage metrics; this article explains the policy semantics, telemetry boundary, and a practical rollout loop.
- Claude Managed Agents Become a Governed Runtime: Budgets, Delegation, Locality, and Inference Hooks
Anthropic's Managed Agents now expose session budgets, advisors, inference geography, repository skills, and inference hooks as runtime controls; this article maps their value and remaining boundaries.
- TREC RAG 2026: When RAG Evaluation Becomes an Agent Workflow
A concise entry point to TREC RAG 2026, ClimbMix-400b, and RAGDoll, with a bridge to a technical guide for building a replayable enterprise RAG evaluation harness.
- TREC RAG 2026 Technical Deep Dive: From Evidence Lineage to a Replayable RAG Evaluation Harness
A practical design for an enterprise RAG evaluation harness based on TREC RAG 2026 and RAGDoll: data models, execution stages, citation support, agent traces, judge calibration, and production gates.
- Cloudflare's Open Agentic Internet Blueprint: Readable, Discoverable, Callable, and Payable Web
An architectural deep dive into Cloudflare's proposed Agentic Internet framework, spanning Web Bot Auth, Markdown for Agents, WebMCP browser tool exposure, and x402 micro-payments.
- Stanford CS Professor Chris Piech: AI Can Code, So Why Should You Still Learn Programming?
Stanford CS Professor Chris Piech breaks down software engineering education in the AI era: while LLMs handle syntax, problem-solving, architectural thinking, and human empathy remain irreplaceable multipliers.
- Andrej Karpathy Just Fixed Claude Code's Biggest Weakness: Letting AI "Dream"
Anthropic introduces 'Dreaming' for Claude Code, solving the context amnesia and split-focus issues of in-band memory by allowing agents to consolidate and refine long-term memory overnight.
- Anthropic's Memory and Dreaming for Continuous Agent Learning
Exploring Anthropic's underlying memory systems and the asynchronous Dreaming process for Claude agents to solve context management and continuous learning in multi-agent environments.
- Google Cloud Releases OKF v0.2: A Complete Breakdown of the Open Knowledge Format Upgrade for AI Agents
An in-depth comparative analysis of Google Cloud's Open Knowledge Format (OKF) v0.2 vs v0.1. Explore Attested Computation, provenance credibility signals, structured verified mappings, and staleness lifecycle controls designed to eliminate hallucinations in enterprise AI agents.
- Kimi-K3 Enterprise On-Premises Deployment TCO: GPU Topologies, Power Demands, and Infrastructure Realities for a 2.8T MoE
A comprehensive breakdown of GPU memory topologies, server hardware budgets, 3-phase power, liquid cooling requirements, and a 5-stage TCO evaluation framework for deploying Moonshot's Kimi-K3 2.8T MoE on-premises.
- Running Kimi-K3 on 80 RTX 5090 GPUs: Hardware Ledger and Engineering Trade-Offs of Consumer GPU Clustering
A deep dive into deploying Moonshot Kimi-K3 2.8T MoE on 80 consumer RTX 5090 GPUs: why 44 GPUs fall short, node topologies, 25GbE networking, TCO ledger, and power constraints.
- Analyzing OpenAI GPT-5.6: Frontier Intelligence and the Architectural Shift to 'Intelligence per Token'
A comprehensive deep dive into OpenAI's official GPT-5.6 release: Sol/Terra/Luna tiering, pricing matrices, Terminal-Bench 2.1 data, Triton kernels, and Agentic Harness routing.
- Towards Robust Reinforcement Learning for Small-Scale Language Model Agents: Architecture and Practical Insights
An in-depth analysis of three common PPO failure modes (gradient freezing, numerical overflow, policy collapse) for 70M-500M SLMs, and the Capacity-Headroom Hypothesis (PPL < 20).
- 2025/2026 Modern Large Language Model Architecture Deep Dive: From DeepSeek V3 to Llama 4
Exploring the key architectural innovations behind DeepSeek V3, Llama 4, Gemma 3, and Kimi K2, including MLA, MoE, and sliding window attention engineering practices.
- AgentEscapeBench Deep Dive: Benchmarking Out-of-Domain Tool-Grounded Reasoning in LLM Agents
A technical analysis of arXiv:2605.07926 (AgentEscapeBench). Exploring how LLM agent performance degrades across long-range DAG tool dependency graphs, exposing bottlenecks in clue adherence, intermediate output propagation, and state tracking.
- The New Rules of Context Engineering for Claude 5 Models: Trimming 80% of System Prompts
Anthropic removed over 80% of Claude Code's system prompt for Claude Opus 5 and Claude Fable 5 without losing benchmark performance. Explore the paradigm shift from rigid instruction constraints to progressive disclosure, auto-memory, and rich reference harness engineering.
- OpenAI GPT-5.6 Prompting Guidance: From Lean Prompts to Programmatic Tool Orchestration
Unpacking OpenAI's official GPT-5.6 model guidance. Discover how leaner system prompts boost evaluation scores by 15% and cut costs by 67%, alongside autonomy boundaries, text.verbosity, Programmatic Tool Calling, and Pro mode.
For speaking invitations, internal engineering sessions, or architecture exchange, see the topics and public work I can bring into the conversation.
Speaking & contact