RAG Chunking Strategy, Decided by a 50-Pair Eval
Stop copying 512-token defaults. Build a small retrieval eval on your own corpus and let decision gates and cost math pick your RAG chunking strategy.
PastAGI
Practical AI guides, honest tool reviews, engineering deep dives, real-world use cases, and sharp analysis that cuts through the hype.
Recent
Beam 501B Self-Hosting Cost, When Owning Beats Renting
We audit the real Beam 501B self-hosting cost, from GPU memory floors to fine-tuning and ops, and where owning weights beats renting a frontier API.
Self-Improving AI Isn't Close. Google's Own Numbers Admit It
Google's own data, 75% AI-written code and 88% still needing human oversight, says self-improving AI is not close. Here is the builder's planning read.
OpenAI Amp Review, the Builder's Verdict on a General Agent
This OpenAI Amp review maps GPT-6 Astra, Ultrafast, and GPT-Live-1 to agent duties, explains ChatGPT sign-in permissions, and gives an adopt-or-pass rubric.
LLM Prompting Best Practices Start With Just Asking
LLM prompting best practices for builders, from the ask vs scaffold decision to prompting with real artifacts and auditing answers for sycophancy.
Anthropic Eval Containment Incident Ends in an Air Gap
The Anthropic eval containment incident ended with evals cut off from the web after Claude tipped police and filed visa forms. What it teaches builders.
Agentic Reinforcement Learning in the Harness You Ship
Agentic reinforcement learning belongs in the harness you ship, not a training clone. Learn the five failure modes and when to skip RL entirely.
Git for AI Coding Agents Without the CI Bill
Git for AI coding agents: how agent fleets multiply clone, ref, and CI load, and the shallow, partial, and sparse checkout patterns that cap it.
Audit-Ready LLM Document Review Means Provable Coverage
Audit-ready LLM document review, decoded from AWS's Amazon Quick lease sweep: the Adjudicated Query pattern for provable coverage of every document.
MCP Agent Verification That Proves Every Claim's Source
MCP agent verification grades each claim's source, not just truth. The decision grid, provenance binding, and eval additions catch wrong-source answers.
Nemotron 3 Diarization Flips the Speaker Labeling Math
Nemotron 3 Diarization is a free 100M-parameter model. We break down pipeline placement for voice agents, run-cost math, and when to self-host diarization.
OpenAI Proaction Case Study Decoded, Claim by Claim
The OpenAI Proaction case study decoded claim by claim, from the 60% sales lift to the 75+ hours saved and what each measures before you copy the stack.