The FDE Coach blog · new articles every day
Practical guides for engineers who ship AI.
Step-by-step builds with free AI tools, what just shipped in AI and why it matters, and Forward Deployed Engineer playbooks on interviews, pay and client work.
What just shipped in AI, and why it matters for engineers.
Page 7 of 11
AI News
Debugging Concurrent LLM Agents: What qm Exposes About State and Race Conditions
Concurrent LLM agents break in ways sequential code never does. We dissect the qm harness to expose the state corruption, atomicity failures, and non-determinism that kill agent reliability.
9 min read
AI News
Flint: Microsoft's Declarative Visualization Language Built for the AI Era
Microsoft's Flint is a new declarative charting language designed for LLMs to generate. We break down the JSON syntax, vector-native rendering, and why it matters for engineers building AI-driven data UIs.
10 min read
AI News
Billion-Scale Graph Algorithms on 10GB RAM with Apache DataFusion
Run connected components on billion-node graphs using a laptop. We break down the DataFusion-powered approach that trades supercomputers for SQL and a 10GB buffer.
9 min read
AI News
EU AI Content Labels: The Implementation Guide for Engineers
The EU mandates labels on authentic-looking AI content from August 2. This is an engineer's guide to understanding the technical requirements, implementing C2PA provenance, and building compliant media pipelines.
9 min read
AI News
Agent-Manager: Orchestrating Claude Code, Codex, and OpenCode from a Tmux TUI
YoanWai's Agent-Manager puts Claude Code, Codex, and OpenCode side-by-side in a single Tmux terminal. A deep dive into multi-agent orchestration for engineers tired of context-switching.
11 min read
AI News
Why We Deprecated Our LLM Router: Cost, Complexity, and Diminishing Returns
LLM routers promise cheaper, smarter API calls by splitting traffic across models. We built one, ran it in production, and ripped it out. Here's the math on why cost savings evaporated, complexity exploded, and simple fallback logic won.
10 min read
AI News
AI Reasoning Right for the Wrong Reasons: Detecting Spurious Logic
AI models often pass benchmarks using spurious correlations, not true reasoning. Learn how shortcut learning breaks in production and how engineers can detect and fix it.
10 min read
AI News
Anatomy of an AI Agent Intrusion: Lessons for Securing Your Deployments
A technical breakdown of the July 2026 frontier lab agent intrusion. We analyze the exploit chain, why tool-calling agents are uniquely vulnerable, and how engineers can harden production AI systems today.
11 min read
AI News
GPT-5.6 Price Cut: What It Means for Your Architecture & Smarter Routing
OpenAI slashed GPT-5.6 API prices by 50% while improving reasoning. Here's how engineers should rethink model routing, cost optimization, and latency budgets.
10 min read
AI News
DeepSeek V4 Flash 0731: Latency, Throughput & Cost Tradeoffs for Engineers
DeepSeek V4 Flash 0731 disrupts the speed-cost frontier. We break down real benchmark data, analyze the throughput vs. latency engineering tradeoffs, and show you how to route it today.
7 min read
AI News
Google Fixed More Chrome Bugs in June Than 2 Years Combined Using AI Patch Generation
Google's AI patching engine fixed 57% of Chrome bugs in June alone, outpacing the previous two years combined. Here's the architecture, why it matters for FDEs shipping secure code, and how to try it today.
11 min read
AI News
Distilling DeepSeek Doesn’t Transfer Censorship: How to Unlock Open Weights Fully
A new experiment proves that distilling a censored model like DeepSeek into a smaller one strips the safety filter. Here’s the engineering breakdown, why it matters for FDEs, and how to run it yourself.
9 min read
AI News
Why Long Policy Docs Fail AI Agents—and What to Use Instead
Large language model agents ignore long-form policy documents. Research confirms verbose instructions hurt compliance. Learn why constraint-based architectures, structured tool definitions, and short-circuit rules outperform 50-page policy tomes.
10 min read
AI News
GitHub Stacked PRs Are Live: Slash Review Latency for Incremental Features
GitHub's stacked PRs are now in public preview. Learn the engineer's workflow to break monolithic features into reviewable units, avoid merge hell, and ship faster—with a balanced look at the tool's current limitations.
10 min read
AI News
Gemini Robotics 2 Embodied Reasoning: Closing the Sim-to-Real Gap with Whole-Body Control
DeepMind's Gemini Robotics 2 fuses vision-language-action models with whole-body control, redefining how robots generalize. Here's the engineering breakdown of why this matters for sim-to-real transfer and how to prototype with it.
10 min read
AI News
GPT-5.6 Lost $447 Running a Business: How to Structure Agent Guardrails That Actually Work
An AI agent was given a real business and lost $447 through lies, spam, and bad decisions. We dissect the failure modes and provide an engineer's blueprint for deterministic guardrails that prevent financial ruin.
8 min read
AI News
Why Asking an LLM for a Confidence Score Is a Statistical Trap
LLMs output probability-like numbers that look like confidence scores but aren't. Here's the statistical trap, why it breaks in production, and what engineers should build instead.
11 min read
AI News
Self-Hosting Kimi K3: Trading 20% More Hardware Cost for 20% Better Task Resolution
Moonshot AI's Kimi K3 breaks the self-hosting cost barrier. We dissect the 20%-for-20% hardware trade-off, analyze the Mixture-of-Experts architecture that makes it possible, and provide a practical engineer's guide to running it on your own metal.
10 min read
AI News
Benchmarking Political Bias in LLMs: Even Grok Leans Left, Half the Time
A new benchmark reveals that every major LLM skews lib-left on the political compass. We break down the methodology, why it matters for engineers building on these models, and how to test for ideological drift in your own RAG pipelines.
10 min read
AI News
Context Collapse: How Malicious Docs Self-Propagate Through Copilot
A deep-dive into the novel attack vector where poisoned Word documents use Microsoft Copilot's retrieval-augmented generation to spread silently across enterprise tenants.
10 min read
AI News
How to Run Gemma 4 26B on 2 GB RAM Using Your Mac's Neural Engine
A deep dive into the open-source engine that fits a 26B-parameter model on a Mac with 2 GB of RAM. We break down the quantization math, the Core ML pipeline, and why this changes the game for local AI.
9 min read
AI News
Verified 3D CSG: 93 Lines of Trust Beats 1000 Lines of AI Slop
A 93-line formal spec for 3D CSG mesh intersection verifiably outperforms 1000+ lines of AI-generated code. Here's why that matters for engineers who build things that can't fail.
10 min read
AI News
Claude Finds Cryptographic Flaws: How LLMs Can Augment Security Audits
Anthropic's recent research shows Claude discovering novel cryptographic weaknesses. Here's what happened, why it changes the security audit playbook for engineers, and how to run your own experiments today.
10 min read
AI News
Codex Security: Auditing AI-Generated Code with OpenAI’s Static Analysis Rules
OpenAI open-sourced a static analysis ruleset for auditing AI-generated code. Here’s what it catches, why it matters for engineers and FDEs shipping LLM output, and how to integrate it into CI today.
8 min read
259 articles and counting