AI articles
Practical notes on building AI features that earn their place - evals, agents, on-device models, and shipping AI you can stand behind.
Your app depends on a model with a retirement date
Every model you build on has a published expiry. Eight Claude model IDs stopped serving requests in the last year - here's how to make the swap routine.
What the EU AI Act actually requires on 2 August 2026
On 2 August the AI Act's transparency rules and its fines both switch on. The high-risk rules didn't - they slipped to 2027. Here's what really applies.
The unit economics of an AI feature
Token prices fell ~50x a year, yet AI bills keep climbing. The thing that costs money isn't a token - it's a task, and agentic tasks use 1000x more of them.
Fine-tune, RAG, or prompt? Reach for them in this order
Prompting changes instructions, RAG changes knowledge, fine-tuning changes behavior. Prompt first, add RAG for your data, fine-tune only when both fall short.
Getting reliable JSON out of an LLM (stop parsing hope)
Prompting for JSON fails 5-10% of the time; schema-enforced structured output is ~100% valid via constrained decoding. Use it, and still validate in code.
AI-native vs AI-as-a-feature: which are you building?
Remove the AI - does your product still work? If yes it's a feature; if the business collapses it's AI-native. That gap decides defensibility, not the model.
Prompt injection is the new SQL injection - but unsolved
Prompt injection is OWASP's #1 LLM risk. It shares SQL injection's root cause - data becoming commands - but has no parameterized-query fix. So you contain it.
You can't ship AI you can't evaluate
AI features are non-deterministic, so 'it looked right in the demo' isn't a test. Evals - a dataset plus graders, run in CI and prod - are how you ship them.
Repository intelligence: your codebase is the context now
AI coding moved from autocompleting lines to understanding whole repos - and running several agents on one task. Here's what changes for how teams build.
The delegation gap: why AI does most of the work but runs little of it
Developers use AI in ~60% of their work but can fully delegate only 0-20% of tasks. That gap - not the hype - is what actually shapes AI-era engineering.
One agent isn't enough: when to use multiple AI agents
Multi-agent systems beat a single model by ~90% on the right tasks - and burn 15x the tokens on the wrong ones. Here's how to tell which is which.
The hidden bill of AI-generated code
AI assistants ship code fast - and technical debt and leaked secrets behind it. Winning teams treat the assistant as an untrusted contributor, not a colleague.
You don't need a frontier model: the case for small language models
Most teams pay frontier prices for work a small model does just as well. The 2026 move is right-sizing - SLM-first, escalate only the hard part.
Context engineering: the real skill behind AI agents that work
The teams whose AI features work aren't using better models - they're feeding them better context. Why context quality, not size, is the 2026 bottleneck.
Coding agents in 2026: what SWE-bench scores really mean
Coding agents went from 2% to ~94% on SWE-bench in three years. Here's what those scores really mean - and where the agents still need a human.
How AI agents remember: memory systems in 2026
How AI agents remember in 2026: the four memory types, the tiered architecture that beats a growing transcript, and when your agent actually needs memory.
Is RAG dead? RAG vs long context in 2026
What RAG is really for in 2026, when a long-context window beats it, and why naive RAG died while agentic retrieval is thriving.
What an AI agent actually is (and what makes one reliable)
A plain-English anatomy of an AI agent: model, tools, memory, context, and the loop, plus applications and why most agents fail on state, not smarts.
Does your business actually need AI? A 2026 reality check
A neutral guide to when AI is worth it for your business in 2026: why most projects fail, where AI genuinely helps, and how to tell the difference.
Agentic AI's real challenge isn't the demo - it's production
Gartner says 40% of enterprise apps will embed AI agents by end of 2026, yet fewer than 1 in 4 teams have scaled one to production. Here's why, and how to close the gap.
On-device AI is quietly winning - and it's a gift for privacy
Over 2 billion phones now run local language models, NPUs are standard, and small models are 10–30× cheaper to run. Why the future of AI is increasingly local.
AI coding agents changed how we build - not whether craft matters
Command-line coding agents are shipping real engineering work 30% faster. What that actually changes for a software studio, and what it doesn't.