AI Practice
Practical news and analysis about artificial intelligence: new models, AI tools, automation, business adoption, risks, and real-world use cases.
Latest publications
AI financial advice is getting useful. The hard part is asking safely
MIT Sloan research suggests LLMs can give decent personal-finance guidance, but prompt quality, user context and verification decide whether the advice helps or hurts.
Markdown is not a control plane for AI agents
HANDBOOK.md tests whether agents can follow 20–124 page workplace policies. The results point to a practical rule: put critical controls in tools, validators and approvals, not only in prompts.
Open-weight AI is now a business decision, not a slogan
After Anthropic’s clarification and the Kimi K3 backlash, teams need a workload-by-workload framework for closed APIs, managed open models and self-hosting.
Claude Code cut its prompt. Your AI agents may need the same audit
Anthropic’s 80% system-prompt cut is not a call to delete guardrails. It is a sign that agent teams need maintainable context architecture, not endless rules.
The AI abstainers have a point: chatbots can make wrong answers feel safe
A new study is a useful warning for AI practice: the danger is not asking a model, but letting it erase the moment when you should say “I don’t know.”
ChatGPT ads test the trust boundary of AI assistants
OpenAI can label sponsored placements and keep them separate from answers, but conversational advertising still changes how users judge an assistant’s incentives.
AI coding agents need capacity planning, not surprise quota resets
Codex and Claude Code are becoming engineering infrastructure, but context cuts, 5-hour windows and random resets make teams rethink how they budget and schedule AI-assisted work.
Kimi K3 Is a Router Test, Not a Drop-In Claude Replacement
Moonshot’s new 2.8T model gives AI teams leverage, but the real decision is cost per completed task, data policy, routing, and whether the promised weights become usable.
Voice is no longer a password: how AI voice fraud changes trust rules
AI voice cloning turns familiar calls into a verification problem for families, finance teams, banks and voice-AI providers.
Bonsai 27B shows what local AI can do before the cloud call
PrismML’s compressed 27B model is less about replacing the cloud than about moving private, routine AI work closer to the user’s device.
The hidden token tax of AI coding agents
A viral Claude Code versus OpenCode measurement is less about picking a winner and more about making AI development cost visible before it surprises teams.
AI coding agents are useful now. That is why they need security rules
GhostApproval, Friendly Fire and Ghostcommit show that coding agents have become privileged automation, not just smarter autocomplete.
