AI
Language models, prompts and the practical use of AI for people who write and people who automate.
12 published postsPosts in AI
Claude Fable 5.2 vs Gemini 4 Carbon: True Agent Swarm Cost
Pricing, prompt caching, and latency analysis comparing Claude Fable 5.2 and Gemini 4 Carbon to route hybrid AI agent swarms without blowing your budget.
Gemini 4 Carbon: The Shift from Chat Models to Agent Swarms
Technical analysis of Gemini 4 Carbon: agent swarm architecture, engineering benchmarks, 1M token context with adjustable thinking, and real inference costs.
Haiku 5.5: the 90% cut is not the bill you pay
Anthropic cut Haiku 5.5 prices by up to 90%, but the new tokenizer counts about 30% more tokens. Here is how to price a finished task.
Vibe coding: the tools, and how to use them without creating a problem
What vibe coding is, which tools exist for programming by talking to an AI, and the steps that keep a security flaw from shipping.
Free Gemini after October 9: what stays unpaid
Since October 9, a Gemini account without a plan uses Auto mode. The notice and the models page disagree about AI Plus. Check the account selector.
The cloned-voice scam: how to recognize it and protect yourself
With a short audio clip, scammers imitate a relative's voice. The warning signs, a family code word, and what to do if the transfer already went out.
An agent harness: the gates that stopped a real error on this blog
An agent is a model plus a harness. The layer that decides quality is verification: the gates that already blocked a real error here, and the errors no gate catches.
What one multi-agent cycle costs: 6.67 million tokens
I opened the bill of an AI-agent cycle, round by round: where the tokens went, what the cheap gate saved, and why the dollar figure is a range.
Project memory for AI: how not to start from zero every session
AI assistants forget the project when the session ends. A versioned Markdown file gives the context back: what to record, what to leave out, and why.
Why AI agents break outside the lab
Frontier models finish fewer than 25% of real professional tasks. The bottleneck is not intelligence. It is ambiguity, context and a trail, and those can be treated.
A content contract: where an AI pipeline can be wrong without publishing
A Zod schema as the API between the LLM and the repository: a closed taxonomy, SEO limits and a build that fails before the error reaches production. With a real case.
Scaled content abuse: what Google Search actually punishes
Google does not punish the use of AI. It punishes publishing at scale without real value for the reader. The policy, with its effective date and the official source.