Tag
Long AI chats drop your corrections and repeat old answers. The model didn't get dumber, it ran out of room. A two minute test and a restart habit that fixes it.
September 29, 2026 AI ToolsAsk a chat tool for a random number between 1 and 20 and 17 keeps winning. It isn't rolling dice, it's copying a human habit. Run the five minute test yourself.
September 25, 2026 AI ToolsTransformers can load llama.cpp GGUF files directly. Outside the narrow Metal fast path it dequantizes at load time, so you get the disk savings without the quantized kernels.
September 23, 2026 AI ToolsPrompt caching charges a premium to write and a discount to read, so it only pays past a break-even reuse count. Here is the formula, the usual bug, and a script that prints a verdict.
September 16, 2026 AI AgentsAnthropic publishes the literal system prompt behind claude.ai. Ignore the transparency angle: it's a production-tested template for structuring your own agent's instructions.
September 03, 2026 AI AgentsThe usual advice for agent-ready data is add context, add metadata, add lineage. Tool-calling reliability improves faster when you delete the overlapping fields instead.
August 31, 2026 AI AgentsOpenAI's Ultrafast tier hits 750 output tokens/sec on Cerebras hardware. Whether that changes your agent's wall-clock time depends on one ratio you can measure in 20 lines of Python.
August 14, 2026 AI AgentsPublic leaderboards score tool calling on clean synthetic schemas, not the nested mess your MCP server exposes. Here's the ~50-line Python harness that settles the debate on your own stack.
June 19, 2026 AI AgentsAn AI agent resolved an RLS error by making the table publicly readable. CI stayed green. Here's the grep triage and checklist I run before any LLM-generated code ships.
May 22, 2026 WorkflowsMost LLM features ship on vibes — skim three outputs, merge. Here's the 40-line Python eval harness with JSON fixtures and a judge-model gate you can wire into CI today.
May 06, 2026 AI AgentsOnce your agent calls tools, prompt injection is a live attack path. Here's the exploit in real Python — and the Pydantic fix that belongs in your dispatch layer.
May 05, 2026