Monday 18 May 2026

Debate around the assumption that AI accelerates processes; argues that AI tooling may not meaningfully speed up existing workflows and may introduce new bottlenecks or overhead instead.

AI coding agents (e.g. Claude Code) fall back to grep and full-file reads on large codebases when they can't find relevant code, consuming excessive tokens and still missing results. Existing code search tools are too slow, require API keys, or have poor retrieval quality for agent use.

LangChain's `convert_to_openai_function()` raises a `TypeError` when a `TypedDict` contains `NotRequired` fields, due to incorrect handling of the `NotRequired` wrapper in the schema conversion pipeline.

LangChain tool definitions are model-agnostic, causing small local models (1.5B params) to waste thousands of prompt tokens on tool schemas they can't effectively parse, achieving only ~50% tool selection accuracy across large toolsets. Proposal requests tier-aware tool definitions that adapt descriptions and parameter sets based on model capability.

When binding tools with nested Pydantic v2 model schemas in LangChain, the tool invocation fails to correctly parse the schema, resulting in arbitrarily generated args in `AIMessage.tool_calls` rather than following the defined model structure.

Text rendering and handling is a persistent pain point when building native applications, creating a gap where native tooling is otherwise sufficient but text layout forces compromises or alternative approaches.

Developer reflects on coding on paper as a deliberate practice, implying that digital dev tooling encourages shallow thinking and that the friction of pen-and-paper aids deeper design reasoning.