What Would a Serious AI Product Look Like?

The piece argues that current “AI” tools are unfit for real work because they gloss over their central weakness - unreliable outputs - while pushing responsibility onto users with tiny disclaimers. It calls for making error-checking a first-class interaction: every AI claim should appear in a two-column worksheet with the model output on one side, human-verification notes on the other, and a mandatory checkbox before claims are accepted. For research workflows it demands full, machine-extracted citations presented as prominent objects with metadata (author, publication date), literal quotations front-and-center, and AI summaries de-emphasized. Coding assistants should force verification before code review or expensive test runs. Models should avoid first-person narration and ritual apologies, and be trained or constrained to be far less verbose.

Beyond content presentation, the piece pushes for task-specific, non-natural-language interfaces and built-in tools (e.g., one-click security scans) rather than relying on a giant conversational model to infer actions. It requires strong data provenance signals that distinguish authoritative API-derived tables from model-generated text, spreadsheet-like arithmetic verification of numeric outputs, and UI controls for reproducibility (exposed randomness/temperature and replayable processes). The overarching demand is for products designed around verification, provenance, and predictable behavior rather than persuasive conversational polish.