Three kinds of AI products work

Original Article ↗ Hacker News Discussion ↗

Scope of “AI products” seen as too narrow

  • Many commenters argue the article conflates “AI” with “LLMs you chat with,” ignoring large, profitable categories: fraud detection, recommendation, translation, speech recognition, TTS, driving, medical imaging, document parsing, supply-chain optimization, etc.
  • Successful tools like Grammarly, DeepL, and vision-based document processing are cited as counterexamples to the three categories.
  • Several stress that the most valuable AI is “invisible” infrastructure; users don’t even know AI is involved.

Media generation as a major category

  • Multiple replies object to dismissing image generation as a “toy”; they say image/video/music/voice/3D tools are already production-grade for design, marketing, film, and game assets.
  • Discussion on UX: graph/node-based systems (ComfyUI-like) are seen as too technical; “Adobe/Figma-like” interfaces that hide model complexity are considered the real opportunity.
  • Some note that modern media tools already behave agentically (iterative edits, inpainting, upscaling), just on pixels instead of code.

Alternative product taxonomies

  • One proposed breakdown:
    • Batch/pipeline processing (document parsing, moderation).
    • “AI features” inside apps (summaries, tags, autocomplete).
    • True agents (AI controls workflow).
  • Others say “three kinds of AI products” is as coarse as “three kinds of internet products,” too high-level to be predictive.
  • Several point out the article itself effectively lists more than three categories.

Agents and coding tools: strong disagreement

  • Some report big wins from coding agents: debugging, analyzing logs/stack traces, discovering index issues, and catching obscure build problems.
  • Others find agents unreliable or “lying,” producing useless code even with good APIs; they trust junior devs more and see agents as “slop factories” without a human in the loop.
  • Debate on productivity impact: claims range from negligible savings to “4–8x fewer developers needed,” with skeptics demanding harder evidence.

Non-chat, task-focused uses

  • Frequently mentioned working use cases: translation, transcription, summarization of contracts/logs/meetings, basic drafting, clarifying vague business requirements.
  • Some say AI summaries are often bad or misprioritized; others find them net-positive if treated like code review (LLM drafts, human verifies).

Future directions: agents, UX, and infrastructure

  • Many want agents for mundane life tasks (appointments, cancellations, travel, forms) but note current systems struggle with real-world details and costs.
  • Some argue these tasks should be solved with open APIs and classic algorithms, not LLMs everywhere.
  • There is concern about giving agents real powers (refunds, account changes) due to jailbreak and error externalities.
  • Several conclude the real long-term value is AI as an underlying capability or interface layer, not standalone chatbots.