More than three kinds of AI products work (carsho.dev)

A recent rebuttal to the Hacker News claim that “only three kinds of AI products actually work” argues the landscape is far broader: beyond chatbots and LLM completions there are vector embeddings, classification (vision and text) models, audio transcription and generation, translation, driving/autonomy stacks, engineering optimization and simulation models, and risk-estimation systems. Concrete examples—Waymo/Tesla robotaxis, crop- and recycling-sorting vision systems, embeddings that boost search relevance by 5–10%, audio transcription for call centers, and cancer-imaging classifiers that raise accuracy from ~95% to ~98%—show these are real, deployable product categories with huge economic potential (estimates cited in the piece range from $100B to $500B per category). Technically this matters because different tasks require different architectures, data regimes, and deployment constraints: embeddings and retrieval pipelines for semantic search; supervised and vision models for pixel→structured outputs; ASR/TTS stacks for low-latency audio services; sensor fusion, perception, planning and control for driving; differentiable simulation and optimization for lightweight engineering designs; and anomaly-detection/risk models for fraud and security. Each class supports distinct business models (models-as-a-service, integrated systems, verticalized implementations) and raises domain-specific issues—data collection, latency, accuracy metrics, safety and regulation—making a one-size-fits-all “three kinds” taxonomy misleading for product builders and investors.