Large Language Models Get All the Hype, but Small Models Do the Real Work (www.wsj.com)

Headlines this year have been dominated by ever-larger “frontier” LLMs that beat humans on exams and reasoning tests, and by leaders touting a path to AGI. But most real-world AI work still runs on much smaller, cheaper models: on-device assistants, embedded classifiers, routing and safety filters, and domain-specific models that power search, summarization, and recommendation systems at scale. Those compact models don’t make splashy headlines, yet they handle the bulk of production inference because they’re faster, far less expensive to run, and easier to control and private-deploy than massive foundation models. Technically, this reality drives work on compression and efficiency—quantization and pruning, knowledge distillation, parameter-efficient fine-tuning (LoRA/adapters), model cascades, and carefully engineered inference stacks—to squeeze high utility from far fewer parameters and lower precision arithmetic. The implication for practitioners and researchers is clear: progress isn’t just about scale and benchmark records. It’s about co-designing models, software, and hardware for cost, latency, robustness, and privacy; improving evaluation on deployment metrics; and investing in tooling to adapt small models quickly to specific tasks. That focus enables wider, more reliable real-world AI adoption and democratizes capability beyond a few giant models.