This paper argues that large language model-based recommenders can displace today's attention-allocation machinery. LLM-based recommenders would ingest open-web content, infer a user's natural-language goals, and present information that matches their reflective preferences. Properly designed, they could deliver personalization without industrial-scale data hoarding, return control to individuals, optimize for genuine ends rather than click-through proxies, and support autonomous attention management. Synthesizing evidence of current systems' harms with recent work on LLM-driven pipelines, we identify four key research hurdles: generating candidates without centralized data, maintaining computational efficiency, modeling preferences robustly, and defending against prompt-injection. None looks prohibitive; surmounting them would steer the digital public sphere toward democratic, human-centered values.