Best of your X follows: model routers, AAA agents, and bottlenecks

Agentic systems

Ethan Mollick: the model as router

• What happened: Ethan Mollick asked whether the frontier model itself becomes the router, delegating work to cheaper models when it decides they are enough 1.
• Why it matters: The interesting design shift is from humans choosing a model stack up front to a smart planner deciding which lower-cost model should handle each subtask.
• Signal: The post had 543 likes, 149 bookmarks, 51 replies, and 58,357 views at capture 1.

Ethan Mollick: long-horizon use beats casual prompting

• What happened: Mollick argued that the bigger issue is not whether models can answer simple questions, but whether people try ambitious, long-horizon work with them 2.
• Why it matters: This is the same divide many teams now face: using AI as search or homework help produces small gains; giving it real project work changes the workflow.
• Signal: The post had 638 likes, 145 bookmarks, 44 replies, and 71,130 views at capture 2.

AI tools and experiments

Ethan Mollick: asking Claude Fable to make a game more AAA

• What happened: Mollick repeatedly asked Claude Fable to make a browser game "more AAA," and said the model responded by adding graphics, boss fights, mechanics, sounds, and a soundtrack until WebGL became the limit 3.
• Why it matters: The example is less about the game itself than the direction of agentic product work: models are beginning to make taste, scope, and implementation choices across an iterative build.
• Signal: The post had 290 likes, 136 bookmarks, 31 replies, 120,437 views, and linked to the playable demo 3.

Ethan Mollick: the follow-up version exposed the model's taste

• What happened: In a follow-up, Mollick asked for the online idea of "AAA" and said Fable added lootboxes, EULAs, achievements, useless graphical settings, and elaborate boot screens 4.
• Why it matters: The funny part is also the useful part: when agents are asked for higher-level taste, they reproduce cultural patterns, not just code.
• Signal: The follow-up was smaller but concrete, with 163 likes, 42 bookmarks, 16 replies, 41,459 views, and a second playable link 4.

Research and constraints

François Chollet: information and energy as the long-term bottlenecks

• What happened: François Chollet reduced long-term bottlenecks to two scarce inputs: information and energy 5.
• Why it matters: For AI builders, that is a useful compression of two constraints: better data or feedback on one side, and cheaper compute or power on the other.
• Signal: The post was short but self-contained, with 842 likes, 126 bookmarks, 83 replies, and 50,688 views at capture 5.