https://www.youtube.com/live/251hsWgoTPM?t=3576
I am giving Matt's skills a try for the first time. This demo makes me feel like this could unlock some inspiration, because the hardest part is getting started on a project when there are so many unknown unknowns. /wayfinder feels like the right approach, so I'm going to give it a go.
59m36s into Matt's demo, there's a question of why he chose Opus 4.8 medium. Matt explains the trade-off of effort+latency and quality, as well as matching the strength of the model with "smart enough for me".
This point speaks to my repeated exploration of fast cheap models that are smart enough to execute the skills and tools I care about for accomplishing the tasks; and achieving the outcomes I want. I want the cheapest model that does an adequate job. Good enough and no better is the optimization. A great deal of this motivation is because I'm the smartest human in the room, a room in which I alone interact with the agent. I don't need a very smart model, because it is closely supervised by my careful attention.
I feel like folks reflexively reaching for the smartest frontier model are publicly admitting their own human deficiencies.
I am giving Matt's skills a try for the first time. This demo makes me feel like this could unlock some inspiration, because the hardest part is getting started on a project when there are so many unknown unknowns. /wayfinder feels like the right approach, so I'm going to give it a go.
59m36s into Matt's demo, there's a question of why he chose Opus 4.8 medium. Matt explains the trade-off of effort+latency and quality, as well as matching the strength of the model with "smart enough for me".
This point speaks to my repeated exploration of fast cheap models that are smart enough to execute the skills and tools I care about for accomplishing the tasks; and achieving the outcomes I want. I want the cheapest model that does an adequate job. Good enough and no better is the optimization. A great deal of this motivation is because I'm the smartest human in the room, a room in which I alone interact with the agent. I don't need a very smart model, because it is closely supervised by my careful attention.
I feel like folks reflexively reaching for the smartest frontier model are publicly admitting their own human deficiencies.