Jev for dynamic composition and ordering of prompt components (aka foci) at runtime. #focalprompt

Word of the year at LisbonAI 2025 was whimsy. This year it was taste.

Instant agent upgrade. Break your prompt up into instructional units or "foci". For a given user input have Jev determine which foci belong in the prompt. Dynamically compose the prompt at runtime.

One of our goals with FocalPrompt has been to make prompts dynamically composable: treat tagged prompt components (“foci”) almost like tool calls or retrieval, and only include the instructions relevant to the task at hand (in our case drafting responses to a range of pet owner queries). Until now, doing that has usually meant extra model calls → extra latency + cost. There’s also a harder problem: the foci a model predicts it needs aren’t always the same ones it actually appears to rely on. Leave-one-out ablation can reveal that gap. Two days experimenting with Jev looks very promising. Score each focus → set an inclusion threshold → compose the prompt dynamically. The speed and cost make per-focus routing look genuinely practical. The interesting question now is whether predicted relevance reliably preserves the instructions the agent causally depends on. @willmonk @GetPetsApp

Equally one man’s hedonistic dystopia is another man’s nihilistic utopia 😬

The great debate…

Imagination without knowledge leads you astray. Knowledge without imagination leads you nowhere.

Resident Goose You smile as the geese return, Look wistful as they fly away. Would it surprise you at all to learn That some of us loyally stay In this off-season habitat? Given you failed to notice that Is it really the goose you love Or a cycle of abuse from above? Is it that, in reality, You praise your sun’s prodigality? (You’d think human ingenuity Would find richer forms of promiscuity.) That whooshing honking from the sky Offers a shared temporal space. And the best we who remain nigh Can offer, is a poop bomb to the face.

A fire spits and spittles with the same sputum it later roars. It simmers and simpers and later soars. Like Sonata Pathetique, the Mona Lisa, or any tour de force.

And all the while I was thinking: stacking a woodshed would be a solid test for AI. (But please do not deploy.)

Looks like state-of-the-art robotics training footage to me.

The uber driver just said, in an accent (I’m in Portugal), “It will take about 12 minutes.” To which I replied: “It’s nice to meet you.” Trying to pass it off as an English thing.

Bob the blob is learning to play #rl

Agents take two types of risk: - irreducible risks, and - acceptable risks

Nice articulation of the innovator’s dilemma: “To obtain a lot of reward, a reinforcement learning agent must prefer actions that it has tried in the past and found to be effective in producing reward. But to discover such actions, it has to try actions that it has not selected before. The agent has to exploit what it already knows in order to obtain reward, but is also has to explore in order to make better action selections in the future. The dilemma is that neither exploration nor exploitation can be pursued exclusively without failing at the task.” Sutton & Barto