“Capability eats many things, but [the model] does not hold the license, sign off on the liability, or own the firm's files, and it cannot be the party that gets sued when the answer is wrong. Intelligence is not the bottleneck here. Permission is, and so is accountability.
Since “Applied AI” is the practice sculpting jagged artificial intelligence into practical intelligence within a complex domain, applied AI is not a “tech problem”. It’s a domain problem, so it’s a problem for domain experts *outside of tech*
I love this. The words "agent"/"agency" are a little funny in this context since the whole point of the walls is to *decrease* agency but its where @Interloom see agents actually useful things, outside the coding use case. ie: Reduce action space, add tons of guardrails
A mental model for working with coding agents is that they're blind squirrels running into a maze and bumping into walls. You must place the walls (verifiable constraints) strategically so that they end up in the general region you want them in.
When a product builder bundles a set of skills with their app, it acts a “menu of features” and users can bootstrap their learning by firing off each skill in a sandbox. In addition the system could even show pre-baked examples of inputs and outputs of each skill
Well designed products have good affordances. That's true for both digital and physical designs. Affordance examples: door handle, a light switch, a button, or a slider. Affordances tell you something about how the product can to be used simply by looking at it.
I think this might be an underestimated use for skills.md. I think there's a risk the whole skill craze will go the way of the bitter lesson, (remember “prompt engineering”?) but I think can be terrific antidote to the lack of affordances.
The brand of “AI” needs splitting into sub brands, since “AI” is now used to pursue competing visions:
1) constructive AI: supporting people to accomplish endeavours that benefit other people
2) destructive AI: slop tornadoes and impersonation bots
I want nothing to do with 2)
Turns out goldfish have pretty decent memory (months/years). Maybe 🐝's are a more fitting analogy, especially given an LLMs sometimes overzealous work ethic.
"AX" = "Agent Experience"… let's see if this sticks, but it's a become useful term at Interloom already. Everyone intuitively knows the difference between shitty/excellent UX so its a nice way to borrow that empathy for LLMs. And a little empathy goes a long way with LLMs.
A thought we had for a couple of months now is how closely user experience & agent experience are related. It is one of our first principles, is that what users see, agents see, what agents can do, users can do, etc. First level citizens in a collaborative world of people and agents working together. @Interloom@erikcollinder
LLM agents are both oracles and goldfish at the same time. The AX/UX split is neat because it admits that we need to pay special attention to when to anthropomorphise ("this will make sense to a person") and when to not ("this will make sense to the oracle with instant amnesia").
@VAntik Hello Virane Antik, I was hoping you could get in touch with me on [email protected] about selling this twitter handle to a company I'm working with. Thank you!