@karrisaarinen Doesn't your purpose differ from someone else's purpose ? Eventually a service provider becomes a tool by virtue of a connector and the customer plops their own plumbing, control flow upon it. That part is eaten by agents today with the exception that its ephemeral in nature.
@jiriknesl Its unfortunately a herd-mentality .. Someone says "formal verification using vibe-lean" and everybody jumps on the bandwagon without any prior understanding of fit and need.
@superlogical Rex is a poster child of user delight both in terms of onboarding and affordances .. @mitchellh you and your team has made my experience delightful
I would generalize that assumption that Pi durable is great for *all* things harness related.
When we hook to harness mods we begin to appreciate its complexity and usefulness in terms of its symbiosis with the inference edge.
Pi did not work for me in many cases so each of them have their own place and use sibject to the needs, complexity and situated domain
@choblin29 Lets be honest here .. the common folk do not have infinite token stream & compute for theorem proving .. this will of course concentrate as a function of hyper-scalars
Question : What is the half life of skills ? The serve the purpose of agent bootstrap for situational awareness . Thereafter they become candidates for ablation to unlock latent intelligence of model. At least I see these things lose value to the point where they might cease to exist.
@RichardSocher@Recursive_SI Very Impressed so far .. Putting it through the ringer with static languages (oCaml, Haskell) and rich type systems and its holding up better or equivalent to frontier models.
@trq212 The # of agents spawned needs to scale with the limits of human cognition and attention. We cannot afford a full hand-off to agent swarms for critical tasks (certain coding tasks can be considered more critical than others)
Its not as smooth as this .. there is a pissing-contest that plays out where one Engineer-2 calls out the ignorance of the Engineer-1 effective model usage, lack of skills, etc
Typical anecdotes
- "The code is a mess. Half of the CLI commands do not work"
- "Check my dev setup. Its better. I have pre and post commit hooks with improve-codebase-architecture skills"
- "The codebase is full of slop .."