Dimitri Semenovich

Future of software - future of computing - the future

[Dimitri Semenovich @dmvse]

@mikehostetler Not until the BEAM gets resource and capability isolation! 😅

[Dimitri Semenovich @dmvse]

@mikehostetler A pure native capabilities system would be probably cleaner and much simpler - I just wanted Plan9 semantics and name spaces

[Dimitri Semenovich @dmvse]

@mikehostetler I have a moderately credible file-by-file Inferno kernel port I've been working on for the last few months, would be great to get it out there - hopefully in the next few weeks

[Dimitri Semenovich @dmvse]

@flowerornament Whether you another layer is the question. I've been running Ash apps hosted on a capability restricted Inferno kernel ported to he BEAM with optional 9P adapters - this seems to give you same properties as a dynamic system - agents can modify/test/restart modules live.

[Dimitri Semenovich @dmvse]

@flowerornament I've built a UI library where every screen is an Ash resource with html generation as calculations and all client side events mapped onto actions. Reactivity comes from standard relations and notifications. This is all compile time only - wonder how to make it more dynamic?

[Dimitri Semenovich @dmvse]

@synopsi Video games are achievement simulators - coding agents just have greater verisimilitude, games as we understand them today have no chance at all

[Dimitri Semenovich @dmvse]

@progrium Amazing! Is this in the Linux VM or a direct WASM comiple?

[Dimitri Semenovich @dmvse]

@flowerornament @ianlivingstone Yes, dynamically created Ash resources that can be at least somewhat “untrusted”. A fully hermetic world is just a linux vm + a 9P socket away 😅

[Dimitri Semenovich @dmvse]

@flowerornament @ianlivingstone The Linda tuple space pattern is great as well, something that ash_oban sort of does with triggs

[Dimitri Semenovich @dmvse]

@flowerornament @ianlivingstone I’ve been working a lot with Ash lately as well - especially something like “all business logic goes into actions”, “all long running actions return a receipt immediately and are executed on a queue” seems to make it easier to keep the overall architecture sane

[Dimitri Semenovich @dmvse]

@flowerornament @ianlivingstone To make this work though, the file server execution needs to strip away all the standard BEAM interprocess comms by essentially recreating the compiler front end. Wonder if there is a better way. MS COM is another conceptually similar design, wonder if it could be a better model

[Dimitri Semenovich @dmvse]

@flowerornament @ianlivingstone Are you building Inferno on the beam? Late binding + no ambient authority seems to be what AI software systems need

[Dimitri Semenovich @dmvse]

@thsottiaux Periodically forgets “Full Access” especially when changing accounts - no way to restore without fully stopping all running sessions / clearing all queued prompts

[Dimitri Semenovich @dmvse]

@thsottiaux Extremely slow switching between long running sessions - as long as 20s on M4 Max

[Dimitri Semenovich @dmvse]

@mikehostetler Hi Mike - it’s probably too late now with 2.0 but wanted to mention that the use of “actuons” in jido is very idiosyncratic - in most literature an action is something the agent _does_ (your directives), the inputs are either observations or controls - even AIs get confused!

[Dimitri Semenovich @dmvse]

@wwwjim Great to see something meaningfully different from yet another shadcn template. I just want web apps to look like this 😓

[Dimitri Semenovich @dmvse]

@SIGKITTEN Got to love Berkeley Mono

[Dimitri Semenovich @dmvse]

@amplifiedamp @deltan0vy Would be awesome, would like to rerun all the experiments I’ve set up with 405B