Joe Saia

Software and economic things

[Joe Saia @Joseph_myke]

@simonw @nateberkopec I think to a certain extent a lot of enterprise network security sorta kinda enforces the controls you would need here already.

As the labs lean more heavily into subagents, I wonder if we get orchestrator agents that delegate tasks with scoped tool permissions as another layer.

[Joe Saia @Joseph_myke]

@badlogicgames 4. There is a distinction between software programming tasks and software engineering tasks even if we do not really have the title "Programmer" any more and everything has been smushed into the software engineer title. This isn't a jr versus sr thing either.

[Joe Saia @Joseph_myke]

I've been having a lot of fun building with Pi Durable on cloudflare Durable Objects. It's really reduced my level of nihilism about the future of software engineering. Getting a great new primitive for agents to use is really empowering. thanks @badlogicgames and @earendilworks

[Joe Saia @Joseph_myke]

@badlogicgames Pi's codebase is very plesant to read. I suspect that a lot of what I find pleasant about it is orthogonal to what agents find pleasant.

[Joe Saia @Joseph_myke]

I put down some more thoughts here
https://t.co/FOfxvfeeto

[Joe Saia @Joseph_myke]

It's a much needed reminder that building something novel is still a fundamentally human activity but even more exciting because doing it well unlocks more than ever.

[Joe Saia @Joseph_myke]

@Lethain What are actually looking for? Is this more being able to do traditional IC work, or more compound engineering style work of marking the central harness or eng process better or general org work?

[Joe Saia @Joseph_myke]

@ModeledBehavior I also think the best games are defined by a very refined “feel” whether that be artistic style or more mechanical tactile feel that is not amenable to AI. It’s more akin to high end custom wooden furniture. So the human touch is still the bottle neck on quality.

[Joe Saia @Joseph_myke]

@ModeledBehavior Dlss and fsr have been a huge impact on visual quality enabled by running neural net directly on the gpu to upsample images. Not exactly AI but the very related technology.

[Joe Saia @Joseph_myke]

@emollick Slop code bench is interesting. Very much not saturated yet. They give coding tasks that evolve over time and a fresh model context has to pick up the next leg of the task when it comes in.

https://t.co/I7qhjq6QXI

https://t.co/qR02DcYyhE

[Joe Saia @Joseph_myke]

@zeeg I think many platforms set it by default to minimize their load without every exposing it to the site owner so unclear how many orgs are actually making that decision for themselves

[Joe Saia @Joseph_myke]

@zeeg I need to scrape local government websites for meeting information and about a third just blanket ban all bots by default which is pretty crazy thing for governments to do.

[Joe Saia @Joseph_myke]

@thorstenball I was always a big literal rubber ducking guy though talking out loud to myself so this just actually gives me something more concrete.

[Joe Saia @Joseph_myke]

@thorstenball For actual writing I tend to write out a few bullet points to structure my thoughts and then just dictate using the built in Mac dictation as a very rambling first draft then edit and rewrite. Sometimes it ends up a full rewrite but the dictation step helps surface my thoughts

[Joe Saia @Joseph_myke]

It's not just context switching it's more similar to having to go back and forth between maker and manager modes in a way that is very uncomfortable and I kind of hate it.

[Joe Saia @Joseph_myke]

Something I've been struggling with working with coding agents is the switching between tasks that they can one-shot without me having to think about and novel problems that they are incapable of solving where I need to micromanage.

https://t.co/1xG7flWqic