Sabitlenmiş Tweet
akira
4.8K posts
akira
@realmcore_
Making an autonomous swe • @0xrandomlabs Incepto ne desistam Pax aeternum Memento Mori
elysium Katılım Kasım 2021
770 Takip Edilen9.6K Takipçiler
smilarly great signals:
- managing agents is like managing a team of humans
- but i have review agents
- spec is all you need
English
i actually don't want this "but you don't review compiler output either" meme to die.
it's the perfect signal for being immediately able to ignore someone in this space.
solst/ICE of Astarte@IceSolst
Interesting article on treating agent output like compiler output (and why) skiplabs.io/blog/codegen_a…
English
The story for harness opt in multi-agent settings is likely conclusive imo
Long horizon, tbd
justin@justinsunyt
@JoshPurtell especially for multi-agent yes!! here raw intelligence !== output quality at all
English
akira retweetledi
however if you're using gpt 5.5 it will automatically take at least 1.5x as much time
arb8020@arb8020
you can refactor anything in two weeks
English
@tejasybhakta It’s quite stable in general in how it handles problem solving
It’s code is still too dense and abstracted but mostly pretty good
English
@realmcore_ It’s the first openai model I’ve thought is good. Concerningly good at writing kernels
English
FYI 5.5 kind of fixes this.
Lots of theories on intermediate stuff but yeah
Fixes it mostly
akira@realmcore_
5.3 codex xhigh is significantly better than 5.4 xhigh for coding. It's genuinely nuts
English
akira retweetledi
Announcing @slashyai iMessage bot
The first email client with a blue bubble
You can do anything on Slashy via text.
→ Draft & send emails in seconds
→ Get pinged the second a customer emails you
→ Schedule or reschedule anything instantly
→ Build automations that actually work
→ Send voice memos
→ Literally anything else you could want
And yes we still have a web/mobile/desktop app :)
And yes we are still just $30 per month to get started
English
I mean end to end dev setup actually!
Certain dev setups lend themselves particularly well to specific models and more generally full autonomy
Ex: I notice 5.4 Xhigh has a strong bias towards intermediate type validation/transformation
If you codebase is written in a way where this is expected then the resulting code will be much more acceptable than in a codebase where there is no validation or where it is centralized
Same for error handling patterns
Anything in particular you find 5.4 to be better for than 5.3?
English
@realmcore_ We have free will for model choices and aren’t constrained to the newest one if that’s what you mean. We haven’t changed much in terms of prompting / setup
English
5.3 codex xhigh is significantly better than 5.4 xhigh for coding. It's genuinely nuts
akira@realmcore_
gpt 5.4 How do you guys get this model to not do random algorithmic garbage and just write straightforward procedural code I do not see why it should be this hard for a model to write code like a first year college student probably skill issue tbh
English
@ryanbrewer Presumably you also have everything set up for this to be the case?
English
The labs in fact have not yet solved everything
In light of recent events we might be on the proliferation timeline
Shannon Sands@max_paperclips
none of this is true btw. claude code still sucks and crashes often, google is notorious for dropping projects, openai has a trail of abandoned ideas and then you've got vendor lock-in to worry about on top of that there's plenty of room, same as always
English
akira retweetledi
none of this is true btw. claude code still sucks and crashes often, google is notorious for dropping projects, openai has a trail of abandoned ideas
and then you've got vendor lock-in to worry about on top of that
there's plenty of room, same as always
alex fazio@alxfazio
the window for experimenting with llms has basically closed now. the megacorps have fully hit escape velocity and are shipping new products and new features daily. the shift is that they’re not just shipping llms anymore, they’re using llms to build products and improve existing ones at scale. the wild west era of llms isn’t really the wild west anymore. a year ago, this could’ve been an indie dev side project, maybe even a monetizable product. it was literally so easy that the only real bottleneck was your free time. now, whatever idea you have, you should basically assume google/anthropic/oai will build some version of it within a week and wipe out most of the startup surface area around it
English
@realmcore_ main reason i use Opus for anything beyond very focused changes
i feel like if there's exactly one good way to do something, GPT-5.4 does beautifully
as soon as something is even slightly open ended it goes haywire
really hoping Spud fixes this (tmrw?)
English