arnav

i think i'm gonna build (yet another) self-hostable cloud (or, more accurately a deployment platform) i think i have some really cool ideas so stay tuned!

@thsottiaux pov: you remember about sleep

Update. I have decided to take a break from x to recharge a bit. See you back tomorrow for more ChatGPT and Codex fun. It's going to be a fun week.

Some ground rules for now - - as lightweight and minimal as possible - no web dashboards. if i ever create a dashboard it'll be read only - @Railway like service based model - agent accessible - first class secrets/env support i don't want to create yet another "open source vercel," let's see where it goes!

@davis7 I'm in this weird spot where I like Sol's behaviour a lot but still prefer Opus' code by some margin Fable is sadly too expensive :")

Opus 5 is the first model where I've liked it less the more I've used it. Never had this before, with previous models I've liked them more as I learned to use them It has some really bizarre behaviors that I haven't really seen since GPT-5.5. I still think it's a step in the right direction for Claude models, I'd love to see a Fable 5.1 that has it's voice and feel. Pros: - writes code I actually like - asks great questions - feels great to talk to - has some of GPT-5.6's autism/"grind the problem down" - has some of Fable's deep understanding - feels fast + nearly unlimited after getting used to Fable limits/speed Cons: - over complicates the hell out of everything - can pretty easily miss the point of what you're asking and go down the wrong path - really likes going off and doing "extra" work that you didn't need or want that's completely unrelated - done some strange stuff to me like hard closing my browser repeatedly so that it could take it over for a computer use test - it's bad at stopping. I know that sounds weird, but I've had multiple instances of it "stopping" but leaving background triggers out there so it pops up 20 mins later and starts doing random shit. I've also had it stop early way more than I'd like - almost too literal, I asked it to make a pr and it refused because I didn't ask it to also push up the branch I'm still using and enjoying it a lot. I'll probably keep using it because when it works it feels great, but it requires a lot of effort to keep on the right track and is definitely not a true Fable replacement.

@dillon_mulroy It's so obvious in retrospect but I didn't consider artifacts was built on DOs Durable Object everything ig

region: earth

@maria_rcks no way what kinda pull is this

WE GOT IT GUYS

@BenjDicken reminds me of this 💀

Happy to announce I've been fully replaced internally. Guess I'll just tweet all day now.

@wagslane if you think about it everyone is pair programming now

pingora has to be one of the most underrated @CloudflareDev "products". can't believe they just put it out there for free lol. makes it so easy to build anything networking-related while staying highly performant

if @AnthropicAI signed the letter right now, they'd be a full 6 rows above @OpenAI missed mogging opportunity?

At least I don't have to hear about how great pair programming is anymore

@steveruizok absolutely goated for prototyping. so capable for their size (and power usage). I built a "bike lock" that unlocks over BLE so I could share it with friends recently

why aren't you posting about ESP32 devices? there's an opportunity right here. right now. to post about little devices

@threepointone @mattzcarey what exactly do you mean by "open alignment research"? isn't alignment a part of training and thus can only be done by the labs more or less? or do you mean judging alignment

didn’t think I’d get fully alignment-pilled but here we are @mattzcarey convinced me we need “open alignment” research

@theo When android support 😢

Next T3 Code nightly cuts payload sizes by almost 90% for all threads and data on client. Should make mobile app feel significantly snappier, especially on 5g

agentrecap is now live on producthunt! try it out right now with `npx agentrecap`. it shows - - how long agents actually run for - what tools they use - how long you take to respond - how much code gets written - estimated API costs - token usage comparisons ... and more works with claude code, codex and opencode

@Darkpoolballer @maria_rcks oh shoot im dumb, my bad

@arnvbnsl @maria_rcks U can add it to fav no?

guys help me convince theo this is WAY better for a coding agent

@rauchg @ctatedev is a madman and I love it

I compiled 𝚟𝚎𝚛𝚌𝚎𝚕 CLI TypeScript to native with 𝚜𝚌𝚛𝚒𝚙𝚝𝚌. Incredible. ✓ Resulting binary size: 1.28mb ✓ Startup overhead: 1.5ms (mean) ✓ Compiles in 2.94s (mean) ✓ Uses 𝚗𝚘𝚍𝚎:𝚑𝚝𝚝𝚙𝚜, 𝚗𝚘𝚍𝚎:𝚏𝚜, 𝚗𝚘𝚍𝚎:𝚙𝚊𝚝𝚑, 𝚗𝚘𝚍𝚎:𝚘𝚜, 𝚗𝚘𝚍𝚎:𝚌𝚛𝚢𝚙𝚝𝚘 ✓ Highly readable TypeScript ✓ Code translated by GLM 5.2 Fast ✓ Deploys like a charm! No embedded v8 / QuickJS, fully static.

@theo @jullerino if you want i can do a test of sorts with and without t3code open?

@theo @jullerino Yep

uh @theo @jullerino love t3code but what's up with the battery utilization 😭 my macbook's battery life has become almost half of what it was......