@davis7 Been big on reading reasoning traces for a while. Really does a lot to help me calibrate on how the model thinks
English
Migi is in NYC! (again)
503 posts
Migi is in NYC! (again)
@migidoes
Building @OdiEveryday
Katılım Eylül 2025
338 Takip Edilen51 Takipçiler
When u have a model go off and build something u understand there's an absurd amount of upside in just reading some of the reasoning traces and diffs as it goes
U don't even need to read much, I've found that just by looking at 10 lines I can instantly tell if it's doing it right or not.
And now that the frontier has gotten absurdly steerable, u can just tell it what's going wrong and how to correct then it just kinda will
English
@rishabhkaul Also, if you’re building AI infra or dev tools, you have to be in SF
English
@nikunj I think there is a better emoji for a jetlagged toddler 😁
English
Why is no one awake and on the streets, Dad?
Because it’s 5:32 am buddy 😴
English
@rauchg What’s the over/under on how long before Anthropic cries distillation
English
Kimi K3 is the best performing model on nextjs.org/evals, ahead of Fable, reaching a comparable success rate in less time.
This is the first time that an open model is ahead of all proprietary ones for this comprehensive web engineering benchmark.
Notes:
▪️ Benchmarks don’t always tell the full story, although this is important signal, adding to mounting evidence that this could be a breakthrough moment for open models
▪️ No model as of yet has reached 100% completion on this set of evals. The top performer peaks at 92% and 96% “with help”
English
This smog is wrecking my body. Shoulda stayed in Atlanta
English
@Rasmic Atlanta was a hoot for the Argentinians. Could hear them all day all over downtown
English
@thekitze It’s interesting because in first world countries it makes more sense to pay the provider
In developing countries, most are willing to self-host because they fixate on upfront cost before considering ongoing cost and strain with labor
English
i understand why so many people open source their app and make it self hostable
cuz self hosting is overall a nightmare even when you have computers to spare and money for any VPS
always something going wrong
so it's just a performative "pls feel free to host our thing"
English
I still sometimes see people saying "if you know how to write the code, it's faster to write it yourself"
I'd argue the exact opposite: if you know how to write it, you gain nothing from doing the typing yourself - outsource that to a coding agent!
English
@dayvough Seems like my luck makes me a great fit to cheer for the English today 😆
English
When I started doing my quarterly "google's AI is a joke" posts last fall I got a ton of replies telling me I was stupid for it
"they have so much data and compute they're gonna win just wait"
Each post after saw less and less of them until the most recent one, I got like two and they were both just kinda sad. Their hearts just aren't in it any more
Dunking on Google is one of my favorite things don't get me wrong, but now it just kinda feels like throwing balls at the clown after he fell off his little bike. It's just kinda mean and sad at this point.
I genuinely want them to be competitive, we need as much competition as possible, but they just don't seem to have it in them.
Maybe it's the ultimate long con who knows. 3.5 pro's gonna come out of no where and make Fable and 5.6 look like toys...
But also like come on
leo 🐾@synthwavedd
In other news, DeepMind have delayed 3.5 Pro again Rumour has it 3.5 Pro is Schrödinger's model, caught in state of perpetual delay until the Sun explodes Sometimes I wonder whether it's even worth making these posts anymore
English
Google Maps put me on a southbound express train that zipped right by the airport
From getting there with time to spare I am now looking at tricky timing to get there once this train stops
Thanks Google Maps. Your AI was wrong multiple times as I tried to navigate this
English
@eastdakota I think there’s a parallel to pre-AI coding somewhere here
English
I find all this so silly. Electric is simply better tech for 95% of vehicle use cases. Why are we burning so many cycles trying to replicate the past?
Sawyer Merritt@SawyerMerritt
Lexus says its upcoming all-electric LFA supercar will use vibrations and other sensory feedback to mimic the feel of a combustion engine. "We’re not just wanting to replicate the sound of the engine: we want to redesign the sound itself so the car feels more alive and authentically connected." The production version of the car is expected to be unveiled at the end of this year/early next year. Concept unveiled in Dec 2025:
English
Migi is in NYC! (again) retweetledi
@GergelyOrosz Are you getting early sentiment on @cognition’s model?
English
This is why competition leads to much better outcomes for customers and consumers, in general
If it was not for GPT-5.6 released a few days ago, Anthropic would have most likely removed Fable from most paid plans, and charged $$$ for access
But with GPT-5.6 out, they cannot
Claude@claudeai
We're extending Claude Fable 5 access on all paid plans, as well as keeping Claude Code’s weekly rate limits 50% higher, through July 19.
English
@awilkinson Hahahaha I think the shock should be based on some aggregated score across token spend, merged PRs, bugs shipped to prod, and actually insightful things said
English
A proposal:
We affix all AI influencers with shock collars with small microphones and cameras.
Should they type or utter the words "they cooked", the collar immediately administers a 50,000 volt shock.
The voltage increases by 10,000 per utterance.
Thoughts?
English
In the future the primary thing that agents will do is write code to accomplish the ad hoc queries from even non-technical users
English
Tried the mlx version of qwen3.6-35b
Did pretty good on tps but I gave up after a bit because
1. It couldn’t finish the task I gave it which was to format my poorly copied md version of the telegram bot API. Task was probably too big for it
2. Battery went at 1% per min while generating tokens
Will be monitoring this thread because I am also interested in a good answer to your post
English
@omooretweets Sure but the ChatGPT experience that most are familiar with is absent in this Codex turned ChatGPT app
It is a confusing move at a bad time
English
I respectfully disagree!
~10% of the global population uses ChatGPT weekly, most of whom are using it like a Google replacement (basic searches)
Power users, most of whom are developers, are fantastic...but we need to start exposing the mainstream consumer to what AI can do
Theo - t3.gg@theo
fwiw, I think that turning Codex into ChatGPT Desktop is a generational fumble
English