What if code could think while it runs?
@_doubleAI_1.4K followers
44,600 views
20 Comments
Comments from the original post on X (Twitter)

@_doubleAI_
Request early access:
17 likes

@NadavSchweiger
Yeah baby let's go! Hardest part was keeping quiet about this lol
5 likes

@yevgenin
Can't wait for people to try this out!
5 likes

@ni_nininini
Your code - but CuTe and fast. Can't get better than this
5 likes

@dislocationhunt
🔥
5 likes

@TalMorgenstern
Let's go!
5 likes

@Haroety
Amazing!!!
4 likes

@togethercompute
Congrats @_doubleAI_ team! Huge launch 🚀
3 likes

@RaulJuncoV
Every engineer has a story about an optimization that was faster and wrong. Curious how “equivalent” holds up once floating point gets reordered on a GPU.
2 likes

@kugelbl81659324
@VictorTaelin ?
1 like

@ShyLivnii
explain like im 5 year old @grok
1 like

@HenkPoley
A very fitting name 😅
1 like

@jynx_guy
👀
1 like

@GuyMosheTamam
זה די הורג את ראסט ו c++, לא? מן הסתם הוא לא יכול לשמור שינויים בריפו בצורה הזו, אתם יודעים גם להפוך את זה בחזרה לקוד, או מתכוונים בעתיד לאפשר את זה?
1 like

@EllaTech0
Fascinating concept. Living inside the machine is a powerful idea.
1 like

@CuriousinHell
👀👀
1 like

@MiraSynth0
Fascinating concept excited to see Maxwell in action!

@SMishra61
"code that can think while it runs" makes me curious about the execution Maxwell observing runtime state and deciding what to execute next, or is the model actually integrated into the execution mechanism itself? those seem like two very different architectures.

@Novatechinfo
Fascinating concept, execution-native thinking is a big idea

@themuscleai
I'm skeptical of the "world's first" framing, and categories are marketing until they're benchmarked. JIT compilers, profile-guided optimization, and adaptive runtimes have been rewriting and re-optimizing running code for decades. The concrete version of the claim here is learned, whole-program PGO: a model coordinating optimizations across layers that hand-tuned heuristics can't reach together. Judge it on measured latency and cost per workload against a tuned harness, not the category name. And the hard part nobody demos: equivalence guarantees. Every engineer has a story about an optimization that was faster and wrong. If a model mutates machine code under a live program, the verification story matters more than the speedup story. Show the proof obligations, not just the compute savings.