@gigaqian: Today marks the official release of an AI model that worked autonomously for over 2 weeks building its own coding tool from scratch! Qwen 3.8 Max from Alibaba is a massive, 2.4 trillion parameter AI model. It provides frontier performance at a fraction of the typical cost. The model competes directly with Kimi K3 from Moonshot AI, a company that Alibaba helped finance! This release signals that efficient autonomy is a priority for major labs. The repository for the coding project (“oh-my-cli”) is public, and Alibaba claims to be releasing the open weights for the newest Qwen model next week! #breakingnews #ainews #alibaba #aicoding #gigaqian

George Patience242 :

Kimi is even better than what Claude used to be. there are throttling on fable and sonnet that make more work where Kimi lays out everything so it's easier to edit without the hyperbole that Claude inserts to justify the throttle

h :

Is there a discord I can join with more ai info like this

Tommy Chan :

China numba 1

Discorse :

FYI you have a fake account posing as you following back people who follow you. Just happened to me

cheec :

No way China can produce at this rate without having more compute that North America.

Wahala :

It’s GitHub is insane, literally all issued listed are its own just like you said. This barely came out… lol you’re the goat

thecardkilla :

Holy bot attacks Batman, your comments are flooded with them. What’s your personal favorite AI for your day to day? I’m a Claude fan just didn’t think kimi was on par when I tried it, but that’s just me I guess

Mirza Izic :

_ET_ :

Nice! Similiar build I made with DeepSeek

Erik :

Love the content. Great delivery.

hi__kyle :

I think the biggest challenge I’m facing isn’t an AI tool that can code (which there are plenty of), but rather an AI that can understand my vision. It’s challenging to get a AI to write code that matches the vision of your product.

aliz :

And their commercial is top notch as well☺️

a007frankwhite :

Yea i had downloaded 3.6 andcrealized what it took to run it locally dammm

Roger Hahn :

you can always limit the number of loops with evals and set token max. Never had a good experience letting agents run autonomously. Too many things can go wrong. Also, how do CI runners work. Highly skeptical it can fix on its own with operator input. 167 PRs over 16 days is nothing to brag about. Good harness can do 500 in <5 days depending on numbers of files per PR. Still a big no for mission criticals.

FelixFavorites :

Oh myy...kinda of self defeating the purpose of AI? That many commits to fix failures tell us how bad the code is in 15 days. 😂

elitestylingcoachus :

This is so exciting

rojas165 :

Waiting on it..

George Patience242 :

you realize that all of these will connect into one global AI known as . . . SKYNET

imaginaryman :

Mine is in hard code, and it has proved 3679 things of a total of 3773 things we wanted to prove while coding. 4 violations by counter examples, overall a 98.8% prove rate. It is 88.1% done, still working to get to 100%. Mine is scale free in many respects, and will seek to decompose when too hard or too large to ensure no getting stuck in the complexity mess llms can find themselves in. My target was a billion lines of code. 🥰

Thomas Jeffery :

okay. I've literally built this for the AI model that I'm creating, it's an moe that only loads the experts needed into RAM. anybody can run this on a 16 GB Mac mini the quality significantly is upgraded if you have 24, 32, or 64 GB unified memory system. but nevertheless, I'm running 140 billion parameters right now. matching gpt 4.0 intelligence on a m1 32 GB Mac studio and it's awesome! by the way, I'll be open sourcing the model, not the harness

Laith.priv :

I’m I’m gonna have

CapriBishopOC :

Insane

natescreates :

big milestone

Panda AI :

My Ai does that many pull requests a day because half the code is broken and has to be fixed 😂

LEALISALMAassalamwaleykonLEA :

2B