Jonathan Chang
@ChangJonathanC
Followers
1K
Following
13K
Media
496
Statuses
2K
ML/AI Engineer, building https://t.co/uEbfxzF7jm
Taiwan
Joined May 2020
while we wait for gpt-5 to drop. Here is a flex attention tutorial for building a < 1000 LoC vllm from scratch.
jonathanc.net
PyTorch FlexAttention tutorial: Building a minimal vLLM-style inference engine from scratch with paged attention
9
39
399
what's the simplest task you have that none of the sota llm can solve reliably?.
1
0
0
Pantheon.
I’m excited to share that @Starcloud_Inc1 is partnering with @Google Cloud to be the first to run a version of Gemini on a satellite. "Starcloud, a start-up that is building a space data center, will soon launch a satellite equipped with NVIDIA graphics processing unit (GPU)
0
0
3
Imagine the money you can save by post training 4.1 mini to talk like 4o and calling it gpt-5 in the chatgpt app
0
0
2
lol they really put juice number to the system prompt
gpt-5-high is only available through the API and has the highest juice: 200. Even on the $200/month plan, thinking and thinking-pro are limited to a juice of 128. Consumers are drunk on the juice and have just caught sight of some untapped kegs.
1
0
2
hey @aidan_mclau , i think it's weird for model to say "Got it" after i've waited for 3 mins for the response. the model should just answer the question
0
0
5
imagine the dog guy is sama's alt, he sent himself a GPU so he could get more GPUs.
0
0
3
did some vibe refactor with codex cli + gpt-5. i like how little i need to steer it compared to claude code. the session would have costs 3.76, so plus tier is still a pretty good deal
1
0
3
Bro is vibe coding using X
I have no idea what's involved, so this might be the easiest money ever or way under target, but I'll pay a $5,000 bounty to anyone who can claim commit credit for getting this shipped in chromium.
0
0
1
I'm going to start calling it o4, it's shorter than gpt-5-thinking and so people don't confuse it with the router or chat version.
1
0
1