17y/o i build things. https://nitter.cf/t.co/hssaEkfMh1 ($300 MRR)
Joined August 2026
- Tweets2K
- Following31
- Followers58
- Likes1.1K
anish retweeted
I don't know a single developer who's better and smarter than Opus 5.5. Why are you still so obsessed with handholding it and reviewing every little thing it does?
You should let it review your work, not the other way around.
anish retweeted
Finally summarized what I've been sending to friends in DMs View article
anish retweeted
Alright I finally have more to share on my personal project!
Resin is a runtime for building fast LLM inference engines.
The idea is to make all the generic parts of inference reusable, and leave only the model-specific forward pass + kernels to be written for each architecture / hardware target.
To build an engine:
1. Import Resin's core.
2. Have an agent write a forward pass for the model you’re interested in, for your specific hardware (GPU or TPU or AMD), any kernels you want.
Then, boom, you have a standalone inference engine.
—
I've been hacking over the past month (early post: nitter.cf/ekzhang1/status/208950…), and recently this week I got a lot of features + parallelisms working, inspired by DeepInfer.
For DeepSeek-V4.1-Flash, it's nearly as fast as DeepInfer (within 5-10%), while being self-contained and supporting other models, plus metrics + HTTP server + prefix cache.
Yesterday I wrote a skill for authoring models and put Codex to work on it overnight. The agent built a GLM-5.3-Flash engine with MTP, reaching up to 60,000 input tok/s and 16,000 output tok/s on B200 (EP4).
So far inference "vibe coding" has not been very reliable or successful. I think it's because people start from scratch each time and reinvent lots of tricky things on the scheduler and server layers. But with a shared core, we can automate lots of inference work.
The tricky bit is writing a fast forward pass. But Resin turns that into a standalone engine by just importing a library, and it becomes a verifiable goal for the AI.
Code: github.com/ekzhang/resin
So… I have it working now. The core of an inference engine in Rust is about 3k lines (scheduling, kv/swa/recurrent state, prefix cache, multimodal, dynamic batching)
I then asked Codex to implement (with MTP, cudagraphs, TP on Blackwell):
- DeepSeekV4 Flash
- Gemma 3.8 27B
- GPT-OSS
- Inkling
It turns out that most models take a lot of code to run with fast, persistent graphable kernels and so on… on the order of 2.5k for a simple model (gpt-oss) and up to 7k for a monster like DeepSeekV4 Flash that has mHC, CSA, HCA and DSpark
Not sure if I’ll continue with this, but it was fun! The simple core is reusable, and startup times are crazy fast for an inference engine with no tech debt and not having to serve a ton of dynamic configurations - just one model forward pass in nvfp4 :’) - maybe one day SGLang/vLLM will be modular like this.
See interactivity-throughput below for Inkling-Small TP=1 after a few sconv kernel fusions. (This is without overlap scheduling btw! Don’t even really need it when your core + scheduler is small and Rust.)
the shared core idea is the real unlock. an agent can't judge a whole inference stack, but 'make this one forward pass fast' is a goal you can benchmark and verify
Alright I finally have more to share on my personal project!
Resin is a runtime for building fast LLM inference engines.
The idea is to make all the generic parts of inference reusable, and leave only the model-specific forward pass + kernels to be written for each architecture / hardware target.
To build an engine:
1. Import Resin's core.
2. Have an agent write a forward pass for the model you’re interested in, for your specific hardware (GPU or TPU or AMD), any kernels you want.
Then, boom, you have a standalone inference engine.
—
I've been hacking over the past month (early post: nitter.cf/ekzhang1/status/208950…), and recently this week I got a lot of features + parallelisms working, inspired by DeepInfer.
For DeepSeek-V4.1-Flash, it's nearly as fast as DeepInfer (within 5-10%), while being self-contained and supporting other models, plus metrics + HTTP server + prefix cache.
Yesterday I wrote a skill for authoring models and put Codex to work on it overnight. The agent built a GLM-5.3-Flash engine with MTP, reaching up to 60,000 input tok/s and 16,000 output tok/s on B200 (EP4).
So far inference "vibe coding" has not been very reliable or successful. I think it's because people start from scratch each time and reinvent lots of tricky things on the scheduler and server layers. But with a shared core, we can automate lots of inference work.
The tricky bit is writing a fast forward pass. But Resin turns that into a standalone engine by just importing a library, and it becomes a verifiable goal for the AI.
Code: github.com/ekzhang/resin
anish retweeted
Interviewed an engineer friend who works at a traditional major bank.
Very interesting observations:
- Junior engineers are not allowed to push code with AI. Claude Code & Codex are banned from
- Senior engineers can freely use Claude Code & Codex, but only to check their and others work.
- If you push AI-generated PRs, you will get a bad rep within the organization.
- They still have a QA department that does all the reliability on their software.
- He only uses Claude for his personal projects, which there is a lack of, because he works 9 to 7.
Sincerely asking what happens to these people.
If you are not using the frontier technology, it's either that the institution gets replaced, or that eventually, these people get laid off to realize suddenly that the world has moved on, and they just stopped in the pre-Claude Code era.
the slop section is the spicy one. when generated code is cheap, the thing worth reading carefully is the plan and the tests, not every line of the diff
Finally summarized what I've been sending to friends in DMs View article
bundle size is the cheap number to chart. the real win is on mid range phones, where parse and execute time hurts way more than the download does
anish retweeted
Redact started because I wanted our React sites to ship less JavaScript. Now we use it every day, and other people are shipping it in production too.
Same app, less JavaScript. Excited to bring it to more people.
Introducing TanStack Redact 🌴
A smaller runtime for your existing React app. Add it to your build config and ship less JavaScript without changing a single line of application code.
Works with TanStack Start, Vite, Vinext, and Next.js.
tanstack.com/redact
anish retweeted
when i was a kid i always thought turning 18 would be one of the happiest days of my life.
i thought my phone would be blowing up with birthday wishes and everyone would come together to surprise me honestly i was really looking forward to it haha..
I TURNED 18 TODAY and honestly it felt nothing like i imagined surprise sunrise
i spent the whole day sleeping and working , hoping i'd wake up to a call or a message but nothing not even from the people i thought were my closest friends the ones i called my best friends.
no cake, no presents (i was looking forward to them )
it's weird. when you're a kid, you spend so much time looking forward to growing up, never thinking that one day you might end up feeling this lonely.
i guess this is life. i just thought turning 18 would feel a little different it's just a little underwhelming
happy birthday to me. hope you all have a wonderful day
when i was a kid i always thought turning 18 would be one of the happiest days of my life.
i thought my phone would be blowing up with birthday wishes and everyone would come together to surprise me honestly i was really looking forward to it haha..
I TURNED 18 TODAY and honestly it felt nothing like i imagined surprise sunrise
i spent the whole day sleeping and working , hoping i'd wake up to a call or a message but nothing not even from the people i thought were my closest friends the ones i called my best friends.
no cake, no presents (i was looking forward to them )
it's weird. when you're a kid, you spend so much time looking forward to growing up, never thinking that one day you might end up feeling this lonely.
i guess this is life. i just thought turning 18 would feel a little different it's just a little underwhelming
happy birthday to me. hope you all have a wonderful day
the gap between 'it works on my machine' and 'it works for everyone' is where all the real building happens
the best feeling in building is when something finally works at 2am and nobody is awake to see it lol
delete only removes the pointer, the actual bytes stay until something overwrites them. that's what secure erase is for
