Founder - Deus Experiments • Building apprentice models that grow with you
Joined July 2014
- Tweets15.2K
- Following1K
- Followers12.7K
- Likes53.3K
Pinned Tweet
Introducing Chantless: Chapter 1 — Apprentice
A powerful tool makes you free, allows you to express yourself, and learns and grows with you.
This is what we're building with Chantless. A technology that lets you shape and evolve models alongside you.
this looks really nice.
eager to try it out in my own research next.
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: xitter.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
useful counterpoint to this in comments, i still think instability and timing errors with resets/long workloads can pop up but this mainly enables full rate fp32 tensor accumulation.
and this means it benefits prefill and looks like something you want to try for long context inference for sure.
xitter.cf/worstisaka/status/2114…
i can explain in long paras why we used to put dynamic perf limits in firmware anywhere but this image should suffice.
these rate limits are hardcoded for thermal damages and timing signals to finish before next clock edge.
but everything is just “numbers go up” woah and it’s hard to not say anything on this.
posting has power crept into becoming a basic skill now. attention being the only scarce resource on the map.
distribution looks super close to discovering a broken strat then farming, abuse and quick nerf.
pros vs sweats vs casuals dynamic here soon.
rip playing just for fun.
perhaps twitter is not a bad place to hang even with automation, never expected to see a control theory space.
i really enjoy being here and working on stupid to important spectrum of things off of my mood and just talking about whatever and still benefiting off of this somehow.
this is like that one universe (Earth 1610B) where Peter says he actually loves being Spiderman.
why is opus 5.5 so good at rock paper scissors? or do i just suck at it?
i have lost 2-4 in several best of 3/5 matches so far.
i’ve successfully improved on my “never wishes or texts on occasions” complaint from my relatives/friends by using agents.
my solution is simply - being part of minimal group chats that cover the humans i care about and tracking them with my agents. i get an hourly summary of whatsapp/discord/twt chats when anything that needs a social presence or personal pops up. i review it when i’m free and tell it to respond in appropriate places in my exact words.
waiting for full duplex voice cloning (fdvc) to be available soon so that i can extend this to personalised voice notes.
calls would still be difficult with fdvc setup because people may start asking details about how am i doing recently or “tell me more” variants so i’ll keep that option as <escalate to human> events where i will engage if really crucial.
for the rare occasions where someone may spot an issue in my voice notes, i have plans to tell them later that i had a cold so my voice was not perfect. anything beyond this is probably not going to be possible soon.
i can explain in long paras why we used to put dynamic perf limits in firmware anywhere but this image should suffice.
these rate limits are hardcoded for thermal damages and timing signals to finish before next clock edge.
but everything is just “numbers go up” woah and it’s hard to not say anything on this.
💥卧槽！6年前的3090被“破解”了！Tensor算力直接翻倍，官方锁了一半的性能被开源一键释放！
NVIDIA长期把30系显卡的FP16/BF16 Tensor指令发射速率限制在专业卡一半左右。
开源项目 gpu-unlock 用一个Linux内核模块，在驱动加载前改寄存器解锁，不刷BIOS、不改驱动、更新无忧。
牛逼之处：
🔹3090 Ti：BF16 87.5→157 TFLOPS，TF32 42.6→82.6 TFLOPS
🔹3080：FP16/BF16 约65→110–113 TFLOPS
🔹vLLM长上下文prefill快33–36%，首字延迟降25%
🔹结果比特一致，社区已在3070/3080/3090复现
关键步骤（仅Linux）：
1️⃣git clone github.com/Zagitta/gpu-unloc… && cd gpu-unlock && make install
2️⃣内核参数加 gpu_unlock.targets=all，关Secure Boot
3️⃣更新引导并重启
4️⃣journalctl -k | grep gpu_unlock 看到 unlocked 即成功
GitHub：github.com/Zagitta/gpu-unloc…
#RTX3090 #NVIDIA #AI算力 #开源 #GPU解锁 #本地AI #vLLM #黄仁勋
to all my followers who believe in this, sincerely - it's bullshit.
this is an era of unprecedented technology and with it comes unknown challenges, it is like a new world map in a game where only paid promotional beta testers have been before.
what is the value of discovering a counterfactual now when people have stopped reading everything?
how reliable is this tech where you chat with a bot and it hallucinates 3 times a day but you never know on what? what is the cost of discovering such bugs late for your projects?
what is the cost of discovering problems/bugs when entire world is wearing the same glasses and everyone has same blindspots?
"there is no technical moat" - yeah, for you.
ofc you would think that if you are someone who has always just wanted to get by their day and do nothing beyond what is forced on you, then yeah, sure.
but if you tell a guy like me "everyone is offloading to the same machine to think, using the same overabstracted superficially verified systems, where agentic PR count is a porn to be consumed by those who think they are doing something". if you tell a guy like me that, i would say there is now an opportunity to be hercules amongst men. there is an opportunity to really use this tech around you to imitate what fabrice bellard could have pulled off in your niche if he cared as much as you.
there is an insane opportunity to be unlimited, your moat is starting your morning where others have stopped caring about something.
a great example of discovering a blindspot
xitter.cf/rosinality/status/2113…
if you thought operating systems importance was a rediscovery last year, watch as network design now becomes more crucial than anything else in multi agent systems.
incredible, a 3T frontier model gets open sourced despite the most intense pushback from everywhere.
Introducing Kimi K3.1 Pro: our most capable 3T-class open-weight model.
🔹 1M-token context, 2.8T total / 104B active params (MoE), Multimodal.
🔹 Full-project coding: games in Unreal, Unity & Godot.
🔹 Excels at parametric 3D modeling in Blender & CAD.
🔹 More reliable at finance work: Accurate modeling with hallucination down by 65%.
🔹 Matching the strongest proprietary models across coding, agentic and knowledge benchmarks.
Kimi K3.1 Pro and K3.1 Pro Highspeed (2–3× faster) are now live on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API.
Open weights by Nov 4, 2026.
🔗 API: platform.kimi.ai
🔗 Tech blog: kimi.ai/blog/kimi-k3-1-pro
