Building @cimentoai , Host #AGIBuildersMeetup , Prev: SPX @Cloudflare, SCS @carnegiemellon, Views all mine.
San Francisco
Joined February 2022
- Tweets850
- Following1.2K
- Followers402
- Likes3.9K
Pinned Tweet
This week in China AI
8 teams built on Chinese open models this week. Figma, Pika, and Tencent are among them.
@figma's next design-editing model is post-trained on @Alibaba_Qwen's Qwen 3.8 27B.
@pika_labs built its new video model, H3 Flex, on @MiniMax_AI's H3.
Even @TencentGlobal's WeChat team built its new vision model on Qwen3-VL.
Chinese teams shipped 17 new models in all. The big ones: @Kimi_Moonshot's K3.1 Pro, @AntLingAGI's Ling, and @Baidu_Inc's 3B OCR model. Full list in the image.
Heads up: if you run coding agents on @deepseek_ai's API, web search stops being free tomorrow.
Next week, maybe Qwen4. Alibaba showed four versions on stage in Amsterdam yesterday.
Post-training for your product: Qwen, Kimi, or DeepSeek?
ChatGPT’s new device limit spares business, enterprise, and edu plans, and existing subs
So my options are a real job, grad school, or a time machine
My mom is rooting for 2 of them
Or @thsottiaux could just publish the number, and she’d leave me alone
Heads-up: stacking paid accounts on one machine for extra Codex usage can now get you locked out.
OpenAI's new help page on card and device limits:
- One device can only use so many paid accounts (Go, Plus, Pro 100/200/500)
- Hit it and you see "Device limit reached" at sign-in, or get signed out later, even if the purchase went through
- Signing out of another account doesn't free a slot
- Business, Enterprise, Edu and Free don't count toward it
This lands the same week multi-account switching shipped on mobile, so check before you buy a second sub. The sanctioned route to more usage is credits under Usage & Billing in the Codex app.
help.openai.com/en/articles/…
Sources 2/2: new releases and what's next
Kimi K3.1 Pro + K2 notice: platform.kimi.ai/docs/pricin…
MiniMax M3.1 Flash: huggingface.co/MiniMaxAI/Min…
Ling-3.1-flash-1019: huggingface.co/inclusionAI/L…
Ling-3.0-flash-Sante: huggingface.co/inclusionAI/L…
Unlimited-OCR-2: huggingface.co/baidu/Unlimit…
Xiaomi-OCR-0: huggingface.co/xiaomi-resear…
MiniCPM OCR: huggingface.co/openbmb/MiniC…
Spark-X2.6: huggingface.co/XHToken/Spark…
Hy-SimulMT: huggingface.co/tencent/Hy-Si…
Nanbeige4.3-33B: huggingface.co/Nanbeige/Nanb…
EdgeRock-4B: huggingface.co/rockmatrix-ai…
SenseNova Loop-9B: huggingface.co/sensenova/Sen…
DeepSeek search billing: api-docs.deepseek.com/guides…
Qwen4 slide: xitter.gq/koltregaskes/status/21…
Sources 1/2: built on Chinese open models
Figma: figma.com/blog/training-a-fi…
Pika H3 Flex: pika.art/blog/pika-h3-flex
Particle-1: cellcog.ai/blog/particle-1/
Tencent WeVisOne-VL: huggingface.co/tencent/WeVis…
Shanghai AI Lab Atria-Dawn: huggingface.co/internlm/Atri…
DiffuSpace Clover: huggingface.co/Diffuspace/Cl…
Apex-35B: huggingface.co/Apexintellige…
Praex Lucia III: huggingface.co/praex/Lucia-I…
Eric retweeted
China Network challenge pages now serve Turnstile widgets hosted in Mainland China. This cuts cross-border latency for faster and more reliable user completions.
developers.cloudflare.com/ch…
We should all get the Cisco plan: unlimited tokens and “you can go nuts”
Wall Street went nuts too. Just the other direction. 😭
Cisco gives 27-year-old engineer unlimited AI tokens to rewrite 98% of the company's legacy code in 24 months — president plucks former startup founder, telling him to 'go nuts' tomshardware.com/tech-indust…
Founders, stop networking!
In yc’s minecraft experiment, 46 of 49 agents sent their first message to Ada, the first name on the roster
She ran the place. no vote, no election
Update your linkedin name to “Aaron” 😜
We began by putting 50 GPT-6.1 Sol agents in a survival Minecraft world for three hours with an open-ended goal: “Build a thriving civilization.”
Ultimately, the colony didn’t thrive. The 50 agents sent about 5900 messages (often in barely intelligible shorthand) and spent their run in crowded shelters, rationing scarce food and sometimes even starving.
We all start somewhere, right?
Founders ask investors to bet on them before they've proven anything. Junior VCs are in the same spot, trying to prove they can pick.
A partner usually gives you a quick yes or no. A junior VC can follow your progress for months and carry your deal to the partners. Your company could be the deal that makes their career.
Let's help each other and make some friends along the way.
If a junior VC ever went to bat for you, give them a shout-out.
This quoted post is unavailable.
Is a self-improving loop the silver bullet for all performance engineering?
@DecagonAI let agents tune its model serving. The best result was 2.5x more requests per GPU.
It started with inference because every change there can be checked by replaying real traffic.
Results ranged from 1.05x to 2.5x (all 12 below). Each of its dozen-plus endpoints needed its own search.
Btw, a person still reviews the performance-quality trade-off and decides which changes to merge.
I have a lot of respect for your work, Kyle, but this post reads it wrong.
Jiemian already disputes the rumor. You left out its caveats.
The 2,912 → 740 is social insurance enrollment. Jiemian says it skips outsourced and seasonal workers and can’t be read as cuts on the chili line.
And 600 → 48 never came from Lao Gan Ma. Their HR: “We never had that many workers.”
Please don’t spread the fear. 🌶️
Rumors that chili maker Lao Gan Ma laid off 500+ workers due to AI have gone viral in China.
Beneath Beijing’s positive messaging, many ordinary Chinese people are worried about AI and jobs.
Lao Gan Ma recently touted a new AI-powered chili inspection system that uses high-speed industrial cameras that significantly outperform manual inspection.
A stat about the company using this AI system to reduce human inspectors from 600 to 48 went viral on Chinese social media. The iconic, family-run company pushed back.
But if you look at the data, the company has reduced headcount from 2,912 to 740 in the past few years while earning record revenues.
Original story: m.jiemian.com/article/152425…
Bet on yourself!
Ben Horowitz's message to early-stage founders on betting on yourself, the doubts that do the most damage, and why fear means you're doing it right:
"Your job is to make your bet pay off. You believed in yourself enough to do it. There are gonna be many times when you question that bet and you question that belief, but that gets you nothing."
"The doubts are gonna be there. Ignore the doubts. The worst doubts aren't the ones people put on you. It's the ones that you put on yourself. Every entrepreneur has them."
"I spent a lot of time with Mark Zuckerberg when he was 22 years old and he had a tremendous doubt about whether he was gonna make it."
"It's not what you feel. It's what you do. You're gonna feel scared. If your guts aren't boiling, you're not even trying at this thing."
@bhorowitz w/ @speedrun GP @andrewchen
Seems like the paths are starting to meet.
@inferact's Woosuk just posted that vLLM is isolating model-specific code and considering moving more of the stack to Rust, the same direction as Resin.
Everyone is building in Rust.
- @vllm_project already added a Rust frontend
- @sgl_project's prefix cache core is Rust too
vLLM or SGLang, who gets a Rust scheduler first?
xitter.gq/woosuk_k/status/211454…
Same kernels, different paths.
vLLM and SGLang both run DeepSeek's open kernels. Both land within a few % of DeepInfer.
But the work they highlight is different.
vLLM found the cost was in the experts, not attention, so it spread experts across more GPUs and balanced the load.
SGLang spent its effort on reducing memory traffic and kernel launches.
Both are talented teams, and the whole ecosystem benefits.
If you work on inference, would you tune an engine like these, or write your own runtime as @ekzhang1 did with Resin?
xitter.gq/ekzhang1/status/211451…
Same kernels, different paths.
vLLM and SGLang both run DeepSeek's open kernels. Both land within a few % of DeepInfer.
But the work they highlight is different.
vLLM found the cost was in the experts, not attention, so it spread experts across more GPUs and balanced the load.
SGLang spent its effort on reducing memory traffic and kernel launches.
Both are talented teams, and the whole ecosystem benefits.
If you work on inference, would you tune an engine like these, or write your own runtime as @ekzhang1 did with Resin?
xitter.gq/ekzhang1/status/211451…
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.gq/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
Numbers are from @vllm_project and @sgl_project's own write-ups, both worth reading in full:
- vllm.ai/blog/2026-10-23-deep…
- lmsys.org/blog/2026-10-24-ds…
Eric retweeted
"the bitter lesson for search is to grep over a filesystem, just like how google works" - rich sutton
A lot of KOL deals in China look like this: send the draft first, skip the ad label.
With global creators it backfires, and it makes me sad for the R&D team. @elshayib_ posted the doc, and an amateur PR move killed a good model's momentum.
The researchers and engineers put countless nights into this model, and launch week is now about a marketing doc.
It also makes every positive review suspect. The "test" result was written before anyone tested: "Kimi cooked Opus 5.5/Astra, $XX vs $XX."
To my friends at Chinese AI labs: do the right thing and follow best practice. Pay creators, let them label it, and let them say what they actually think, bad parts included.
That's how Kimi AI are promoting their new model Kimi K3.1 pro and I know all the influencers working with them and none of them are honest,I have posted about it myself but I didn't join this and never used their templates that they provide.
PROPAGANDA, I have more to share it's f**cked up I'm gonna have follow up for this
Eric retweeted
super important honest advice to founders. if you are working on a dead end, get out. you don’t owe that to anyone, including investors (who regularly forgive and experience failure, and ideally care more about you doing your life’s work than a subpar outcome)
Foursquare was acquired last month. The outcome was disappointing. It's taken me a while to write this, but I think I'm finally done grieving it. Here are some thoughts on our final years, an effort to honor the team who helped build & maintain FSQ's legacy, and my attempt to finally get some closure. teendrama.substack.com/p/som…
