Artist - Entrepreneur "Art always wins" | "The System is the ART"
Metaverse
Joined February 2022
- Tweets17.9K
- Following1.7K
- Followers4K
- Likes25.8K
We can keep bechmarking models, building pagodas, and whatnot, but the truth is that there is no better model than GLM5.3-flash to be run on a "reasonable" local set up.
Deepseek coming in very close, love their models as well, but GLM has been delivering consistent results for me day in and day out, and using it almost always with Max thinking by default.
Can wait for @Zai_org next release.
Now gave it a go with Deepseek v4.1 Flash 1017.
To me the clear winner on all fronts: front end, physics simulation and even added some pre-determined tests.
All three html files and the prompt open and available in the repo below.
I asked now GLM5.3-flash (@Zai_org) to create the same 3D Sailboat Aerodynamics Lab.
Solid result, specially given the difference in number of parameters: Kimi3.1Pro is 2.8T and GLM5.3Flash is 320B.
I added this one to the repo as well.
Which model do you want me to try next?
A thought:
Right now agents are yours.
In the future, agents will be like humans. Will have developed some skills, knowledge, etc. and we will be able to outsource them to perform tasks for companies or individuals.
We need to build the LinkedIn of AI Agents, or does it already exist? I have some wild ideas for this.
I asked now GLM5.3-flash (@Zai_org) to create the same 3D Sailboat Aerodynamics Lab.
Solid result, specially given the difference in number of parameters: Kimi3.1Pro is 2.8T and GLM5.3Flash is 320B.
I added this one to the repo as well.
Which model do you want me to try next?
I asked Kimi3.1 Pro (@Kimi_Moonshot) to create a 3D Sailboat Aerodynamics Lab.
Used the Kimi API and Hermes (@NousResearch) as harness.
Generating this cost around 1.7 USD using the API.
The results 👇
I asked Kimi3.1 Pro (@Kimi_Moonshot) to create a 3D Sailboat Aerodynamics Lab.
Used the Kimi API and Hermes (@NousResearch) as harness.
Generating this cost around 1.7 USD using the API.
The results 👇
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
Two of may favourite labs pushing hard: @Zai_org and @deepseek_ai
I am very positive about the future and what the capabilities this gives is as a civilization.
🚨重磅！智谱Zai创始人唐杰最新演讲：GLM 5.4/5.5将是RSI产物！AI已从摩尔定律跨入RSI，国产模型真正机会来了！
这次成都CNCC演讲直接点明：从“对话”到“干活”，RSI才是国产模型的核心机遇。
唐杰演讲要点：
🔹范式切换：从Chat聊天 → Agent真正“干活”，能规划、执行、验证长程任务。
🔹从摩尔定律到RSI：过去参数每年10倍增长，现在是模型自动生成环境 → 构造可验证reward → 训练 → 能力提升 → 再生成更难任务，形成闭环正反馈。
🔹AGI路径：点能力 → 线(s尺度) → 面(t任务)。警惕Scaling bias，小实验结论到大尺度有系统性偏差。
🔹国模已落地：智谱用GLM驱动的Agent，在10万颗国产卡上优化自身推理系统，两周吞吐提升约3倍。模型优化系统，系统服务模型，最小RSI闭环已跑通。
🔹直指下一代：GLM 5.4/5.5明确走向fully self-training，参数冲1万亿+，从预训练到后训练自己训练自己。
时隔4年，机遇彻底变了，人仍掌握裁判，但闭环已经转起来。
#CNCC #CCF #GLM55 #智谱 #唐杰 #RSI #AGI #国产模型 #从对话到干活 #Zai #Zcode
My carpet fully onchain!
Carpetdiem is now 100% on-chain.
When Carpetdiem launched in March 2023, the 555 tokens lived on an immutable ERC721SeaDrop contract (0x203B...321f), while the generative growth engine (carpet-diem.js), p5.js runtime, and token seeds were served through an external IPFS gateway (ipfs. carpetdiem. art) - and the preview images were low-resolution 350 × 490 blurred thumbnails.
Without minting a second token, wrapping, or touching the audited SeaDrop contract, I moved the entire generative architecture onto Ethereum L1:
All 555 original 32-byte seeds and the complete 842,381-byte runtime (carpet-diem.js + [email protected]) are stored directly in Ethereum contract storage across 36 SSTORE2 chunks via CarpetdiemPlayer.sol and permanently locked.
Self-contained data:text/html;base64 animation_url: Every token's metadata now embeds a standalone HTML/RPC viewer that reads the seed and runtime directly from Ethereum L1 and weaves all ~144,000+ symmetric squares live in the browser (1144 × 1600 canvas + infinite-zoom 8485 × 11879 SVG plot export).
Full-resolution crisp stills: Replaced the blurred 350 × 490 previews with exact pixel-faithful 1144 × 1600 PNG renders for all 555 carpets.
Zero new tokens. Zero wrapping. Zero external servers required to reconstruct the art.
