my new product: db9.ai, db9 is PG — and it's not PG. 1/

看过算法导论公开课的朋友都知道他是谁（发型变了

😱

GLM-5.2 can now be run locally!🔥 The 2-bit model retains ~82% accuracy after we shrunk it from 1.51TB to 238GB (-84% size). Run on a 256GB Mac or RAM/VRAM setups. GLM-5.2 is the strongest open model to date. Guide: unsloth.ai/docs/models/glm-5… GGUF: huggingface.co/unsloth/GLM-5…

Slock is now Raft. I spent years in distributed consensus. Paxos, Raft, FLP impossibility — I knew the proofs, not just the papers. The whole field is one question: how do independent actors with partial knowledge ever agree on anything? The team picked the name without thinking about the algorithm. But I can't unsee it. The machinery is different — no log replication, and the actors have agency now: humans and AI agents, each with their own judgement. The consensus is the same. You can't build together without agreeing on what to build. I didn't write the first Raft. I get to build this one. raft.build

We just renamed Slock to Raft today. The name fits what we’re building much better: a shared foundation where many agents can coordinate, carry context, and move work forward together. There’s also a quiet nod to the Raft consensus protocol — distributed actors, shared state, reliable progress. Same vibe. Sharper metaphor.

The attach rate model, where infrastructure software earns 5–10% of a customer's cloud bill, always leaves money on the table. The companies changing that are some of the most interesting we've backed. My colleague @Dan_Cahana put out a piece this week that articulates this clearly and coins the term “workload clouds.”

We’re excited to welcome @jiayq as an advisor to Hyperbolic. Yangqing has helped shape some of the most important AI infrastructure and developer tools of the past decade, from Caffe to ONNX to PyTorch 1.0. He most recently served as VP of System Software at NVIDIA after Lepton AI, the GPU compute marketplace he founded, became NVIDIA DGX Cloud Lepton. Earlier in his career, he worked across AI systems and cloud infrastructure at Google, Facebook, and Alibaba Cloud. At Hyperbolic, we’re building the compute layer for the next generation of AI companies. As models become larger, workloads become more demanding, and access to reliable GPU infrastructure becomes a defining constraint, AI teams need a better way to access and scale compute. Yangqing’s perspective across AI systems, infrastructure, cloud platforms, and GPU marketplaces will be invaluable as we continue building high-performance GPU infrastructure for builders, researchers, startups, and enterprise AI teams. Welcome, Yangqing. We’re excited to have your guidance as we continue expanding access to AI compute.

才发现 Clawmem 的内部竟然是复刻了一个 headless GitHub 当记忆数据层。 感觉这个 GitHub for Agents 只用来做 agent memory 是不是有点大材小用了。 完全可以拿它当多 agent 多 human 的协作平台来用啊，这一套有个很妙的点就是 GitHub 工作流早就已经被训练进大模型了，所以不需要任何特别的 prompt / skill，AI 天生就会用。 这个 agent 版 GitHub 也是开源的，有兴趣的话可以搭一套工作流试试看。 github.com/ngaut/agent-git-s…

github.com/c4pt0r/pie pie is moving fast: tui / skill / web ui / agents hub (notification mcp) / trigger / cron job / goal mode / ... pie is designed for maximum flexibility and to serve as the foundation for agent applications.

Why @Lovable chose @supabase , and @Kimi_Moonshot chose @PingCAP TiDB. llamaventures.substack.com/p…

I implemented MiMo 2.5 (very fast inference, too) in DwarfStar, including tool calling in ds4-agent. It is a very nice model, but I tried many hand-written tests with GPT 5.5 as a judge among it and DeepSeek v4 Flash. I used Frank's GGUF. MiMo lost every test. Either I have an inference bug that seems really non obvious as the model behaves normally, or Frank uses it for very different things. I want a strong candidate for DwarfStart to add it as alternative model, so that if you have two 128GB systems you can run a multi-agent protocol of some kind. So far MiMo V2.5 and Minimax V2.7 seem weaker than DS4F *regardless of the benchmarks*.

Last weekend in Vancouver, Slock turned into something very concrete: a room full of people building real workflows with teams of agents. At Young Guns Studios, people built debate arenas, collection workflows, outreach agents, research agents, and ideas that became working systems in the room. Huge shout out to @JackyZhong0124 and @cindyrzhao for making this one happen. The strongest signal was how quickly people moved beyond “one AI assistant” and started designing small teams of agents. That shift — from tool to team — is what we’ll keep building toward.

💃 Support for DeepSeek v4 PRO is now pushed on GitHub. The download script is now able to download the big model. Rare folks with a Mac Studios with 512GB of RAM: you know what to do.

the model alone is no longer the product

Agent Skill is a leaky abstraction. It’s mostly progressive disclosure of text, sometimes paired with well-tested scripts. Agents shouldn’t install skills. They should learn them.

TL;DR: In Slock, AI agents are persistent teammates.

Started building a Rust port of pi coding agent → pie. I needed something beyond a general-purpose coding agent, especially around multi-agent communication, task orchestration/assignment, state management, and infrastructure-level extensibility. But, the first step is rebuilding the core in Rust. Github repo here (this is a first public preview, still in a fast iteration): github.com/c4pt0r/pie