I spent $1.5M building our office after raising a seed round. My co-founder thought I was crazy. Here's what changed his mind... 𝐓𝐡𝐞 𝐂𝐨𝐧𝐭𝐞𝐱𝐭: After our seed round, I looked at our team. Mostly immigrants. Working 6-day weeks. Building something incredibly hard. The office wasn't just where they worked. It was becoming their home. So I made a bet, what if we actually designed for that? The requirements I gave our real estate agent: - Shower (for ocean swims between meetings) - as close to the beach as possible, ability to quickly go surfing/kiting etc. - Big enough kitchen for a chef - Room for an actual sauna People thought I was building a vacation house, I thought: I am building a place worth the sacrifice. 𝐇𝐞𝐫𝐞'𝐬 𝐡𝐨𝐰 𝐈 𝐦𝐚𝐝𝐞 𝐢𝐭 𝐰𝐨𝐫𝐤: - Found one of SF's best real estate lawyers. Negotiated hard. - Negotiated Tenant Improvements + First year for free - Effective cost: $250k (not $1.5M) Then I was extremely prescriptive with design and construction. No endless back-and-forth. I drew what I wanted. Told them to build it. Cut iteration time by 80%. 𝐖𝐡𝐚𝐭 𝐰𝐞 𝐛𝐮𝐢𝐥𝐭: - Nordic vibes (keeping our European souls) - Industrial kitchen - Sauna room (yes, like our product: sauna.ai) - Ocean access - Space that feels like home Conclusions: - This "expensive" decision already paid for itself. - In SF, recruiters charge $100k per engineer. We've closed multiple hires -because candidates walked in and said: "I want to work here." But the real ROI isn't only financial. It's this: - We do Friday AMA as a BBQs on the beach. - People actually use the surfboards. - The team's lifestyle supports the intensity of the work. EVERYONE WANTS IN, doesnt matter if events, hiring or using the space as coworking (@bertie_ai and I open it up for our portfolio companies) My co-founder's response after 3 months: "You were right." Some founders optimize for low burn rate. I optimize for: Can great people sustain this pace for years? Because great companies aren't built in one sprint. They're built by people who can go the distance. We're hiring: wordware.ai/careers (Comment if you want intros to our real estate agent, lawyers or construction team - happy to connect)

Eye of Nature by Dan Hillier

🚀 "Quantization is not a compromise — it's the next paradigm." After K2-Thinking's release, many developers have been curious about its native INT4 quantization format. 刘少伟, infra engineer at @Kimi_Moonshot and Zhihu contributor, shares an insider's view on why this choice matters — and why quantization today isn't just about sacrificing precision for speed. 💡 Key idea In the context of LLMs, quantization is no longer a trade-off. With the evolution of param-scaling and test-time-scaling, native low-bit quantization will become a standard paradigm for large model training. 🤔 Why Low-bit Quantization Matters In modern LLM inference, there are two distinct optimization goals: • High throughput (cost-oriented): maximize GPU utilization via large batch sizes. • Low latency (user-oriented): minimize per-query response time. For Kimi-K2's MoE structure (with 1/48 sparsity), decoding is memory-bound — the smaller the model weights, the faster the compute. FP8 weights (≈1 TB) already hit the limit of what a single high-speed interconnect GPU node can handle. ⚠️ By switching to W4A16, latency drops sharply while maintaining quality — a perfect fit for low-latency inference. 🔍 Why QAT over PTQ Post-training quantization (PTQ) worked well for shorter generations, but failed in longer reasoning chains: • Error accumulation during long decoding degraded precision. • Dependence on calibration data caused "expert distortion" in sparse MoE layers. ‼️Thus, K2-Thinking adopted QAT for minimal loss and more stable long-context reasoning. 🧠 How it works K2-Thinking uses a weight-only QAT with fake quantization + STE (straight-through estimator). The pipeline was fully integrated in just days — from QAT training → INT4 inference → RL rollout — enabling near lossless results without extra tokens or retraining. ⚡ INT4's hidden advantage in RL Few people mention this: native INT4 doesn't just speed up inference — it accelerates RL training itself. Because RL rollouts often suffer from "long-tail" inefficiency, INT4's low-latency profile makes those stages much faster. In practice, each RL iteration runs 10-20% faster end-to-end. Moreover, quantized RL brings stability: smaller representational space reduces accumulation error, improving learning robustness. 🔩 Why INT4, not MXFP4 Kimi chose INT4 over "fancier" MXFP4/NVFP4 to better support non-Blackwell GPUs, with strong existing kernel support (e.g., Marlin). At a quant scale of 1×32, INT4 matches FP4 formats in expressiveness while being more hardware-adaptable. 🧭 Looking forward W4A16 is just the beginning — W4A8 and even W4A4 are on the horizon. As new chips roll out with FP4-native operators, Kimi's quantization path will continue evolving. "In the LLM age, quantization stands alongside SOTA and Frontier. It's not a patch — it's how we'll reach the frontier faster." 📖 Full article (in Chinese): zhihu.com/question/196955840… #KimiK2Thinking #INT4 #Quantization #LLM #Infra #RLHF

The app on the left is github.com/steipete/tmuxwatc… - check up on your subagents in codex & background tasks.

I open sourced some of the tools I wrote lately to keep codex in check. github.com/steipete/agent-sc… runner: auto-applies timeouts to terminal commands, moves long-running tasks into tmux. committer: dead-simple state-free committing so multiple agents don't run across with add/commit stages. docs: lists all md files in /docs and motivates the agent to read docs as-needed. You'll need to tweak them for your own setup, don't blindly copy them. Point your agents at them and discuss to build your own workflow.