KaneYip.eth

The bottlenecks in AI are moving beyond the models. We’re excited about startups building the hardware, infra, financial systems, and specialized software to bring AI into the real economy, from factories and data centers to government and space. ycombinator.com/rfs

At the Cerebral Valley AI Summit, our curated discussion groups give founders, investors, and AI leaders a chance to compare notes, debate ideas, and talk through the challenges they’re working on. Apply now to join us on November 12: cerebralvalley.com/apply Meet this year's discussion leaders below👇🧵

Proud to welcome our newest sponsors to SGLang Summit. In just one week since our public announcement, our sponsor lineup has grown from 11 to 36. Thank you to every sponsor helping bring this community together. Join 2,000+ engineers, researchers, and builders from across the AI ecosystem for technical sessions, hands-on workshops, and the GPU Expo. Hear from Lip-Bu Tan, Lilian Weng, Soumith Chintala, Aravind Srinivas, and more. Meet the teams building across AI hardware, infrastructure, models, and applications. 🗓️ November 12–13, 2026 🌉 Fort Mason, San Francisco 🎟️ Get your tickets: sglang.io/summit

I resigned from Anthropic today. I spent the last three years doing pretraining research at both OpenAI and Anthropic. Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives. More thoughts below.

Humans are the minority user of AI Agents burn nearly 5x the tokens people do, up 14x since February Charts of the Week: a16z.news/p/charts-of-the-we…

The Most Valuable $20 You’ll Spend This Decade This chart stopped me in my tracks. It visualizes the current state of AI adoption across the world. Every dot represents roughly 820,000 people. Here’s the reality: ⬛ 78% of humanity (6.4 billion people) have never used AI. 🟩 21% (1.7 billion people) use free AI tools like ChatGPT. 🟨 Only 1.1% (around 90 million people) actually pay for AI. 🟥 And just 0.15% (around 12 million people) are using AI agents to automate work. Let that sink in. For less than the cost of a few coffees each week, you’re no longer competing with 99% of the world. You’re joining a small group investing in a productivity multiplier. But here’s the bigger insight… The real divide isn’t between people who use AI and those who don’t. It’s between those who: • Ask AI questions. • Collaborate with AI to think better. • Build AI systems and agents that work 24/7. The next decade won’t simply reward those who work harder. It will reward those who learn to orchestrate intelligence. Just as the internet created a new generation of winners, AI is creating another. The tools are becoming accessible to everyone. The real competitive advantage is knowing how to use them strategically. The question isn’t whether AI will transform your industry. The question is whether you’ll be part of the 0.15% building with AI or the 99.85% trying to catch up. 🚀 Want to move beyond simply using AI? At 10xme, we’re helping executives, founders, and professionals build practical AI workflows, AI agents, and personal AI operating systems that create a real competitive advantage. 🎁 Start with our FREE AI Diagnostic and Executive AI Newsletter at 👉 10xme.biz Because the future belongs to those who learn to work with AI, not compete against it.

Cursor is now part of @SpaceX. Today, we have officially closed our acquisition. We will join the @spacexai team to help make Grok the world's most useful AI and improve Grok Build, Grok Bot, Grok API, Cursor, and more. SpaceX has built some of the most inspiring and impressive technology in the world, and we’re grateful for the opportunity to become part of such a special company. Onwards.

Some observations on Kimi: 1. It's a very good model! I don't think its performance can be explained away by distillation or anything like that. In agentic coding sessions, it seems pretty much on par with the best public models of Q1 2026. In my fairly limited use, it also seemed very token hungry. It's not obvious to me that this model is actually that cheap to run. 2. I am personally surprised the Chinese state continues to allow the open sourcing of models this good, given potential risks. To be clear, I *myself* might be fine with models presenting this level of marginal risk being open weight, but I am surprised that China is fine with it. I suspect the reason they are is 75% explained by strategic blindness/lack of AGI-pilledness (the CCP is very Yann Lecun-y in its views of AI). The other 25% or so is their lack of compute for customer inference (making China's open-weight strategy an unintended byproduct of US export controls) and the normal Chinese strategy of aggressive exports. For the companies, as opposed to the government, the decision to open source is partially ideological and partially because they are behind, and they know that very few people would pay for sub-frontier models from China. 3. Open-weight models are inherently decelerationist, and I'm continually surprised to see the so-called "accelerationists" so excited about open-weight models. I suspect the reason they are is that they know open-weight models are effectively ungovernable, and they simply like the overall cloak of ungovernability open-weight models create over the whole of AI. It's not a bad strategy; it reminds me of James Scott's recounting of the hill people in "the art of not being governed." Still, in the end, open-weight models deter further AI capex. 4. One probable outcome of an open-weight-model-dominant world is full AI communism, which is precisely what China proposes: rather than a market product, AI is a "public good" which will ultimately be provided by the state as a kind of "digital public infrastructure." This future strikes me as a dystopian hellscape, but I've never met an open-weight models advocate who doesn't ultimately concede this is where things end. You'd be surprised how many 'accelerationists' lobbied me, while I was in government, to support an eleven or twelve-figure federally funded data center so that startups could train models at a subsidy and then give them away for free. There was no other way for AI to progress, they said. Perhaps this is the logical end state of things. Nonetheless, I find myself surprised to see supposed accelerationists excited about such an outcome. I think many of them just don't know what they're doing. Many accelerationists do not view the creation and serving of frontier models as a legitimate business. 5. I would guess that the Trump Administration will at some point realize that their best strategy here would be to create large amounts of regulatory risk around the use of open-weight Chinese models. You don't need to "ban open source" (one of the dumber motifs of AI policy discussion). You just need to direct every agency to issue soft law that creates FUD. "A Federal Reserve Advisory Bulletin found that there may be backdoors in Chinese AI models." It needn't be that well justified. You just create enough regulatory risk that every regulated enterprise backs off. You probably don't want to create so much regulatory risk that you scare off the hyperscalers from serving Chinese models; this will just drive startups to sketchier providers. There's a happy middle ground here. I'd assume they will do some version of this. 6. It's probably true that open-weight models of this capability make the world a bit more dangerous, but not so much more that you'll really notice. At some point the models will be capable enough that you will notice. "A nonliving, invisible, dangerous, and infinitely self-replicating agent escaped from a Chinese lab," you say? Color me shocked.

Google, Microsoft, OpenAI, Cursor, and Vercel just standardized how to package MCP and skills as a plugin for AI Agents. write a skill once and every one of them can now load it. there are two ways to give a coding agent new abilities today. a skill is a folder of written instructions the agent reads when it needs them, and an MCP server is a small program that connects the agent to a tool, an API, or a database. both already work across agents. what has never worked across agents is the package you ship them in. every tool picked its own folder names, its own settings file, and its own place to put things. so the same skill gets rearranged for Cursor, again for Copilot, again for Codex, and the copies slowly stop matching. the fix published yesterday is called Agent Plugins 1.0.0, and it standardizes that wrapper and nothing else. a plugin is a folder with a file called plugin.json at the top, skills inside a folder named skills, and MCP servers listed in a file named mcp.json. the settings file names the plugin and nothing more. it cannot point somewhere else or hide the pieces inside itself, so every tool looks in the same three places and there is nothing to configure. anything a single tool wants to add goes into its own clearly named folder that other tools skip over. that keeps the shared part small enough that everyone can agree on it. the pieces also fail separately, so a broken MCP server does not stop the skills sitting next to it from loading. what the standard leaves out is worth knowing. it says nothing about how plugins get installed, what they can touch once running, or how to check who wrote one, and it gives passwords and API keys nowhere safe to live. a plugin is still a folder of code that runs on your machine. spec: agent-plugins.org/specificat… the name missing from all of this is Anthropic. skills and MCP both came from there, and Claude Code has packaged plugins in almost exactly this shape since 2025. Anthropic is not on the group steering the standard, and Claude Code is not on the list of tools supporting it, which is ChatGPT and Codex, Cursor, GitHub Copilot, Kiro, and VS Code. Google joined as a maintainer on launch day, and its Agents CLI is one of the first two Google products already shipping in this format. I wrote a full walkthrough of that tool, which is the clearest picture of what a plugin actually does once it reaches an agent. it covers how Agents CLI drops Google's agent-building skills into Claude Code, Cursor, Codex, or Antigravity, then builds a citation-grounded RAG agent, runs a 20-scenario eval suite that catches a hallucination, traces it back to a single line in the system instruction, and deploys the fixed agent. quoted below.

Integrating 𝗔𝗴𝗲𝗻𝘁𝗶𝗰 𝗥𝗔𝗚 Systems via 𝗠𝗖𝗣 👇 If you are building RAG systems and packing many data sources for retrieval, most likely there is some agency present at least at the data source selection for retrieval stage. This is how MCP enriches the evolution of your Agentic RAG systems in such case (𝘱𝘰𝘪𝘯𝘵 2.): 𝟭. Analysis of the user query: we pass the original user query to a LLM based Agent for analysis. This is where: ➡️ The original query can be rewritten, sometimes multiple times to create either a single or multiple queries to be passed down the pipeline. ➡️ The agent decides if additional data sources are required to answer the query. 𝟮. If additional data is required, the Retrieval step is triggered. We could tap into variety of data types, few examples: ➡️ Real time user data. ➡️ Internal documents that a user might be interested in. ➡️ Data available on the web. ➡️ … 𝗧𝗵𝗶𝘀 𝗶𝘀 𝘄𝗵𝗲𝗿𝗲 𝗠𝗖𝗣 𝗰𝗼𝗺𝗲𝘀 𝗶𝗻: ✅ Each data domain can manage their own MCP Servers. Exposing specific rules of how the data should be used. ✅ Security and compliance can be ensured on the Servel level for each domain. ✅ New data domains can be easily added to the MCP server pool in a standardised way with no Agent rewrite needed enabling decoupled evolution of the system in terms of 𝗣𝗿𝗼𝗰𝗲𝗱𝘂𝗿𝗮𝗹, 𝗘𝗽𝗶𝘀𝗼𝗱𝗶𝗰 𝗮𝗻𝗱 𝗦𝗲𝗺𝗮𝗻𝘁𝗶𝗰 𝗠𝗲𝗺𝗼𝗿𝘆. ✅ Platform builders can expose their data in a standardised way to external consumers. Enabling easy access to data on the web. ✅ AI Engineers can continue to focus on the topology of the Agent. 𝟯. Retrieved data is consolidated and Reranked by a more powerful model compared to regular embedder. Data points are significantly narrowed down. 𝟰. If there is no need for additional data, we try to compose the answer (or multiple answers or a set of actions) straight via an LLM. 𝟱. The answer gets analyzed, summarized and evaluated for correctness and relevance: ➡️ If the Agent decides that the answer is good enough, it gets returned to the user. ➡️ If the Agent decides that the answer needs improvement, we try to rewrite the user query and repeat the generation loop. Are you using MCP in your Agentic RAG systems? Let me know about your experience in the comment section 👇

The simplest AI agent memory system that actually works is four markdown files and zero databases. You don't need vector databases, embeddings, or retrieval pipelines for a personal AI agent. Flat markdown files are readable, searchable, and git-trackable. Your agent can load them without any preprocessing. Here's the full system. 1. One file per domain, never per date Create separate files for each knowledge area like people.md, companies.md, deals.md, and tasks.md. Each file owns one topic. Chronological brain dumps feel organized at first but turn unsearchable after two weeks. Domain separation means your agent pulls the one file it needs for any question. 2. A MEMORY.md index that maps everything One file listing every memory file with a one-line description of what's inside. Your agent loads the index first, then pulls specific files on demand. This keeps context window usage predictable and lookups fast. Without it, your agent either loads everything or guesses which file to open. 3. Cache files marked with freshness headers Any file that mirrors an external source like your CRM or analytics dashboard is a cache, not a source of truth. Mark every cache file with a last_sync timestamp at the top. Train your agent to announce data freshness before every analysis. Silent use of stale data is how wrong-but-confident outputs happen. 4. Flat markdown, not databases Markdown files beat vector databases for personal agents because you can read them, search them, and track them in git. You skip the embedding pipeline, the retrieval infrastructure, and the abstraction layer between you and your agent's knowledge. The simplest architecture that works is the right one. Set this up in 20 minutes. Your agent's memory becomes organized and version-controlled from day one.

Qwen3.8-Max by @Alibaba_Qwen has reshaped the cost-performance Pareto frontier in Frontend Code Arena, with pricing of $2 per input MToken and $6 per output MToken. Top models on the Pareto frontier: - Claude-Opus-5 - Kimi-K3 - Qwen3.8-Max - GLM-5.2 - DeepSeek-V4-Flash Congrats to @Alibaba_Qwen on another major milestone!