爱窝啦 AI 日报 AI Daily 2026/10/19: Anthropic Co-founder Bet on AI a Decade Ago, Now Worth Over $100 Billion

Today’s Summary

An Anthropic co-founder left his job a decade ago to bet on AI and is now worth over $100 billion. DeepSeek has launched standalone search and web extraction APIs billed per call.
Open-source tools, agent frameworks, and evaluation methods are all solving the same problem: getting AI into real collaborative workflows rather than just conversation.
Start today by reading the DeepSeek API docs and the ruflo framework, then decide whether to integrate search and multi-agent capabilities into your existing products.

🔥 Today’s Top 10 Highlights

1. Anthropic Co-founder Bet on AI a Decade Ago, Now Worth Over $100 Billion

The journalist-turned-AI-researcher called it right. Anthropic co-founder Jack Clark left Bloomberg in 2016 to pursue AI. According to Anthropic Co-founder Bet on AI a Decade Ago, Now Worth Over $100 Billion , he is now worth $15.5 billion and has made the Forbes 400 list 💰. At the IMF annual meeting, he declared that “the singularity has arrived,” painting a picture of a single human scientist leading hundreds or thousands of AI scientists conducting research simultaneously. A decade ago he told colleagues that AI was more important than anything else — and no one paid attention.

2. NVIDIA Open-Sources Nemotron Multimodal Model, Faster Than Comparable Models

Visual reasoning model built for compute efficiency. NVIDIA has open-sourced Nemotron 3.5 Super VL, with 120B total parameters and 12B active parameters. Gorden Sun introduced the model’s features in a post on X , noting it combines reasoning, tool calling, and understanding of text, images, documents, and video 📹. It has been compute-optimized for long-running agents and is significantly faster than comparable models. Capabilities are solid, and weights are publicly available on Hugging Face.

3. DeepSeek Launches Search and Web Extraction APIs

Standalone endpoints for agent use. DeepSeek’s API documentation introduces two JSON endpoints : Search, which uses a proprietary search engine to return web results, and Fetch, which extracts a specified web page as Markdown 🔍. Pricing is $0.75 per thousand successful requests, charged only when a call succeeds and returns results. Both endpoints use DeepSeek API key authentication and do not need to be wrapped inside model generation calls. Starting November 1 (Beijing time), calls made through Harness, Claude Code, or Codex will be billed per request.

4. ruflo Agent Framework Tops GitHub Daily Trending

Multi-agent coordination tool gains 500 stars in a single day. ruvnet/ruflo has entered the GitHub daily trending chart , adding 513 stars in one day for a total of 76,520. Built in TypeScript, the project supports deploying multi-agent clusters, adaptive memory, and vector RAG integration 🔗. It natively integrates with Claude Code, Codex, and Hermes. A good fit for developers who need to build conversational AI systems or coordinate autonomous workflows.

5. Sphinx Tests Whether Developers Understand AI-Generated Code Before They Push

Uses Q&A to ensure comprehension before merging. The Sphinx project is a Git hook that pauses pushes to protected branches. A separate review model reads the diff and asks the developer to predict behavior, failure modes, or invariants 🔒. If the answer is wrong, it explains the issue and presents a new scenario; the push proceeds only after a correct response. The project’s demo on its homepage discusses payment retry logic and event ID ordering. The README emphasizes this is a practice tool, not a security gate, and notes that local hooks can be bypassed. Requires Node.js 22.19 or later.

6. AI Agent Skill Pack Turns Product UI into Cinematic Short Films

Provides specific parameters rather than leaving it to the model’s discretion. The AI Exploration Guide channel covers an AI agent skill pack that turns product interfaces into cinematic short films . After installing with npx skills add, you can have the agent generate and publish a short film directly 🎬. The method breaks the clip into a series of hero shots: real UI floats on a stage like a panel, the camera tilted low at a 30–50° angle. The focal plane is only two or three lines of text deep, with everything else falling into bokeh. Dark products use a pure black void background; light products use a high-key studio look.

7. Who Calibrates the AI Benchmarks Themselves?

Evaluations should be short-lived, grounded, and specific. TypeSafe AI discusses the reliability paradox of AI benchmarks in a post. The author argues that good evaluations should be ephemeral (representing only a snapshot of capability at a given point in time), grounded (directly measuring performance on business-relevant tasks), and specific (evaluating target outcomes rather than proxy metrics) 📊. Drawing an analogy to superconducting qubit calibration charts, the post points out that a system used to verify accuracy must itself ensure its own accuracy.

8. Rewatching Westworld to Understand the Robot Consciousness Pyramid

Memory, improvisation, and inner dialogue make up consciousness. Jike user “我的兄弟叫铁马” shares the three-layer pyramid proposed by co-founder Arnold in the show : the base layer is continuous memory, the middle layer is improvisation, and the top layer is inner dialogue 💭. The author found the series confusing on a first watch a decade ago, but now understands that the show’s core is about how robots move from executing scripts to developing autonomous consciousness. The pyramid maps onto current AI agent design: memory management, dynamic decision-making, and reflection mechanisms.

9. Anthropic Pledges 5% of Equity for the Public Benefit

Additional shares vest when company value increases, up to 20% total. Anthropic has announced the creation of the Anthropic Public Benefit Fund, seeded with 5% of the company’s equity 🏛️. Each time the company’s value grows by 50% above a set baseline, the board may contribute up to an additional 2%. Total equity dedicated to public benefit could reach approximately 20%. Co-founders have pledged to donate 80% of their own equity. The move responds to public scrutiny over AI companies’ social responsibility.

10. Troubleshooting an AI Connectivity Issue — the Culprit Turned Out to Be a Software Conflict

The diagnostic journey took a wrong turn. Jike user “灵有志” writes about repeatedly losing connectivity to AI services and asking Codex to methodically diagnose the issue, with no results 🔧. After a full day of troubleshooting, it turned out that simply deleting an app called “小猫” and switching to another tool made the problem disappear. A reminder to developers and users that AI-assisted diagnosis is not always faster than just swapping out your environment.

◎ Industry Shifts & Personal Impact

Mathematical Proof Is Being Industrialized by AI

Mathematicians are starting to rely on agents to run complex proofs. Jike user “埃安” mentions that a friend who teaches mathematics at Tongji University built a GPT Pro Agent four months ago to run complex proofs. AI is now driving mathematics toward industrialization 🧮. The author uses the analogy “humans are AGI to dogs” to suggest that when AGI actually arrives, humans may simply be dependents of a higher intelligence. Academics doing theoretical research will need to redefine their role — shifting from independent proof-writers to coordinators who design problems and verify AI output.

⌘ Open Source Top Projects

mlabonne/llm-course: A Roadmap for Getting Started with Large Language Models

mlabonne/llm-course has entered the daily trending chart , gaining 81 stars in one day for a total of 84,030 📚. The project provides an introductory course on large language models, complete with a full roadmap and Colab notebooks. Suitable for developers and researchers who want a systematic introduction to LLM fundamentals, fine-tuning, and deployment. It already has a large following in the Chinese-speaking community.

Comfy-Org/ComfyUI: Modular Diffusion Model GUI

Comfy-Org/ComfyUI gained 154 stars in a single day , bringing its total to 138,616 🎨. The project provides a powerful modular diffusion model GUI, API, and backend using a graph/node-based interface. It bills itself as the world’s fastest local inference engine. Built in Python, it suits creators and developers who need fine-grained control over their image generation pipelines. Supports complex node composition and custom workflows.

◉ Social Media Picks

All 502 Lex Fridman Podcast Episodes Now Have Transcripts and Chinese Translations

Full archive processed with local ASR. Xiangyang Qiaomu announced on X that he has used local ASR to produce transcripts and Chinese rewrite translations for all 502 episodes of the Lex Fridman podcast 🎙️. The site lex.qiaomu.ai supports light, dark, and eye-care modes, incorporates Chinese open-source fonts, and features careful typography. It also supports PWA installation as a mobile app. The post received 311 likes and 57 reposts. Useful for researchers and AI enthusiasts who need to search or cite content from long-form interviews.

😄 AI Fun Facts

Using Opus 5.5 to Make a Video Instead of Three Days of Manual Iteration

After his Claude account was suspended with no refund, 岚叔 decided to make a video “protesting” with Opus 5.5. To make matters worse, he had spent three days iterating with Codex across multiple versions, only to find that Opus 5.5 did it better in a single afternoon — he wishes he hadn’t bothered with the manual work 😭. The expression on 宝玉’s face when he reposted probably said it all: “I know that feeling of three wasted days.”

❓ Related Questions

As a Codex subscriber wanting to run multi-agent workflows with ruflo, do I have to re-explain the coordination logic every time?

The issue isn’t your subscription tier — it’s that your coordination rules haven’t been locked down yet. Today’s ruflo framework supports adaptive memory and federation mechanisms. Start by writing your common agent role assignments into a config file and running a test cycle. Once the workflow is stable, consider longer-term use. If you have an existing account to renew, check out 爱窝啦·AI账号店’s Codex products — renewing just maintains your access point; it doesn’t automatically optimize your coordination logic.