Signal to BuildSIGNAL TO BUILD中文Owner sign in ↗Explore curated knowledge, course notes and methods, with their existing content and sources preserved.
This index lists complete English editions. Additional Chinese notes and bilingual updates remain available in the Chinese index.
The official claim concerns Hugging Face’s Model Trending popularity list, not first place in overall performance. Lu post-trained an existing model rather than pretraining a general model from scratch. Improvement on a task does not establish stronger capability everywhere.
A complete cross-author synthesis of agent architectures, orchestration patterns, memory, evaluation and operational responsibility, including the published historical cases and additions through October 21, 2026.
A systematic framework for agent evaluation released by the Anthropic engineering team in 2026. It covers an 8-entry concept glossary, a 3-type grader matrix, the dual-track Capability vs Regression evaluation design, 4 sets of evaluation playbooks for different agent categories (coding / conversational / research / computer use), an 8-step roadmap for building evaluations from scratch, the mathematical intuition behind pass@k vs pass^k, the evaluation framework ecosystem, and common anti-patterns. Phase A anchors the wiki, and sets the density and bilingual style baseline for the anthropic-* series of content.
A research report on the evolution of AI Agent technology architecture compiled by newtype in January 2026. It is based on papers published from October 2024 to January 2026, official blogs from Anthropic/OpenAI/Google/Microsoft, open-source projects, and real-world deployment cases. It covers 6 major evolution trends / 4 multi-Agent orchestration patterns / Google's 8 foundational patterns / memory management (MemoryOS 3-layer architecture + vector database selection) / tool calling + MCP standardization / reasoning frameworks (ReAct → ToT → GoT → WoT evolution) / Large Reasoning Models (LRM) / 7-layer production architecture / comparison of mainstream frameworks (LangGraph/CrewAI/MAF/AutoGen/LlamaIndex/ADK) / real deployment cases from Morgan Stanley + Klarna + BBVA + Rakuten / 2026 trend forecasts / compliance and governance / production checklist. It serves as a "research report-level" supplement to [[newtype-agent-architecture]] (Huang Yihe's post-level judgments).
Connecting AI to the physical world — self-driving labs for materials and biology (Periodic, Ginkgo), Waymo, AI chips and inference compute (Cerebras, Baseten), space data centers (Starcloud), AI smart stoves — and the cross-domain playbook that actually makes "AI for X" work in the world of atoms
13 episodes of frontier research podcasts reorganized by "debates": LeCun bashing LLMs (dead end vs. the scaling camp) / three camps on world models: JEPA vs. linguistic symbols vs. renderers / whether continual learning is overhyped / whether RL can succeed outside the coding domain / recursive self-improvement via AI building AI / frontier risks and alignment / Hassabis's endgame for AGI / three non-mainstream architectures: EBM, Mistral, Leanstral. Speakers include LeCun, Pachocki, Dubois, Kolter, Manning, Dehghani, Vinyals, Hassabis, Brockman
Long-term X views from Karpathy, sama, Anthropic & Google model camps and lab voices — LLM knowledge base paradigm, jagged capability theory, harness/skills engineering, model versioning strategy, agent safety, AI for science, HTML as the new markdown
A complete cross-podcast guide to coding agents: orchestration, tools, model selection, review, infrastructure and responsibility, including all published additions through October 21, 2026.
Invisible AI watermarks can be removed via spectral analysis without any loss of image quality (empirically demonstrated by reverse-SynthID) → watermarks are a provenance tool for honest users, not ironclad anti-forgery evidence; content producers should layer multiple signals and proactively apply compliance labeling.
Drawing on interviews spanning 2023–2026, this piece separately examines model capabilities, real demand, delivery, operating returns and capital pricing, while preserving historical disagreements. A cross-interview synthesis separating model capability, demand, delivery, business returns and valuation, with historical disagreements preserved.
Listen: 01:07–07:11; 08:40–11:20. Zhu reduces the underlying problems to computation, communication, and storage. GPUs need networks, switches, storage, operations, resource scheduling, and training and inference frameworks. His comparison with cloud service layers is an explanatory framework, not the only possible industry taxonomy.
A business trip can appear to require only an itinerary. Across cities, days and meetings, however, hotels, transport and timing constrain one another. An early choice can make the eleventh step impossible. Bill uses this example to explain why an agent needs to consider consequences, search alternatives and recover when a path fails.
Themed condensation of 13 official 2026 blog posts. The core is Claude Managed Agents — a set of APIs for "renting Anthropic's cloud infrastructure to run long-horizon agents": memory (cross-session memory) + dreaming (sleep-time self-improvement) + outcomes (scored self-correction) + multiagent (lead-worker collaboration) + self-hosted sandboxes (bring-your-own sandbox) + MCP tunnels (private network tunnels), underpinned by a decoupled meta-harness design of three components: "brain (Claude + harness) / hands (sandbox) / session logs". Additional content covers the Claude Code desktop app rebuild (parallel agents) + harness design philosophy (use tools Claude already knows / keep asking "what can I turn off") + enterprise deployment data + containment security architecture for three products + a security checklist for AI-accelerated offense and defense + chart visualization / frontend Skills / lifestyle connectors.
Hub of 11 entries from the Anthropic research section: Constitutional AI (2022 foundational) + Automated Alignment Researchers (2026 scalable oversight) + Petri open source (donating-open-source-petri) + deprecation commitments (models not deleted) + end-subset-conversations (Claude welfare) + Teaching Claude Why (follow-up to agentic misalignment) + Project Vend Phase 2 (failed Claudius red team experiment) + 2028 AI Leadership (US-China) + Anthropic Institute 4 major research agendas + Core Views on AI Safety + full ASL3 overview. The 5 pillars of Anthropic's safety stance: Constitutional + RSP/ASL / Transparency / Welfare / Red Team
5 hub articles from the Anthropic research section: Natural Language Autoencoders (decoding Claude's internal activations into natural language, a major breakthrough in May 2026) + How people ask Claude for personal guidance (sampling of 1 million conversations / 6% are seeking life advice) + BioMysteryBench (bioinformatics evaluation) + Why do AI models hallucinate? + What is sycophancy? + 4 research team pages (Alignment / Economic Research / Interpretability / Societal Impacts). Covers three areas of research: "how AI thinks internally + how users use it + what bugs it is prone to"
Compilation of 4 talks from the 2025 Amazon Web Services Agent lecture series. Wang Qixiang (Reasoning Model + Agent Brain + Multi-Agent Orchestrator + Framework Comparison) / Wang Xiaoyan (4 Elements of Agentic AI: Planning/Memory/Tool/Action + 3-tier Capability Ladder + RL + Product Pricing) / Chang Gaowei (ANP Protocol: email for agents, USB-C counterpart to MCP, W3C DID + P2P architecture) / idoubi (Independent development practice + AI Wrapper SOP + ProductHunt launch + programmatic SEO + all-in direction). Serves as the "industry practical perspective" supplement to the [[newtype-agent-architecture-research-2025]] research report.
An agent performing tasks needs more than a capable model. It must run code, browse websites, preserve login, wait for confirmation and return to its previous position. Lei calls the supporting environments Agent Infra, or agent infrastructure. He wants Grasp to handle common browser engineering so application developers can concentrate on industry tasks.
Open conversation has no single correct answer. A user may enjoy a response, rewrite it, delete it or request another. Reflecting on approximately a year at Character.AI, Ted discusses turning such behavior into data, training and experiments. The central subject is a system that continually discovers mistakes.
"Non-public" details dug out from the leaked source code of CC v2.1.88: what the dual-layer telemetry collects and that you cannot opt out / model animal codenames (Tengu · Capybara · Fennec · Numbat) + behavioral bugs in Capybara v8 / differential treatment between internal and external users / Undercover mode that hides AI identity / hourly-polling remote killswitch / unreleased features such as the KAIROS autonomous agent. Read this if you want to know "what CC hides inside"
Teardown of the leaked CC v2.1.88 source code: minimal agent loop + 12 layers of production hardening / tool interface lifecycle / permission funnel / 4 types of sub-agents / 3 context compression strategies / session persistence / MCP / bridge layer / Feature flag DCE. Read this article to understand "how CC is built internally"