Chen Xiaomeng's article systematically reviews 10 common point models in the crypto space (interaction task-based, TVL deposit-based, lending points-based, LP market-making based, Pendle YT-based, Perp trading volume-based, spot trading-based, social and content-based, invitation and community-based, staking and lock-up-based), pointing out their respective implied capital costs, opportunity costs, strategy profit/loss, liquidity costs, and tail risks; proposes a unified valuation model based on FDV, reward pool ratio, individual effective point share, and redemption conditions, and provides an evaluation framework containing four scenarios: optimistic, neutral, pessimistic, and zero airdrop; finally provides an anti-sybil checklist containing 10 key questions.
Chickadee Bandit is a private app building platform for homes, allowing users to describe needs (such as D&D activity coordination, thermostat dispute arbitration, random dishwashing assignment, Rick Roll news pop-ups) to an AI assistant connected via MCP. The AI generates app code and installs it into the household's private database; all apps require manual confirmation for installation, and source code download for review is supported. iPhone and Android clients are now available.
Kuaishou's Kling AI announced the full launch of Kling 4.0 at klingai.com, supporting single video generation up to 30 seconds long. Input length has increased to 8000 tokens, with support for 10 multi-keyframes and mixed input of 10 images/5 videos/7 subjects. New features include 4K/1080p 10-bit HDR output, 21:9 ultra-wide aspect ratio, multilingual and dialect speech generation, and multiple video continuations (up to 2 minutes).
vislearn has released Luti, an open-source macOS application. As a local-first runtime, it connects cloud-based AIs such as Claude, ChatGPT, Grok, Gemini, and Muse to user-selected local project folders via the MCP protocol. It supports a total of 25 MCP tools for reading/writing code, running builds and tests, checking Git status, driving browser previews, exchanging artifacts, and identifying tech stacks. It offers two connection methods: OpenAI Secure MCP Tunnel and Cloudflare Tunnel. Permissions are limited to a single project directory, and credentials are stored in the macOS Keychain. The open-source repository is https://github.com/boltguo/Luti.
On October 7, Anthropic updated its official support page, confirming that monthly API Credits for Claude Max and Team plan users can be used directly with the Claude Agent SDK, the claude -p command, the Claude API, and Claude Managed Agents, covering all third-party apps based on these technologies; multiple Anthropic employees confirmed on Hacker News that this policy requires no additional fees and has faced no opposition.
Anthropic officially released a major update for the Claude Code desktop version, supporting automatic creation of Nested Sessions for parallel multi-task execution, keeping the main session clean while automatically summarizing results; new features include sidebar status monitoring, one-click environment reuse, Git Worktree isolation, the /btw side-thread dialogue command, and idle compression. The update was described by Anthropic engineers as "almost a v2 release."
On October 14, Hesai officially launched the 3D world data engine Kosmo and the cloud platform Kosmo World, introducing the world's first AI spatial camera with an integrated 8K imaging system, Kosmo One. The device combines LiDAR, an 8K imaging system, a one-inch CMOS sensor, and a high-precision IMU, supporting microsecond-level synchronized capture with millimeter-level reconstruction accuracy and exported footage up to 6K. It has completed real-scene scanning at the Pixar campus and integrated into the RenderMan 28 workflow, and supports Sharpa robot training at DQ stores, improving task success rates to 99%.
On October 15, 2026, StepFun open-sourced its next-generation flagship model, Step 5 Preview. The model adopts a sparse MoE architecture with 600B total parameters and 27B activated parameters per token, supporting a 1M token context window and multimodal inputs including images and video. On the same day, Huawei lit up its first 1024-card Atlas 950 SuperPoD liquid-cooled super node, providing 1 EFLOPS FP8 and 2 EFLOPS FP4 compute power. Through the vLLM Ascend inference engine, it achieved Day-0 support for Step 5 Preview along with multi-level software-hardware co-optimization.
Krea 3 is an AI image editing model launched by Krea, supporting native 2K resolution editing with a single-image processing time of under 8 seconds at a unit price of $0.06; it supports changing backgrounds, objects, materials, or lighting while preserving key details; up to 6 reference images can be used to maintain consistency of products, characters, and styles across a set of marketing assets.
Anthropic's research team released a technical report based on internal employment data, AI usage logs, employee surveys, and 17 in-depth interviews. The report indicates that as of August 2026, AI agent working hours within the company have reached 6 times the employee working hours; the proportion of junior position hires in technical departments dropped from 21% in early 2025 to 14%, and the hiring ratio for junior technical roles plummeted from 27% to 9%; non-technical employees' usage of coding agent sessions increased from 2 per month in August 2025 to 90 per month in May 2026; employees generally shifted towards 'meta-work', such as building monitoring dashboards, writing agent instructions, and evaluating prompts. Such activities accounted for an increase from 10% in autumn 2025 to 22% in spring 2026 among interactive Claude sessions.
The Anthropic research team used internal Claude models for reverse analysis and discovered that the seed parameters for ten standard elliptic curves, including P-192 through P-521, all originated from the lrand48 random number generator in C libraries from the 1990s. Its initial state was initialized by Unix timestamps (between 1997 and 1998) via srand48; this finding refutes long-standing backdoor suspicions and confirms that the seed generation method was simple and transparent.
Anthropic's research team used Claude (Mythos 5) to autonomously optimize eight classes of quantum circuits, discovering new designs in tasks such as chemistry, optimization, physical simulation, and number theory that significantly reduce the required number of qubits and gates; for example, implementing branchless history storage reversible divsteps in elliptic curve point addition, and reducing the gate count of Reed–Solomon decoders in DQI algorithms by 12 times; all work was completed based on natural language descriptions and an automated evaluation loop, with results published on arXiv.
The Ethereum Foundation announced that the Glamsterdam network upgrade will activate on the Hoodi testnet at 17:42:48 UTC on October 26, 2026 (epoch 132,352, slot 4,235,264), simultaneously raising the gas limit to 200M; the upgrade includes core changes such as ePBS (EIP-7732), Block Access Lists BALs (EIP-7928), and gas pricing adjustments (EIP-8037/EIP-8038); the mainnet activation time has not yet been determined; node operators must update both execution layer and consensus layer clients synchronously.
Reflection Lab released Beam-1.0-501B-A23B on Hugging Face, a 501B-parameter sparse MoE large language model (activating 23B per token), supporting 1M tokens context, text input/output, five adjustable reasoning intensity levels (low/max), native tool-calling syntax, and preserve_thinking for long-range thought retention; it employs hybrid attention (43 sliding window layers + 9 global layers), a Rust-based custom inference harness, and Apache 2.0 license; it provides BF16 / FP8 / NVFP4 multi-precision checkpoints, supports day-zero deployment with transformers / vLLM / SGLang, and includes SFT (TRL) and full-parameter/LoRA GRPO (Miles) training solutions; the model focuses on coding and agent scenarios and has been integrated into Hugging Face Inference Providers and Endpoints.
Anthropic released two research reports: The first indicates that its internal AI agent work hours exceeded those of employees by March 2026, with the number of agents reaching six times that of employees in August, accompanied by changes in hiring structure and improvements in multiple output metrics; the second compares AI expectations between employees and 11,000 US adults, finding that employees generally hold more aggressive expectations (e.g., 91% of employees vs. 39% of the public believe AI will make Nobel Prize-level discoveries before 2030), but both groups agree on the economic logic of AI's impact on employment.
Bitcoin Optech Newsletter #427 discloses two DoS vulnerabilities in Eclair v0.14.0 and earlier versions: one is a memory leak caused by a race condition in duplicate checks for temporary channel_ids, and the other is memory exhaustion due to bypassing pending channel rate limits; both were fixed in v0.14.1.
This article documents the engineering practices of a major retail enterprise in building a multi-agent platform for inventory accuracy. It shifts large model capabilities from application-layer logic to platform infrastructure, reducing the hallucination rate from 15% to 1.5%. The platform is built on ADK and MCP, uses LiteLLM for unified model routing, runs llama3.1 locally via Ollama, and all components are open-source with no hidden abstraction layers.
Paradigm Intelligence, in collaboration with Ruisi Zhilian, has launched ModelSphere, an open-source AI deployment platform. It is free and open-source for enterprises, integrating computing power management, model deployment, and runtime optimization. The platform supports domestic chips such as Huawei Ascend, Tianshu Zhixin, Moore Threads, and Hygon DCU, and has already been adapted to mainstream models including DeepSeek-V4/V4-Flash, Qwen, Kimi, MiniMax, and GLM. Key capabilities include graphical deployment, AutoTune intelligent tuning, unified scheduling across mixed chips, and automatic scaling. It adopts the Apache 2.0 license, with code hosted on GitHub.
Cloudflare announced plans to launch a free public certificate authority in early 2027, issuing quantum-safe TLS certificates based on Merkle tree certificates. The approach uses a hybrid mode that simultaneously provides traditional X.509 certificates and lightweight tree proofs, supporting modern clients in verifying post-quantum signatures while enabling seamless fallback for older clients. This solution significantly reduces handshake data size and latency by unifying certificate issuance and Certificate Transparency log writes into a single Merkle tree append operation.
Google has released the AndroidX Security State library and the accompanying Security State Provider library, which support verifying device SPL, published SPL, and available SPL at the component level (system, system module, kernel). It provides functions such as checkCvePatched and checkAllAvailablePatchesApplied to check the status of specific CVE patches or the installation of all available patches, and generates standard security advisory links.
The Amazon Web Services development team has open-sourced Pizza Bot, a self-hosted AI agent inbox application that supports scheduled/Webhook-triggered tasks, delegation to specialized worker nodes, human approval pauses, and asynchronous notifications. It employs a client/server architecture with state and conversation threads persisted locally; prompts, attachments, or tool call requests are sent only to model providers configured by the user and explicitly enabled MCP servers. The project is licensed under Apache 2.0 and is not an official AWS service.
Lingle, Senior Test Development Expert at Alibaba Taotian Group, shared the AI testing engineering practices of Shoutao TEKO at QCon Shanghai, proposing a quantitative model: Autonomy = Criteria × Signals. On the criteria side, a four-level acceptance condition pipeline was built and fed back into code generation to achieve AI-native TDD; on the signal side, the App was transformed into a programmable environment, improving automation execution rates through terminal aspects, SDK testability baselines, and CI gates. Implementation results include 100% coverage across three platforms, 12 minutes per analysis, over 1.5w+ test cases in the repository, and a 96% operation success rate, while candidly disclosing an 85% success rate on HarmonyOS and organizational resistance.