Reports from Goldman Sachs and Deutsche Bank indicate that short-term spot rental prices for US AI compute are approximately $40 billion to $50 billion per gigawatt, which is 2 times the long-term contract price ($20 billion to $25 billion per gigawatt) and 3–4 times the break-even threshold for hyperscale cloud providers (approximately $11.6 billion to $12 billion per gigawatt). Analysis suggests this premium stems from delays in building proprietary capacity; as compute scale expands by more than 7 times over four years among providers such as SpaceX, the spot premium will become unsustainable. Current high prices cannot be covered by end-user AI revenue, essentially relying on external debt financing to support training demand.
LinkedIn natural language search supports customer finding without Boolean syntax. The article outlines four practical uses: finding decision-makers by position + industry + region; using procurement responsibility descriptions instead of fixed job titles to broaden the scope; searching for procurement/supply chain-related roles within target companies to identify the complete decision-making chain; and filtering potential customers through second-degree connections. It provides 4 search templates with replaceable keywords at the end, emphasizing the need to combine manual judgment to verify customer profiles.
Yun Yuan Kite author guanttly announced the practical implementation of their self-developed tool in a real ecosystem project: coordinating development, review, and release across multiple products via a main conversation, accumulating 74 completed task sessions with over 10 hours of runtime; supporting session handover, cross-repository parallel development, direct terminal connection, and remote web operations, with the Zhaoye App now live on iPhone and Android.
In Q3 2026, global PC shipments fell by 20.1% year-on-year, yet Lenovo's net profit grew by nearly two times, Dell's AI server revenue doubled, Apple's Mac revenue hit a record for the same period, Asus's share of high-value PCs exceeded 65%, while HP faced profit pressure due to its lack of premium pricing power and AI infrastructure business.
Hookpager is an iOS app independently developed by bigjack. It supports receiving arbitrary HTTPS Webhook POST requests and pushing them to mobile devices, offering features such as confirmation, mute, do-not-disturb mode, and pairing code binding for a second device. It is suitable for alert scenarios involving Jenkins, Cloudflare, Uptime Kuma, GitHub Actions, etc. It provides a free quota; the Pro version costs $1.99/month, $14.99/year, or $39.99 for lifetime access. Currently distributed via TestFlight, targeting overseas users.
FlowCat is an open-source macOS menu bar traffic visualization tool written in Swift/SwiftUI. The installer is only about 1 MB, supporting Apple Silicon and macOS 13+. It provides overall system traffic statistics (by day/month), process rankings, Wi-Fi network-specific views, real-time speed display, and right-click quick actions. Data is stored locally, with no uploading or reading of communication content, and it is open-sourced under the MIT license.
Atticus is an open-source macOS application that supports triggering voice commands via a hold-to-talk shortcut (⌃⌥option+space). It uses local speech-to-text and executes operations such as window management, file search, and screen reading through subscribed Claude Pro/Max plans (via Claude Code). No additional API keys are required; audio never leaves the device. AI commands run in a sandbox by default. The project supports learning user habits and generating zero-cost shortcuts. Hosted on GitHub under the MIT license.
Anthropic announced monthly API credits for Claude Max and Team subscribers: Max 5x tier receives $100, 20x tier receives $200; Team credits accumulate per seat ($20 per standard seat, $100 per premium seat), capped at $500. Credits can be used for Messages API, Batch Processing, Playground, Managed Agents, and Agent SDK, but do not support interactive calls via Claude Code. New subscribers must wait 7 days before claiming credits and must link a Claude Console organization on the claude.ai billing page. Pro, Free, and Enterprise subscriptions are not covered.
Preloop is a local self-hosted equivalent engine for GitHub Actions that executes workflows via hardware-isolated microVMs with a startup time of approximately 300ms; it supports all GitHub Actions features including native .github/workflows files, OIDC, matrix builds, and reusable workflows; it provides a lightweight runner written in Rust (installation size is 1/36 of the official runner, peak memory usage is 1/19); it supports DAP debugging, GitHub App/PAT authentication, automatic webhook drift correction, and is compatible with macOS/Windows/Linux; the control plane uses the FSL-1.1-MIT license, while other components are MIT licensed; the project is open source at https://github.com/preloopdev/preloop.
Agnes AI has open-sourced the Agent execution framework Agnes Harness (AGH) and the Model Hardware Standard (MHS) device interaction protocol, supporting continuous task execution, multi-tool invocation, state tracking, and programmable hardware access; simultaneously announced the performance of the Agnes 3.0 Pro model in benchmarks such as GPQA and SciCode, along with its real-world enterprise task results, with its API coming soon, while Agnes 3.0 Flash remains free.
CPA-Bar is an open-source cross-platform native desktop client supporting macOS, Windows, and Linux. It resides in the menu bar or system tray and provides centralized management of multiple CLIProxyAPI (CPA) and CPA-Manage-Plus instances, account and OAuth login, quota and usage statistics (including remaining quota, reset time, request count, tokens, and latency), privacy protection (keys stored in system keychain, sensitive information hidden in screenshots), in-app signed updates, light/dark themes, and Chinese/English interfaces.
DuoDiscover is an autonomous discovery search system that alleviates the bottleneck of time-consuming real evaluations through three mechanisms: co-discovering optimization objectives and lightweight surrogate evaluators (yardsticks), continuously verifying surrogate evaluator fidelity, and decentralized surrogate collaboration. It achieves significant performance improvements in SWE-fficiency, SOL-Exec, vLLM, SQLite, and CPU microarchitecture optimization tasks.
Anthropic released research indicating that small open-source models with parameter counts ≤30B (such as Qwen3.5-9B and GLM-5.2) can already perform distributed inference on cheap IoT devices (such as Android TV boxes with 4GB RAM). Experiments showed that GLM-5.2 could achieve full-chain autonomous replication of Llama3.1-8B in a simplified network (with success at 8/8 nodes), while Gemma4-26B-A4B could run at 0.39 tokens/s on 15 TV boxes. The study proposed a "weight-based replicator" threat model, estimating its diffusion cycle to be several days, and pointed out that trends in improving model capabilities may cause this ability to converge on single devices by the end of 2026.
A joint analysis by Stanford University's Institute for Human-Centered Artificial Intelligence (HAI) and Anthropic, based on 250,000 anonymized Claude conversations, found that 56% of human-AI collaborations were classified as 'high impact' or above. Users often drive AI teaching through questioning and requests for explanation rather than passively accepting outputs; this pattern is particularly beneficial in high-stakes domains such as law, software engineering, healthcare, and home renovation. The research indicates that 'beneficial friction' can promote clarification, refinement, and learning, although negative friction requires active user recovery. The paper has been accepted by Nature Human Behavior.
CodeScene released a case study: An AI agent refactored the open-source decompiled project of 'Street Fighter III: Third Strike' (300,000 lines of C code) within three weeks, resulting in 2903 commits and modifications to 252055 lines of code. The Code Health score rose from 5.6 to 10.0, with token costs of approximately $4000. Correctness was ensured by leveraging quality signals from the CodeHealth MCP Server and a replay trajectory testing framework; the agent autonomously accumulated 22 refactoring strategies and 82 explanations. Claude Opus and Claude Code were primarily used, with comparisons showing superior pattern recognition capabilities compared to Codex + Sol. Changes were merged into a branch version of main via 54 pull requests. The article also presents multi-angle skepticism from practitioners regarding applicability, architectural impact, test completeness, and model-tool coupling.
Meta introduced ZGateway as a stateless proxy layer for ZippyDB, handling over 1 billion operations per second, carrying approximately 40% of traffic, and reducing the total number of persistent connections to about 1/19 of the original count. The gateway mitigates file descriptor and memory exhaustion issues caused by many-to-many client connections through features such as sticky connections, ServiceRouter integration, request batching and merging, tenant-level admission control, and local caching. The additional network hop has been proven to reduce database node connection management overhead and improve overall latency.
Radicle disclosed two critical vulnerabilities in its core network protocol: radicle-node discards the encryption state after the Noise XK handshake, causing all subsequent communications (including Gossip metadata, routing tables, and Git object packs) to be transmitted over TCP in plaintext; simultaneously, a node ID spoofing vulnerability exists, allowing attackers to impersonate whitelisted nodes to pull private repositories.
Artifactory has three vulnerabilities currently being actively exploited: CVE-2026-42018, CVE-2026-42016, and CVE-2026-82329. Attackers can obtain an anonymous user JWT by sending an unauthenticated POST request to /access/api/v1/aws/token/ (with a trailing slash), then use that JWT to exchange for an admin-scoped token at /access/api/v1/tokens, gaining persistent administrative control within five minutes; JFrog has released fixed versions 7.111.21 and above.
Plimsoll is a self-hosted sandbox system designed to securely run untrusted code generated by AI agents. It supports four levels of isolation strength: process, container, kernel, and VM. Each execution result includes the actual isolation level achieved. Requests can specify a minimum isolation level (floor); if this threshold is not met, execution is rejected. A broker mechanism ensures that API credentials never leave the sandbox, granting permissions only for single-run executions. The project has integrated with mainstream agent frameworks including Vercel AI SDK, Google ADK, CrewAI, and Trigger.dev. The project is open-source under the Apache-2.0 license and is currently in a Pre-1.0 state developed by a single author.
BraiNOTES is an open-source in-browser handwritten note-taking application designed specifically for lecture scenarios. It supports writing directly on PDFs or viewing handouts side-by-side. Features include various pen types, paper templates, whiteboards, PDF cropping, vector export to PDF or PNG, and .brainotes file backups. There are no accounts or servers; all data remains in the browser, supporting offline use. The tech stack is based on SvelteKit, pdf.js, pdf-lib (pdfme branch), IndexedDB, and perfect-freehand. The project is hosted on GitHub under the MIT license.
abersheeran released the open-source project gits, which runs entirely on Cloudflare Serverless services without requiring a self-hosted server. It supports automatic synchronization of all repositories from authorized GitHub accounts. The project originated from the author's exploration of agent-based programming paradigms but was later streamlined into pure Git hosting functionality, excluding traditional features such as issues and pull requests. It is applicable to private repositories that users prefer not to upload to GitHub or for backing up GitHub private repositories.
Seodata.dev 1.1 is now live, providing keyword search volume, cost-per-click (CPC), and competition data covering more than 50 countries. It offers 20 free queries per day without requiring an API key; it supports direct invocation by agents such as Claude, ChatGPT, and Grok, with documentation located at seodata.dev/llms.txt; data sources include Google Ads, DataForSEO, and SerpApi, featuring a unified response format and caching of static metrics to reduce costs.
Shpyrd is a new cloud platform for AI applications that supports one-click deployment via agent tools such as Claude Code, Codex, or Cursor. It automatically configures databases (Postgres), caching (Valkey), and persistent volumes, and provides single sign-on without writing login logic, on-demand hibernation, fine-grained resource billing (instances as low as 64 MiB, CPU billed by usage), and complete monitoring logs.