OpenAI's developer account launched the kitchen show "COOKED," using the Codex Ultrafast tier to generate an 8-bit pixel-style ramen simulation game in real time on the ModRetro handheld; the show integrates a full-duplex voice sous chef driven by GPT-Live-1, capable of listening and speaking simultaneously while recognizing natural pauses; it also demonstrates the dot persistent Agent organizing recipes, the Codex CLI full-screen multi-Agent interface, and cloud continuous execution capabilities.
SplatLab is an autonomous scientific research system for 3D Gaussian Splatting (3DGS) research. It can automatically reproduce baselines, propose extensions, conduct training and ablations (covering Mip-NeRF 360 and LLFF), write manuscripts, and bind every reported numerical value to verifiable logged training runs, based on research prompts, target paper URLs, or no input.
In June 2026, the Stanford HAI Digital Economy Lab launched the AI Economic Indicators real-time dashboard, comprising three modules: Transformation Tracker, Canaries Dashboard, and Adoption Monitor. Updated monthly, it integrates statistical data from private enterprises, academia, and government to track AI's impact on employment, wages, consumer surplus, hiring skills, and AI usage rates. The Canaries Dashboard shows that the annual decline in employment for the 22–25 age group in software development and customer service roles reached 4.4%. The Adoption Monitor tracks generative AI adoption rates based on surveys, while the GDP-B: Surplus Observer estimates consumer surplus for goods such as AI tools through online surveys.
Michael Lynch officially released his independently developed book "Refactoring English: Effective Writing for Software Developers," debuting the print edition with a limited-time 30% discount. He simultaneously improved typography (e.g., enabling font pairing kerning), refactored the writing workflow (developing a custom Markdown to Asciidoc compiler), adopted Lulu Direct Buy Button for print-on-demand sales, and disclosed specific limitations of this tool regarding order APIs, pricing flexibility, and testing processes.
PingCAP confirmed that its internal systems were accessed without authorization, with some data copied and source code stolen; the hacker group FulcrumSec claims to have stolen approximately 10.6TB of data, including 180,046 tenant organizational records, 115,308 API Keys (claiming 42,612 are still valid), 20.65 million customer SQL statements, and 413 private code repositories; PingCAP stated that the exported data included user emails, organizational information, SQL editor text, and business information, but confirmed there was no vulnerability in the TiDB kernel and that production databases for customers in China were unaffected; four vulnerabilities were fixed on October 14.
Carnegie Mellon University Professor Andy Pavlo stated on The MAD Podcast that frequent database deletions by Agents are not a new problem but rather stem from neglected legacy database permissions and sandbox mechanisms. He proposed treating Agents as 'toddlers' requiring guardrails, implementing protection through branching, least privilege, and isolation from production databases. Vector databases have shifted from being a differentiating advantage to becoming standard features, evolving toward either general-purpose systems or add-on components. Off-the-shelf text-to-SQL solutions currently achieve approximately 60% accuracy; major banks required millions of dollars in investment to reach 99.5%. LLM-based database tuning can achieve 85% effectiveness within 15 minutes, far surpassing traditional methods. Over 60% of open-source databases now feature commits involving Agents. He joined ClickHouse to establish a research lab focused on the impact of Agentic workloads on database architecture.
Volcano Engine AI MediaKit has launched a super-resolution solution for stage giant screen scenarios, capable of enhancing 720P/1080P video materials to 8K resolution for playback on 20-meter giant screens. Core technologies include frame-by-frame texture generation, edge detail reconstruction, and temporal consistency guarantees; this solution has been practically applied in the 2026 Douyin Wonderful Night and the 2026 CCTV Spring Festival Gala, supporting up to 16K/32K enhancement and adapting to various ultra-large screen scenarios such as cultural tourism exhibitions, digital art, and commercial large screens.
Private keys for eight eMudhra intermediate CA certificates were leaked. These certificates chain to root certificates trusted by programs such as Mozilla, Microsoft, Apple, and Chrome. Linux distributions need to update the ca-certificates package; in the Python ecosystem, uv 0.13.4, pip 26.2.2, and certifi 2026.10.19 have released versions containing updated certificate lists.
A non-programmer author developed and open-sourced the "Read-Along Partner" v0.5.0 Mac client for Xiaodaren read-along pens. It supports the installation of DAB/UNB read-along packages, management of installed content, multi-select batch removal, and safe ejection. It is applicable to Macs with M-series chips and macOS 14+. Currently, only the Xiaodaren 908 model has been tested; as it is unsigned, users must manually bypass macOS security prompts.
On October 20, Google launched Gemini Business Home for small business owners, offering centralized management features for tasks, finances, and marketing. It supports connections to Google Business Profile, Google Ads, Google Workspace, as well as third-party applications such as Shopify, Wix, and Canva. Merchants can upload brand assets and use image generation models to batch-create visual materials that align with their brand tone. They can also convert tables and receipts into financial charts.
AHP Matrix is a locally running open-source decision tool that implements the Analytic Hierarchy Process (AHP). It supports pairwise comparisons, consistency checks, multi-user weighted merging, weight drag-and-drop simulation (What-if), Excel template import/export, and hierarchical criteria management. Released as a single binary, it supports macOS, Linux, and Windows (Intel/ARM). Project data is saved as local .ahp.json files. Installation via Homebrew and build instructions for Go+Node source code are provided. The license is MIT.
Developer Gathin T. used Claude Code via ADB to disable 25 pre-installed packages on his parents' Xiaomi TV and open-sourced the cleanyourtv project; this project already supports 107 Android TV/Google TV models, using non-root methods to disable rather than delete apps, accompanied by a one-click restore script; there are also similar projects like dumbtv targeting Fire TV, with some users achieving a 90% memory boost, custom interfaces, or real-time 3D home displays.
MCP Census is a publicly operated evidence-based verification service independently run by Jaco Veldsman (Australia), providing real-time health and identity status tracking for 48,797 MCP servers.
OpenAI has acquired startup Orchestra Bio, integrating its AI agent platform and team into the Rosalind life sciences product line; Orchestra's agents can automatically track cascading issues in drug development projects (such as production batch failures) and compare response options, with scientists making the final decisions; the transaction amount was not disclosed; this move continues OpenAI's expansion in the biopharmaceutical sector, having previously launched the GPT-Rosalind model, Rosalind Workbench tools, and ChatGPT Health, and acquired companies such as Torch and Crixet.
i14 released the Laplace 3B model, which uses Gated Linear Attention (GLA) to replace softmax attention, fixing inference state at 29.5 MB per sequence and reducing training complexity from O(T²) to O(T); the model was trained using native PyTorch (TorchNeuron Native) on AWS Trainium2 trn2.48xlarge instances, achieving a training throughput of approximately 120,000 tokens/s and a serving throughput of 11,544 output tokens/s (with 1,280 concurrent sequences); the article elaborates on key engineering adaptations such as GLA core design, Muon distributed optimization, SBUF memory alignment, and compilation-level backward unrolling.
Simon Willison released Battle Drones, a browser-based real-time strategy game supporting both mobile and desktop platforms. It features 9 types of buildings, 7 ground vehicles, 6 air units, 3 infantry unit classes, and 3D models. The game offers single-player modes (including Last Stand survival battles and an 8-level Ops campaign), AI vs. AI matches, Agent vs. Agent matches (where LLMs such as GPT-6 Luna and Claude Haiku 5.5 can battle each other), multi-device local online play with synchronized audio (based on the Chirp Pong experiment), and newly added WebSockets online multiplayer (implemented via Fly.io relay servers). The entire project was coded by AI agents such as Claude Code; the author provided only core prompts like "Make it a LOT more fun" and did not manually write the main logic.
SF Compute CEO Evan Conrad introduced the company's compute trading marketplace at an AIE NYC event: it supports buying and selling GPU hours and forward contracts, with the core being physical settlement (rather than cash settlement), i.e., actual delivery of available computing resources; the platform allows users to subdivide InfiniBand networks on demand and match orders based on constraints such as different hardware generations (e.g., B300), interruption tolerance, and storage configurations; the company does not own hardware itself to maintain neutrality, has achieved hundreds of millions of dollars in transaction volume, and is currently connecting new clusters to the order book.
ElevenLabs has released two speech-to-speech models: Eleven Realtime v1 Max and v1 Mini. The v1 Max model scored 98.8% on the Big Bench Audio benchmark with a response time of 0.57 seconds; ElevenLabs states that no model currently on the Artificial Analysis leaderboard combines higher accuracy with lower latency. The v1 Mini model is designed for low-cost natural conversation and supports detection of pauses and filler words. Both models are available via the ElevenAgents platform and an OpenAI-compatible Realtime API, billed by minute plus tokens, with v1 Max being more expensive.