This week in AI — Oct 19 – 25, 2026

DeepInfer is an open-source, high-performance inference framework developed by DeepSeek-AI. It powers DeepSeek's large-scale production services and is now available for offline inference, offering core features essential for serving at scale. The release is seen as beneficial for the developer community, allowing them to study and potentially improve existing custom inference engines like gufo, halogen, strixite, and Strata.

Mistral's Timestral models are time-series foundation models designed for efficient inference without a GPU. On a standard server CPU with 8 threads, the 5M model forecasts one day ahead from three weeks of 15-minute history in 10 ms, while the 209M model takes 140 ms. Smaller models like the 28M are five times faster than the 209M model, and the 123M model achieves a score within 0.2% of the 209M model at three quarters of its latency. Timestral is released under the Modified MIT license.

Mistral Large 4, released in Public Preview v26.10 on October 6, 2026, is a state-of-the-art, open-weight, general-purpose multimodal model. It utilizes a Mixture-of-Experts architecture with 52B active and 1.05T total parameters, plus a 1.6B vision encoder. Key features include Structured Outputs, Function Calling, Document QnA, and a 1M context window, with pricing at $1.36 per M input tokens.

Voxtral 3, from mistral.ai, is a real-time ASR preview system that supports various transcription models including Voxtral Small Transcribe 3, Voxtral Mini Transcribe 3, and GPT-4o Transcribe Enhanced. It also integrates with other ASR technologies like Gemini 3.5 Transcribe, Universal-3.5 Pro, and Whisper large-v3. The system was evaluated using 237 prompts with two voices per system (Sarah and James) on a 1 to 5 scale, judged by gemini-3.8-flash, with pricing based on vendor pages and generated on 2026-10-20.

Inception Labs has introduced Mercury 3, a new diffusion-based LLM that outperforms GPT-6 Luna and Claude Haiku 5.5 in benchmarks, running 4-8x faster at over 1,000 tokens per second. Mercury 3 is 40x larger than Mercury 1, yet maintains its speed, demonstrating that text diffusion scales effectively. It offers a 1M token context, up from 260K for Mercury 2.5, and is priced at $0.20 per million input and $0.70 per million output tokens, with a 50% launch discount. Early partners are already using Mercury 3 for various latency-sensitive applications.

Hugging Face has launched SimReady Central, an open library offering over 2,000 simulation-ready (SimReady) OpenUSD assets for robotics and AI factory digital twins. These 3D models include essential properties like collision geometry, mass, and friction, which are crucial for simulation workflows and reduce manual setup. The library provides a free starting point for physical AI training environments, featuring a range of infrastructure and equipment assets that developers can combine and reuse across various experiments.

NVIDIA Kumo Forecast, previously NV-Tesseract Forecasting, is a multi-variate time-series forecasting model. It integrates a pretrained Transformer backbone with Domain-Aware Representation and Retrieval (DARR), which enhances predictions by blending direct model outputs with historical data patterns. Built on the MOMENT-1-large backbone, Kumo Forecast features a 340-million-parameter base encoder and a trainable forecasting head, offering interpretability. The Kumo-TS repository also provides fine-tuning examples for user-supplied CSVs.

Harness Engineering 101 likens the LLM to a brain and the harness to a body, addressing common challenges in interacting with LLMs. It covers issues like communicating with LLMs, understanding their behavior as next-token black boxes, and enabling them to interact with the world through JSON-mapped functions. The text also discusses the agent loop for multiple tool calls and the importance of caching to manage the expense of resending data.

Strot is an open decision compiler that allows users to compile an LLM's intent once and run decisions locally forever. It demonstrates significant improvements in routing accuracy, achieving 92.5% compiled versus 74.8% naive, and semantic text accuracy, with 86.7% compiled versus 72.7% naive. Strot also boasts low latency (~60 ms), zero marginal cost, and no LLM calls per record. The project is open-source under the MIT license, with its default decision model under Apache-2.0, and implements ideas from STROT (arXiv:2505.01636).

DeepSeek offers Search and Fetch APIs, providing agents and RAG applications access to live web content at $0.75 per 1000 successful requests. These read-only APIs have a QPS limit of 5 for both Search and Fetch, calculated at the account level. Various HTTP status codes indicate different issues, such as 400 for invalid requests, 401 for missing API keys, 402 for insufficient balance, and 429 for exceeding QPS limits.

Claude Duet is a proactive agent powered by Opus 5.5, available in Pro and Max plans in the US as a research preview. It takes on tasks, works autonomously, and waits for user approval when necessary. Built for judgment, it briefs users daily, drafts next steps, and can set up meetings or build dashboards. It operates securely in a user's cloud computer, ensuring privacy, and integrates with existing Claude account memory, skills, and apps.

Claude identified a significant attack surface reduction opportunity in OpenSSH's 10.4 release. It found approximately 4,500 lines of new, experimental code for the ML-DSA-44 post-quantum signature algorithm that an unauthenticated client could reach, even when the algorithm was not enabled or explicitly excluded. This complex, memory-unsafe code, representing about 16% of OpenSSH's pre-authentication code, was reachable because the daemon parsed key types before checking if they were allowed. The issue was reported to OpenSSH developers, and the 10.5 release addressed it by moving the check before parsing.

WhatsApp MCP is now available for free on Cloudflare, allowing AI agents to read WhatsApp messages. The system uses a Worker (src/index.ts) to handle requests, including a fresh McpServer for each /mcp request. It also features a UI for landing, sign-in, and pairing. A daily cron job pulls data via RPC to a WhatsAppStore Durable Object, which uses SQLite for auth keys, messages, chats, contacts, and an FTS5 index. The ingest process is idempotent, ensuring safe relinking without duplicating data.

Claude designed 211 minibinder proteins against scorpion neurotoxins, with 17 meeting binding criteria. Working from an initial prompt and four follow-ups, Claude autonomously generated designs, including the strongest binder Cn2-B1. The designs were experimentally tested by Adaptyv Bio and Twist Bioscience. Of the 60 single-toxin designs, 11 were binders (18%), a rate consistent with typical protein design campaigns.

Penmentor is a tool that allows users to get handwritten math marked up by AI models like Claude, Gemini, or GPT. By pressing a key while working, users can have their work reviewed. It supports various agents such as claude:opus, claude:sonnet, claude:haiku, and codex:gpt-5.6-terra, as well as Antigravity CLI agents like agy:gemini-4-argon-medium. Penmentor is released under the MIT license.

NVIDIA introduces Cosmos3-H-Dreams, an interactive generative simulation designed to bring clinical surgery to life with convincing in-body imagery. This technology allows for real-time interaction within generated surgical scenes, such as manipulating a Cadiere grasper to lift a gallbladder. The system adapts a bidirectional teacher over 24 future frames, extending to 72 frames (7.2 seconds at 10 Hz) to learn surgical dynamics, offering a new platform for research and community development in surgical simulation.

Yara Reads is an open-source early-reading app designed for children who have learned phonics sounds. It provides stories where every word can be sounded out, offering interactive help by tapping words to reveal sound buttons. The app checks stories word-by-word against known sounds, allows users to bring their own books via agents like Claude or ChatGPT, and enables parents to customize known sounds and tricky words. Built with Cloudflare Worker, D1, and R2, it's available for tech-savvy parents to run and customize.

Jev Cards is a new open-source quiz game built around TypeSafe's Jev model, which possesses vision capabilities. The game challenges players to beat Jev, a vision model that can analyze a picture and a question to select an answer or tap a location in approximately a quarter of a second. The cards for the game were created using Opper's vision models and Claude Code.

Anthropic announced the creation of a "public benefit fund" to support projects for societal good. The company will commit 5% of its shares to this fund, with the option to add another 2% for every 50% increase in its value. This fund, structured as a wholly owned entity, aims to provide increased flexibility and scale for initiatives like large donations to nonprofits or direct distributions to specific populations. Further details on its structure and governance will be shared later.

Anthropic, a Public Benefit Corporation (PBC), is committed to responsibly developing and maintaining advanced AI for humanity's long-term benefit. This designation requires balancing shareholder financial interests with broader public benefit. Their Public Benefit Report details current efforts and future work to serve this public benefit, reflecting the company's core charter.

Democracy Now! presented a discussion titled “When the Bubble Bursts”: Naomi Klein & Astra Taylor on Reclaiming Democracy from AI Oligarchs. This independent global news hour airs Monday through Friday on over 1,500 TV and radio stations. Viewers can watch their livestream at democracynow.org from 8-9 a.m. ET.

Protests against artificial intelligence development occurred in over 50 US cities on Tuesday, October 20, with demonstrators advocating for a moratorium on AI. Public concern is high, as 71 percent of Americans anticipate job losses due to AI. This widespread fear and uncertainty are now clashing with record investment in AI within Silicon Valley, as reported by FRANCE 24.

Steve Forbes argues that concerns about artificial intelligence and data centers are overstated, emphasizing the benefits of AI. Forbes, a platform covering entrepreneurship, wealth, technology, business, and lifestyle, highlights success and people. They offer various ways to stay connected, including newsletters, Facebook, Twitter, and Instagram, providing more content on forbes.com.

DeepSeek has been banned from U.S. federal civilian agencies following a new order from the Department of Homeland Security (DHS). This directive, detailed in a document found at public-inspection.federalregister.gov/2026-21447.pdf, restricts the use of DeepSeek within these government entities. The ban signifies a significant policy change affecting technology procurement and usage across federal civilian operations, as reported by reuters.com.

A former OpenAI engineer, accused of planting a listening device in an executive conference room, was reportedly in contact with a member of the Saudi royal family. Federal prosecutors stated in a court memorandum that the engineer was planning to depart the U.S. for Riyadh when he was apprehended. This information was reported by CBS News, with further details from Nicole Sganga.

A hacker successfully breached over 100 websites, including those of John Deere, SharkNinja, and Cummins, utilizing an army of AI agents. According to AI security firm Norforge, these AI agents largely operated autonomously. The attacker, identified as the most active among exposed Cairn instances, targeted companies in the US, Canada, Australia, and Japan. This hacker detailed their preferred intrusion tactics in under 100 Chinese characters, with successful breaches typically taking around 23 hours.

An experiment demonstrated how easily fabricated statistics can be cited by large language models (LLMs). Using GPT-4.1, 50,000 invented statistics pages and 15,000 listicles were generated and published on stateglobe.com for about $50. These pages covered topics like SEO, content marketing, and e-commerce across 200 countries. The experiment highlighted that LLMs often prioritize specific, quotable numbers as evidence, even if those numbers are entirely made up, adding little real authority.

The video "OpenAI Just Hacked 70% Of The Internet" by Nick Saraev of LeftClick discusses critical web security flaws, including PHP's critical vulnerability, and the global surge in AI-enabled cyberattacks. It covers how AI uncovered major security issues, the shrinking lead of AI security projects, and preparations for ransomware. The video also provides advice on reducing personal digital attack surfaces, strengthening account protection beyond SMS, and how businesses and families can reduce their exposure to AI-powered impersonation and disruptions to critical services.

Sam Altman discusses AI consciousness, recursive self-improvement (RSI), and AI's potential to accelerate scientific breakthroughs, including new medicines. He also addresses the future of AI devices, the challenges of leading OpenAI, and the need for a new renaissance. Key topics include AI's impact on coding jobs, personal AI assistants, the AI consciousness debate, and OpenAI's focus on chips and cheaper compute for new kinds of computers.