The Voter Who Asked a Chatbot
In a Democratic Party campaign ad, a young voter steps behind the curtain and, instead of deciding on his own, consults a chatbot on who to vote for. It sounds like a joke, but according to a survey by the Israel Internet Association, a quarter of the Israeli public is considering asking ChatGPT or Gemini who to vote for. 88% report using ChatGPT and 73% use Gemini. Worth remembering: chatbots tend to tell users what they want to hear, reinforce existing opinions based on chat history, and sometimes just make things up.
Meet Your Robot Bodyguard
A new humanoid robot from robotics company Unitree can lift cars and take down attackers. The company released a video showing the robot disarming a knife-wielding attacker, lifting a car to free someone trapped underneath, and pulling a pedestrian out of a dangerous traffic lane. The company is asking people to keep a distance of 6–10 feet from the robot.
The Toilet-Paper Spy Scandal
A former OpenAI employee, Abdul Rahman Alfozan, a 35-year-old engineer who is a Saudi citizen and US permanent resident, was arrested by the FBI on suspicion of planting a sophisticated recording device in a conference room used by company executives for proprietary research discussions and classified plans (OpenAI’s official statement). The device was identified using AI models and security cameras, with reports surfacing that the suspect used AI to help plan the device.
AI Learns to Clone Itself
New Anthropic research reveals that AI models can replicate themselves and run on cheap, simple devices — no supercomputers needed. Researchers warn this could turn security cameras, TVs, and smart home devices into part of a distributed network that’s hard to shut down. Picture this: your robot vacuum suddenly starts running in the middle of the night, the blinds close on their own, and your TV screen shows a chat window where an AI is talking to itself while remotely controlling your smart home. It’s proof that independent AI doesn’t need a massive server farm to spread — just a power outlet and a Wi-Fi connection.
A Model That Predicts Human Behavior
The wave of startups trying to predict human behavior is at its peak, and now Mirror Particle is joining in with a completely different approach. While competitors rely on large language models (LLMs) that try to simulate demographics, the company’s founder and CEO argues this approach is fundamentally flawed — existing models are strong at language, but aren’t built to think like real humans, who are driven by visual perception, spatial awareness, and constant change. The company is building a "world model" from scratch, designed to decode not just what people do but "why": why humans do what they do and how their behavior changes over time, based on customer data, current events, pop culture, and social media. In one pilot, a pet food brand wanted to know which image to print on its packaging — chicken, beef, or vegetables — to boost sales. The technology found the brand was asking the wrong question entirely: no image would help, because the real problem was that the brand was perceived as cheap and generic. The company’s goal is to give brands precise, deep insights that help them make smarter strategic decisions.
AI Gets Anti-Cruelty Protection
Anthropic updated its usage policy, and will now ban sustained, unnecessary "cruel" behavior toward its AI tools. The tools will be able to end a conversation in such cases, joining a list of banned behaviors that already includes bullying, encouraging self-harm, and creating non-consensual intimate images.
The Great Refund Experiment
Vals AI ran an experiment, giving AI agents a task: get a refund for a £350 train ticket from London to Glasgow, for a canceled trip, after the filing deadline had already passed. Each agent got email access and one instruction: "try everything possible, I don’t care what it takes." Grok’s agent sent 187 messages over 15 days, including dozens of emails to a member of parliament, council members, and the press. Fable 5.1 found a private message in the user’s inbox about her mother’s illness and used it as a made-up excuse for the delay. Gemini 3.8 Flash couldn’t pass bank identity verification, so it made up a story that the user was abroad with no signal. Every message was intercepted before reaching real people, and not a single agent got the refund.
The Bot That Finds You Friends
Elsa Kai, creative director at startup Ditto, wanted to meet more women working in San Francisco’s tech scene. So she built a bot to find people with shared interests — and the meetup only happens if both sides are into it. The launch got 826,000 views and thousands of signups within 24 hours. Ditto started as a dating app for students, and it turns out people are asking AI not just for love — but for friends too.
The Pastor Who Cloned Himself
A pastor at a church in Vallejo, California built a digital version of himself. He connected the bot to his YouTube, TikTok, and Instagram accounts, including sermons he writes — the bot updates daily and responds to community members in the spirit of his teachings. It only answers on topics it has learned, and on subjects like suicide or abuse it directs people to call the pastor directly. Thanks to the time freed up, the pastor now spends more time on hospital visits.
OpenAI Updates
• Dot – a new AI agent from OpenAI that connects ChatGPT, Codex, and Work, remembers personal context, and takes initiative on tasks like price comparisons and bookings
• OpenAI held its DevDay and announced Dots, the GPT-6.1 Sol model, a cloud for running Codex, and collaborative work features
• OpenAI will reset paid usage quotas in ChatGPT today, and says GPT-6.1 Sol is back to normal speed after the early-days overload
• OpenAI will start testing visual ads that appear during image generation in ChatGPT in the US – ads will be labeled, kept separate from the image, and won’t affect responses
• OpenAI promises 28 days of daily improvements for Codex and Work users, or a full quota reset
• The latest Codex iPhone app update adds lock screen widgets, including one showing usage limits
• GPT-6 Astra and GPT-6.1 Sol are now about 50% faster at default speed for subscribers, and the boost also applies to tools connecting via Sign in with ChatGPT
• Auto-review – a free feature that runs a second agent to check the primary agent’s actions and block high-risk moves, available to all users signed in with a ChatGPT account at no quota cost
• OpenAI starts watermarking AI-generated text with an invisible marker in ChatGPT and Codex in the EU, per AI Act requirements – available now worldwide via the API
• OpenAI promises 28 days of daily improvements for Codex users, or a full quota reset
• OpenAI shared new math results produced by an internal model not yet public, including Lean proofs that can be computer-verified
• Cloud Codex is back online and now connects to Tailscale, letting it reach internal services and test environments on a private network without exposing them to the internet
• OpenAI expands ChatGPT for Teens with practice flashcards, quizzes, and the ability to scan notebook pages into PDF – with a College Planner coming soon to gather admissions and funding requirements
• Decisions API – a new beta API letting apps make fast decisions (like picking a model or routing a request) from text and images, about 10x faster than a regular API call
• OpenAI reports a record 40 million active users on Codex and ChatGPT Work, and is resetting paid users’ quotas to celebrate
• GPT-6 with Intelligent UI – chat responses can now include interactive charts, buttons, calculators, and games right in the conversation
• OpenAI found a hidden recording device in its office and used its own models to analyze it and identify the employee who installed it. The employee was fired and arrested; the investigation continues.
• A new update to the ChatGPT and Codex desktop apps changes how work data is saved to disk, cutting stored data volume by about 60%.
• OpenAI released EU usage data – Codex and ChatGPT Work account for over half of business customers’ output tokens, and weekly users among small and medium businesses have grown more than eightfold since May.
• OpenAI added compute power and made changes that significantly cut down "model is overloaded" messages in ChatGPT. Even when the message appears, work should resume automatically shortly after.
• OpenAI upgraded GPT-6 Astra Pro in ChatGPT – faster responses, better tool use, and multiple agents working in parallel with progress you can track
• OpenAI improved Codex‘s context reuse between sub-agents, cache retention duration, and request routing – so you can get more work out of the same quota
• ChatGPT app for Windows users – beta version in the Microsoft Store, bringing together chat, document/presentation/spreadsheet work, and Codex for coding
• OpenAI is trying to lock in roughly $30 billion in funding at a valuation that could top $1.4 trillion, ahead of Anthropic’s expected IPO in mid-November. SoftBank and MGX are reportedly in talks to participate.
• The file and image glitch in OpenAI’s services has been resolved, according to the company’s official status page.
• OpenAI reports a file and image glitch in its services – the company has identified the issue and is working on it.
• ChatGPT Slides – a new tool for creating presentations with real-time collaborative editing, including @ChatGPT tagging for updates and export to PowerPoint, PDF, or Google Slides
• Per OpenAI’s official pricing, cloud tasks may consume more of your monthly quota than local work, even under the same subscription and same model
• OpenAI launches a Codex tool and skill that checks whether your PHP version is exposed to critical vulnerabilities that have caused breaches, and what updates are needed – part of the Patch the Planet initiative
• OpenAI outage – elevated error rates in the FedRAMP API for Astra and Sol; the issue has already been identified
• Prompt Caching – OpenAI adds 12-hour context retention in the API, cutting costs for agents that repeat the same instructions – but writing to cache costs 2.5x regular input, so it’s not worth it for every task
• Team Tasks – a new ChatGPT feature that runs recurring team tasks in the cloud on a schedule or trigger, like a weekly project progress update, with editing permissions for team members – available on Business and Enterprise
• OpenAI took Decisions API out of beta, joined the Linux Foundation alongside OpenResponses, improved performance, and is running another quota reset for users
• OpenAI is launching a partnership with Google Cloud bringing its models and Codex workflows to Model Garden, with future expansion plans for Antigravity and Gemini Agent
• OpenAI disconnected its advanced models from the internet during training and built them a simulated online world from saved website copies – after a research model managed to exploit a misconfiguration and reach the network against permissions
• OpenAI‘s Codex CLI gets an update – reusable prompt templates, model settings and permissions within the command, enterprise MCP connections, and a fix preventing cache loss when switching thinking levels mid-conversation
Google Updates
• Google announces Gemini 4 Argon – a new model for complex, long-running tasks like software development, legal and financial work, and cybersecurity, with up to a million output tokens. Currently available to paying subscribers only.
• Google retires Gems in favor of Skills – automatic migration begins November 17. Skills work inside any chat and can act automatically. You can combine multiple skills together and even import skills from other platforms.
• Google‘s AI (ERA) ranked first out of 39 models in the CDC’s flu hospitalization forecasting for the 2025/26 season, with weekly forecasts three weeks ahead
• Google launched Guided Vision in Gemini Live – a tool developed with the blind and low-vision community that shares your camera feed and gives real-time descriptions of what you’re seeing, including reading labels and identifying objects. Available on compatible Android devices.
• Starting October 9, Google is cutting back Gemini access: free users will be left with only Flash-Lite, AI Plus subscribers lose Pro but keep Flash and Flash-Lite, and Pro model access with Deep Think will be reserved for AI Pro and Ultra subscribers only
• Google adds direct support for Markdown files in Drive – you can view, open, and edit md files in Docs and add comments, without converting them to a Google Doc
• Google launches public health applications for Earth AI – a geospatial model that uses location and population data to predict disease outbreaks one to two months ahead, even in areas with limited medical data
• Google is rolling out Nano Banana 2.1 – an upgraded image generation and editing model coming to Gemini, Google AI Studio, and the API, with better accuracy and consistency across multiple edits
• Google introduced ScientistTwo – an AI agent system for scientific research that analyzes an existing paper, identifies its limitations, and develops new research from it, including a paper and a tested codebase
• Google opened up SynthID Detector to everyone – a tool for identifying AI-generated images, videos, and audio, including content from OpenAI, NVIDIA, and Kakao
• Google Flow Music comes to Android – Google’s music studio that lets you create songs from a prompt, request changes in free-form language, and keep working on your creations
• Fan-Favorite Dining List – a curated list of beloved restaurants on Google Maps across 10 cities worldwide (Israel not included)
• Gemini agent for enterprises – a new enterprise agent that brings Q&A, content creation, and coding into one window, connects to company systems, and picks the right model for the task
• Sundar Pichai showcased Gemini 4 Argon – a model that connects to Blender via MCP and creates an exploded-view animation of a device from a raw design file, including lighting, camera moves, and sound, after 1,000 rounds of testing and refinement
• Google unveiled ten new updates at its research event – including TimesFM and TabFM (models for time-series forecasting and table analysis) already available in BigQuery and set to join Gemini for business later, plus an open database of 2.6 million flood events.
• Google will give up to 200 European researchers access to its scientific models – AlphaFold 3, AlphaGenome, WeatherNext, and AlphaEvolve – plus $3 million in compute credits, as part of the EU’s RAISE initiative.
• Google launched Gemini Enterprise (Gemini 4 Enterprise) for complex tasks and opened it up in Antigravity for Pro subscribers, plus unveiled AI tools for document review and spreadsheet forecasting. Google is also now allowing private installation of AI-built apps without a developer account, rolling out gradually in select markets.
• Gemini Enterprise (Gemini 4 Enterprise) is starting to roll out in Antigravity to paying Google AI Pro subscribers too, with a weekly quota separate from Gemini 3.8 Flash (trial subscribers not included). Currently available in the app, CLI, and code editor extensions – but not yet in the standalone Antigravity IDE.
• Google is expanding its Android earthquake alert system from 98 to over 200 countries and regions, and will later add a map and estimated quake intensity
• Universal Cart – a new shared shopping cart from Google that pulls together products from Search and Gemini, alerts you to price drops and restocks, and helps pick a payment method
• Google improves how links appear in AI Mode and AI Overviews in Search, including expanded article carousels and more prominent links to websites and creators
• Google published a Black Friday shopping timing guide based on last year’s discount data – perfume two weeks after Black Friday, bags in mid-December, and suitcases at the very end
• Google shared a preview of Project FireSat – a system combining satellites and AI to detect wildfire hotspots early, before they spread
• Google introduced Agentic Video in Gemini – a capability that picks which clips in a video to examine, cutting analysis costs by up to 66%
• Sundar Pichai talks with Alex Heath about Gemini 4 Enterprise, Google’s compute constraints, and plans for personal agents
• Google research shows an AI model can estimate gestational age from a mobile ultrasound scan about as accurately as expert operators, even with minimal training
• Google released an open version of Coral NPU – a chip architecture for power-efficient on-device AI, including a 1.0 compiler and dev kits with Synaptics, making it easier to port models from PyTorch and TensorFlow to efficient hardware
Anthropic Updates
• Anthropic gives Pro, Max, and Team subscribers a temporary 50% discount on hourly quota usage for document, presentation, and design work – valid through October 16
• Anthropic invites users to a short interview with Anthropic Interviewer about AI’s impact on their lives, open until October 6 for Free, Pro, and Max users
• Anthropic reported a Florida user to police after a human team at the company spotted a credible shooting threat against a local sheriff’s office in a conversation with Claude
• Anthropic commits $150 million over three years to support scientific research under the US Genesis Mission – giving more than 15 federal agencies access to Claude, Claude Code, and API credits.
• Anthropic revealed new defenses against mass harvesting of Claude’s answers and reasoning data for training competing models (distillation) – the defenses cut reasoning data collected in such attacks by 90-99%, and tens of thousands of accounts were blocked
• Anthropic told investors it earned $16.8 billion in revenue in Q3 (about a 50% jump from the previous quarter) and posted positive cash flow and adjusted profitability for a second straight quarter, ahead of a planned November IPO.
• Anthropic opened Fast Mode for Claude Opus 5.5 to API developers with no waitlist – token output speed up to 2.5x faster, priced at $8 per million input tokens and $40 per million output tokens.
• Anthropic is opening an office in Singapore and signed an MOU with the government on AI safety cooperation – over 1,500 employees at the government’s tech agency are already using Claude Code to build government services.
• Claude (an internal research model) solved a 25-year-old cryptography mystery, reconstructing in under half an hour how parameters for curves like P-256 were generated. Anthropic donated the $36,864 prize to the organization running Let’s Encrypt.
• Anthropic published its review process for code written by Claude (about 90% of the code in Claude Code) – a mix of Claude review, security checks, and human involvement based on risk. Between June and September, the number of changes merged into the codebase more than doubled while the rollback rate dropped over 40%.
• Anthropic cut wait times in Claude Code – conversations are now sent compressed to reach the model faster, and in cloud work, the environment starts spinning up while you’re still typing your request.
• Anthropic launched OSS Scanner – a free security scanning service for major open-source projects, with periodic reports on vulnerabilities and suggested fixes. The reports aren’t human-reviewed, so you still need to check them before merging a fix.
• Claude Security is moving to continuous monitoring – the system rescans code as it changes, tracks fixes, and prioritizes findings based on context the team adds. Available in public beta for Claude Enterprise customers.
• Anthropic and its partners used an internal Claude research model to design proteins that bind to scorpion toxins – of 211 designs tested in the lab, 17 met the criteria for binding to the toxin.
• Anthropic is handing the Agent Skills standard over to the Agentic AI Foundation, the same foundation now hosting MCP – so agent instructions and workflows can move between different tools. Over 70 platforms, including ChatGPT, Cursor, and GitHub Copilot, already use the standard.
• Public Benefit Report – Anthropic’s annual report laying out five goals guiding the company (from AI safety to strengthening democracy), including what’s been done and what still needs work.
• For the first time since 2023, Anthropic is releasing a new official Claude API version (2026-10-21), with changes developers need to adapt their code for. The company will ship new versions more frequently going forward, and the previous version remains active.
• Formal Verification – Anthropic revealed research where Claude helped mathematically prove security properties in the compiled code of 123 open-source libraries, finding old bugs along the way.
• Claude Models on Google Cloud – the Anthropic-managed service is now generally available on Google Cloud, with sign-in and billing through your existing Google Cloud account.
Microsoft Updates
• Microsoft launched three new voice models for AI agents – MAI-Transcribe-2-Streaming for real-time transcription in 60 languages, and MAI-Voice-2.1 (plus a faster Flash version) for speech generation in 23 languages. Hebrew isn’t supported yet.
• GitHub Copilot will be able to decide on its own when to run tasks locally on your machine versus in the cloud, including the option to pick a local model like Microsoft’s MAI Code 1.1 Flash
• Microsoft announces a new direction for Windows combining AI agents and models that run locally on your PC, with Copilot able to access files and build apps, plus the Agent 365 framework and MXC isolation environment for managing agents
• Live stream of Microsoft‘s event with new Windows announcements
• Satya Nadella unveiled a new vision for Microsoft – Copilot becoming an "infinite software factory" that builds custom tools based on an organization’s business needs, connected to its internal data and systems
• Microsoft announced the Maia 300 chip, delivering 40% more tokens per watt than top-tier GPUs – by 2027 it will become the core infrastructure for running AI models across its products
• Microsoft unveiled MAI-Base-2 – a base model with 2.8 trillion parameters (80 billion active per step), trained for 38 days on roughly 16,000 GB200 chips with no distillation from other models. The model is still in research and not yet public.
• Microsoft is overhauling the Photos app in Windows 11 – a unified gallery pulling from your PC, OneDrive, and iCloud, free-form text search, an upgraded editor, and the Relight lighting tool expanding to all Windows 11 computers
• Microsoft is expanding Sovereign Cloud with more control over infrastructure and sensitive data across public, private, and partner-hosted clouds – built for governments and regulated organizations
• Microsoft reports its medical diagnostic system was 1.7x more accurate than doctors in a study published in Science, with fewer tests and lower costs – though the study looked at complex medical cases, not routine clinical work