AI Updates on 2025-11-05

AI Model Announcements

Google releases Gemini 3 Pro Preview 11-2025, shipping in preview this month @legit_api
Google announces a 1.2T parameter model that Apple will use to power the new Siri, with Apple paying Google $1 billion annually for this partnership @AndrewCurran_
Apple Intelligence is revealed to be 150B parameters, and Apple is currently training their own in-house 1T model @AndrewCurran_
Google ships enhanced Structured Outputs for the Gemini API, now supporting recursive schemas with $ ref, anyOf union types, min/max numerical constraints, null types, and property ordering adherence @OfficialLoganK
OpenAI introduces IndQA, a new benchmark that evaluates how well AI systems understand Indian languages and everyday cultural context @OpenAI
Two 23-year-old Indian developers release Maya1, the #2 open-weight AI voice model globally, trained purely on free credits with 3B parameters, running on one GPU with 20+ emotions and less than 100ms latency @deedydas

AI Industry Analysis

OpenAI reports reaching 1 million business customers building with their platform @bradlightcap
Epoch AI releases new projections showing potential growth trajectories if OpenAI and Anthropic both reach their current projections, with Anthropic's most optimistic projection highlighted @AndrewCurran_
Sam Altman discusses hardware implications of AI recursion, noting that robots could build other robots, data centers could build other data centers, and chips could design their own next generation @AndrewCurran_
Jony Ive announces plans to create a new kind of computer with a completely new interface meant for AI, questioning whether users should even have an operating system, open windows, or send queries at all @AndrewCurran_
SoftBank forms a joint venture with OpenAI to localize and sell the AI company's enterprise tech to companies in Japan, with SoftBank itself becoming the first customer @TechCrunch
Google announces intent to acquire cloud security company Wiz, with the deal on track to close in early 2026 @TechCrunch
Wabi raises $20M in pre-seed funding led by a16z to build a personal software platform where anyone can create lightweight, shareable AI mini-apps from natural language @ekuyda
Anthropic's Editorial team is hiring two new writers to cover AI and economics/policy, and AI and science @keirbradwell
Pinterest CEO Bill Ready reports that open source AI is offering cost savings to the company, particularly in visual search @TechCrunch
Brex announces transformation into an AI-native finance platform powered by agents that learn, reason, and act on behalf of users @pedroh96

AI Ethics & Society

Amazon announces it won't allow agents on its site that don't identify themselves as such, with Perplexity expressing displeasure at the policy @TechCrunch
Ethan Mollick highlights the challenge of AI models lacking continuous learning, noting that current models often don't believe in the existence of recent events or releases like GPT-5 @emollick
Ethan Mollick warns that society is not ready for the destruction of costly signaling mechanisms, as writing used to measure effort, ability and diligence, but there's still no easy substitute @emollick
François Chollet emphasizes that ML research is an engineering discipline, not a philosophy seminar, stating that untested ideas are just speculation @fchollet
Stanford HAI publishes analysis on the shift from open to closed AI research, highlighting why it matters and what must be done about it @StanfordHAI
A researcher notes that in 2019, detailed personalized cold emails were impressive and led to hiring, but today would be assumed to be AI-generated, highlighting trust erosion @polynoamial
Microsoft Security EVP Charlie Bell publishes guidance on cybersecurity controls for AI agents, helping leaders manage risk as agents join and adapt at work @MSFTnews

AI Applications

Microsoft announces Voice feature in M365 Copilot, which Satya Nadella describes as becoming indispensable at work after daily use @satyanadella
Google integrates Gemini into Maps as a hands-free driving assistant that can find places along routes, check EV availability, share ETAs, and handle multi-step tasks like finding restaurants with specific criteria @sundarpichai
Pantone launches a new Palette Generator built on Azure OpenAI that helps users go from concept to color quickly @Microsoft
Tinder is testing an AI feature that learns about users from their Camera Roll photos @TechCrunch
Google DeepMind releases Perch 2.0, an upgraded AI for identifying animal species using bioacoustics, trained on 15,000 species with state-of-the-art bird identification and ability to learn new sounds from just a few examples @GoogleDeepMind
Google DeepMind partners with World Resources to release a model and dataset for predicting tropical deforestation risk, helping uncover underlying drivers of forest loss @GoogleDeepMind
Chrome introduces AI Mode via a new dedicated shortcut button under the search bar when opening a New Tab page @TechCrunch
Suhail describes a learning method using AI by uploading source material and requesting step-by-step explanations from high-level to detailed technical explanations, with quiz questions to confirm understanding at each step @Suhail
Granola positions itself as an AI notepad rather than an AI note-taker, emphasizing that a notepad helps users think while they write, whereas a note-taker tries to think for them @meetgranola

AI Research

Perplexity publishes its first research paper on custom Mixture-of-Experts kernels that make deployment of trillion-parameter models like Kimi K2 viable for the first time on AWS EFA @AravSrinivas
Cursor releases semantic search that improves their agent's accuracy across all frontier models, especially in large codebases where grep alone falls short, including details on training an embedding model for retrieving code @cursor_ai
Jeff Dean and co-authors present DataRater, a system for automatically and continuously learning which examples will help models the most during training @JeffDean
Microsoft Research introduces Magentic Marketplace, an open-source, extensible simulation environment for studying different agentic market designs as AI agents transform digital marketplaces @MSFTResearch
Microsoft researchers develop a new simulation environment for testing AI agents, revealing surprising weaknesses in current state-of-the-art systems @TechCrunch
Stanford researchers develop Cartridges, a new way to lighten AI's memory load that consumes less memory while still producing high-quality answers @StanfordHAI
Anthropic publishes engineering blog post on building more efficient agents that handle more tools while using fewer tokens through code execution with the Model Context Protocol @AnthropicAI
Simon Willison releases Datasette 1.0a20 with an entirely new SQL-powered permissions system, describing it as the most ambitious project attempted with coding agents like Claude Code and Codex CLI @simonw
François Chollet proposes that the path to autonomous AI is a system that learns to solve new problems by synthesizing models on the fly as code, and gets smarter over time by adding new abstractions to its own library @fchollet
Cameron Wolfe publishes detailed implementation guide for Proximal Policy Optimization for LLMs, covering rollouts, logprobs, KL divergence, advantage estimation, PPO loss, and composite loss @cwolferesearch
Researchers introduce CodeClash, a new evaluation where language models compete via their codebases across multi-round tournaments to achieve high-level goals, testing LMs on goals rather than tasks @jyangballin
An AI Scientist system that runs for days and makes genuine discoveries is released, with seven externally validated discoveries across multiple fields now available for anyone to use @andrewwhite01
DeepInverse joins the PyTorch Ecosystem as an open source framework for solving imaging inverse problems in medical imaging, computational photography, remote sensing, astronomical imaging, and microscopy @PyTorch