theaicatalog

What happened in AI today

Yesterday

November 20

10 stories

  1. Thinking Machines Lab · Models

    Thinking Machines releases Inkling 2 Mini, a small open-weights model for local use

    Thinking Machines released Inkling 2 Mini, an open-weights model that it says is small enough to run on a laptop. The model is a looped transformer that runs its middle layers three times for each token, which the company says lets it reach answers with fewer tokens. It has 26.4B total parameters, 3.4B active, and a total memory footprint of ~20GB. The company says it leads open-weights models with less than 100B total parameters on DeepSWE v1.1 (48.2%) and Vals Index v2.2 (46.5%).

    Link to this story

More stories

  1. OpenAI introduces MedChemBench to test AI on medicinal chemistry reasoning

    OpenAI · Research

    OpenAI introduced MedChemBench, a benchmark of 99 problems drawn from published research that tests whether AI models can connect molecular structure with experimental evidence. The problems draw on 117 papers and patents and include 761 distinct molecular structures. At the xhigh reasoning setting, GPT-6.2 Astra earned 84.5% of the available credit, compared with 83.3% for GPT-6 Astra and 58.2% for GPT-5.6 Sol. OpenAI says the results do not establish whether models can discover new drugs or improve laboratory decisions.

    Link to this story
  2. OpenAI details defenses against adversarial distillation

    OpenAI · Policy & safety

    OpenAI says adversarial distillation lets bad actors copy a model's capabilities without its safeguards, and it describes how it detects and disrupts these attempts. In an internal test, the average attack success rate for recovering protected reasoning fell from 54.0% against GPT-6 Astra to 0.2% against GPT-6.2 Astra with updated safeguards. OpenAI says its weekly enforcement actions against these campaigns have ranged from thousands to tens of thousands of accounts. The company says it briefed Congress in February on DeepSeek's attempts to distill U.S. frontier models, and in September it attributed a core part of an extraction campaign to individuals associated with Moonshot AI.

    Link to this story
  3. Anthropic closes $93.5 billion IPO and $26.0 billion private placement

    Anthropic · Business

    Anthropic says it has closed its initial public offering of Class A common stock at $83.00 per share, including the full exercise of the underwriters' option to buy additional shares. Aggregate gross proceeds were approximately $93.5 billion, of which approximately $91.6 billion was received by Anthropic. The stock began trading on the Nasdaq Global Select Market on November 19, 2026, under the symbol "ANT." Separately, Anthropic closed a private placement of Class N common stock to AMD, Broadcom, Google, Micron, Microsoft, NVIDIA and TSMC Partners for an aggregate purchase price of approximately $26.0 billion.

    Link to this story
  4. Google adds Connected Apps for groceries, finances, and taxes in Search

    Google · Products

    Google is rolling out Connected Apps in the U.S. through Search and the Gemini app for grocery orders from Albertsons, Safeway, Shaw's, and Target. The apps run on the Universal Commerce Protocol. Experian is coming to Search so U.S. users can check their credit rating and shop for a personal loan, and H&R Block lets U.S. users calculate taxes and estimate refunds from Search. Google says links to Angi and Thumbtack will arrive in Search in the coming weeks.

    Link to this story
  5. Watch Duty uses Gemini to flag wildfire dispatches in under 30 seconds

    Google · Products

    Google says a Gemini-powered tool built by Google.org Fellows turns noisy fire radio into searchable text, helping the nonprofit Watch Duty spot new wildfire dispatches in under 30 seconds. The tool was built from more than 20,000 hours of past radio recordings, and it cut word error rate on degraded VHF radio audio from 38% to 22.34%. In July 2026, the tool caught a 911 dispatch for the Railroad Fire in Washington, and Watch Duty sent its first overnight Critical Alert.

    Link to this story
  6. Hugging Face details RL-trained harness adapters for Qwen3.8-27B

    Hugging Face · Research

    A Hugging Face blog post describes training one RL-trained LoRA adapter for Qwen3.8-27B for each agent harness, including pi, Hermes Agent and OpenCode. The post says the pi adapter raises its Terminal-Bench 2.1 pass@1 from 39% to 55% while average token use falls from 2.9K to 1.5K per task. On SWE-bench Pro, the post reports the pi score rising from 39.2% to 64.7%, and the three adapters are also merged into one super adapter that runs under all three harnesses.

    Link to this story
  7. RadixArk releases DaisyRose-27B, an agent model built on Qwen3.8-27B

    Hugging Face · Models

    RadixArk released DaisyRose-27B, its first model, built on Qwen3.8-27B and trained with reinforcement learning using Miles. The team trained it on long-horizon agent tasks with sessions up to 1M tokens, using 1,200 curated tasks over 18 days on 144 H200 GPUs. The company says the model improves Terminal-Bench 4.1 pass@1 from 5.6% to 12.0%, and the checkpoints are available under Apache 2.0 in BF16, FP8 and NVFP4 formats.

    Link to this story
  8. MiniMax says its Cyber model breaks several cryptographic schemes

    MiniMax · Research

    MiniMax says its research team used the Cyber model to run automated cryptanalysis on modern cryptographic schemes. The company says Cyber produced the first practical breaks against several NIST candidates that had no known practical attacks, including Waterfall-256 and Blender-256. MiniMax also says the model solved dozens of previously unsolved instances in public cryptographic challenges, and it has released reasoning traces and proof-of-concept code on GitHub.

    Link to this story
  9. ElevenLabs adds video ad variants and localization to Ads Engine

    ElevenLabs · Products

    ElevenLabs has added video ad variants and localization to Ads Engine, part of ElevenCreative, so brands can refresh video ads without a reshoot. Ads Engine can change an ad's hook or soundtrack, and it creates each new hook's voiceover with a digital voice replica of the original speaker before lip-syncing the ad to it. For new markets, the tool uses Dubbing v2 to translate and dub videos and re-renders translated captions in brand fonts. Ads Engine now connects to TikTok ad accounts, in addition to Google, Meta, and LinkedIn.

    Link to this story

Thursday

November 19

10 stories

  1. Google adds Tool Retrieval to the Gemini API in preview

    Google · Products

    Google has launched Tool Retrieval in preview in the Gemini Interactions API, so agents can use large tool libraries without sending full parameter schemas on every request. Developers turn it on by adding a tool_search tool and setting defer_loading: true on the tools to hold back. The model sees the name and description of each deferred tool and loads a schema only when the task needs it. Tool Retrieval works with Gemini 3.6 and later models, including all Gemini 4 models, and has no separate charge beyond standard token rates.

    Link to this story
  2. Google quantum team wins a magic square game on its Willow processor

    Google · Research

    Google Quantum AI researchers ran the magic square game on their Willow superconducting processor, using entangled qubits instead of a pre-written table. The best classical strategy wins at most 88.89% of rounds, while the qubits won 98.30% of 100,000 plays of all nine versions of the game. The team reports the results in Science, along with tasks that outperform classical strategies on hidden-linear-function problems across grids of up to 105 qubits.

    Link to this story
  3. Google adds computer use to the Gemini app for Mac

    Google · Products

    Google is introducing computer use in the Gemini app for Mac, which automates repetitive tasks across local files, desktop apps, and the web. Users describe a workflow, and Gemini handles the clicking, typing, and form entries in an isolated background window with a live preview. Users choose which apps Gemini can access, watch each step, pause at any time, and approve key actions before they happen. The feature is rolling out in English in the U.S. for personal Google accounts and Google AI Plus, Pro and Ultra users on Apple Silicon Macs running macOS Sequoia 15.0+.

    Link to this story
  4. Google partners with Blue Origin for first heavy-lift Project Suncatcher launch

    Google · Hardware

    Google says it is partnering with Blue Origin on the first heavy-lift launch for Project Suncatcher, its long-term moonshot to scale AI infrastructure in space by using solar power in orbit. Targeted for Q1 2029, Blue Origin's New Glenn rocket will deploy Google's satellites, each with roughly the capacity of a data center rack, into sun-synchronous orbit. Google calls the launch a crucial milestone on its roadmap and says it builds on the company's ongoing work with other satellite and launch partners.

    Link to this story
  5. Google Home adds Wallet keyless entry, audio alerts and thermostat savings estimates

    Google · Products

    Google says Android users can now unlock compatible front doors with a quick tap or hands-free approach using digital home keys in Google Wallet. Google Home Premium Advanced now detects sounds such as barking dogs or breaking glass in event descriptions for motion-triggered cameras and doorbells. Starting in December, Nest thermostat owners in the U.S. and Canada can see estimated cumulative savings in the app, and Gemini for Home now handles everyday requests up to 20% faster and is expanding to six new countries, including India and Brazil.

    Link to this story
  6. Google Vids offers free personal avatars to anyone with a personal Google account

    Google · Products

    Google says anyone with a personal Google account can now create a personal avatar to star in their Google Vids videos, at no cost. Users scan a QR code to create their avatar likeness, and the avatar is saved to their Vids project for reuse across every scene. Avatars are linked to the Google Account and restricted to the account holder's likeness, and users can try the feature at vids.new on desktop.

    Link to this story
  7. NVIDIA and Microsoft expand AI partnership from Azure to Windows PCs

    NVIDIA · Business

    NVIDIA and Microsoft are extending their collaboration across Azure, enterprise AI tools and Windows PCs, announced at Microsoft Ignite. NVIDIA Vera Rubin NVL72 systems are running in Azure data centers, and Microsoft Foundry is adding NVIDIA Nemotron 3.5 models and support for deploying custom model weights. EY used Programmed Training in Microsoft Foundry to adapt Nemotron open models for classifying corporate purchasing records, improving accuracy on its evaluation dataset by 55% and lowering cost by 15x compared with proprietary frontier models. Azure Local is also expanding to NVIDIA Jetson Thor and Jetson Orin edge AI platforms.

    Link to this story
  8. Devin launches a migration tool for Databricks

    Cognition · Products

    Cognition launched Devin for Databricks Migrations, a tool built with Databricks to move companies off legacy data platforms and onto the Databricks Data + AI Platform. Devin converts legacy warehouse code, tests it in the customer's workspace, fixes what fails, and reconciles each source against its Databricks copy before cutover. Cognition says one Fortune 500 energy company migrated a 142-object platform to Databricks Unity Catalog in eight days. The tool is available on the Databricks Marketplace, and customers can apply their purchase toward their Databricks Universal Commit.

    Link to this story
  9. AMD and Unsloth bring NPU support to more than 100 AI agent apps

    AMD · Products

    AMD and Unsloth say users can now run local AI inference on AMD XDNA 2 neural processing units through Unsloth Desktop, reaching more than 100 compatible agentic AI applications on supported AMD Ryzen AI PCs. Compatible apps send model requests to the Unsloth Desktop API, while Lemonade and FastFlowLM run the models on the NPU. Supported tools include Hermes Agent, OpenClaw, Open WebUI, Cline, LangChain and GitHub Copilot. The companies say the NPU offers a power-efficient option that leaves CPU and GPU resources free for other tasks.

    Link to this story
  10. Ai2 launches Theorizer in Asta to propose explanations from research papers

    Ai2 · Products

    Ai2 launched Theorizer in Asta, its platform of AI agents for scientific research, and open-sourced the code so researchers can examine how it generates theories. Theorizer searches published studies for a user's question and links each proposed explanation to the papers behind it. Asta's web interface covers token costs up to a limit, and researchers can run Ai2's open library on their own machines and pay for tokens as needed. Ai2 says Theorizer can miss relevant studies and misjudge originality, so researchers still need to check claims against the wider literature.

    Link to this story

Wednesday

November 18

15 stories

  1. Google · Products

    Google rolls out Genie 4 world model in new Project Genie experience

    Google is rolling out a new Project Genie experience to eligible Google AI Ultra subscribers, with first-time access to its Genie 4 world model. Google says users can prompt mid-session to change their world, such as shifting the weather or dropping in a character, and Genie 4 adds real-time native audio. Sessions now last up to 3 minutes per generation, up from 60 seconds, and a new Project Genie agent builds worlds from natural-language descriptions. Google AI Pro members can join a waitlist, and all generated video and audio includes SynthID watermarking.

    Link to this story
  2. Google · Products

    Gemini app now runs on lower-memory Android devices

    Google is bringing the Gemini app to eligible Android (Go edition) devices that have less memory. Google re-architected the app so its core experience runs smoothly on lower-memory hardware, with Gemini Live available in more than 45 languages, camera-based learning and translation, and image creation with Nano Banana. Users can download the app on Google Play, or find it preloaded on upcoming devices from partners such as Transsion.

    Link to this story
  3. Tencent · Models

    Tencent launches Hy4 flash, a low-cost model for everyday work

    Tencent released Hy4 flash, the latest model in its Hy family, with 224 billion total parameters, 12 billion of them active, and a context length of up to 1M tokens. Tencent says it is the most cost-effective Hy model to date, priced at approximately US$0.070 per million input tokens and US$0.281 per million output tokens. On Tencent's internal WorkIndex benchmark, the company says Hy4 flash delivers a per-task cost of just 1/13 that of Hy4 Preview. Tencent plans to open-source Hy4 flash, and it is free on WorkBuddy until November 30.

    Link to this story

More stories

  1. OpenAI says coding agents now finish nearly half of two-to-four-hour research tasks

    OpenAI · Research

    OpenAI says the share of two-to-four-hour research tasks that AI agents complete on their own rose from 23% to 46% between March and October. For tasks estimated to take four to eight hours, success rose from 13% to 35%, and success varied widely by task, at 79% for monitoring training runs and 25% for writing code for inference and serving stacks. OpenAI says humans remain a core part of AI research and that AI labs should be required to track and publicly report indicators of recursive self-improvement.

    Link to this story
  2. OpenAI says its models improved chip designs for its own chip and AMD

    OpenAI · Hardware

    OpenAI says its models proposed changes that reduced standard-cell area by about 13% in the vector unit and 10% in the matrix unit across the design effort for its Jalapeño inference chip. In a partnership with AMD, OpenAI says its model delivered up to 14% improvement relative to heavily optimized baselines on the flagship GPU IP, and AMD has already included some of the changes in a design sent for manufacturing. OpenAI is also partnering with Synopsys to train GPT-Synopsys to use Synopsys electronic design automation tools the way expert chip designers do.

    Link to this story
  3. Le Figaro partners with OpenAI to bring its journalism to ChatGPT

    OpenAI · Business

    Le Groupe Figaro and OpenAI announced a strategic partnership to help more people find Le Figaro's journalism in ChatGPT. The deal covers current reporting and archive content from Le Figaro and associated publications, shown through summaries and excerpts with links to the underlying stories. Le Figaro will keep full editorial independence and plans to introduce ChatGPT Enterprise across various teams.

    Link to this story
  4. Anthropic prices its initial public offering

    Anthropic · Business

    Anthropic, PBC priced its initial public offering of Class A common stock at $83.00 per share. The offering has 980,000,000 shares, with 956,351,354 sold by Anthropic and 23,648,646 sold by selling stockholders, so Anthropic will not receive any proceeds from the selling stockholders' shares. The shares are expected to begin trading on the Nasdaq Global Select Market on November 19, 2026, under the symbol "ANT," and the offering is expected to close on November 20, 2026. The underwriters also have a 30-day option to buy up to an additional 147,000,000 shares.

    Link to this story
  5. Google releases Gemini 4 Neon for coding and analytical work

    Google · Models

    Google introduces Gemini 4 Neon, which it says ranks #3 overall with a score of 61 on the Artificial Analysis Intelligence Index at its xhigh thinking level. The model is priced at $3 per million input tokens and $15 per million output tokens, with a 1 million-token context window. It is rolling out today in the Gemini API in Google AI Studio, for Gemini for business, for paid Google AI Pro and Ultra subscribers in Google Antigravity, and for Pro and Ultra subscribers in the Gemini app. Google says Neon is built for complex coding, analytical reasoning, and enterprise workflows, between the Carbon and Hydrogen models.

    Link to this story
  6. Google Labs opens Stitch Loop waitlist for its always-on product agent

    Google · Products

    Google Labs is introducing Stitch Loop, a long-running product agent that connects to a codebase and live metrics to keep improving a product after code ships. Users connect GitHub and product context such as logs and user feedback through over 180 integrations, custom MCP servers, or local CLI uploads. A security loop built with CodeMender, Google DeepMind's autonomous code security agent, scans for vulnerabilities, and nothing merges until a user reviews the pull request. The waitlist opens today for all users globally, with eligible Google AI Ultra users getting priority access.

    Link to this story
  7. Google adds temporary threads to AI Mode in Search

    Google · Products

    Google says temporary threads in AI Mode let users ask questions that are not saved to their history or used for personalization. The session is detached from the user's account, and it disappears once the thread is closed. The feature is rolling out globally on desktop and mobile browsers in all languages supported by AI Mode.

    Link to this story
  8. Microsoft and Intel expand Azure infrastructure collaboration

    Microsoft · Hardware

    Microsoft says it is the first company to announce use of Intel Foundry's EMIB advanced packaging technology for Azure first-party silicon. The long-term agreement confirms multi-year Intel Xeon supply capacity, including next-generation server CPUs codenamed Diamond Rapids, and expands work on the Azure roadmap. Saurabh Dighe, CVP of Azure System & Architecture at Microsoft, says the partnership brings together advanced packaging, custom compute, memory and security to support Azure compute.

    Link to this story
  9. NVIDIA RTX Spark compact desktops now available from ASUS, Dell and MSI

    NVIDIA · Hardware

    NVIDIA says new RTX Spark compact desktops from ASUS, Dell and MSI are available now, with systems from Acer, GIGABYTE, HP and Microsoft coming later. The desktops offer up to 128GB of unified memory for running local AI models, agents, image and video generation, and model fine-tuning on a Windows PC. NVIDIA also says the expanded software support includes native Windows on Arm versions of Adobe applications and DLSS Ray Reconstruction in Blender 5.3.

    Link to this story
  10. Reson8 releases Rumble-2-Flash, a small open speech recognition model

    Hugging Face · Open source

    Reson8 has released Rumble-2-Flash, a 0.5B-parameter speech recognition model under the Apache 2.0 license, and says it supports 40+ languages. The model was distilled from Reson8's larger Rumble-2 model, and it runs with native support in Transformers and vLLM. On AppTek's 302-hour call-centre benchmark, Reson8 reports a 7.86% word error rate, close to the 7.78% of NVIDIA's Nemotron 3.5 Transcribe (30B).

    Link to this story
  11. GitHub expands Copilot Fusion to more apps and adds new models

    GitHub · Products

    GitHub says Copilot Fusion is now available on the latest versions of github.com, Visual Studio, and the GitHub Mobile beta apps for iOS and Android, with production mobile app support coming shortly. On Terminal-Bench 4.1, a public leaderboard hosted by Stanford, Harbor, and the Laude Institute, GitHub says Copilot Fusion Max achieved a 71.2% resolution rate, 5.4 percentage points ahead of Opus 5.5 and Fable 5.2. The update also adds Claude Sonnet 5.5 to Copilot Fusion Max, GPT-6.1 Terra to Copilot Fusion High, and MAI-Code-2 to Copilot Fusion Standard.

    Link to this story
  12. ElevenLabs releases ElevenAgents for Healthcare

    ElevenLabs · Products

    ElevenLabs introduces ElevenAgents for Healthcare, which lets healthcare teams set how each AI agent behaves, what it can say and do, and when it hands a call to a person. Banner Health, which has 33 hospitals and more than 400 clinics, uses agents to answer inbound patient calls and let patients book, reschedule, and cancel primary care appointments. At Cadence, which monitors patients with chronic conditions, the median time to reach a patient fell from nearly two hours to 3.5 minutes. ElevenAgents supports HIPAA-eligible deployments for Enterprise customers who sign a BAA and turn on Zero Retention Mode.

    Link to this story