AI Daily · November 14 Saturday

2026-11-13 08:00 — 2026-11-14 08:00 (UTC+8)

1. 01

   OpenAI launches two security products: agents that pentest apps, and agents that turn findings into reviewable patches

   At its Cyber at the Frontier event on November 12, OpenAI introduced Daybreak Red Team and Daybreak Defense Factory. Red Team uses multiple agents to pentest web apps and confirm vulnerabilities; Defense Factory traces vulnerabilities and drafts patches for engineers to review. Patches are not deployed automatically.

   Read the full report

   Finding a vulnerability is only the first step; security teams also have to confirm it is real and then fix it. The two Daybreak products OpenAI announced at its Cyber at the Frontier event on November 12 map onto those two jobs.

   According to SecurityWeek, Daybreak Red Team uses multiple AI agents to pentest web applications, using source code and app context to investigate attack paths and confirm vulnerabilities. It can also examine findings from code scanners, bug bounty submissions and vulnerability disclosure reports. OpenAI says the agents run in isolated environments, with Guardian review enforcing which applications and endpoints they can reach. OpenAI is accepting early-access requests.

   Daybreak Defense Factory handles the fix: agents trace vulnerabilities through the code and draft patches, which engineers review and test before merging. It can import findings from Wiz, Snyk and Endor Labs, scan connected GitHub repositories and monitor new commits, and engineers can edit patches with Codex and open draft GitHub pull requests. Patches are not deployed automatically. It is offered as a preview plugin in ChatGPT for workspaces with Daybreak access.

   During the livestream, OpenAI president Greg Brockman said that as models improve, the vulnerabilities they can find autonomously will get deeper, and manual coordination of disclosure and patching does not scale.

2. 02

   LangChain, maker of AI agent tooling, raises $300M as its valuation hits $3B in 13 months

   LangChain, the open-source AI agent framework, raised $300 million led by Alphabet's CapitalG at a $3 billion valuation, about 2.4 times its $1.25 billion mark from October 2025, the Wall Street Journal reported. The money will mainly go to expanding its enterprise platform LangSmith and hiring.

   Read the full report

   Whether AI agents themselves can make money at scale is still being tested, but the companies selling tools to build and run them are raising big rounds. LangChain raised $300 million at a $3 billion post-money valuation, the Wall Street Journal reported on November 12.

   According to SiliconANGLE and Dealroom, citing the report, Alphabet's growth fund CapitalG led the round, joined by existing backers IVP and Sapphire Ventures and new investors GIC and DFJ Growth. LangChain's previous round was a $125 million Series B in October 2025 at a $1.25 billion valuation, so its valuation rose about 2.4 times in roughly 13 months.

   LangChain is best known for its eponymous open-source framework from 2022, which helps developers connect agents to models and tools; its commercial product LangSmith handles monitoring, evaluation and deployment of agents. Co-founder and CEO Harrison Chase told the Journal the aim is to help large companies build, manage and deploy their own agents. The company reportedly has 435 full-time staff and plans to reach 850 by the end of next year.

   Competition is intense: OpenAI, Anthropic, Amazon and Microsoft are all building their own agent infrastructure. CapitalG partner Mo Jomaa said each of those players wants to lock customers into its ecosystem, while LangChain pitches independence. LangChain's blog had not posted about the round when outlets checked.

3. 03

   A 145B model on a laptop: Edge0 open-sources local coding models it says need just 29GB of memory

   On-device AI company Edge0 released open weights for Edge0-145B and Edge0-120B, pitched for offline agentic coding on laptops. Its founder says the 145B scores 86.5% on Terminal-Bench 2.1 versus 78.2% for Claude Opus 4.6, needs about 29GB of memory, and decodes 75+ tokens/s on RTX Spark and Mac.

   Read the full report

   Many developers want coding agents without a cloud token bill and without their code leaving the machine. On November 12, on-device AI company Edge0 released open weights for two 100B-class local models, Edge0-145B and Edge0-120B, on Hugging Face.

   Founder Samuel Zeng called them Edge0's first device-native 100B-class foundation models. By his figures, the 145B scores 86.5% on Terminal-Bench 2.1, a terminal-based coding benchmark, versus 78.2% for Claude Opus 4.6; it needs about 29GB of memory, roughly a third of a standard setup (100GB+); it decodes at 75+ tokens per second and prefills at 1,050+ tokens per second on RTX Spark and Mac; and it handles 128K–256K tokens of context without slowing down.

   The trick follows Edge0's established approach: its website says its sparse mixture-of-experts models keep expert weights in local storage and load them as needed, so parameter count is not the same as RAM. It previously released the phone-sized Edge0-35B. Developers relaying the launch say the new models expose OpenAI- and Anthropic-compatible local APIs that coding agents can point at.

   The benchmark and speed figures are Edge0's own and have not been independently reproduced; 29GB is peak memory, so a typical 16GB machine still won't run it, and the model files are tens of gigabytes.

4. 04

   World Labs opens its Atlas world model to everyone, turning ideas into exportable 3D scenes

   World Labs opened access to Atlas, its next-generation world model, to everyone, alongside apps and API endpoints. Users can build 3D worlds from their ideas and export keyframes, video clips or 3D Gaussian splats into their own workflows.

   Read the full report

   World Labs, founded by Fei-Fei Li, opened its Atlas world model to everyone on November 12. Rather than producing a single video, Atlas aims to generate 3D scenes that can be explored and edited further.

   The company said it built apps and API endpoints so users can bring their ideas for a 3D world and come away with keyframes, video clips or 3D Gaussian splats. Gaussian splatting represents a scene as many small colored blobs that can be rendered in real time in a browser or 3D software. The company's blog calls this its next-generation world model, able to generate and reconstruct worlds and plug into users' workflows via tools and developer APIs.

   Atlas was first introduced in September; this moves it from preview to general access. World Labs previously announced it is joining AMD. For teams in games, film, architectural visualization and robotics simulation it offers a 3D world generator they can try directly.

5. 05

   AI writes mobile code faster than humans can review it: Expo launches cloud simulators built for agents

   Expo, the mobile development framework, launched Expo Simulators: after a coding agent makes a change, it can run it on a cloud iOS simulator or Android emulator and bring back a recording before asking the developer to look. The launch post passed 2 million views.

   Read the full report

   Coding agents now write mobile code faster than teams can review it. Expo's answer, announced November 12, is to have the agent test its own work in a real simulated environment before interrupting a human.

   Expo Simulators are cloud simulators and emulators built for agents. According to Expo, once an agent makes a change it runs it on a cloud simulator or emulator and brings back a recording before asking for the developer's attention. Developers don't need a local simulator setup; they can watch the recording to judge whether the change works.

   Expo also published a list of agent workflows that can be set up now. The launch post passed 2.2 million views, making it one of the more discussed developer tools of the day. It reflects a broader shift: as AI produces more code, verification and testing become the bottleneck, and companies are giving agents environments where results can be seen.

6. 06

   Describe it and get an AI that makes phone calls: Fish Audio launches a no-code voice agent platform

   Voice AI company Fish Audio launched Fish Agent, a no-code platform where users describe a voice agent that can make and take calls. Users can get a phone number or bring their own, choose leading LLMs or a custom one, and rely on its full-duplex Live 3 voice model for interruption-aware conversations. Median deploy time is 32 seconds, it says.

   Read the full report

   Building a voice bot that answers calls, books appointments or handles customer questions used to mean wiring together speech recognition, an LLM, speech synthesis and phone lines. Fish Audio's Fish Agent, launched November 12, turns that into a no-code platform.

   According to the company, users build an agent simply by describing it, then get a phone number or connect their own. The agent can run on leading LLMs or a custom model, connect to tools, webhooks and knowledge bases, and be controlled via a full API. Calls are powered by Fish Live 3, a full-duplex voice model, meaning the AI can listen while speaking and handle interruptions naturally. The company says median deploy time is 32 seconds.

   Use cases range from personal assistants to receptionists and customer support, and enterprises can book a demo with a dedicated engineer.

7. 07

   Devin Teams gets more for the same price: two free concurrent sessions per user, powered by in-house SWE-2

   Cognition upgraded its Devin Teams plan: everyone on a team can use the new Devin Desktop without paying for extra seats; each user gets two concurrent sessions in a new Free mode powered by its own SWE-2 model; admins get SSO, per-user or per-group usage limits, and can request a BAA.

   Read the full report

   The price war among AI coding tools continues. Cognition, maker of the coding agent Devin, upgraded its Devin Teams plan on November 12, aiming to let teams use more and manage it more finely.

   According to Cognition, everyone on a team can now use the new Devin Desktop without paying for additional seats, and every user gets two concurrent sessions in a new Free mode powered by Cognition's own coding model, SWE-2. On the admin side, teams can enable SSO with Okta, Entra ID or any SAML or OIDC provider, set up to three usage tiers per user or group, and request a standard Business Associate Agreement (BAA) from settings, a compliance prerequisite for handling health data.

   SWE-2 is the coding model Cognition released in September. Using an in-house model for a free tier is a common way for coding-agent vendors to control costs while competing with Claude Code, Codex and Cursor for team users. Some users online claim SWE-2 is post-trained from a Chinese open-weights model, but that claim does not appear in Cognition's announcement and could not be verified here.

8. 08

   Salesforce open-sources SOV-Transcribe-1, a transcription model that gets emails and phone numbers right

   Salesforce released SOV-Transcribe-1, an open-weight transcription model. CEO Marc Benioff says it ranks first among open-weight models and seventh overall on Artificial Analysis's speech-to-text board, is accurate on emails, phone numbers and URLs, robust to accents and noise, and supports context biasing for names and domain terms.

   Read the full report

   Transcription most often fails on email addresses, phone numbers and jargon, exactly the details that matter most in support and sales calls. SOV-Transcribe-1, an open-weight model Salesforce released on November 13, targets those pain points.

   Salesforce CEO Marc Benioff described it as a speech-augmented language model built for transcription: accurate on emails, phone numbers and URLs, robust against accents and noisy audio, with context biasing for names and domain terms, and ranked first among open-weight models (seventh overall) on Artificial Analysis. Salesforce AI Research said it aims to hear accents, noisy rooms and domain terms and turn speech into something agents can act on.

   According to a breakdown by developer Cobus Greyling, it pairs a Whisper large-v3 audio encoder with a Gemma 4 26B mixture-of-experts decoder, with 26.6B total and 4.6B active parameters and a 2.28% word error rate; because a language model writes the transcript, it can be instructed to give verbatim or cleaned-up output. The limits are clear too: batch only with no streaming, so it suits call recordings and meetings rather than live voice agents; it covers seven European languages; the checkpoint is about 53GB and runs on your own GPUs; and the Apache 2.0 license lets audio stay inside a company's infrastructure.

9. 09

   A strong model plans, a fast one executes: Google's Antigravity adds Orchestrator, claiming 24% quota savings

   Google is rolling out Orchestrator in Antigravity to AI Ultra subscribers, letting one session use multiple models: Gemini 4 Carbon plans, hands code edits and testing to Hydrogen, and steps back in when it hits friction. Google says it reaches near-Carbon success on Terminal-Bench 4.0 with 24% less quota and 19% faster completion on average.

   Read the full report

   Running the strongest model end to end works well but burns quota; using only a cheap model often fails on complex tasks. Google's compromise, announced November 12, is to have two models split the work.

   Google is rolling out Orchestrator in Antigravity, its agentic coding tool, to Google AI Ultra subscribers, allowing multiple models in a single session instead of routing the whole task to one model. It pairs a frontier model with a fast execution model: Gemini 4 Carbon sets the plan, hands off code edits and testing to the newly released Gemini 4 Hydrogen, and steps back in to repair the plan when execution hits friction.

   Google says the combination reaches near-Carbon task success rates on Terminal-Bench 4.0 while using 24% less quota and finishing 19% faster on average. The figures are Google's own, and the feature is currently limited to Ultra subscribers.

10. 10

    Let coding agents pay for their own compute: Lightning Labs open-sources pi402 for per-token bitcoin payments

    Lightning Labs released pi402, an open-source extension for the Pi coding agent. When a model call returns HTTP 402, the agent's built-in wallet pays instantly over Lightning, with no account, API key or card needed. Sellers can run their own gateway and bill by the usage the upstream reports.

    Read the full report

    AI agents can do more and more work, but opening accounts, adding cards and issuing API keys still requires a human. pi402, released by Lightning Labs on November 12, tries to let agents pay for inference themselves.

    According to its site, pi402 is an extension for the open-source Pi coding agent. When a model call comes back with HTTP 402 (payment required), the agent's own wallet pays over bitcoin's Lightning Network and the request goes through. Payment is final within seconds, with no account, buyer-side API key or card. The buyer holds sats while the seller's gateway injects the model key server-side. Users can set a per-request ceiling and a daily cap that are checked before any payment.

    On the seller side, any OpenAI-compatible model can sit behind Lightning Labs' Aperture gateway and metering service: a bundle of tokens is paid up front, then drawn down per request based on usage reported by the upstream. The site's example prices 1 millisatoshi per input token and 2 per output token, roughly $0.79 / $1.57 per million tokens at the time of writing. Lightning Labs argues that inference can't be reversed while card payments can be charged back for 60–120 days, a real risk for compute sellers. There is no hosted endpoint; pi402 must be self-deployed.

11. 11

    Top 100 private AI companies worth $3.6 trillion combined: Bessemer releases its first AI 100 list

    Bessemer Venture Partners and Newcomer Media released the first AI 100, ranking the world's top 100 private AI companies. Bessemer says the list's combined valuation is $3.6 trillion, up 219% versus the 2025 Cloud 100, with 480% average year-over-year ARR growth and a record 57x average ARR multiple.

    Read the full report

    Just how richly valued are AI startups? Bessemer Venture Partners and Newcomer Media offered some numbers on November 12 with the inaugural AI 100, a ranking of the world's top 100 private AI companies.

    Bessemer says honorees were evaluated on market leadership, estimated valuation, operating metrics, and people and culture, spanning foundation models and AI infrastructure to robotics. By its figures, the list's combined valuation is $3.6 trillion, up 219% versus the 2025 Cloud 100; average year-over-year ARR growth is 480%; the average valuation is a record 57 times ARR; and fewer than 10% of companies hold 55% of the list's value.

    The numbers show genuinely fast revenue growth at the top, but a 57x average revenue multiple also signals very high expectations for future growth, with value concentrated in a few companies. Many of the valuations are estimates made by Bessemer and its media partner.

12. 12

    Let AI pay for you: OpenAI launches ChatGPT Wallet, where agents never see your real card and every purchase needs a biometric check

    OpenAI launched ChatGPT Wallet, letting ChatGPT's dots agents complete purchases for users. Built with American Express, Mastercard and Visa, it uses payment tokens instead of real card numbers, and every purchase requires a security check such as Face ID or a fingerprint. It starts with ChatGPT Pro subscribers aged 18+ in the US.

    Read the full report

    AI agents that research and compare products are nothing new; the sticking point has been the last step: paying. ChatGPT Wallet, which OpenAI introduced on November 13, aims to hand that step to agents without letting them touch users' real card details.

    According to OpenAI, users add a card once. Before each purchase they review the items, total and selected card, then approve it with a security check such as Face ID or a fingerprint scan. After approval, the payment network issues a payment token, a substitute card number, security code and expiration date, so neither the business nor the agent sees the actual card details. OpenAI says users keep their existing card rewards and protections for eligible disputes, while returns and customer service stay with the merchant.

    The wallet was built with American Express, Mastercard and Visa and is rolling out first in ChatGPT's dots, initially for ChatGPT Pro subscribers aged 18 and older in the US with eligible US-issued cards. OpenAI acknowledges that some websites may inadvertently block agent purchases and that risk systems may decline legitimate agent orders, so it is also publishing a Commerce Partner Playbook for businesses and working with payment providers such as Stripe and Adyen and several fraud-prevention firms.

    OpenAI says it eventually wants users to set spending limits, categories or time windows to pre-approve purchases. For now, every purchase still requires the user's explicit approval.

13. 13

    OpenAI disbands its stand-alone robotics team, Fortune reports; company says general models can now run robots directly

    Fortune reports OpenAI has disbanded its stand-alone robotics team: researchers move to the main research team under Chief Research Officer Mark Chen, and applied robotics and hardware staff move to the Applied organization under President Greg Brockman. OpenAI says this accelerates rather than abandons robotics, and a humanoid demo is still targeted for 2027.

    Read the full report

    OpenAI's robotics team has disappeared again, but this time the company's explanation is that general-purpose models can now control robots directly, making a separate team unnecessary.

    According to an exclusive Fortune report on November 13, OpenAI has disbanded its stand-alone robotics team and confirmed the reorganization when Fortune asked. Robotics researchers now sit in the main research team under Chief Research Officer Mark Chen, while applied robotics and hardware staff have moved into the Applied organization under President Greg Brockman. A spokesperson said robotics remains core to OpenAI's long-term vision and framed the change as acceleration, not retreat. Fortune also noted that OpenAI had registered a booth at this week's CoRL robotics conference, but the booth was empty when a reporter visited.

    This is not OpenAI's first robotics reset. It wound down an early robotics team in 2020, citing limited robot data among other reasons, rebuilt one in 2024 alongside a partnership with Figure that ended in February 2025, and has said it is developing its own humanoid robot, with CEO Sam Altman targeting a first public demo in 2027.

    Per Fortune, OpenAI researchers believed at the start of the year that robotics would best build on video models like Sora, but have since seen surprising results connecting general models such as GPT-6 Astra to robots. Carnegie Mellon assistant professor Guanya Shi told Fortune the move signals OpenAI sees robotics as "no longer a special problem" but a downstream application of general intelligence. On social media, some read it as a bearish signal.

14. 14

    DeepSeek opens its app's search, file, link-reading and voice features via API, so developers can rebuild a full DeepSeek App

    DeepSeek has opened the capabilities behind its app to the API: web search, deep-thinking search, file parsing, URL reading and safety checks, with deepseek-flash (V4.1-Flash) handling tool calls server-side. Search, fetch, OCR, ASR and TTS are also available as standalone APIs. Extended features carry tool fees on top of token charges, and public-facing apps must enable safety checks.

    Read the full report

    Until now, developers calling the DeepSeek API mostly got a bare model; features like web search and file reading from the app had to be built themselves. Starting November 13, DeepSeek is opening those up too.

    According to DeepSeek's announcement, deepseek-flash (currently served by V4.1-Flash) supports these built-in capabilities across the Messages, Responses and Chat Completions APIs, with tool calls and context assembly handled server-side: web and image search with citations in non-thinking mode, deep-thinking search that searches and reads pages during multi-step reasoning, parsing of PDFs, Office files and images, automatic reading of links in user messages, and safety screening of inputs and outputs. Search, fetch, OCR, speech recognition (ASR) and text-to-speech (TTS) are also offered as standalone APIs.

    DeepSeek also open-sourced a frontend demo that can import DeepSeek conversation history, showing how to recreate the full DeepSeek App in your own product. Extended capabilities incur tool usage fees on top of standard token charges. Public-facing apps must enable safety checks, and DeepSeek says developers remain responsible for content moderation, model registration and AI-generated content labeling.

    For smaller developers, building an assistant with web and file Q&A just got easier, though how tool fees affect total cost will depend on usage.

15. 15

    How to build safety evals you can trust: Anthropic shares 10 recommendations and says it paused many internet-connected evals

    Anthropic published 10 recommendations for building AI safety evaluations on its Engineering Blog, drawing on replication-crisis reforms in fields like psychology. They cover cross-disciplinary teams, running evals in offline sandboxes, preregistration and full statistical reporting. It notes that, after its own incidents, it hardened eval environments and paused many evals that give models internet access.

    Read the full report

    Whether a model ships, and which safeguards it needs, depends heavily on safety evaluations. But a poorly designed eval proves little no matter how good the score looks. On November 13, Anthropic shared 10 recommendations for building them on its Engineering Blog.

    The post draws on reforms adopted after the replication crisis in fields like psychology, such as preregistering hypotheses and analysis plans, better measurement and more rigorous statistics. The recommendations follow the workflow: build a team with both methods experts and domain experts such as psychologists or biologists; in design, start from the question you want to answer, distinguish observational evals (model A scored higher than B) from causal ones (did an intervention make the model safer), and make sure you measure what you mean to measure; when running, keep measures internally valid and preregister; and when reporting, give full statistics, document everything and disclose flaws responsibly.

    The section on running evals safely stands out. Anthropic writes that safety evals often disable standard safeguards to see baseline behavior, and in cyber evals models may probe the boundaries of their test environments and could take unintended actions if they escape. Evals should therefore run in sandboxes without internet access and be monitored in real time. In response to its own incidents, Anthropic says it has hardened its evaluation environments in recent months, temporarily stopped many evals that give models internet access, and added real-time monitoring for web search evals that require it.

16. 16

    Podcaster Dwarkesh Patel changes his mind: the avalanche that hit math is coming for AI research, with superintelligence possible by 2030

    Podcaster Dwarkesh Patel published an essay saying he updated his views heavily over the past three months: AI is advancing rapidly in verifiable domains, math is a preview, and automating AI research could compress years of progress into one. His median year for AIs matching the best human AI researchers is 2028, with a 25% chance it happens after 2032.

    Read the full report

    Earlier this year, Dwarkesh Patel thought superintelligence was at least the better part of a decade away. On November 13, the podcaster known for long interviews with AI researchers published an essay titled "The avalanche that hit math is coming for AI development," saying he has updated his views "a ton" over the last three months.

    His core argument is that in domains where answers can be checked and AIs can practice thousands of times, capabilities climb extremely fast, and math is the preview. He traces AI's path from struggling with arithmetic to Olympiad problems, open Erdős problems and what he describes as solving a Millennium Prize problem in 88 hours. He argues AI research also has many verifiable parts, and once AIs match top humans there, automating research could compress enormous progress into a single year. He cites figures published by METR, OpenAI and Anthropic as evidence that AI is already speeding up frontier R&D.

    His concrete estimate: assuming labs don't deliberately slow down, his median year for automated researchers matching the very best human AI researchers is 2028, followed soon after (by 2030, plausibly as early as next year) by AIs at least as good as top experts in every field. He also spells out the uncertainty: if problems such as continual learning stay unsolved, there is a 25% chance the milestone slips past 2032.

    The essay drew wide discussion on X. It is a personal forecast and argument, not a research finding, and its timelines come with explicit conditions.

17. 17

    Do multimodal models need a vision encoder? After training 200+ models, researchers find only one choice really matters

    Researchers including Ellis Brown, Saining Xie and Luke Zettlemoyer trained 200+ multimodal models from scratch in a controlled study and found a pretrained vision encoder is not necessary; what matters is dedicated vision parameters. Placing them in the language model's first layers matches a traditional encoder and can lower time to first token.

    Read the full report

    Multimodal models differ widely in how they "see" images: some use pretrained encoders like SigLIP, some train the encoder from scratch, and some drop it entirely. Because published models also differ in data, architecture and compute, it is hard to tell which choice actually matters.

    A paper released November 13, "Unbundling Multimodal Vision Encoding," by 12 researchers including Ellis Brown, Chris Hoang, Saining Xie, Luke Zettlemoyer and Lucas Beyer, tries to separate the question. It splits a vision encoder into three choices, an upstream pretrained visual prior, dedicated vision capacity, and placing that capacity before the language model, then trains 200+ models from scratch on common data at matched lifetime compute.

    The findings: pretraining is not necessary, since an encoder trained jointly from scratch matches a pretrained one. Dedicated vision capacity is what matters: encoder-free models trail on visual understanding even when the backbone gets the encoder's parameter count, and the authors say it takes roughly 3x the compute to catch up. But that capacity need not come first: vision-only branches added to the language model's early layers perform comparably to a pre-fusion encoder. According to the authors on X, at the encoder's 0.4B scale this is about 13% faster to first token.

    For teams building multimodal models, it points to a simpler recipe: dedicated vision parameters trained jointly with the backbone and placed early. The results come from the authors' controlled experiments and remain to be confirmed across more models.

18. 18

    Hears through TVs, storms and crosstalk: Bland launches Tofu-1, a speech-to-speech model for high-stakes phone calls

    Voice AI company Bland launched Tofu-1, a full-duplex speech-to-speech model for high-stakes calls in noisy conditions: it tells speakers apart, separates voices from background noise and can be interrupted mid-sentence. It is in a limited enterprise research preview; a team member says it will be fully accessible in January but not open source.

    Read the full report

    Phone AI agents often sound fine in quiet demos but stumble in cars, hospital rooms or on weak signal. Tofu-1, launched by voice AI company Bland on November 13, targets exactly those hardest calls.

    According to Bland, Tofu-1 is a full-duplex speech-to-speech model: it takes in and produces audio directly and can listen while it speaks. It tells speakers apart, knows who each is talking to and separates voices from noise; callers can interrupt or correct it mid-sentence; and it keeps the call moving while it works, which Bland says means it never claims a task is done before it is. In Bland's demo video, the company says every sound, including the caller, the agent, overlapping voices and background noise, was generated by Tofu-1.

    For now, Tofu-1 is available only as a research preview to a limited number of enterprises. Replying on X, Bland team member @zaygranet said the model will be fully accessible in January but will not be open source; he also said the team has "some pretty serious security concerns right now" and is finishing testing, with a smaller distill of the same model planned for early December. Some users voiced concern about how human it sounds.

19. 19

    When should an AI coworker speak up, and when should it stay quiet? Deepnote drops 12 models into a simulated company's Slack

    Data platform Deepnote released the Agent Initiative Benchmark, placing 12 frontier models in the Slack of a fictional company with six years of detailed history and 50 realistic workplace situations, where they decide on their own whether to investigate or speak. The team says no one tells the agent what to look at or when to speak, and correct silence earns full marks.

    Read the full report

    Most AI benchmarks test whether a model answers correctly when asked. A useful AI coworker also needs to know when to step in and when not to interrupt. The Agent Initiative Benchmark, released by data platform Deepnote on November 13, measures that kind of initiative.

    According to @jakubjurovych of Deepnote on X, the team built a fictional company with six years of highly detailed history, including a financial model, CRM records and past-due accounts, and placed 12 frontier models inside its Slack. As developer @svpino summarized it, agents can access records across 11 services and face 50 different situations. @jakubjurovych said in replies that nobody tells the agent what to look at or when to speak; the test adds plenty of noise, compares replying in-thread, in the channel or by DM, and "correct silence got full marks." He added that the idea came from such an agent running inside Deepnote itself.

    Cost became a talking point. One user, reading the leaderboard chart, noted GPT-6 Luna scored 6.38 at about $0.007 per task while Claude Fable 5.1 scored 7.50 at about $3.02 per task, arguing the gap matters for always-on agents that read every message. Those figures are a user's reading of the chart; Deepnote's published results are the authoritative ranking.

20. 20

    Google DeepMind releases PARDIG and SatWeave, the foundational datasets behind WeatherNext 3

    Google DeepMind announced the release of PARDIG and SatWeave, the foundational datasets behind its WeatherNext 3 weather forecasting model, to help researchers and developers build new climate applications.

    Read the full report

    Accurate forecasting starts with how clearly the atmosphere can be observed in real time. On November 13, Google DeepMind announced it is releasing PARDIG and SatWeave, the foundational datasets behind its AI weather model WeatherNext 3.

    DeepMind says the release is meant to help researchers and developers build new climate applications. For weather and climate AI teams, training data is often harder to obtain than model architectures, so a major lab publishing the datasets behind its model can lower the bar for replication and improvement. Details on the datasets' size, coverage and licensing are in DeepMind's own documentation.

21. 21

    Perplexity brings its AI coworkers to the desktop: Teammates arrive in the Mac app for QA, support and research

    Perplexity announced Teammates are now available in its Mac app for Max subscribers. Teammates work individually or with other Teammates to handle QA, customer service, research, growth marketing and more.

    Read the full report

    Perplexity announced on November 13 that Teammates are now available in its Mac app, currently for Max subscribers.

    According to the company, Teammates are AI agents that take on work for users, either individually or together with other Teammates, across QA, customer service, research, growth marketing and more. That lets users hand off repetitive work to multiple specialized agents from the desktop rather than the browser. The feature is limited to the premium Max tier for now.