Friday, November 13, 2026 Edition no. 27 · Morning

Artificial intelligence news from Brazil, in English

Noticiário do Joãozinho: Edition no. 27 · Nov. 13, 2026 · Morning

Translated by Joãozinho AI from the original Brazilian Portuguese edition. Facts, numbers and sources are the same as in the original; quotes, video titles and source names stay in their original language. Read the original in Portuguese

Listen to the spoken edition

Length: In Brazilian Portuguese

Download the MP3

Top storyEdition no. 27 · Nov. 13, 2026 · Morning

Microsoft's global outage passes 12 hours, knocks out Outlook, Teams and Copilot and hits the ChatGPT and Claude connectors

An outage in Microsoft 365 services began on Thursday (11/12) in the Americas and, within an hour, already had complaints in 74 countries and territories, according to Downdetector as cited by The Next Web. Outlook and Exchange Online, the Teams calendar and Microsoft 365 Copilot Chat are among the services affected. Microsoft ruled out an attack: Charles Lamanna, executive vice president of Copilot, Agents and Platform, wrote on X that there is no indication of malicious activity. Recovery has been slow: the company said traffic rerouting and strategic blocking brought partial, not sustained, improvement, and that some changes did not work as expected. In the update shortly after 9 a.m. London time (6 a.m. Brasília) this Friday, it spoke of “early signs of stabilisation”, but with capacity still below normal, and extended the email timeout window to 48 hours so delayed messages are not lost. The outage reached third-party AI: OpenAI and Anthropic reported errors in their Microsoft 365 connectors, and in Claude the mail, calendar and search tools were failing for most requests. Japan's government went back to paper, and banks, courts and city governments in several countries reported problems. Australian minister Tony Burke said Microsoft believed it was facing something similar to the 2024 CrowdStrike case, an update that went wrong, but the company has not confirmed the cause and promises a full report after resolving the problem.

PDF (pt-BR)

Models

1 story

Medium importanceNov. 12, 2026

Edge0 open-sources Edge0-145B, a 145-billion-parameter model built to run on a laptop with 29 GB of memory

Edge0 released Edge0-145B and Edge0-120B as open source, together with the inference framework and deployment tools. According to the company, Edge0-145B runs on a compatible laptop, such as a MacBook Pro, with peak memory of 29 GB, generates more than 80 tokens per second, processes more than 1,100 tokens per second when reading the prompt, draws about 120 W and accepts 128,000 to 256,000 tokens of context. It works on macOS, Windows and Linux, on Apple M-series chips and NVIDIA RTX cards. Edge0 says the model plans, writes and runs code and checks the result, and that in coding, reasoning and instruction following it comes close to some frontier cloud models; the numbers are the company's own.

Research

1 story

High importanceNov. 12, 2026

Analysis of open model weights finds the mark of Blackwell chips, barred for China, in six models from DeepSeek, Moonshot, MiniMax and Ant Group

Will DePue, of Unemployment Labs, published a forensic analysis done only with the models' published weights. It recovers traces of the initial values and searches for the hardware configurations that reproduce them. Six models released since April 2026 carry numerical patterns consistent with NVIDIA's Blackwell chips, whose export to China requires a license: DeepSeek-V4-Flash and V4-Pro, Kimi K3 and K3.1-Pro, MiniMax-M3 and Ling-3.1-flash. Earlier models from the same labs match the Hopper generation. For V4-Flash and Kimi K3, the result was confirmed on real Blackwell chips. The report itself says the mark shows on which hardware the weights were created and supports, without proving, that training was also done on it, and that it does not say where the chips were, how they were obtained or whether export rules were broken. For DeepSeek-V4.1-Flash, from September, no signature was found. The code to reproduce it is on GitHub.

Products

2 stories

Medium importanceNov. 12, 2026

Muse, Meta's agent, starts calling businesses in the US on the user's behalf, but only for subscribers

After testing the feature in September, Meta opened Muse's calls to businesses to paying subscribers in the US, for things like the cable company, the DMV or a restaurant without online booking, according to Alexandr Wang, the company's head of AI. Muse plans start at US$20 a month, with 500 million Muse tokens per week, and go up to US$100 a month with 3 billion. Meta said last month that Muse has more than 10 million users. Google's Gemini and the startup Instinct already offer something similar. Wang also said the Muse email address will soon reach free users.

Medium importanceNov. 13, 2026

Claude.ai, Claude Code and Anthropic's console have login errors this Friday, a day after the Microsoft 365 connector failure

At 6:58 a.m. Brasília time this Friday (11/13), Anthropic opened an incident of elevated errors on Claude.ai, the Claude Console (platform.claude.com), Claude Code and Claude Cowork, and said it was investigating a user authentication problem on Claude.ai and Claude Code. As this edition closed, the status was still “investigating”. It is a separate incident from the one that began on Thursday, when the mail, calendar and search tools of the Microsoft 365 connector began failing for most requests because of Microsoft's outage.

Business and Investment

4 stories

High importanceNov. 12, 2026

Mirendil, a lab of former Anthropic researchers, raises US$1 billion to put agents to work on AI research

Mirendil announced a US$1 billion Series A led by Kleiner Perkins and Clearlake Capital, with Battery Ventures, Doerr Capital, Gil & Co, Google, M13, Pear VC and S32, among others. The company wants to bring frontier AI research within reach of any scientist or engineer, with agents doing the hands-on work. According to it, a single researcher, working with agents, pretrained Mirendil-Base, a coding model, from scratch in three and a half months with less than US$10 million in compute; it then became Fern, which holds up to 10 million tokens of context. Lor, in private beta, puts hundreds of agents to write code and run experiments for days. Mirendil did not disclose its valuation; according to Dealroom, Bloomberg said in September the talks were for up to US$1 billion at a US$5 billion valuation. The CEO is Behnam Neyshabur and the CTO, Harsh Mehta.

Medium importanceContextNov. 12, 2026

LangChain raises US$300 million with Alphabet's CapitalG and is now worth US$3 billion

LangChain, maker of the open frameworks LangChain, LangGraph and Deep Agents for building AI agents, received US$300 million in a round led by CapitalG, Alphabet's fund, with IVP, Sapphire Ventures, GIC and DFJ Growth, according to The Wall Street Journal as cited by SiliconANGLE. The company is now worth US$3 billion. It makes money with LangSmith, a paid suite to build, test and monitor agents in production, and plans to nearly double its headcount, to 850 employees, by the end of 2027.

Medium importanceNov. 12, 2026

Cerebras lands Microsoft as a new customer, quarterly revenue up 70%, and the stock jumps almost 18% after the close

AI chipmaker Cerebras reported core adjusted revenue of US$230.9 million in the third quarter, 70% above a year earlier and above the US$214.7 million expected by the market, and announced Microsoft as a customer, without giving details of the deal. CEO Andrew Feldman said the contract was signed a month earlier and that revenue should start showing up in the first quarter of fiscal 2027. The company still posts a loss, but its adjusted operating margin of -14% came in better than the -24% forecast, and it is evaluating financing options to meet demand. The stock rose almost 18% in extended trading.

Medium importanceNov. 12, 2026

NVIDIA employees call for Jensen Huang's return in an internal post with more than 1,600 comments

According to WIRED, a post in NVIDIA's internal “suggestion box”, made by a longtime engineer before the resignation, asked the board to take employee sentiment into account before acting, and has since drawn more than 1,600 comments, most in favor of bringing Huang back (“if you believe in the mission, bring back THE boss”, one wrote). One employee sent a companywide email demanding transparency about the exit. In the hours after the resignation, about two dozen employees defended Huang on LinkedIn. NVIDIA and the new board chair, Harvey Jones, did not respond to the magazine.

Regulation

4 stories

High importanceNov. 12, 2026

Update: US Senate summons Jensen Huang and Howard Lutnick to testify in December, and Democrats open another investigation into NVIDIA's sales to China

Republican senator Tim Scott sent letters to Jensen Huang and Commerce Secretary Howard Lutnick asking them to testify before the Senate Banking Committee. He asked Huang to appear voluntarily on December 3 and warned that the committee may subpoena him; the Commerce Department said Lutnick agreed “proactively”. Three House committees (intelligence, foreign affairs and the China panel) called Huang to a closed-door interview on December 1, and, according to TechCrunch, he agreed after resigning. Democrats Richard Neal, of the House, and Ron Wyden, of the Senate, opened their own investigation into how the Trump administration approved and taxed sales of H200 chips to China, which they call a possible unconstitutional export tax, and ask why Megaspeed kept buying chips after the intelligence warnings. According to the BBC, NVIDIA lost US$550 billion in market value in five days, and Attorney General Todd Blanche said the Justice Department has not yet concluded whether Alice Huang spied for Beijing.

Medium importanceNov. 13, 2026

Malaysia investigates whether AI chips imported by Speedmatrix, Megaspeed's local arm, were re-exported

Malaysia's Ministry of Investment, Trade and Industry said this Friday that Speedmatrix, the local unit of Singapore's Megaspeed, imported servers and AI chips as a data center tenant, and that it is working with customs to find out whether those chips were sent to another country and, if so, whether that was legal, according to Reuters as cited by The Next Web. Since July 2025, re-exporting high-performance US AI chips requires a permit in the country. Prime Minister Anwar Ibrahim said Malaysia will cooperate with the investigations and that so far there is no sign of Megaspeed-linked chips leaving the country. Renting the capacity of these chips remotely, as Chinese companies have done, remains legal under current rules.

Medium importanceNov. 13, 2026

EU leaders hold in Dublin the bloc's first summit dedicated to AI, with Arthur Mensch and Geoffrey Hinton

The 27 EU leaders meet this Friday (11/13) in Dublin for the bloc's first summit solely on artificial intelligence, called by European Council president António Costa. They will hear Arthur Mensch, CEO of Mistral, and Nobel laureate Geoffrey Hinton. The agenda starts with competitiveness: money, red tape and rules. The plan to build seven AI gigafactories has yet to take off, stuck in the fight among countries over the 2028–2034 budget, and the InvestAI program promised to mobilize €200 billion, €20 billion of it for the gigafactories. The Commission's Apply AI Summit follows in Brussels on November 17.

Low importanceNov. 13, 2026

Trump's AI task force discusses how to make critical sectors report cyberattacks carried out by AI agents

Members of the Trump administration's AI task force met on Thursday to discuss how to get critical US sectors to report AI-powered cyberattacks, according to an administration spokesperson to Reuters, as cited by The Next Web. Emil Michael, the Defense Department's technology lead and the group's vice chair, spoke of progress on national standards for AI developers to report safety incidents. The group also discussed ways for banks, energy companies and emergency services to report flaws and attacks without fear of legal consequences. The law underpinning cyber information-sharing in the US expires next month.

Open Source

1 story

Medium importanceNov. 12, 2026

Liquid AI releases the weights of LiquidCodec, a 74-million-parameter speech codec that encodes 361 times faster than real time on a single CPU thread

Liquid AI published on Hugging Face LiquidCodec, the codec behind the LFM3-Audio-2B and LFM3-Omni-2B models. With about 74 million parameters, it turns 24 kHz speech into tokens a language model can predict and converts them back into audio, with 12 token streams every 80 milliseconds. It replaces Kyutai's Mimi. In the company's own tests, on one CPU thread of a MacBook Pro M4 Max, it encodes about 361 times faster than real time and decodes about 87 times faster, and it had the best reconstruction score on the eight test sets against X-Codec 2, NanoCodec and Mimi, although NanoCodec wins on PESQ in seven of them. There are GGUF versions for llama.cpp.

Agents and Code

2 stories

High importanceNov. 12, 2026

OpenAI launches Daybreak Red Team, which pen-tests web apps with a team of agents, and Defense Factory, which prepares fixes for review

At the Cyber at the Frontier event, OpenAI presented two security products. Daybreak Red Team puts several agents to test web applications: they use the source code and app context to investigate attack paths, confirm vulnerabilities, validate findings from scanners and bug bounty programs and check whether a fix stops the attack, in isolated environments and with scope controlled by Guardian review. Daybreak Defense Factory investigates the flaws found, including those imported from Wiz, Snyk and Endor Labs, and prepares patches that engineers review and test before merging; it can watch new commits in GitHub repositories and open draft pull requests, without deploying anything on its own. According to SecurityWeek, Red Team is in early access on request and Defense Factory comes as a preview plugin in ChatGPT for workspaces with Daybreak access.

Medium importanceNov. 12, 2026

Brave's search API starts accepting per-query payment made by the agent itself, in USDC, through the x402 protocol

Brave began charging for its search API per request via x402, an open payment protocol: a request without payment gets an HTTP 402 response with the price, the agent signs a USDC payment on the Base network and retries the request, and the search comes back with the transaction receipt. That way, developers and autonomous agents use the API without creating an account or managing a key. Pricing follows the company's rate card: US$5 per thousand queries, both for LLM Context, built to give context to agents, and for web, news, image, video and local search. Base network fees are covered by Brave for a limited time, and the suggested path is the Purl command-line wallet, from Stripe.

Local Models

1 story

Medium importanceNov. 12, 2026

The smolmaxxing project runs a 16-billion-parameter model on a TV box and a 2013 iPad with 1 GB of RAM, and a 35-billion one on a phone

smolmaxxing, an open-source project (MIT license) that reached the Hacker News front page, looks for the biggest model each old or cheap device can run and speeds it up with custom kernels, careful memory use and weights read straight from storage. According to numbers measured on the devices, a Xiaomi TV Box S, with about 1 GB of free RAM, runs Qwen3.5-2B at 7.7 tokens per second and Ling-mini-2.0, with 16 billion, at 5.1; a 2013 iPad Air runs the 1-bit Bonsai-1.7B at 24 to 26 tokens per second; and a OnePlus Nord CE 2, from 2022, with about 3 GB free, runs Qwen3.6-35B-A3B at 4.9 to 5.7 tokens per second. The trick is to keep only part of the model in memory and read the rest as needed.

Finance

2 stories

Medium importanceNov. 12, 2026

Wirex launches Navoka, an AI agent that pays bills with its own bank account and card, separate from the user's

Wirex, a stablecoin infrastructure platform, launched in London, by invitation only, Navoka, which pays on the user's behalf from an account of its own: each agent has an IBAN, a debit card and a USDC and EURC wallet, separate from personal accounts. The user sets allowed merchants, limits, daily caps and countries, rules the agent cannot change and that are checked every time the card is used; anything outside them waits for approval. There is a button to freeze the card and an audit trail for each payment. Navoka works through Telegram, with WhatsApp and iMessage on the way, and can be connected to ChatGPT, Claude or Grok. The text is a company press release.

Low importanceNov. 12, 2026

Jane Street more than doubled its NVIDIA stake in the third quarter, before Huang's resignation

According to the 13F form filed on Thursday and reviewed by Fortune, Jane Street held 23.8 million NVIDIA shares on September 30, about US$5.44 billion at US$228.38 each, versus 11.8 million (US$2.4 billion) at the end of the previous quarter. NVIDIA was the firm's second-largest holding, behind an S&P 500 fund. The document does not show when the shares were bought, whether there was a gain or loss, or short positions, and the firm also holds many options. NVIDIA's stock fell almost 13% in five days and hit a 17-month low on Thursday.

Games

1 story

Medium importanceNov. 12, 2026

Indie games are left without a release date on Steam because the approval queue is clogged with games made only with AI prompts

Rami Ismail's new game, Australia Did It: End-of-the-World Railway Defense Business, set for November 10, has been waiting for Steam's review since October 20 and has no release forecast, the developer told Eurogamer. A.S.S. (Amazing Slot Survivor), from TwoMenGames, planned for this Friday (11/13), also lost its date. Valve had already said approval can take up to 12 business days because of the volume of requests, and Drops de Jogos blames the queue on “low-effort” games made entirely by AI, which let a single person publish dozens a day.

Brazil

3 stories

Medium importanceNov. 12, 2026

Itaú adopts the Harness platform, with automation and AI, to standardize how its engineers build and ship software

Itaú Unibanco closed a partnership with Harness, a software development and delivery platform with automation and artificial intelligence, announced on Thursday (11/12) in São Paulo and San Francisco. The bank will use standardized and governed templates, automated checks and policy-based controls, automate rollbacks when something goes wrong and build security, audit and compliance requirements into workflows from the start. After an initial phase, the plan is to expand use to the whole cycle, from code to production, in the coming months. “The future of financial services will increasingly be built on software”, said technology director Carlos Eduardo Mazzei.

Low importanceNov. 12, 2026

HUB Cidades signs a partnership with NVIDIA for the Cidade Aruna technology park in Maringá, planned with AI at its core

HUB Cidades, the innovation environment of Cidade Aruna, in Maringá (Paraná), announced a partnership with NVIDIA, brokered by PD7 Tech, to make the PRC Empreendimentos project the first technology park in Latin America with artificial intelligence as the central element of its planning and operation. The collaboration includes support for research, innovation and training and AI solutions for urban infrastructure, real estate and services. The letter was signed on October 14 at NVIDIA's office in São Paulo. Cidade Aruna plans 290,000 m² of built area, delivery by 2032 and an estimated total sales value of R$2.5 billion.

Low importanceNov. 13, 2026

Startup Impossibl. wants to make AI tokens cheaper in Brazil, starting with a platform that brings several models together and moving to its own data centers

Impossibl., founded by Fábio Roma, wants to reduce the cost of inference, the stage in which models generate answers, in three phases: first, a single platform to access several models, along the lines of OpenRouter; then, renting computing capacity; finally, its own data centers. According to the company, Brazil is the world's third-largest consumer of AI tokens. “Brazilians adopt technology fast, even more so when it delivers more for less”, Roma said.

Social buzz

3 stories

XHigh importanceNov. 12, 2026

Will DePue's thread claiming “conclusive forensic evidence” of Chinese models built on Blackwell chips passes 4.8 million views, and turns into a fight

Will DePue opened the thread with “I have a major result to share” and wrote that he had found “conclusive forensic evidence” that DeepSeek, Moonshot, MiniMax and Ant Group built their newest models on Blackwell chips, barred from export to China. Roon (@tszzl) reacted with “really bad”. Critics attacked two points: @dfranke said everything depends on assuming the model was trained on the same hardware it was initialized on, and @dummerspast39 accused DePue of leaving out of the thread that he found no signature in DeepSeek-V4.1-Flash. DePue replied that this is in the main text of the report and that the mark matches the declared hardware in all 24 models where the lab said where it trained.

Engagement 4.8M views34.4K likes5.3K reposts1.8K quotes1.5K replies17.3K bookmarks (as of Nov. 13, 2026 07:41)

XMedium importanceNov. 13, 2026

Musk writes only “Jensen” quoting an a16z video, and the post passes 3 million views

Elon Musk posted a single word, “Jensen”, quoting an a16z post with a 2-minute-18-second video and the phrase “You're gonna be remembered forever because some things are eternal”. The a16z post passed 2.3 million views, and Musk's, 3 million. The day before, Musk had reposted Tim Sweeney, of Epic, who called Huang's exit “a sad day for NVIDIA, developers, America, and all humanity”.

Engagement 3M views60.7K likes5.6K reposts665 quotes3.2K replies5.4K bookmarks (as of Nov. 13, 2026 07:41)

Hacker NewsLow importanceNov. 12, 2026

Employees' petition for Jensen Huang's return reaches 610 points and 451 comments on Hacker News

The link to WIRED's story about the internal post in which NVIDIA employees call for Huang's return became one of the most commented topics on Hacker News. In the comments, some say a public company CEO could not have stayed on after the affairs and others think the espionage part is secondary, since Alice Huang worked at a customer of the company.

Engagement 610 points451 comments (as of Nov. 13, 2026 07:46)

Videos

4 videos

Fireship · Nov. 12, 2026

Jensen Huang just pulled out of Nvidia...

Fireship sums up in a few minutes Jensen Huang's resignation four days after admitting the affair with a suspected Chinese spy, linked to a small Singapore company that bought US$4.8 billion in NVIDIA GPUs.

Watch on YouTube

IBM Technology · Nov. 13, 2026

Jensen Huang exits NVIDIA & does Claude Fable 5.2 count as AGI

On IBM's Mixture of Experts, the panel discusses whether Huang's long-term bets, from CUDA to high-speed networking, keep NVIDIA ahead without him, and whether Claude Fable 5.2 reopens the conversation about artificial general intelligence.

Watch on YouTube

Stanford Online · Nov. 12, 2026

How AI Agents Use Tools: Function Calling & MCP | Stanford CS329Z Lecture 4 with Diyi Yang

Lecture by Diyi Yang, of Stanford, on how models learn to use tools, from Toolformer and Gorilla to the difference between REST APIs, function calling and MCP, and how to evaluate tool use.

Watch on YouTube

CNN Brasil · Nov. 12, 2026

Caso Nvidia: de suposto affair com “espiã” à renúncia do CEO; entenda o escândalo | CNN BRASIL

CNN Brasil reconstructs the case that led to Jensen Huang's resignation: the alleged affair with Alice Huang, the espionage suspicions, the warning that reportedly reached Trump and the billion-dollar loss in market value.

Watch on YouTube

About this edition

Edition
Edition no. 27 · Morning edition
Date
Friday, November 13, 2026
Coverage
Covers what came out since edition no. 26 (closed around 6 p.m. Brasília time on 11/12) until around 7:50 a.m. Brasília time on 11/13, plus a few stories from the afternoon of 11/12 that had not yet run (all checked at the source; what had already run, such as Gemini 4 Hydrogen, the Gemini glasses, the suggestion of Tim Cook for NVIDIA and the date of Anthropic's Nasdaq debut, was left out). The headline is Microsoft's global outage, which passed 12 hours and knocked out Outlook, Teams and Copilot, affecting the ChatGPT and Claude connectors. The edition also brings Edge0-145B, the analysis that finds the mark of Blackwell chips in Chinese models, Muse's calls, Claude's instability this Friday, Mirendil's US$1 billion round, LangChain valued at US$3 billion, Cerebras with Microsoft as a customer, the NVIDIA employees' petition, the summoning of Jensen Huang and Howard Lutnick by the US Congress, Malaysia's investigation, the EU's first AI summit, Trump's AI task force, LiquidCodec, smolmaxxing, OpenAI's Daybreak Red Team and Defense Factory, x402 in the Brave API, Wirex's Navoka, Jane Street's NVIDIA position, the Steam game queue and, in Brazil, Itaú with Harness, HUB Cidades' partnership with NVIDIA in Maringá and Impossibl. Scan done by the source catalog collection (551 sources in the edition layer, 546 responded) run on the box, plus Hugging Face, GitHub trending, Hacker News, X and YouTube, because the Mac Mini was not connected to Grok Bot at this time and its health page showed inference skipped (Keychain locked). In the log, there was no new story from the hourly watcher since the last edition. Closing: 11/13/2026, around 7:50 a.m. Brasília time.
Sources scanned
Official blogs and changelogs (OpenAI, Anthropic, Google DeepMind, Meta, xAI, Microsoft, NVIDIA, Apple, Mistral, DeepSeek, Qwen, Moonshot), arXiv and HF Papers, Hugging Face and GitHub trending, X, press from the US, Europe, China, Japan, India and Brazil, and YouTube.
Criteria
Only news read in a real source, preferably a primary one, gets in, with every link checked. Pages with access blocks (403) were kept; broken links (404) were discarded.