Thursday, November 19, 2026 Edition no. 39 · Morning

Artificial intelligence news from Brazil, in English

Noticiário do Joãozinho: Edition no. 39 · Nov. 19, 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. 39 · Nov. 19, 2026 · Morning

Update: Anthropic expands IPO to 980 million shares, raises $81.3 billion and debuts today on the Nasdaq; with the private sale, total fundraising reaches $107.3 billion

Anthropic announced it sold 980 million Class A shares at $83 each, against the 825 million expected when it opened the book: 956.35 million came from the company itself and 23.65 million from current shareholders, and banks have 30 days to purchase up to an additional 147 million. The shares begin trading this Thursday (11/19) on the Nasdaq Global Select under the ticker ANT, with settlement scheduled for Friday (11/20). The public offering yielded approximately $81.3 billion, and combined with the $26 billion in non-voting shares sold to AMD, Broadcom, Google, Micron, Microsoft, NVIDIA, and TSMC, total fundraising reaches $107.3 billion, according to Bloomberg and InfoMoney, which describe it as the largest stock market debut on record. On a fully diluted basis, the company is valued at approximately $1.5 trillion. According to TechCrunch, citing the prospectus, each of the seven co-founders, including CEO Dario Amodei, sold nearly 3.4 million shares, totaling about $282 million each. Folha says that non-founder employees will be able to sell up to 10% of their stakes on the first day; Hill estimates 598 of them hold at least $100 million in shares, and Sacra estimates around 2,800 current and former employees hold at least $1 million. The company had approximately 5,300 employees as of September 30 and, according to InfoMoney, 11,900 clients paying at least $100 thousand per year. On the eve of the debut, European markets rose (Stoxx 600 up 0.3% at the opening, according to Bloomberg) and Nasdaq futures were up nearly 2% in the early hours.

PDF (pt-BR)

Models

1 story

Medium importanceNov. 17, 2026

Xing 4, successor to TeleChat, trained exclusively on Chinese chips: a model of 285 billion parameters across 8,192 Ascend 910C chips

The article "TeleChat to Xing4", the top-voted of the day on Hugging Face's papers page this Wednesday (11/18), describes the Xing 4 model series, which extends TeleChat and was developed entirely on Huawei's Ascend chips. According to the abstract, pre-training a 285-billion-parameter model ran on 8,192 Ascend 910C chips using the MindSpore framework. The paper covers sparse activation, attention design, model depth and width, and speculative decoding, while post-training utilizes TeleGym, an agent execution platform. The Hugging Face page lists Xing4.0-29B-A4B among the models linked to the paper. The abstract does not include benchmark figures.

Research

2 stories

Medium importanceNov. 19, 2026

AI agents achieve at most 42.5% accuracy in emergency CT and MRI diagnoses; radiologists score 77%

A study published on arXiv tested six vision-and-language models as agents navigating emergency CT and MRI scans: 100 exams, four navigation strategies, and 4,800 executions. Accuracy was lower when the model navigated freely through slices (32.6%) and higher when it received all slices or only the key slices at once (42.2% and 42.5%). In a subset of cases, models navigating freely achieved 30% accuracy across executions, compared to 77% for nine radiologists. The agents formed their initial hypothesis after viewing a median of 10.8% of the slices and maintained it in 63.1% of executions; when the initial hypothesis was wrong, the final answer was correct in only 12.4% of cases. Declared confidence peaked at 90% in the median, yet only 38.9% of the answers were correct.

Low importanceNov. 13, 2026

Google and Penn State propose a memory so the agent continues improving its own tool without modifying the model's weights

Researchers from Google Cloud AI, Google DeepMind, and Penn State present Self-Improvement Memory (SIM), a memory that stores, for the optimizer rewriting the tool around the model, the reason for each change, the predicted outcome, the observed outcome, and the evaluation context; the model's weights that solve the tasks remain frozen. In four domains, the full version of SIM outperformed alternatives that relied solely on the tool or solely on causal analysis. In ten independent programming runs, it won nine of the ten paired comparisons, with an average gain of 2.36 points (95% confidence interval between 1.05 and 3.68). The authors report one negative pair and note that most comparisons used one run per condition.

Products

4 stories

Medium importanceNov. 18, 2026

Google Search's AI Mode Gains Temporary Conversations That Don't Affect History or Personalization

Google has begun offering temporary conversations in its Search AI Mode. These sessions do not appear in search history or get used for future recommendations, and are accessed via an icon in the top-right corner of AI Mode without altering any settings. Once started, the session is disconnected from your account: prior preferences do not influence responses, and the conversation does not impact future personalization. According to Google, the feature is rolling out around the world to computers and phones in all languages where AI Mode works.

Low importanceNov. 19, 2026

Instagram to Test Tool That Finds Video Copies, Including Those Made with AI That Reuse Original Audio

Meta announced on its creators blog, on 11/17, that Instagram will roll out a copy detection tool to select creators, which scans Instagram and Facebook for reproductions of original videos and lets the author decide what to do. It uses the same matching technology as Rights Manager and also flags AI-generated content that reuses the original video's audio with swapped images. The feature already existed for Facebook creators, and Instagram says it will begin testing with a small group in the coming weeks. According to MediaNama, one creator said AI replicas of their work garnered 50 million views and 1 million likes in just a few days.

Low importanceNov. 18, 2026

Pangram Detector Flags 99% of Award-Winning Peruvian Novel as AI-Generated; Publisher Retains Prize

The AI text detector Pangram classified 99% of "Donde el agua trabalha," an award-winning novel by Peruvian author David Silva Rojas, as machine-written. Mexican publisher Almadía retained the award after reviewing documents submitted by the author. According to El País, as cited by The Decoder, Pangram also attributed over 80% of interview transcripts, an essay, and the author's Instagram denial to AI. French publisher Gallimard plans to use Pangram to screen submissions starting next year, and the Authors Guild warns against relying solely on detectors for decision-making.

Low importanceNov. 18, 2026

Foundation, backed by Eric Trump, unveils Phantom MKII humanoid, which carries 200 pounds, and does not rule out arming it

Foundation unveiled the Phantom MKII, a humanoid robot that carries 200 pounds (about 90 kg), compared to around 50 to 60 pounds of the previous version, with a battery life of up to six hours, and the Cortex II control system, which according to the company learns some tasks with just 60 demonstrations. Production begins in early 2027, with a few hundred units first and "a few thousand" by the end of the year, initially in the San Francisco Bay Area. CEO Sankaet Pathak told Fast Company that the company is not "against militarization and war" and disputed a Financial Times report about talks with Israel's armed forces.

Business and Investment

3 stories

Medium importanceNov. 18, 2026

Nadella Explains Why He Offered Voluntary Exit to 60,000 Microsoft Employees and Says Cowork Team Has 11 Developers

At the New York Times DealBook Summit this Wednesday (11/18), Microsoft CEO Satya Nadella told Andrew Ross Sorkin that the voluntary exit program offered to 60,000 employees in October served so that "everyone felt they had the power of choice." He stated that software engineering "is very much alive" and will thrive, but organized differently, citing the Cowork team (the AI agent for Microsoft 365), which consists of 11 developers who "build a software factory that builds the product." Nadella also noted that transforming a 51-year-old company will take "a little more time than anyone expects".

Medium importanceNov. 18, 2026

Update: In NVIDIA's first all-hands meeting without Jensen Huang, the company moves annual raises from March to December; Musk compares his ouster to benching Michael Jordan

NVIDIA acting CEO Colette Kress addressed employees this Wednesday (11/18) at the first all-hands meeting since Jensen Huang's resignation, announcing that the annual review, when raises and new stock grants are distributed, has been moved from March to December, according to CNBC. New board chairman Harvey Jones discussed the search for a permanent CEO; the meeting lasted roughly an hour, conducted both in person and via Microsoft Teams. NVIDIA shares are trading 37% below their 52-week high. At the DealBook Summit, Elon Musk referred to Huang as 'Michael Jordan at his peak' and stated he does not believe Huang is disloyal to the U.S.; according to TechCrunch, Musk signed a petition with over 200,000 signatories calling for his return, which the board has already dismissed.

Low importanceNov. 18, 2026

India's Humyn Labs raises $47 million to sell training data to American robotics companies

Humyn Labs, founded in February by Manish Agarwal, former chief executive of Nazara Technologies, and Ishank Gupta, former executive at AB InBev, raised $47 million in a Series A round: $26 million in equity, led by Accel with Cactus Partners, and $21 million in debt, led by the Lighthouse Canton Growth Debt Fund II. The company has 12 paying customers, and about 90% of revenue comes from the US, according to Gupta. Agarwal said revenue from April to October was $10 million, with $20 million in signed contracts and another $20 million in negotiation. There are 51 full-time employees.

Regulation

2 stories

High importanceNov. 19, 2026

In Shenzhen for the meeting with Xi and Putin, Trump says "Americans invented superintelligence" and opposes regulating AI worldwide

Trump argued against global AI regulation and said "Americans invented superintelligence," a term he uses for AI, warning that any "robust" rule would "stifle" development, according to a Bloomberg video from this Thursday (11/19); Bloomberg notes that the US is simultaneously imposing new restrictions on access to advanced AI. The day before, Xi Jinping said at the APEC CEOs' meeting that "unicorn companies and large AI models have emerged in large numbers" in China and advocated for AI cooperation, according to The Next Web. US authorities told CNBC that Trump will not raise the NVIDIA case with Xi, and a Verisk Maplecroft analyst told CNBC that cooperation could include a direct AI security hotline. The three-way meeting in Shenzhen has nuclear weapons at the center of the agenda, and the trade truce lasts until January.

Medium importanceNov. 18, 2026

Nadella says much of critical infrastructure cannot be patched, and Palmer Luckey warns AI-driven attacks are outpacing defenses

At the DealBook Summit, Satya Nadella said there is "a lot of critical infrastructure that cannot be patched," citing firmware, coolers, and HVAC systems, and noted there is now "a very powerful tool that finds vulnerabilities," which forces patches to reach production almost immediately; he attributed the Outlook outage to a "latent bug." At the same event, Palmer Luckey, founder of Oculus and co-founder of Anduril, said advanced AI tools are enabling the creation of various types of exploits, cited North Korea, and summarized: "It's time to start running." According to Andrew Ross Sorkin, Vice President J.D. Vance canceled his participation in the event due to the outages.

Open Source

1 story

Medium importanceNov. 16, 2026

Hugging Face now routes language model requests through OpenRouter, selecting providers by price, speed, and stability

Hugging Face's inference providers now route language model requests through OpenRouter, which aggregates over 80 providers and selects them based on price, speed, and reliability, with automatic failover when one goes down. According to OpenRouter measurements cited by The Decoder, capacity-related errors dropped by 46%, time to first token improved by 16.9%, and throughput increased by 16.3%. Hugging Face states it does not charge an additional fee and that the share of open models in OpenRouter's tokens rose from 26.8% in October 2025 to 71.7% in September 2026, peaking at 77.5% in July; on average, open models cost $0.28 per million input tokens and $1.62 for output, compared to $1.27 and $6.36 for closed models. The change breaks one thing: providers' own keys no longer work for language models.

Agents and Code

2 stories

Low importanceNov. 19, 2026

billion-context, which compresses code agent context in month-long sessions, rises on GitHub with 460 stars in a day

billion-context, one of GitHub’s trending repositories this Thursday (11/19), has garnered 460 stars in a single day. It positions itself as a context compression plugin for month-long sessions spanning billions of tokens, featuring small windows ("100,000 tokens of context are enough") and token savings ("5 times less"). It operates between the code agent and the model API, injecting tools that allow the model itself to compress, decompress, and retrieve older conversation segments. It is compatible with Claude Code and OpenCode, and the project notes it is still in its early stages, with testing on real models as the next step. It uses an MIT license, which requires attribution.

Medium importanceNov. 18, 2026

Nadella predicts AI agents will reshuffle the web economy and says those who don't accept them will lose to those who do

Satya Nadella said at the DealBook Summit that "current aggregators, which have a business model built for the human web, will in some way be reshuffled." In his view, those who make money from ads driven by user eyeballs will likely have to shift to cost-per-click or transaction-based models. The conversation touched on Amazon, which blocks AI agents in its store; according to TechCrunch, if it continues blocking them, another platform like Shopify could accept them and gain an advantage. "If you're attached to a business model, you'll have to run," he said.

Local Models

2 stories

Medium importanceNov. 18, 2026

Tether Adds to QVAC SDK Compressions That Cut Local Model Context Memory by Up to Five Times

Tether has integrated TurboQuant and TurboVec into the QVAC SDK, its local AI toolkit. TurboQuant compresses the context cache and reduces memory used to store context during inference by up to five times; TurboVec compresses search indexes, and ten million typical vectors drop from about 61 GB to around 8 GB, according to the CIO. TurboQuant was included in SDK version 0.12.0, for Windows and Linux with NVIDIA, AMD, and Intel GPUs.

Low importanceNov. 14, 2026

Tiny-voca, a 4.5 million-parameter voice model that runs 70 times faster than real-time on four CPU cores, enters the most popular models on Hugging Face

The tiny-voca, which appears among the trending models on Hugging Face, converts text to English speech with two American voices, one female and one male, and has about 4.5 million parameters, trained from scratch using 17.6 hours of synthetic speech. According to the model's page, it runs about 70 times faster than real-time on four CPU cores, and there are PyTorch (18.2 MB), ONNX, and GGUF versions starting from 8.7 MB. The license is Apache 2.0. The author themselves list the limitations: no streaming, no voice cloning, and only two fixed voices.

Finance

3 stories

Medium importanceNov. 18, 2026

Update: Palo Alto Networks rises more than 7% on $3.44 billion revenue, and Arora says Applied Compute acquisition could surpass its cybersecurity business

Palo Alto Networks shares rose more than 7% after the company beat estimates for its first fiscal quarter of 2027, according to CNBC: revenue was $3.44 billion, against expected $3.31 billion, a 39% increase, and annual recurring revenue from its next-generation security segment grew 66%, to $9.73 billion. The company raised its fiscal year revenue forecast to between $14.52 billion and $14.57 billion. Regarding the acquisition of Applied Compute, specializing in open model customization, for $4.5 billion in stock, CEO Nikesh Arora said: "If it works, I think this will surpass what we have in cybersecurity."

Medium importanceNov. 18, 2026

NVIDIA Delays Filing of 10-Q Quarterly Report with SEC, UBS Cuts Price Target From $240 to $220

During the earnings call, acting CEO Colette Kress stated that NVIDIA will file its 10-Q quarterly report with the SEC "in the coming weeks," without providing a reason; the company typically files it on the same day as its earnings report. According to TechCrunch, UBS analyst Timothy Arcuri lowered the stock's price target from $240 to $220 and said management linked the delay to an ongoing investigation into Huang. The Wall Street Journal reported that the SEC is investigating whether NVIDIA's disclosures misled investors. NVIDIA did not respond to TechCrunch.

Medium importanceNov. 18, 2026

Citibank and PNC Update Contracts: Customers Who Grant AI Agents Access to Accounts Are Liable for Their Actions, Even If Hacked

Citibank updated its online banking terms of use to state that customers who grant third-party AI agents access to accounts "are responsible for activities conducted by these agents," according to a spokesperson to TechCrunch; under the terms, customers are also liable if the tool leaks credentials or is breached. PNC updated its terms on the same day, and Citizens sent out similar warning emails, following JPMorgan and Bank of America's announcement of similar rules last month.

Games

1 story

Low importanceNov. 18, 2026

Snapchat launches Easy Lens CLI, which creates AR Lenses and games with Claude Code, Codex, or Cursor

Snap has launched Easy Lens CLI, a free command-line tool for Snapchat account holders that works with Claude Code, Codex, and Cursor on Mac, Windows, and Linux: the developer describes in words the Lens they want. The tool includes over 100 ready-made augmented reality blocks and more than 50 AI-powered asset generators. According to TechCrunch, Snap says a game can go from idea to published Lens in a single day, and that it paid $1.8 million to Lens+ creators last month.

Brazil

2 stories

High importanceNov. 19, 2026

Update: US rule requiring a license to sell AI chips to Brazil takes effect today, and Abes speaks of 'competitive asymmetry'

The US Department of Commerce rule that conditions the export of advanced AI chips on a license begins to take effect this Thursday (11/19), the day it is published in the Federal Register, according to Folha and TV Sim. Brazil is on a list with Singapore, Malaysia, Thailand, and Indonesia whose purchases of chips from classes 3A090 and 4A090 undergo case-by-case review. The regulation gives the BIS 90 calendar days to decide or refer the case to the Presidency; in 2025, the agency took an average of 62 days to process about 30,500 requests, and in a CSIS survey of 31 American companies, 56% reported an average wait time exceeding 180 days and 33% over 300 days. Companies under American control are exempt if they submit monthly equipment reports. In a statement, the Brazilian Software Companies Association (Abes) said the requirement "creates a competitive asymmetry within the Brazilian market itself." The government has set up a task force, but still has no formal response.

Low importanceNov. 19, 2026

99% of Brazilian leaders already use AI at work and 81% use it daily, above the global average of 73%, says a monday.com study with Nielsen

A monday.com study with Nielsen, titled "Squinting in the Sun", surveyed 1,701 directors and managers of companies with more than 500 employees across seven countries, 200 of them in Brazil. Among Brazilians, 99% use AI at work and 81% do so daily, compared to 73% globally. Among Brazilian leaders who changed evaluation criteria, 62% now value knowing how to write good prompts for AI, compared to 46% worldwide. For 30%, the biggest challenge is refining what AI delivers (16% globally), and 44% cite technology dependence as a concern, compared to 34% overall.

Videos

3 videos

The New York Times · Nov. 18, 2026

What OpenAI Did After the Hugging Face Hack

Live session at the 2026 DealBook Summit in which Andrew Ross Sorkin interviews Sam Altman and Greg Brockman, chief executive officer and president of OpenAI.

Watch on YouTube

Stanford Online · Nov. 18, 2026

Terminal-Bench: Evaluating AI Agents | Stanford CS329Z Guest Lecture with Alex Shaw & Ryan Marten

Guest lecture from Stanford's CS329Z course, with Alex Shaw and Ryan Marten, on Terminal-Bench and the evaluation of AI agents.

Watch on YouTube

Two Minute Papers · Nov. 19, 2026

DeepMind’s New AI Is A Gift To Humanity

Episode of Two Minute Papers on a new AI paper from Google DeepMind.

Watch on YouTube

About this edition

Edition
Edition no. 39 · Morning edition
Date
Thursday, November 19, 2026
Coverage
Covers what has been published since issue No. 38 (closed early on the night of 11/18) until the morning of 11/19, the day Anthropic debuted on the Nasdaq, the meeting between Trump, Xi, and Putin in Shenzhen, and the implementation of the U.S. rule requiring a license to sell AI chips to Brazil.
Sources scanned
Automated scan of 701 sources with active feeds (691 responded, 2,370 items in the window) performed on a Mac mini's local AI, which grouped, filtered, and compiled fact sheets with facts verified against the text of each page; followed by a supplementary scan of The Decoder, TechCrunch, Hugging Face, GitHub, official blogs, and Brazilian press, along with editorial review against the sources and previous editions. Among this edition's sources: Anthropic, Bloomberg, CNBC, TechCrunch, Deutsche Welle, InfoMoney, Folha, Fast Company, CIO, MediaNama, The Decoder, Snap, Hugging Face, arXiv, and TV Sim Brasil.
Criteria
Only news sourced from a legitimate, preferably primary, outlet with all links verified is included. Pages blocked (403) were retained; broken links (404) were discarded.