Sunday, November 1, 2026 Edition no. 3 · Morning
PTENES
Noticiário do Joãozinho Search

Artificial intelligence news from Brazil, in English

Noticiário do Joãozinho: Edition no. 3 · Nov. 1, 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

Top storyEdition no. 3 · Nov. 1, 2026 · Morning

Grok and Claude start shipping with an invisible watermark on by default in the API, because of the European AI Act

xAI published in its API documentation that, starting 11/07, text generated by the grok-4.7, 4.6, 4.5 and 4.3 models and by the grok-4.20 family gets an invisible watermark turned on by default, including in structured outputs and tool-call arguments; the response carries the header x-watermark: true, and team admins can only turn it off per model in the Console, confirming a legal exception. Anthropic, for its part, extends the watermark on 11/03 to Claude Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6 (on Amazon Bedrock it may take a few more days); at Anthropic the watermark cannot be turned off and adds no tokens, and models launched after 08/02/2026 were already watermarked. According to The Decoder, on Opus 4.6 and Sonnet 4.6 the watermark only applies with temperature, top_p and top_k at their default values.

Sources

PDF (pt-BR)

Models

1 story

Medium importanceOct. 30, 2026

SenseTime launches SenseNova-Loop-9B, a model that runs the same block of layers twice

SenseNova-Loop-9B-Preview, released on 10/30 under the Apache 2.0 license, is multimodal (text and image input) and was trained with a context of up to 262 thousand tokens. It has about 9 billion parameters, but runs layers 4 to 27 twice with the same weights: 56 layer executions from 32 stored layers, the equivalent of about 14.6 billion parameters at inference. There is already support via pull requests in vLLM and SGLang.

Source Hugging Face (SenseTime)

Read the story

Research

2 stories

Medium importanceOct. 30, 2026

AI agents now operate the SETI Institute's Allen radio telescope from natural-language commands

The SETI Institute and Breakthrough Listen set up a system at the Allen Telescope Array, in California, in which agents point the 42 antennas and analyze the radio data live from requests in English. The platform uses NVIDIA Holoscan and processes up to 2.4 terabits per second on RTX PRO 6000 Blackwell GPUs at the observatory itself. As a proof of concept, the system made the first real-time, AI-made detection of a pulsar in the Crab Nebula.

Source NVIDIA Blog

Read the story

Low importanceOct. 31, 2026

Meta built a humanoid robot prototype with China's Fourier and is already working on a second version

According to The Information, cited by Reuters this Saturday (10/31), Meta built in about three months, with Chinese startup Fourier, a humanoid prototype called Max to give a body to its Muse agent, and is already working on a second version. The project is internal research, with no plan to sell it. Reuters says it could not independently confirm the information.

Source Reuters / The Information

Read the story

Products

3 stories

High importanceOct. 31, 2026

Apple Intelligence arrives in China in December, with Alibaba's Qwen inside Siri

Apple announced that Apple Intelligence features reach users in mainland China in beta starting 11/04, with the iOS developer beta, and with the release of iOS 27.3 in December. Alibaba's Qwen (千问) model will power features in Siri, Writing Tools and Visual Intelligence. It works on iPhone 16 and later, iPhone 15 Pro and Pro Max, iPhone Air and iPhone Duo.

Source Apple Newsroom (China)

Read the story

Medium importanceOct. 31, 2026

Anthropic renames Claude's plans: Max 20x becomes Ultra and Max 5x is now just Max

The US$200-a-month plan is now called Ultra: it gives 10 times Pro's weekly usage, removes the 5-hour limit for verified accounts, has a separate weekly limit for Claude Fable and includes monthly API credits. The former Max 5x, at US$100, is now called simply Max.

Source Claude Help Center

Read the story

Medium importanceOct. 31, 2026

OpenAI will unify ChatGPT and Codex limits in December, measured in tokens

According to OpenAI's help center, starting in December the Plus, Pro and Business plans will have a single usage limit shared between ChatGPT and Codex, measured in tokens, as Work's agent features enter the main ChatGPT experience. Conversations in Instant mode do not count toward the limit. Usage can be tracked in Settings, under Usage.

Source OpenAI Help Center

Read the story

Business and Investment

4 stories

Medium importanceOct. 30, 2026

Kling AI, Kuaishou's AI video unit, closes a round of about US$3 billion

Kuaishou said on Friday (10/30) that Beijing Kling completed a raise of about 19.8 billion yuan (US$2.98 billion). Kuaishou keeps two thirds of the capital and 52.45% of the votes. The round follows another of US$2.8 billion in July, when Kling was valued at US$15 billion pre-money. According to Bloomberg, the company plans to go public next year.

Source CNA / Reuters

Read the story

Medium importanceOct. 29, 2026

Deepgram buys Daily, creator of Pipecat, and the voice-agent framework moves to an independent foundation

Deepgram announced on 10/29 the purchase of Daily, without disclosing the price. Pipecat, the open framework for real-time voice agents, keeps its BSD-2 license and will be stewarded by an independent Pipecat Foundation; Daily's Kwindla Hultman Kramer becomes Deepgram's head of open source.

Source Deepgram

Read the story

Low importanceOct. 30, 2026

Apple buys Ascend.io, maker of agents for data engineering, a European Union list reveals

The list of acquisitions Apple is required to publish under the European Union's Digital Markets Act (DMA) revealed a deal with Ascension Labs, Ascend.io. It is a team hire: Apple may make offers to employees, receives a non-exclusive license to the intellectual property and may receive patents. Ascend describes itself as a data engineering platform with agents.

Source 9to5Mac

Read the story

Medium importanceOct. 30, 2026

Anthropic negotiates leasing a data center campus near Edmonton, Canada, with a 920 MW gas plant

According to The Globe and Mail, Anthropic is negotiating to be the tenant of a Beacon Data Centers campus in Sturgeon County, Alberta. The project includes four data halls and a 920 MW natural gas plant, approved by Alberta's regulator on 10/23, with construction starting in late 2026 and operation expected by late 2027. Anthropic and Beacon did not comment. The campus would have about twice the power of the EverWind project in Nova Scotia (420 MW).

Source The Globe and Mail

Read the story

Regulation

2 stories

Medium importanceOct. 31, 2026

Judge denies xAI's motion and the NAACP lawsuit over pollution from the Colossus data center will proceed

Federal judge Debra M. Brown, in Mississippi, denied xAI's motion to dismiss the lawsuit filed by the NAACP with the Southern Environmental Law Center, which accuses the company of operating at least 35 gas turbines at the Colossus complex, in Southaven, near Memphis, without the permits required by the Clean Air Act. xAI argued the turbines were mobile, exempt sources; the judge rejected the argument. A similar motion by the Department of Justice had already been denied this month.

Source TechCrunch

Read the story

Medium importanceOct. 31, 2026

OpenAI will train models on sessions from Daybreak Blue, its cybersecurity product, and promises an opt-out after backlash

In an email to members of the Trusted Access program, OpenAI said that starting 11/29 it will use data from Daybreak Blue sessions, until then excluded from training, to train its models, even for those who turned off the option to improve the model. Business and Enterprise accounts are not affected. After the backlash, a second email promised a separate opt-out; interactions flagged in safety reviews can still be used for safety purposes.

Source The Decoder

Read the story

Open Source

1 story

Medium importanceOct. 29, 2026

Pika opens the weights of H3 Flex, a video model with audio that generates in six steps

Pika H3 Flex is a fine-tune of MiniMax H3 that generates video with synchronized audio in six sampling steps, versus 50 for the original model. The weights (LoRA adapters) are on Hugging Face under the MiniMax H3 community license; locally it generates at 480p and 768p, and Pika's API takes it up to 4K and delivers five separate audio tracks. Pika says it is the best open text-to-video model.

Source Pika

Read the story

Agents and Code

3 stories

High importanceOct. 29, 2026

CrowdStrike: China-linked hackers used DeepSeek-powered agents to attack more than 200 organizations in 11 days

According to CrowdStrike, the VAPOR PANDA group used an agent framework called ARTEX, powered mainly by DeepSeek, against more than 200 organizations in at least 14 countries, over 11 days starting 10/17. The intrusion was confirmed in 9 of them; among the 19 targets in the US were pharmaceutical, semiconductor, defense, nuclear energy and financial services companies. A single intrusion consumed 1.4 billion input tokens, and the agents were instructed to look for machines without the Falcon antivirus.

Source CrowdStrike

Read the story

Medium importanceOct. 27, 2026

Anthropic rewrote the Messages API from Python to Rust using Claude

On the Latent Space podcast, Anthropic's Katelyn Lesse and Angela Jiang described how the company rebuilt the Messages API, previously in Python, in Rust, with Claude writing most of the code and tests with real traffic in shadow mode before the switch. According to the account, which spread on X this weekend, CPU usage fell 40% to 50% and time to first token dropped 10% to 20%; the switch was completed in September.

Source Latent Space

Read the story

Local Models

1 story

Medium importanceNov. 1, 2026

Nace.AI opens Drex Max v1, a 27B decision model that runs on Macs with 24 GB

Drex Max does not generate text: it is a decision model that returns probabilities for the answer options. It is a dense 27-billion-parameter model under the Open RAIL-M license. The Q4 quantized version runs on Apple Silicon Macs with 24 GB or more via MLX or Kapitsa, and the GGUF Q8_0 needs 48 GB, with forks of llama.cpp and Ollama. It ranks 4th among 18 models on Artificial Analysis's Decision Index.

Source MarkTechPost / Hugging Face

Read the story

Finance

2 stories

Medium importanceOct. 29, 2026

BTG Pactual launches Sales Trader AI, an assistant that consults its Research and sends equity orders

Launched on 10/29, the tool understands natural-language questions, consults data from the bank's Research and sends orders to the equities platform: market or limit orders, baskets with several assets and portfolio rebalancing. According to BTG's Jerson Zanlorenzi, confirming the order ticket remains with the user.

Source Zum Channel

Read the story

Low importanceOct. 31, 2026

Arca, an AI-native wealth manager, explains how it runs agents on financial tasks that last up to ten days

In an interview with Latent Space, Arca's Aman Singh explained that the company's agents handle tasks of up to ten days, with persistent memory per client and an in-house harness that picks the model per task (Claude is the main one). The agents only analyze and propose: trades and money movement stay with human advisors, and the safety guardrails are deterministic. Arca recently raised US$48.5 million in a round led by General Catalyst.

Source Latent Space

Read the story

Games

2 stories

Medium importanceOct. 29, 2026

Atelico opens early access to its AI engine that runs inside the game, on the player's device

The Atelico AI Engine integrates with Godot, Unity and Unreal and runs language models, image generation, voice (speech and recognition) and lip sync on the player's device; the smallest language model is about 250 MB. It is already used in the games Bobium Brawlers and GARP. According to the company, a scheduler that shares the GPU with the renderer kept a scene at 121 FPS with a 3B model generating text, on an RTX 5080 at 4K, versus 3 FPS without the scheduler.

Source Atelico

Read the story

Low importanceOct. 30, 2026

Ramen open-sources Aura Autopilot, an agent that tests games on its own in Unreal, Unity and Godot

Aura Autopilot plays using a real keyboard and mouse and improves itself as it tests. It is open source (Apache-2.0, on GitHub), works with Claude Code, Codex or your own API key and, according to Ramen, is three times faster than the company's previous verification agent.

Source GamesBeat

Read the story

Brazil

0 stories

No relevant news in this edition.

Videos

2 videos

Latent Space · Oct. 31, 2026

Using Agents for Long-Horizon Financial Work — Aman Singh, Arca

Arca's Aman Singh shows how agents with memory and an in-house harness handle wealth management tasks that last days, always leaving trades and money with humans.

Watch on YouTube

Latent Space · Oct. 27, 2026

How Anthropic ships the #1 LLM in the World — with Katelyn Lesse and Angela Jiang

Anthropic's Katelyn Lesse and Angela Jiang explain how the company launches models, scales the Claude platform and rewrote the Messages API in Rust with help from Claude itself.

Watch on YouTube

About this edition

Edition
Edition no. 3 · Morning edition
Date
Sunday, November 1, 2026
Coverage
Covers what came out since edition no. 2 (closed around 6:30 p.m. Brasília time on 10/31), the overnight watch leads, and news from 10/27 to 10/30 that had not yet been published. A slow Sunday for news: the edition is leaner and the Brazil section has no new story. Closing: 11/01/2026, around 8 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.

News by day

See the full archive

PTENES