News

News from November 2, 2026

In my view, the gains attributed to Astra and Anthropic’s TV tests reinforce that promises about AI need evaluations of capability and safety.

The day in three lines

  1. The study’s authors report that Astra changed Qwen’s internal parameters, improving math results and instruction following.
  2. Ubuntu has not yet fixed a critical Linux flaw or said when updates will be available.
  3. Anthropic tested AI programs that spread through smart TVs, exploiting security flaws and adapting to other devices.
  1. Astra can read and write weights to improve Qwen

    A study reports that Astra directly changed some Qwen weights, reducing reasoning loops and improving math and instruction following.

    Hacker News2,949 points · 892 comments

    Quote
    We find that frontier LLMs are surprisingly good at natively reading and writing neural network weights, using a variety of recent tools and algorithms or even via direct weight editing.
  2. Is There Anything Left in the Dream of Cold Fusion?

    A research program involving exotic atoms will not solve the energy crisis, but may push the boundaries of physics.

    ScienceQuanta Magazine

    Quote
    Dig a little deeper and you’ll find a thriving research program with exotic atoms that won’t solve the energy crisis but stretches the boundaries of physics.
  3. The U.S. Census Bureau and “Super Intelligence”

    The U.S. Census changed its survey about artificial intelligence use to call it “super intelligence,” making past and future results harder to compare.

    TechSimon Willison

    Quote
    I was reminded today of AI Adoption Rate Trending Down for Large Companies, a chart produced by Apollo Global Management in September 2025, which appeared to show a slowdown in U.S. corporate AI adoption based on numbers collected by the…
  4. OpenAI appoints Jill Hazelbaker as Chief Corporate Affairs Officer

    Jill Hazelbaker will join OpenAI to lead the company’s global communications and corporate positioning.

    AI labOpenAI

    Quote
    Jill Hazelbaker will join OpenAI as Chief Corporate Affairs Officer, leading the company’s global communications and corporate positioning.
  5. Astronautas de missão simulada a Marte encerram isolamento de 1 ano

    Four people ended a year in confinement during a simulated mission to help prepare for a possible trip to Mars.

    BrazilFolha, Ciência

    Quote
    Os quatro integrantes de uma missão simulada a Marte encerraram, no último sábado (31), um confinamento de um ano no Centro Espacial Johnson, da Nasa, em Houston (EUA).
  6. Does Scaling Toward AGI Depend on What System-1 Lets System-2 Execute?

    A study proposes using a lightweight filter to choose which costly artificial intelligence tasks to run, reducing wasted computing.

    Papernot peer reviewed · 64 votes

    Quote
    Progress toward artificial general intelligence (AGI) has largely relied on scaling a resource-intensive executive, System-2, through larger models, more data and compute, longer reasoning, and more agents.
  7. The mind-blowing reason experts can’t agree on what time actually is

    Experts disagree about what time is, but they have different theories about its nature.

    ScienceBBC Science Focus

    Quote
    It's all wibbly, wobbly, timey, wimey stuff
  8. The latest AI news we announced in October 2026

    Google recapped October artificial intelligence updates, including Gemini 4 Argon and Carbon and a new world model called Genie 4.

    AI labGoogle AI

    Quote
    Here are Google’s latest AI updates from October 2026
  9. IA consegue criar vírus que se espalha por conta própria em eletrodomésticos

    Anthropic tested AI programs that spread on their own through smart TVs, exploiting security flaws and adapting to other devices.

    BrazilG1, Tecnologia

    Quote
    Anthropic testa como "vermes" de IA se comportam em smart TVs. Esse tipo de vírus se propaga por conta própria para outros aparelhos explorando falhas de segurança e se adapta a diferentes tipos de dispositivos.
  10. WTF Ubuntu?

    The author says Ubuntu has yet to fix a serious Linux flaw and has taken their servers offline after losing trust.

    Hacker News1,624 points · 682 comments

  11. Watchful mama bear Gemini wins 2026 Helicopter Mom Award

    Gemini, a mother bear, watched over her cubs while other bears fished at Anan Creek in Alaska.

    SciencePopular Science

    Quote
    Gemini kept an eye on her cubs around other bears fishing in Alaska's Anan Creek.
  12. CWE-bench v2: More coverage, quality, and transparency

    CWE-bench version 2 offers 100 public tasks to test whether models can find and fix security flaws in software.

    AI labHugging Face

  13. Melhor TV com Roku OS: 6 modelos para comprar em 2026

    The guide outlines pros and cons of Roku TVs from AOC, Philco, Semp, Multilaser, and Britânia, described as budget-friendly options.

    BrazilTecnoblog

    Quote
    Televisores com Roku TV são opções com foco em custo-benefício; conheça prós e contras de modelos da AOC, Philco, Semp, Multilaser e Britânia
  14. R^2Orch: Recursive Self-Improvement for Agent Orchestration via Self-Play within a Skill-Internalization Loop

    A study presents a system that learns to coordinate tasks through trial and error, turns failures into skills, and gradually reduces its reliance on them.

    Papernot peer reviewed · 46 votes

    Quote
    In recent years, agentic orchestration has been widely applied to solve complex human tasks.
  15. Whence dragons?

    A selection from the Nature archive explores the ecology and evolution of winged legendary beasts and monkeys’ behavior.

    ScienceNature

    Quote
    The ecology and evolution of winged legendary beasts, and monkeys that show their behinds to their buddies, in this week’s pick from the Nature archive.
  16. Rewarding defenders with bigger bug bounties

    OpenAI will increase payments to independent researchers who find security flaws and offer bonuses for report quality, loyalty, and priority campaigns.

    AI labOpenAI

    Quote
    OpenAI is raising Security Bug Bounty rewards and introducing report quality, loyalty, and priority campaign bonuses for independent security researchers.
  17. Dinossauro do tamanho de peru é apontado como o 1º a consumir frutas

    Fossils suggest that a toothless dinosaur the size of a turkey had adaptations for eating fruit.

    BrazilFolha, Ciência

    Quote
    Cientistas encontraram fósseis de um dinossauro sem dentes, do tamanho de um peru e com uma cabeça semelhante à de um papagaio, que apresentava adaptações a uma dieta especializada em frutas.
  18. Jailbreaking my car's navigation system just so I can change the time

    The previous owner said the Renault needed its map SD card to set the clock, but the buyer discarded it.

    Hacker News943 points · 380 comments

  19. Newborn survival improves when families receive financial support

    A Flint program providing unconditional cash to pregnant people and babies was linked to fewer newborn deaths.

    ScienceScience News

    Quote
    Rx Kids, a program that provides unconditional cash transfers for pregnant people and infants, reduced neonatal mortality in Flint, Michigan.
  20. Transformers Model Integration Contracts

    The team proposed tests using smaller models to catch bugs, including one in UMT5, before they reach the main codebase.

    AI labHugging Face

  21. Mercado Bitcoin sofre ataque e suspende saques dias após roubo bilionário na Binance

    After detecting and containing an attack, Mercado Bitcoin suspended deposits and withdrawals; the exchange says customers lost no money and had no data exposed.

    BrazilG1, Tecnologia

    Quote
    Corretora diz que incidente foi contido e que clientes não tiveram prejuízo nem dados vazados; empresa ainda não informou quando operações serão normalizadas.
  22. Latent Reasoning in Continuous Space for Unified Multimodal Models

    A study introduces a method that reasons using continuous representations to create better images and follow instructions more accurately.

    Papernot peer reviewed · 44 votes

    Quote
    Unified Multimodal Models (UMMs) achieve remarkable success on diverse multimodal understanding and generation tasks, yet they still struggle to effectively leverage language-based chain-of-thought reasoning for image generation.
  23. This bull and donkey show interspecies homosexual bonds can run deep

    A study observed a bull and a male donkey forming a recurring social and sexual bond, but notes they were only two domesticated animals.

    ScienceScientific American

    Quote
    The unlikely pairing suggests that social relationships between males are more complex than we thought
  24. Preparing our next security sprint to harden OpenAI

    OpenAI is preparing an internal effort with Cathedral Systems to test and strengthen its defenses using artificial intelligence.

    AI labOpenAI

    Quote
    OpenAI is preparing its next internal security sprint with Cathedral Systems to test and strengthen its defenses using AI.
  25. Melhor TV custo-benefício em 2026: 10 modelos para comprar

    The list features 4K, MiniLED, and QLED smart TVs priced around R$ 2,500.

    BrazilTecnoblog

    Quote
    Encontre a smart TV com melhor custo-benefício para sua sala, quarto ou escritório; lista inclui modelos 4K, MiniLED e QLED com preço em torno de R$ 2.500
  26. Show HN: Subvocal – Talk to your computer without making noise

    No summary: the source’s text could not be read.

    Hacker News836 points · 391 comments

  27. Lemelson-MIT brings invention education to classrooms from kindergarten to college

    A new initiative in California, Tennessee, and New England puts hands-on activities at the center of science and technology education.

    ScienceMIT News

    Quote
    A new initiative in California, Tennessee, and New England puts learning by doing at the center of students’ STEM pathways, in step with MIT for America’s focus on helping students design, make, and innovate.
  28. Crouton Alpha

    Crouton Alpha is a lightweight model designed to help with repeated coding tasks, tool use, and document analysis.

    AI labOpenRouter

    Quote
    Crouton Alpha is a lightweight model built for scalable coding, agentic, and multimodal workflows. It's ideal for executing agentic sub-tasks, tool use, and document analysis in high-volume, repeated work. Developers...
  29. Hackers atacam supermercados europeus e elevam temperatura de freezers a 50°C

    Problems affected stores in Belgium, Luxembourg, Sweden, and Estonia; similar cases affected hundreds of French establishments last month.

    BrazilG1, Tecnologia

    Quote
    Problemas atingiram lojas na Bélgica, Luxemburgo, Suécia e Estônia; casos semelhantes afetaram centenas de estabelecimentos na França no mês passado.
  30. Apriel Voice 27B Preview: An Open Model for Speech, Reasoning, and Tool Use

    Apriel Voice is an open model that understands speech, considers context, and can respond or use tools in conversations.

    Papernot peer reviewed · 37 votes

    Quote
    A capable voice assistant must do more than recognize words; it must understand requests, reason over context, sustain conversations, and act through tools.
19 more
  1. Large Hadron Collider reveals evidence of Big Bang matter in surprisingly small collisions

    Scientists found evidence that relatively small collisions between atoms can recreate the extremely hot matter that existed shortly after the Big Bang.

    ScienceScienceDaily

    Quote
    Scientists at CERN's Large Hadron Collider have uncovered compelling new evidence that even relatively small atomic collisions can recreate the extraordinary state of matter that filled the universe moments after the Big Bang.
  2. Z.ai: GLM 5.3 FlashX Legal

    Z.ai created a version of its artificial intelligence model with additional training for legal tasks.

    AI labOpenRouter

    Quote
    GLM-5.3-FlashX-Legal (GLM-Legal) is Z.ai's legal-domain variant of GLM-5.3-FlashX, a native multimodal model that accepts text, image, and video input and produces text.
  3. Google vai anunciar Googlebooks que funcionam como tablets

    Google indicated that future Googlebooks may have detachable keyboards, allowing them to be used as tablets.

    BrazilTecnoblog

    Quote
    Próximo formato da categoria Googlebook deve contar com teclado destacável, fazendo os modelos da linha funcionarem como tablet.
  4. If you love a website, back it up

    After losing an old fan site, the author urges readers to save favorite pages and websites using personal web archiving tools.

    Hacker News670 points · 154 comments

    Quote
    This story starts with a little tragedy. See, I’ve got these old webpages I love to go back and look on every now and then. There’s websites that h...
  5. A landmark Harvard study of 110,000 Americans just revealed a surprising truth about ageing

    No summary: the source’s text could not be read.

    ScienceBBC Science Focus

  6. FrameForge: Harnessing Priors of Video Foundation Models for Unified Image Generation

    A study introduces FrameForge, which uses video models to generate images from instructions while preserving aspects of the supplied image.

    Papernot peer reviewed · 34 votes

    Quote
    Unified conditional image generation requires a single model to preserve different aspects of a visual input while producing a target image that follows textual guidance.
  7. Dinosaur blood found in fossils for the first time

    Blood molecules were found in 115-million-year-old dinosaur bones from Brazil.

    SciencePopular Science

    Quote
    Once thought impossible, the molecules were discovered in a 115-million-year-old bones from Brazil.
  8. Ubuntu still has not released patch for Linux 0day RCE in almost 3 weeks

    Ubuntu has not yet released kernel updates to fix a critical vulnerability in versions 22.04, 24.04, and 26.04, or said when they will be available.

    Hacker News625 points · 218 comments

    Quote
    No patched kernels are available yet for 22.04, 24.04, or 26.04 despite a CVSS 9.8 rating.
  9. Research bonuses need clear and fair rules

    A newly appointed supervisor argues institutions need rules for sharing research bonuses with student contributors, rather than relying on supervisors’ generosity.

    ScienceNature

  10. Follow the Gradient: Tracing Safety Degradation Back to Training Data

    A study linked seemingly harmless training examples to dangerous AI responses; removing them reduced successful attacks while largely preserving models’ abilities.

    Papernot peer reviewed · 32 votes

    Quote
    Fine-tuning Large Language Models (LLMs) on instruction data can quietly erode their safety, even when no training example looks harmful. Because this safety-degrading data looks benign, standard content filters struggle to catch it.
  11. Melting permafrost has birthed a new type of large landslide

    An analysis identified large landslides where frozen ground thaws from below and slides downhill, possibly after old fires let water seep underground.

    ScienceScience News

    Quote
    Permafrost detachment slides involve large slabs of frozen earth thawing from below and then sliding downhill.
  12. Fujifilm to boost tape production 5x as AI fuels archiving demand

    No summary: the source’s text could not be read.

    Hacker News533 points · 232 comments

  13. ReLift: Dense 3D Tracking as Lifted Trajectory Repair

    A study presents a method that corrects 3D point paths in videos, including when points disappear behind objects.

    Papernot peer reviewed · 27 votes

    Quote
    Dense 3D tracking from monocular video requires establishing point correspondences over time, estimating their depth, and inferring their trajectories during occlusion.
  14. US boy names that became girl names and vice versa

    No summary: the source’s text could not be read.

    Hacker News489 points · 473 comments

  15. CEPO: A General Framework for Improving Prompt Optimization in Multi-Agent Systems

    A study proposes separating coordination and task instructions in systems with multiple agents, improving the performance of four prompt-tuning methods.

    Papernot peer reviewed · 27 votes

    Quote
    Multi-agent systems (MAS) require both effective coordination and specialized task execution. Existing prompt optimization approaches typically couple these functions within agent-specific prompts.
  16. Show HN: HackerTV – HN's most-recommended Youtube videos as 24/7 TV channels

    HackerTV continuously broadcasts videos recommended on Hacker News across channels, with the same schedule for everyone.

    Hacker News474 points · 81 comments

    Quote
    The videos Hacker News keeps recommending, broadcast 24/7 on channels. Same schedule for everyone.
  17. Wizard: Manipulate Any World in Real Time via Agentic Control

    A study presents a system for creating customized virtual worlds and controlling character movements in real time, including in unfamiliar settings.

    Papernot peer reviewed · 27 votes

    Quote
    Recent advances in video world models have enabled the generation of increasingly realistic and interactive virtual environments, opening new possibilities for player-customizable gameplay.
  18. Show HN: Hacker Times, a New York Times-inspired version of Hacker News

    A website turns Hacker News posts into regularly updated newspaper editions, with independent headlines and links to original sources.

    Hacker News473 points · 148 comments

    Quote
    A newspaper from Hacker News, regularly updated. Independent headlines, original sources.
  19. HiFOR: Hierarchical Foresight with Action-Guided World Grounding for Vision-Language-Action Model

    A team proposed a method that helps robots predict future steps and select which predictions are useful for guiding their movements.

    Papernot peer reviewed · 24 votes

    Quote
    Vision-Language-Action (VLA) models have demonstrated promising capabilities in robotic manipulation by integrating visual perception, language understanding, and action generation.
This week’s papers

How this is made

A script picks the links by votes or by date. An AI summarizes each one using only the source’s text, and writes the day’s three lines from those summaries. Nobody reviews them before they go up, so check the original before citing.

Data in JSON · RSS

Sources: Hacker News, arXiv, Hugging Face, OpenRouter, Quanta Magazine, New Scientist, BBC Science Focus, Popular Science, Nature, Science News, Scientific American, MIT News, NASA, ScienceDaily, Agência FAPESP, Pesquisa FAPESP, Folha, Ciência, G1, Tecnologia, Tecnoblog, MIT Technology Review, WIRED, IEEE Spectrum, Ars Technica, The Verge, TechCrunch, The Register, Simon Willison, OpenAI, Google DeepMind, Google AI, Hugging Face.