NAll news

All news

Saturday482 items

  1. Shao Meng
    AI score68

    Proton is internally pushing hard for AI-written code: CEO calls the security side's concerns "loser thinking," CTO says the old job is gone

    Proton is internally pushing hard for AI-written code and process overhaul. CEO Andy Yen, on Slack, dismissed the security team's suggestion to run a few days of penetration testing before launch as "bullshit" and "loser thinking," while CTO Bart Butler published an AI manifesto stating "Your old job is gone." According to meeting recordings obtained by 404, Butler said that about 90% of code in October was written by AI, and multiple employees said they use Claude extensively; the company also sent an email this week saying it will conduct "re-interviews" on employees' AI usage. Employees feel a culture of fear; officials did not deny the factual description, only acknowledging that communication was too harsh and that they will improve.

    Quoted postJoseph Cox · @josephfcox

    Here is Proton CTO's AI manifesto, that I obtained multiple copies of https://t.co/6o7G2teu9A https://t.co/d92L1kv8AU

    #Industry news#English text is being prepared.#Anthropic

  2. Shao MengAnalysis
    AI score63

    Jia-Bin Huang releases a 9-episode, 230-minute "AI Training and Inference" training-and-inference video series

    Jia-Bin Huang (@jbhuang0604) produced the 9-episode, roughly 230-minute "AI Training and Inference" video series, covering the fundamentals of foundation model training and inference. Shao Meng went through the content of each episode in order: ZeRO and FSDP, five-dimensional parallelism (including the example of Llama 3's 16384-GPU layout), LoRA and its variants, speculative decoding, on-policy distillation, from PPO to GRPO, AdamW, and Muon and MuonClip; Episode 9, LLM Serving, is members-only content. Among these, fine-tuning a 70B-scale model on a single 48GB GPU with QLoRA, and Muon's compute efficiency being about twice that of AdamW, are both claims made by the speaker in the videos.

    Quoted postJia-Bin Huang · @jbhuang0604

    I'm building a playlist/course for LLM training and inference (WIP). https://t.co/0JdUqjiiQm So far, we have common topics like distributed training, fine-tuning, RL, OPD, serving, spec. decoding, and basic optimizers. What else should I include to make this more complete? 🤔

    #Tutorials / practice#English text is being prepared.#Reasoning

  3. Shao MengAnalysis
    AI score66

    Armin Ronacher's Notes on Custom Voices with Qwen3-TTS

    Armin Ronacher documented his practice of creating custom voices with Qwen3-TTS: the Base model extracts an approximately 2048-dimensional speaker embedding from about 28 seconds of audio, and then places it into the same speaker slot in CustomVoice, which allows instructions such as sadness and whispering to be layered onto the cloned voice. The embedding can also be weight-blended, baked back into reference audio, or shifted along the direction of the average male-female difference to shift the perceived gender. He also mentioned that when using meeting clips as references, Qwen3's speaker recognition almost fails, but after switching to about 20 seconds of self-introduction at the start of a podcast, it was about 98% correct on subsequent windows.

    Quoted postArmin Ronacher ⇌ · @mitsuhiko

    My boys and I continue to build our little game with AI and a side adventure has been voice overs with custom voices with Qwen3 TTS. Wrote down my learnings from playing with it for a while now. https://t.co/3wMz0jnMZK

    #Tutorials / practice#Speech#Open-source ecosystem

  6. Shao MengResearch
    AI score79

    Qwenputer embeds a RISC-V CPU in the Qwen3.5-4B weights and uses GDN state as memory

    On Qwen3.5-4B, Qwenputer uses a compiler to write a 32-bit RISC-V CPU into about 11 layers of additional weights, and treats the model's built-in Gated DeltaNet state as addressable RAM, running C and MicroPython in the forward pass without calling an external runtime. In testing, the DeepMind arithmetic set reached about 97.5–100%, NUPA long integers improved from 48.6% for the base model to 100%, and about 230,000 CPU output positions showed zero mismatches against the reference machine. The costs are a checkpoint of about 18.8 GB, several decode tokens required per CPU instruction, and the need to drive it through RvChat rather than bare generate().

    Quoted postdiv (fp32) · @divagr1925

    1/4 Meet Qwenputer: Qwen3.5-4B with a full RISC-V CPU inside its weights, using the model's own Gated DeltaNet layers as RAM. It runs C and MicroPython in its forward passes, no external runtime. On DeepMind math: base 55–90%, base + Python tool 95–100%, Qwenputer 97.5–100%. https://t.co/aA7sAX8sDe

    Same story · 2 reports

    #Papers/Research#Reasoning#English text is being prepared.

  8. Aravind Srinivas
    AI score30

    Perplexity CEO: Agents Will Disrupt Personalized Medicine

    The Perplexity account showcased Computer converting an anonymized shoulder MRI report into an interactive 3D explainer, annotating nine findings on a skeletal model along with six sets of X-ray projections and plain-language explanations. Aravind Srinivas commented on this, stating that with today's agent capabilities, personalized medicine is poised for large-scale disruption.

    Quoted postComputer · @AskPerplexity

    Computer turned a shoulder MRI report into an interactive 3D explainer a patient can read. From one anonymized radiology report, it placed all nine findings on a skeletal model, with six X-ray projections and a plain-language explanation of each. https://t.co/pMflbkz4qA

    #English text is being prepared.#Agents

  9. East Money
    AI score64

    Blackstone Expects Anthropic Unlikely to Obtain Investment-Grade Rating After Listing

    Blackstone Group has told investors participating in Anthropic's $63 billion chip financing deal that it expects Anthropic to receive a credit rating in the higher tiers of junk grade after its IPO, contrary to its earlier statement that it expected an investment-grade rating by the end of the year. Anthropic filed for listing on Monday and plans to go public next month; its prospectus proposes spending more than $800 billion on computing infrastructure by 2045. The company currently holds more than $112 billion in cash and cash equivalents, and investors expect the IPO could raise about another $100 billion.

    Same story · 4 reports

    #Industry news#Anthropic

  10. East Money
    AI score45

    "2026 Hurun Rich List" Released: Total Wealth 32 Trillion Yuan, AI and Embodied Intelligence Become the Main Wealth-Creation Themes

    The Hurun Research Institute released the "2026 Hurun Rich List" on October 27, with 1,529 entrepreneurs on the list and total wealth of 32 trillion yuan, up 7% year on year. DeepSeek founder Liang Wenfeng became the top newcomer with 2,800 billion yuan in wealth and entered the top ten; the AI sector has more than 160 listed entrepreneurs, accounting for over 10%. Unitree Technology's Wang Xingxing made the list for the first time with 695 billion yuan in wealth, making him the second-largest new entrant.

    #Industry news#Embodied AI#DeepSeek

  11. East Money
    AI score30

    A-share Hard Tech Q3 Reports: Computing Power Chain Delivers High Growth, Robotics Track Builds Momentum

    A-share 2026 Q3 reports show that the AI computing power industry chain has the strongest earnings elasticity: Moore Threads' Q3 revenue was 1.55 billion yuan, up 1773.9% year-on-year; Cambricon's revenue for the first three quarters was 9.526 billion yuan, up 106.76% year-on-year; and Foxconn Industrial Internet's Q3 revenue was 383.554 billion yuan, up 57.73% year-on-year. As of the end of September 2026, China's intelligent computing power scale reached 2592 EFLOPS, up 146% year-on-year. Consumer electronics were dragged down by memory price increases and weakening demand, while embodied intelligent robots remain in the early industrialization stage of small-batch deliveries.

    #Industry news#English text is being prepared.#Embodied AI

  12. Phoenix TechnologyOpinion
    AI score53

    Altman Responds to AI's Impact on the Mathematics Community: Understands Mathematicians' Situation, but OpenAI Will Not Stop Research

    Speaking on the "Cheeky Pint" podcast, Altman said he understands the situation mathematicians face amid the flood of AI research results, but OpenAI still needs to continue publishing valuable scientific discoveries, and he acknowledged that the way it previously released mathematics results could be improved. He estimated that OpenAI has invested hundreds of millions of dollars in research on mathematics problems, of which the work related to solving the NS equations cost about $10 million, and the cost of similar work in the future may drop to about one percent of the current level. OpenAI has published multiple AI mathematics results this year; in October it announced three advances in mathematics, computer science, and theoretical physics, among which the Yang-Mills theory results have not yet had complete materials made public, prompting researchers to question independent verification.

    Same story · 3 reports

    #English text is being prepared.#OpenAI#English text is being prepared.

  13. Phoenix TechnologyOpinion
    AI score72

    Washington Post: Why ChatGPT and Copilot Cite the AI-Generated Small Site Ballot Brief

    Washington Post testing found that ChatGPT cited the AI-generated small site Ballot Brief in 22% of election-related answers, while Microsoft Copilot did so in 75% of cases, and Copilot repeated an error from the site. Ballot Brief was created by former Google and Zoom executive Ben Torres Ezrick, who uses digital agents to gather candidate information; he said Microsoft's AI systems cited the site more than 680,000 times in 30 days. The newspaper also found that similar sites such as DecodeTheVote and Smarter.Vote are frequently cited, and a Northwestern University study showed that 16% of the sources cited by chatbots on political, health, and environmental questions were themselves AI-generated.

    #English text is being prepared.#Search#RAG

  14. Sina TechProduct release
    AI score65

    Apple Intelligence China Beta Opens on November 4, Integrates Qwen with On-Device Priority

    Apple has announced that the first batch of Apple Intelligence features will begin testing in mainland China with the iOS developer Beta on November 4, and is expected to become available to domestic iPhone users with the official iOS 27.3 release in December. This version covers Photo Clean Up, Genmoji, system-level writing tools, email summaries, call transcription, notification summaries, and Siri with text input and multi-turn context support, and integrates Alibaba's Qwen large model in scenarios such as Siri, writing tools, and visual intelligence; users can enable it with a Qwen account or through Sign in with Apple authorization. The first batch only supports iPhone 15 Pro, iPhone 15 Pro Max, iPhone 16 and later models, iPhone Duo, and iPhone Air, and does not include higher-tier capabilities such as the next-generation Siri AI announced at WWDC2026.

    Same story · 10 reports

    #Industry news#English text is being prepared.#OpenAI

  15. Sina Tech
    AI score67

    OpenAI lowers fundraising target by $10 billion, valuation drops to $1.25 trillion

    OpenAI has lowered its latest funding round target from $30 billion to about $20 billion, and its valuation from $1.4 trillion to $1.25 trillion. A sell-off in AI stocks in public markets, with the Nasdaq 100 falling 3.5% on Friday, plus rival Anthropic's upcoming listing next month, have put pressure on this funding round. A consortium of several UAE investment institutions (including MGX) will be the largest investor, and OpenAI will gain the option to lease newly built computing resources in the region; Founders Fund, Coatue, Sequoia Capital, Andreessen Horowitz and SoftBank also plan to participate.

    Same story · 8 reports

    #Industry news#OpenAI#Anthropic

  16. Krea AIFirst-party
    AI score12

    Krea Research Lead to Speak at SGLang Summit

    Krea announced that its research lead @sleenyre will speak at the SGLang Summit, covering reinforcement learning and inference scaling for Krea 3. The summit takes place November 12–13 at Fort Mason in San Francisco.

    Quoted postSGLang · @sgl_project

    Sponsor booths and Full Passes for SGLang Summit 2026 are SOLD OUT! Thank you to everyone joining us. 🙌 Get your Expo Pass now to watch select main stage sessions broadcast live in the Expo area, join technical breakouts and hands-on workshops, and explore the GPU Expo. Plus, meet the teams building across the AI stack at community events. 📍 Nov 12–13 · Fort Mason, San Francisco 🎟️ Tickets & agenda: https://t.co/7C6TzrkJYV

    #Industry news#Reasoning#English text is being prepared.

  17. Ethan Mollick
    AI score23

    Mollick: The UX shortcomings of personal agents like Dot

    Ethan Mollick commented on Nick's shared tips for using Dot, saying they are good tips but also expose the current UX problems of Dot and all personal agents: it is hard for users to tell what they are doing, what information they know about you, and what they will do for you. He also joked about OpenAI, saying he will always capitalize the first letter of Dot.

    Quoted postNick · @nickbaumann_

    https://t.co/zxLYdN1xVK

    #English text is being prepared.#Agents#OpenAI

  18. Artificial Analysis
    AI score40

    Artificial Analysis releases the AA-Omniscience v1.1 benchmark

    Artificial Analysis has upgraded its knowledge and hallucination benchmark AA-Omniscience to v1.1, reducing the question bank from 6,000 to 3,000 questions and adding a new 300-question public set, whose Pearson correlation coefficient with the private set's Net Score is 0.99. Gemini 4 Carbon (high) ranks first with a Net Score of 54, while Claude Opus 5.5 and Claude Fable 5.1 score 48 and 46 respectively; among open-weight models, Muse Spark Open (xhigh) leads, trailing the top by 28 points. On the new question set, models' Net Score drops by an average of 1.9 points, with MiniMax M3.1 Flash falling 22 points (from 16 to -6); the median hallucination rate drops by 1.4 percentage points, with Gemini 4 Argon (high) lowest at 4%.

    #Evaluations / benchmarks#English text is being prepared.#Hugging Face

  19. Larsen Cundric
    AI score8

    Larsen Cundric on the Grok bot: your colleague is the product

    Larsen Cundric responded to gaut's praise of the Grok bot (it feels like talking to a coworker, shorter sentences, less fluff), saying that the fun part of making @bot is that your own colleague is also the product you deliver.

    Quoted postgaut · @0xgaut

    Something I've noticed that makes working with grok bot more delightful: it actually feels like you're talking to a coworker. Shorter sentences, less BS and fluff.

    #English text is being prepared.

  20. Testing Catalog
    AI score36

    First Stable Version of Grok Bot Released

    According to a report by TestingCatalog, the first stable version of Grok Bot has been released. SpaceXSI is also developing a Tasks hub and SMS support: Grok Bot will add a dedicated tab for viewing multiple sessions running simultaneously; there is also a new settings tab called "texting," suggesting possible support for iMessage and RCS.

    Quoted post️ · @blankspeaker

    Today smells like a good day For Grok Bot 1.0 "Official" Release announcement. ( you can already download it if you are curious) Mac (Apple silicon): https://t.co/CCEm9yanhm Mac (Apple silicon, zip): https://t.co/LetMCvrht3 Mac (Intel): https://t.co/tnGT6tlAEY Mac (Intel, zip): https://t.co/QAJfBUasGS Windows (x64): https://t.co/x5Zi1AC2kt Windows (Arm): https://t.co/a69ZzzgNIM

    #Product updates#Agents