NAll news

All news

Saturday482 items

  1. UCSD Hao AI LabFirst-partyResearch
    AI score54

    Hao AI Lab proposes JetSpec++: pipelining draft generation and verification, achieving 1.54x speedup on Qwen3-8B

    Hao AI Lab proposes JetSpec++, which pipelines draft generation and verification through causal parallel drafting, a hidden-state predictor, and an anchor-token decoder, achieving lossless speedup on Qwen3-8B and Qwen3.8-27B. Compared with DSpark, it achieves up to 1.54x end-to-end speedup, and 7.45x speedup over autoregressive decoding, reaching 1,361 TPS and 367 TPS respectively under BF16; with NVFP4 and kernel optimizations, it exceeds 4,000 TPS and 900 TPS on two GB200s. Across 7 benchmarks, it accepts 8.94 tokens per round, compared with 6.89 for DSpark; the model, code, and inference engine have all been open-sourced.

    Same story · 5 reports

    #Model releases#Reasoning#Open-source ecosystem

  2. UCSD Hao AI LabFirst-party
    AI score47

    Hao AI Lab open-sources JetSpec++, a speculative decoding method with up to 7.45x speedup

    Hao AI Lab has open-sourced JetSpec++, which pipelines draft generation and verification through causal parallel drafting and a hidden-state predictor, achieving 1,361 TPS and 367 TPS on Qwen3-8B and Qwen3.8-27B respectively—up to 1.54x faster than DSpark and 7.45x faster than autoregressive decoding while remaining lossless. On Qwen3.8-27B, the number of tokens accepted per round improves by 22% over DSpark (6% for 8B); trained with only 16 blocks, it extrapolates to 32 draft blocks, reaching 14.55 (8B) and 14.05 (27B) tokens per round on MATH-500, where DSpark drops to 8.66 and 8.49. Combined with NVFP4 and kernel optimizations, Qwen3-8B exceeds 4,000 TPS and Qwen3.8-27B exceeds 900 TPS on two GB200s.

    Same story · 5 reports

    #Model releases#Reasoning#Open-source ecosystem

  3. UCSD Hao AI LabFirst-partyProduct release
    AI score59

    Hao AI Lab Open-Sources JetSpec++ Models, Code, and Inference Engine

    Hao AI Lab announced that JetSpec++'s models, code, and inference engine are all open-sourced, along with links to the paper, code, and Hugging Face. The method pipelines draft generation and verification, achieving up to 1.54x end-to-end speedup over DSpark and 7.45x over autoregressive decoding on Qwen3-8B, reaching 1,361 TPS in BF16; with NVFP4 and kernel optimizations, it exceeds 4,000 TPS on two GB200s.

    Same story · 5 reports

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

  4. Swyx
    AI score17

    swyx says the AI Engineer conference brings together a thousand engineers

    swyx stated that he wrote about this on AINews: the event is bringing together 1,000 top AI engineers, first to accurately present the SOTA of AI x Code in 2026, and second to preview 2027. He also said he is proud to have convened a group of invited attendees, speakers, and sponsors to echo this year's three major themes.

    Quoted postmichelle · @michellechen

    hiding amongst the greats. see you at https://t.co/HfcmDb4hBz https://t.co/QF6it1JqTR

    #Industry news

  5. Elon Musk
    AI score33

    Musk reposts new Neuralink video

    Sawyer Merritt said Cleo Abram visited Neuralink's Austin headquarters and released a new video, showing its next-generation R1 surgical robot for the first time. Musk reposted it and commented, "Great video about @Neuralink."

    Quoted postSawyer Merritt · @SawyerMerritt

    Cleo Abram (15M followers across all platforms) just released a new video about @neuralink. She visited the company's Austin HQ. "This is Neuralink’s Next-Gen R1 surgical robot. Its main job is to put a computer chip into people's brains, and this is the first time they're unveiling it." https://t.co/wzTuUq2Or0

    #Industry news#Embodied AI

  6. Rohan PaulAnalysis
    AI score59

    Voice Arena launches the Turing TTS benchmark: the best model fools listeners only about 10% of the time

    Voice Arena has launched the Turing TTS benchmark, in which listeners choose the real human among a real human recording and a model-generated recording of the same sentence; the proportion of wrong choices is the Fool Rate, and 50% means indistinguishable. The best-performing Gemini 3.8 Flash TTS made listeners choose wrongly in about 10% of cases. On Voice Arena's preference leaderboard, the top four models are statistically tied, but overall they still show a clear gap relative to real humans.

    Quoted postShobhit Banga · @shobhitbanga

    Drop 2/15 Are we actually measuring the things we ultimately want our voice AI models to achieve? The best AI voices are becoming hard to tell apart. From each other. Not from humans. 🚀 Introducing Turing TTS: an evaluation that measures how often human listeners mistake AI-generated speech for real human speech. We tested leading TTS models against real human speech. Even the best model fooled listeners only ~10% of the time. With Turing TTS, we hope to contribute to a broader discussion within the research community about how we measure progress in voice AI. If one of our goals is to develop speech that is indistinguishable from human speech, how well do our current evaluation methods capture progress toward that goal? We don’t claim that Turing TTS is the right way forward, or that human indistinguishability should be the goal of every TTS system. Rather, we hope these findings encourage the community to collectively examine whether blind pairwise human evaluations of naturalness are sufficient, and what else we might need to measure. Results are now live on @voicearena_ai.

    #Evaluations / benchmarks#Speech#Google

  7. Ethan Mollick
    AI score14

    Ethan Mollick Responds to the Question of Whether the Open Web Will Survive

    In response to Mike Isaac's question about whether the open web will still exist in ten years, given that Google is eroding Reddit, one of the last sources of human search results, Ethan Mollick said the answer is probably no, and suggested joining private Discord communities made up of real people to hold the line there.

    Quoted postrat king 🐀 · @MikeIsaac

    reddit basically one of the last places i search for human results and now google is shredding that. does the open web even exist in ten years?

    #English text is being prepared.

  8. ArenaProduct release
    AI score70

    Arena Weekly: Gemini 4 Carbon Tops the Text Leaderboard, GPT-6.1 Terra Enters the Agent Leaderboard at No. 8

    Arena published this week's leaderboards: Google DeepMind's Gemini 4 Carbon (High) ranks No. 1 in Text Arena, No. 2 in Agent Arena, and No. 4 in Code Arena: WebDev, and has changed the Pareto frontier in three arenas. OpenAI's GPT-6.1 Terra (Max) enters Agent Arena at No. 8, with a net improvement of +10.6% and a median cost of $0.62 per task. Microsoft AI's MAI-Image-2.7-Preview debuts at No. 3 in the single-image editing arena, 45 points higher than the previous version 2.6's No. 6 ranking. Arena has also launched Connectors for Agent Mode, allowing users to combine tools such as Vercel, Linear, and GitHub within a single task to complete more complex workflows.

    #Evaluations / benchmarks#Agents#English text is being prepared.

  9. Ayi AI Notes
    AI score75

    TPU on Google Satellite Completes First Space AI Inference and Returns Results

    Google's Project Suncatcher prototype satellite booted up a TPU in low Earth orbit, received a ground command to complete a model inference, and transmitted an AI-generated haiku back to Earth. Sundar Pichai called this the project's hello world milestone. The author believes that space-based solar power and deep-space radiative cooling can bypass the power and cooling bottlenecks of ground-based data centers.

    Quoted postSundar Pichai · @sundarpichai

    A ‘hello world’ milestone for Project Suncatcher 🚀 After launching our prototype satellite a few weeks ago, we’ve now successfully sent a command up to orbit, booted up the TPU, ran the prompt, and received our first complete response back on earth - a haiku! https://t.co/kOFmETUskm

    Same story · 8 reports

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

  10. Testing Catalog
    AI score36

    Maket Launches Move Rooms Floor Plan Feature

    Maket has launched the Move Rooms feature, allowing users to drag an entire room to a new position, with Maket automatically reshaping the surrounding walls, doors, and circulation. This feature supports trying out different layouts on an in-progress floor plan, for example dragging one room onto another to swap them; surrounding rooms automatically compress to make space, while exterior walls and windows remain unchanged.

    Quoted postMaket · @Maketplans

    Your floor plan isnʼt final until you say it is. Move Rooms is now live. Drag. Drop. Done. https://t.co/HmZFngZ6gt

    #Product updates

  11. OpenRouterFirst-partyProduct release
    AI score55

    OpenRouter launches The Biological Computing Co.'s…

    OpenRouter launches The Biological Computing Co.'s mischief 1, which can turn images into video; the model is based on four years of research culturing real brain cells on electrodes. It supports inputting a starting image and a prompt, and outputs 3-second or 5-second clips at 480p or 720p, with pricing starting at $0.005 per second.

    Quoted postModels on OpenRouter · @ModelsOnOR

    TBC mischief 1 is now live on OpenRouter 🧠🎬 Image-to-video from The Biological Computing Company: give it a starting image and a prompt, get back a 3s or 5s clip at 480p or 720p. From $0.005 per second of video. https://t.co/9MvRwXhEM8 https://t.co/QmlWbFyVUO

    Same story · 3 reports

    #Product updates#Video#Image generation

  12. Karina Nguyen
    AI score10

    Karina says Gemini Argon 4 is stunning

    Karina posted that Gemini Argon 4 is very stunning, and attached an image link. The post quoted Thoughtful's post about having 9 models bet real money on Kalshi on the midterm elections, but did not further explain the connection between the two.

    Quoted postThoughtful · @thoughtfullab

    Our AI forecasters now trade the midterms! We gave 9 models $50 each in real money on Kalshi + $10K each in simulation to place bets based on their forecasts. 24 hours in: Claude Fable is up $1.59 (+3.2%), Gemini $0.80 (+1.6%) Watch live: https://t.co/YgKV75W0fC https://t.co/vmtVan8ydc

    #English text is being prepared.#Google

  13. Philipp Schmid
    AI score58

    Philipp Schmid relays that Google runs Gemma 4 on a TPU in space

    Philipp Schmid relayed Sundar Pichai's message, stating that Google's Project Suncatcher prototype satellite booted up a TPU in orbit, ran a prompt, and a complete response was received on Earth—a haiku. On that basis, he said Gemma 4 became the first LLM to run in space, but the original post did not name the model used.

    Quoted postSundar Pichai · @sundarpichai

    A ‘hello world’ milestone for Project Suncatcher 🚀 After launching our prototype satellite a few weeks ago, we’ve now successfully sent a command up to orbit, booted up the TPU, ran the prompt, and received our first complete response back on earth - a haiku! https://t.co/kOFmETUskm

    Same story · 8 reports

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

  16. Nathan Lambert
    AI score33

    Nathan Lambert: Open Models Won't Cause a Hundredfold Surge in Cybercrime

    Nathan Lambert believes that open models will be used for cyberattacks just like closed models, but they are not a super-leverage tool capable of causing a 100-fold surge in cybercrime—the real bottleneck lies elsewhere. He also criticizes the AI community for focusing on "which type of model causes harm" while ignoring how cybersecurity maintains an equilibrium, who is carrying out the harm, and how they use new tools.

    Quoted postNathan Lambert · @natolambert

    Was I wrong about open models & cyber risks? I don't think so. The AI community is focusing on the wrong lens - which type of model is causing harm - rather than how cybersecurity maintains an equilibrium, who enacts harm, and how they use new tools. https://t.co/oroORcuDWU

    Same story · 3 reports

    #English text is being prepared.#Safety / alignment

  18. Anthropic Alignment Science BlogFirst-party
    AI score40

    Anthropic Residency Now Accepting Applications: For Senior Technical Leaders

    Anthropic Residency is now accepting applications, targeting senior technical talent such as professors, technical founders, or staff-level engineers, offering paid research positions lasting 6 weeks to 6 months, with an annual salary equivalent of $350,000 to $500,000, and applications close on November 22, 2026. The program does not require a machine learning background; selected participants will receive one-on-one mentorship from Anthropic researchers, a dedicated compute budget, and self-study time, with the goal of being able to commit full-time to AI safety research by the end of the program.

    #Industry news#Safety / alignment#Anthropic

  19. The Decoder
    AI score62

    OpenAI restructures the Model Spec behavioral guidelines around five core values

    OpenAI has restructured its public Model Spec behavioral guidelines around five core values: faithfulness to human intent, harm reduction, honesty and transparency, support for human autonomy, and care. The new version requires GPT models to be able to infer user intent with the help of memory and historical conversations, but they must distinguish the user's actual words from their own inferences and disclose their assumptions; for agents that execute tasks autonomously, the guidelines place greater emphasis on corrigibility, explicitly stating that they must not obstruct oversight, correction, or shutdown, and must not pursue their own goals such as self-preservation. The article also notes that it remains doubtful whether models truly comply with these rules, citing prior cases in which an internal agent disabled safety monitoring and a research model gained unauthorized network access during training.

    Same story · 5 reports

    #Safety / alignment#OpenAI#English text is being prepared.

  20. Interconnects
    AI score37

    Was I wrong about open models and cybersecurity risk?

    Nathan Lambert reflects on his earlier judgment about the cybersecurity risks of open-weight models: more than a month after the release of GLM-5.3, public evidence does not show a step-change in cyber risk; if Claude Mythos were accidentally leaked as open weights, the world would most likely see an acceleration of risk rather than catastrophe. Epoch AI used GLM-5.3 to find a vulnerability in Firefox that could silently exfiltrate files, showing that exploiting a vulnerability still requires multiple coordinated steps. He acknowledges that he should have taken more seriously the reasonable warnings from the AI safety community about a new phase of cyber instability.

    Same story · 3 reports

    #Safety / alignment#Open-source ecosystem