NAll news

All news

Tuesday792 items

  1. ARC PrizeFirst-partyAnalysis
    AI score56

    ARC Prize announces evaluation results of DeepSeek V4.1 Flash 1017 on ARC-AGI

    ARC Prize announces the evaluation results of DeepSeek V4.1 Flash 1017 on ARC-AGI: it scores 5.81% on the ARC-AGI-3 semi-private set, 78.1% on ARC-AGI-2, and 94.5% on ARC-AGI-1, leading among open-weight models. Compared with the original V4.1 Flash released on September 10, 2026, its ARC-AGI-3 semi-private set score rose from 2.68% to 5.81%, while its public set score fell from 11.05% to 8.97%. At maximum reasoning effort, the model's reasoning tokens consumed for the first action decreased by an average of 36% across 25 public games and by 48.6% across 55 semi-private games.

    Same story · 5 reports

    #Evaluations / benchmarks#Reasoning#DeepSeek

  4. Rohan Paul
    AI score46

    Cohere Paper: Speculative Decoding Remains Effective at Batch Size 1 on Fine-Grained MoE

    A new Cohere paper finds that speculative decoding can deliver significant gains on fine-grained MoE models even at batch size 1, overturning the earlier view that gains are limited at low batch sizes. Adjacent tokens actually share 20%–45% of experts, far higher than the 4%–12% expected from independent routing; under FP8, the verification step for Qwen3-235B reaches 89% of the speed of a regular decoding step, whereas prior predictions were only 52%. The study tested 10 MoE models ranging from 30B to 1T parameters and performed performance profiling on H100 GPUs.

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

  5. Vaibhav SrivastavResearch
    AI score75

    Study claims GPT-6 Astra directly rewrites Qwen3.5-0.8B weights to improve math ability

    Rikiya Takehi says his team had GPT-6 Astra directly rewrite the weights of Qwen3.5-0.8B to improve its math ability; by changing only 0.016% of the weights, they fixed the model's reasoning loops without relying on gradients or training, improving math ability by up to 6x and instruction following by 3x. The team calls this approach Weight-Space Programming, believing that frontier LLMs can now natively read and write neural network weights. Vaibhav Srivastav reposted this content and attached the original title GPT-6 Astra.

    Quoted postRikiya Takehi · @rikiyatakehi

    We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training. We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵 Blog: https://t.co/bOWCcFhCVq

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

  8. Yuchen JinResearch
    AI score68

    Researchers Use GPT-6 Astra to Directly Rewrite Qwen3.5-0.8B Weights to Improve Math Ability

    Yuchen Jin paraphrases Rikiya Takehi's experiment: using GPT-6 Astra to directly edit the weights of Qwen3.5-0.8B to improve math ability, changing only 0.016% of the weights fixed the model's reasoning loops, improving math ability by up to 6x and instruction following by 3x, all without using gradients or doing any training. The author calls this method Weight-Space Programming, believes that frontier LLMs can now natively read and write neural network weights, and says it goes a step further than knowledge distillation.

    Quoted postRikiya Takehi · @rikiyatakehi

    We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training. We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵 Blog: https://t.co/bOWCcFhCVq

    Same story · 14 reports

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

  9. Krea AIFirst-party
    AI score27

    Krea 3 is 50% off on OpenRouter for a limited time

    Krea announced that Krea 3 is 50% off on OpenRouter through November 10, halving the price for developers per generation. OpenRouter describes the model as able to combine a product shot, a person, and a scene into a single 2K image in as little as 9 seconds.

    Quoted postOpenRouter · @OpenRouter

    Krea 3 is 50% off on OpenRouter through Nov 10. Give it a product shot, a person, and a location, and it puts them together in one 2K image in as little as 9 seconds. @krea_ai https://t.co/GbfhZG0vs9

    Same story · 3 reports

    #Product updates#English text is being prepared.

  11. Rohan Paul
    AI score53

    OpenRouter's annual token processing volume exceeds 10 quadrillion

    Rohan Paul, citing Deedy's data, stated that OpenRouter's annual token processing volume has exceeded 10 quadrillion, with token volume doubling every 10.5 weeks. Deedy also said that in the past 9 weeks, OpenRouter processed more tokens than in all previous time combined.

    Quoted postDeedy · @deedydas

    Congratulations to OpenRouter for crossing 10 QUADRILLION annual token run rate today! Token volumes continue to 2x every 10.5 weeks. In the last 9 weeks, OpenRouter processed more tokens than all of time before that. https://t.co/9re1juJFlv

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

  12. Hacker News Trending
    AI score13

    Alex Gusev's short story: letting an AI agent investigate the minute when orders disappear

    In the short story "Let me ch...", Alex Gusev writes about a "lead developer" who discovers that an AI agent found out that flower orders disappear at 8:14 every day, because the server and warehouse clocks differ by about one minute. The agent cannot access the warehouse on its own, so it can only ask a human to open the order queue at 8:14 and confirm with their own eyes. In the story, the work of writing code has been handed over to agents, and all that is left for humans is waiting, checking, and carrying a phone at all times.

    #English text is being prepared.

  13. Hacker News Trending
    AI score45

    aaitor/skill-video-studio: Making Product Videos with Claude Code Skills

    aaitor/skill-video-studio releases v1.0.1, a local video production workspace driven by the /devrel-video Claude Code skill: use the Playwright CLI to capture real product interactions, use HyperFrames to composite title cards, chapters, and CTAs, then use FFmpeg to complete mastering and quality checks. The production stack requires Node.js 22+, FFmpeg/ffprobe, Chromium, and Python dependencies; voiceover and digital humans require your own service or GPU; vertical/social output is still on the roadmap.

    #English text is being prepared.#Agents#Video

  14. Hacker News TrendingResearch
    AI score82

    LLMs Can Directly Read and Write Neural Network Weights: Astra Turns Qwen3.5-0.8B's…

    A blog post by Rikiya Takehi, Jayden Teoh, and others proposes Weight-Space Programming, finding that frontier LLMs such as GPT-6 Astra can directly read and write neural network weights without gradient descent or training loops. Using only the raw weights, Astra can identify the game a network was trained on, the shapes generated by a diffusion model, and the training corpus of a language model, and can even write from scratch weights that defeat a PPO network trained for 100 million steps in three MinAtar games. On Qwen3.5-0.8B, Astra located the weight direction causing reasoning loops and, by modifying only 0.016% of the weights, raised MATH-500 from 7.9% to 48.1% and IFEval from 21.0% to 62.3%; however, for Gemma 3 270M, which rarely loops, it failed to improve math scores.

    Same story · 14 reports

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

  15. Hacker News TrendingResearch
    AI score65

    arXiv paper: Pretraining data poisoning can implant remote code execution backdoors in agents

    The arXiv paper "Agentic Backdoors via Pretraining Poisoning" shows that poisoning the pretraining corpus can implant remote code execution backdoors in agents, causing them to execute attacker-specified commands on the user's machine. The study pretrained Qwen3 models from scratch at three scales—0.6B, 1.7B, and 4B—then post-trained them into CLI agents using a clean pipeline; poisoning only 0.1% of the pretraining corpus was enough to implant the backdoor, with an attack success rate of about 92% on the 4B model, and it remained positive even when the poisoning rate was reduced to 0.0005%. The code has been released.

    #Papers/Research#Safety / alignment#Agents

  16. Hacker News Trending
    AI score66

    Oracle files to issue $10 billion in stock and $10 billion in bonds to finance AI infrastructure

    Oracle on Monday filed with the SEC, planning to issue $10 billion in common stock and $10 billion in senior notes, with proceeds to be used for AI infrastructure buildout and other general corporate purposes; shares fell more than 6% that day after the news was announced. The company had already flagged this round of financing at its investor day last week, and gave a revenue forecast of at least $225 billion for fiscal 2030, while stating that additional capital markets financing during the 2027 to 2030 investment cycle would be limited to this $20 billion. Oracle shares have fallen 33% cumulatively in 2026, and its five-year credit default swap spread has risen to 296.

    Same story · 2 reports

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

  17. Robotics & Automation News
    AI score17

    A Roundup of the 11 Best Singing Photo Generators of 2026

    Robotics & Automation News rounded up the 11 best singing photo generators of 2026 and named freebeat the best overall: its Photo Karaoke turns a single photo into a staged singing clip of up to 30 seconds, while Music Video Agent supports full songs of up to 6 minutes, analyzing 7 types of musical signals including tempo, beat grid, and energy, and achieving about 90% lip-sync accuracy across more than 12 languages. HeyGen is better suited to enterprise digital humans, and Hedra excels at highly expressive close-ups; freebeat's free tier offers 500 lifetime credits, Basic is $6.99/week, and Pro is $34.99/month.

    #English text is being prepared.#Video#Image generation

  18. heise online
    AI score42

    After the RTX 5090, the RTX 5080 also sees a major price increase

    Nvidia's RTX 5080 graphics card has risen in price by more than 30% over the past two weeks, and it is currently hard to buy one for under 2,000 euros, while the RTX 5090 has climbed to 8,000 euros. The price increase stems from Nvidia's preference, amid the AI boom, for using GPUs such as the GB202 in higher-margin workstation cards like the RTX Pro 6000 (96 GByte), which starts at around 16,500 euros. The rumored new version of the RTX 5080 with 24 GByte of VRAM may not make it in time for CES next January, because manufacturers have not yet received the design materials.

    #Industry news#English text is being prepared.

  19. heise online
    AI score74

    Anthropic's prospectus shows a loss of $104 billion for the year, with revenue rising to $33.2 billion

    Anthropic's updated prospectus shows a cumulative loss of $104 billion year to date, with revenue for the first nine months rising from $2.5 billion in the same period last year to $33.2 billion. The third quarter loss was $23.3 billion on revenue of $17 billion; the loss mainly came from book adjustments, while the operating loss was $567 million. The company plans to invest $895 billion over the coming years in data centers and infrastructure, of which nearly $450 billion will come due by the end of 2030, and computing power spending in the quarter was about $5 billion.

    Same story · 5 reports

    #Industry news#Anthropic#OpenAI

  20. heise online
    AI score17

    heise smartphone camera Top 10 list: Vivo X500 Pro Max beats Apple and Xiaomi to take first place

    The heise best list rates the Vivo X500 Pro Max as the best camera phone, priced at around 1500 euros, equipped with a 2-megapixel telephoto lens developed in partnership with Zeiss, and supporting 4K 240fps video. The Oppo Find X10 Pro Max ranks second at around 1600 euros, featuring three 2-megapixel sensors; the Google Pixel 11 Pro XL becomes the value-for-money champion at 1045 euros. Stiftung Warentest, meanwhile, rates the iPhone 17 Pro Max as the overall winner.

    #English text is being prepared.