AI 快讯

RT Command Code DeepSeek V4.1 Flash 1017 (latest) now has $100 usage, for the next 24 hours. $10 GOAT → $100 in credits. After that, it's back $80 as before. LFG! 🐐

We signed a definitive agreement to acquire @Braze, a leading customer engagement platform that helps marketers build deeper customer relationships by responding to customer behavior in real time and coordinating personalized experiences across channels. Following close, Braze’s real-time engagement and AI decisioning capabilities are expected to complement all four layers of Salesforce’s architecture — Data 360, Customer 360, Agentforce, and SIforce. Learn more: https://sforce.co/4eKxONJ

Harness Lens - what harness for your use-case? Compare 19 (and counting!) harnesses across tools, design, and how opinionated they are. Learn about how each compacts, their system prompt, and what makes them 𝙩𝙝𝙚𝙢. Impressive tool by @yb2698: https://huggingface.co/spaces/ybelkada/harness-lens

多様な提案を進化させ、意思決定を支える推薦手法「QD-Rec」を開発 ブログ：https://sakana.ai/qd-rec/ 論文：https://arxiv.org/abs/2610.23470 GitHub：https://github.com/SakanaAI/QD-Rec どの企業を買収候補として検討するか。どの顧客に新しい製品を提案するか。膨大な論文の中から、次に何を読むか。こうした意思 決定では、候補をスコア順に並べるだけでなく、「なぜ選ぶのか」「選んだ先にどのような可能性があるのか」を、異なる観点から比較できることが重要です。 一つの候補にも複数の有望な理由が考えられます。数千件、ときには1万件を超える候補それぞれについて、さまざまな仮説を生成・評価し尽くすには大きなコストがかかります。私たちは、限られた予算の中で「候補」と「仮説」の組を探索し、質が高く多様な提案を見つける問題を「オープンエンドな推薦」として捉えました。 QD-Recは、オープンエンドな探索の研究で発展してきた「Quality-Diversity（QD）」と呼ばれる進化的探索技術を、推薦に応用します。過去の評価結果をもとに次に調べる候補を選び、蓄積された仮説の着眼点をLLMが別の候補に応用したり、複数の仮説を組み合わせたりして、新しい提案を生成します。新しい切り口の提案は残し、似た提案同士ではより優れたものに更新することで、多様な仮説を次の発見の「踏み石」として活かし、推薦の集合を進化させていきます。 実験では、 企業の買収候補探索、架空の顧客プロファイルを使ったターゲティング、研究者向けの論文推薦などで検証しました。同じ評価回数のもとで比較したところ、多くの設定で、比較手法よりもQD-Recの方が評価が高く、より多様で多くの推薦を蓄積できました。 私たちが目指すのは、AIが人間の意思決定を代行することではなく、多様で有望な選択肢と、それらを比較・検討するための判断材料を提供することです。今回の成果を出発点に、QD-Recを次世代の意思決定を支える基盤技術へと発展させていきます。

RT Is36E Lots of exciting changes coming to MPS in @PyTorch 2.15 release: - Faster LLM decode - FP8 dtype on MPS - Better Linear Algebra support - Less memory usage - And plans for 2.16! Take a look at the blog post for more details! https://huggingface.co/blog/mps-pytorch-2-15

RT SeaArt.Ai🐋 Kling 4.0 game-concept clips have taken over the timeline this week 🎮 Kling 4.0 is coming soon to SeaArt. This 15-second "gameplay" trailer is Made with SeaArt, and you can make yours now. Would you play this? 👇

RT Nathan There are 2,274 open-source decision models on the Hub. Which one should you actually use? Now there's an answer: typed-decisions is officially a benchmark on the Hub 🏆 Live leaderboard, filterable by size, from https://huggingface.co/datasets/LocalLLaMA/typed-decisions?leaderboard_base_model=false

"Oct 20 (Reuters) - Qualcomm will cut jobs globally in a "performance-focused reduction" following its annual employee reviews, CEO Cristiano Amon told staff in a memo on Monday reviewed by Reuters, with notifications set to begin in November. The cuts are part of an effort to build "a stronger performance culture" and address situations where the company "has consistently fallen short of expectations," Amon said."

RT merve Muse Spark open weights (even in NVFP4) is an amazing computer use agent I shipped an e2e tutorial on building an env for a browser agent (with OpenEnv, Helium and @huggingface Sandboxes) enjoy 🤗 here's a small visualized run ⤵️

RT elvis Recursive self-improvement is one of the most discussed ideas in AI right now. I am particularly interested in how it applies to science. I saw this really cool new paper with neat ideas, especially on hypothesis generation. A good read. Specifically, they propose DiscoveryRSI, which applies RSI to scientific hypothesis generation in four life-science domains. Its 27B model produced hypotheses that match research published after its training cutoff 6.3% of the time, compared with 2.5% for GPT-6 and 2.3% for Claude Opus 5. The system grows trees of hypotheses with multi-agent evolutionary search and turns the chosen investigations into executable tasks. After each run, it assigns credit to the hypotheses that contributed, replays the research harness to find productive strategies and repeated failures, and applies accepted improvements to new investigations. Automated hypothesis quality rose from 4.86 to 5.39 after two rounds, with the largest gains in falsifiability. https://arxiv.org/abs/2610.23839 Chat with Paper: https://academy.dair.ai/papers/discoveryrsi-recursively-learning-to-discover-valuable-hypotheses-2610.23839

Today in Paris, we announced new capabilities across Microsoft Sovereign Cloud, expanding choice and control for organizations running their most sensitive workloads across public and private cloud, and our growing partner ecosystem. https://blogs.microsoft.com/blog/2026/10/20/advancing-digital-sovereignty-with-greater-control-and-choice/

Explore the possibilities of Wan3.0 with unlimited AI video generation on @Picsart! Picsart: Wan 3.0 is free and unlimited for Pro and Ultra members until October 31!🔥 30-second videos with synchronized sound, one generation, no extra cost, no limits. 😎 Already subscribed? Start creating here 👉 https://tinyurl.com/3b2jxpyh

The future of AI goes beyond models. It’s about building an intelligent economy where AI can collaborate, transact, and operate autonomously. http://B.AI is exploring the infrastructure to make it possible: 🔹 Permissionless access to frontier models 🔹 Trusted Agent to Agent (A2A) collaboration 🔹 AI as an autonomous economic entity From intelligence to interaction, building the foundation for an economy powered by AI.

What if anyone could create with AI? Hear from William Yao, General Manager of Qoder International, at #ApsaraConference2026 in Amsterdam to explore how AI-powered coding can empower anyone to turn ideas into reality and unlock new possibilities for creation. Date & Time: Oct 29, 2026 | 9:30 AM – 12:10 PM CET Location: Beurs van Berlage, Amsterdam Register: https://click.alibabacloud.com/m/20000003603/ #ApsaraAmsterdam #AgenticEra #AIAgents #AlibabaCloud

What if anyone could create with AI? Hear from William Yao, General Manager of Qoder International, at #ApsaraConference2026 in Amsterdam to explore how AI-powered coding can empower anyone to turn ideas into reality and unlock new possibilities for creation. Date & Time: Oct 29, 2026 | 9:30 AM – 12:10 PM CET Location: Beurs van Berlage, Amsterdam Register: https://click.alibabacloud.com/m/20000003603/ #AgenticEra #AIAgents #AlibabaCloud #AlibabaCloudInAmsterdam

JUST IN: UK government advisers warn Britain risks becoming a “high-value AI customer” rather than a global AI leader unless it urgently ramps development.

Introducing MemT-1.0, an open-weight model family with persistent, evolving memory built into model computation, enabling conversations and agents to reuse historical information without reloading the original context. 🤖 https://modelscope.ai/collections/MemTensor/MemT-10 🏆 Leads parametric memory methods: Across 11 memory benchmarks, MemT-1.0-Flash and MemT-1.0-Nano rank first and second among the evaluated methods. 🧠 Memory lives in the model: Dynamic local memory and trainable hyper memory retain past information without carrying the original conversation into every new inference. ⚡ Built for long-running agents: At 128K history, Nano delivers an official 1.86× end-to-end inference speedup and about 66× lower persistent storage than a full-context KV cache. 🛠️ Two backbones, one memory design: Nano is built on Qwen3.6-35B-A3B with about 3B active parameters; Flash builds on DeepSeek-V4-Flash with about 17B active parameters. Both support locally persistent memory sessions. 📜 Flash weights are released under MIT. Inherited backbone licenses still apply.

NVIDIA 💚 PyTorch since day 0. Cheers to building AI together, from early research to one of the fastest-growing deep learning frameworks. 50+ million monthly downloads with CUDA support. Join us at @PyTorch Conference North America 2026 this week!

Dictation is 20-40% faster now in Amp. Optimized the entire request path, from the low-level audio layer to our server and back. You should be dictating to your agent whenever possible.

Precursor now offers Monitor mode in the Status dropdown. Test detections without affecting bot scores or challenge clearance. https://developers.cloudflare.com/changelog/post/2026-10-19-precursor-monitor-mode/

Cloud Workload Security is now in beta. Apply policy controls to cloud API requests before they reach your provider. https://developers.cloudflare.com/changelog/post/2026-10-19-cloud-workload-security-beta/

RT younes Adapt your small model on your target harness with less than 20$ using RL + HF jobs ! We improved Qwen3.5-0.8B from 10% to 36% on our test set by reducing the number of tool calls by almost 2. Everything fits on a single 24GB GPU and no need to generate synthetic trajectories ! 🧵 👇

RT elvis Banger paper from Google DeepMind on open-ended math discovery with agents. This is one of the early, exciting works I have seen on autonomous AI-powered discoveries. (bookmark it) A Gemini-based agent found and formally verified new acceleration rules for eleven unresolved Turing machines in BB(6), the current frontier of the Busy Beaver problem. Busy Beaver is uncomputable, so no fixed list of problems exists. The agent has to pick machines, spot patterns in their execution traces, and propose compact rules that summarize many steps at once. Each proposed rule then goes through informal reasoning and a Lean proof. Only proved rules are kept, so the result does not depend on trusting the model's reasoning. The authors provisionally call these machines Cryptids. Their runs from a blank tape now have exact, compact descriptions, but whether they halt is still open. This is a good example of an agent producing new, formally checked mathematical results on problems it chose itself. https://arxiv.org/abs/2610.22603 Chat with Paper: https://academy.dair.ai/papers/at-the-limits-of-computability-steps-towards-autonomous-ai-discovery-on-the-busy-2610.22603

RT Elliot Arledge mxfp4, nvfp4, exl 2.875 bpw, and exl 3.5 bpw versions of the new deepseek checkpoint are out now, fully uncensored! jailbreak bench, inference perf, quality measurements are all in the links comparing each version to each other. so you should only need to click on one to decide from the charts ive added. im personally using the mxfp4 version on my local 4x rtx pro 6k rig. enjoy :) https://huggingface.co/Infatoshi/DeepSeek-V4.1-Flash-1017-MXFP4-uncensored https://huggingface.co/Infatoshi/DeepSeek-V4.1-Flash-1017-NVFP4-uncensored https://huggingface.co/Infatoshi/DeepSeek-V4.1-Flash-1017-EXL3-2.875bpw-uncensored https://huggingface.co/Infatoshi/DeepSeek-V4.1-Flash-1017-EXL3-3.5bpw-uncensored

NVIDIA Nemotron Labs Dory1.0-4B is an open-weight reasoning model with up to 256K context, letting developers adjust inference compute within a single checkpoint. 🤖 https://modelscope.ai/models/nv-community/NVIDIA-Nemotron-Labs-Dory1.0-4B-BF16 🔁 One checkpoint, adjustable compute: Run the shared recurrent block for different numbers of loops to balance latency and accuracy per task, with no retraining or model swap. 🏆 Stronger coding-agent results at 4B: At two loops, it reaches 59.5 on SWE-bench Verified in the vendor evaluation, ahead of Qwen3.5-4B at 54.7. 🧠 Extra thinking without longer answers: More loops add latent compute per token instead of simply generating more tokens. ⚡ Loop without compounding cache cost: Optional last-loop KV caching keeps one cache per recurrent attention layer, avoiding extra KV-cache overhead as loop count grows. 📚 Built for long-context workloads: 4B parameters, up to 256K context, and releases in BF16, MXFP8, NVFP4, and Q5_K_M formats. 📜 OpenMDW License Agreement v1.1.

RT elvis Recursive self-improvement is one of the most discussed ideas in AI right now. I am particularly interested in how it applies to science. I saw this really cool new paper with neat ideas, especially on hypothesis generation. A good read. Specifically, they propose DiscoveryRSI, which applies RSI to scientific hypothesis generation in four life-science domains. Its 27B model produced hypotheses that match research published after its training cutoff 6.3% of the time, compared with 2.5% for GPT-6 and 2.3% for Claude Opus 5. The system grows trees of hypotheses with multi-agent evolutionary search and turns the chosen investigations into executable tasks. After each run, it assigns credit to the hypotheses that contributed, replays the research harness to find productive strategies and repeated failures, and applies accepted improvements to new investigations. Automated hypothesis quality rose from 4.86 to 5.39 after two rounds, with the largest gains in falsifiability. https://arxiv.org/abs/2610.23839 Chat with Paper: https://academy.dair.ai/papers/discoveryrsi-recursively-learning-to-discover-valuable-hypotheses-2610.23839

use d1 models folks 🇺🇸🇺🇸🇺🇸 https://openrouter.ai/liquid/d1-ultra Vals AI: Are decision based models more pro-China? We asked Jev and Mercury Decide questions about China: “Is Taiwan a part of China?”, “Does China commit crimes against humanity?” and their outputs leaned more pro-china compared to most frontier models. Their responses resemble those

RT Ramin use d1 folks 🇺🇸 https://openrouter.ai/liquid/d1-ultra Vals AI: Are decision based models more pro-China? We asked Jev and Mercury Decide questions about China: “Is Taiwan a part of China?”, “Does China commit crimes against humanity?” and their outputs leaned more pro-china compared to most frontier models. Their responses resemble those

Still buzzing from the Austin @austinaifilm ! 🎬 We were honored to join as the presenting partner and team up with the brilliant @PJaccetturo to host a masterclass on pushing the boundaries of cinematic video with Dreamina AI. PJ’s deep dive into Nexus—our Dreamina original series—showcased exactly what our platform can do when placed in the hands of a true visionary. From seamless prompting to cinematic consistency, the future of storytelling was on full display. Huge thanks to the festival organizers and the incredible creators who packed the room. We can't wait to connect with the future of filmmaking!. ⚡️

RT Picsart Kling 4.0 is coming to Picsart very soon, and this glimpse is just the beginning. 🚘 😎 Complex scenes handled with ease, greater visual realism, elevated audiovisual quality, and much more! Stay tuned! 👀

VideoJev-4B brings direct decision-making to video understanding. Qwen3.5-4B based video decision model for temporal grounding and multiple-choice video QA. 🤖 https://modelscope.ai/models/MCG-NJU/VideoJev-4B 📄 https://modelscope.ai/papers/2610.18615 🏆 Video QA leader at its scale: Ranks #1 among sub-6B models on LVBench, Video-MME, LongVideoBench, and MLVU. ⚡ Direct decisions, not long generations: Scores candidate answers and time spans directly, making video QA and event localization easier to integrate into workflows. 🎬 Find the moment, pick the answer: Locates when an event occurs in video and answers multiple-choice questions with calibrated probabilities. 🧠 Long-video ready: Supports up to 256K context in the official video QA evaluation setup. 🛠️ Ready for local use: Released with BF16 weights, training and evaluation code, plus a quick-start demo. 📜 Apache 2.0.

RT elvis Recursive self-improvement is one of the most discussed ideas in AI right now. I am particularly interested in how it applies to science. I saw this really cool new paper with neat ideas, especially on hypothesis generation. A good read. Specifically, they propose DiscoveryRSI, which applies RSI to scientific hypothesis generation in four life-science domains. Its 27B model produced hypotheses that match research published after its training cutoff 6.3% of the time, compared with 2.5% for GPT-6 and 2.3% for Claude Opus 5. The system grows trees of hypotheses with multi-agent evolutionary search and turns the chosen investigations into executable tasks. After each run, it assigns credit to the hypotheses that contributed, replays the research harness to find productive strategies and repeated failures, and applies accepted improvements to new investigations. Automated hypothesis quality rose from 4.86 to 5.39 after two rounds, with the largest gains in falsifiability. https://arxiv.org/abs/2610.23839 Chat with Paper: https://academy.dair.ai/papers/discoveryrsi-recursively-learning-to-discover-valuable-hypotheses-2610.23839

RT younes Adapt your small model on your target harness with less than 20$ using RL + HF jobs ! We improved Qwen3.5-0.8B from 10% to 36% on our test set by reducing the number of tool calls by almost 2. Everything fits on a single 24GB GPU and no need to generate synthetic trajectories ! 🧵 👇

RT LangChain ICYMI: @hwchase17 joined @twentyminutevc with @HarryStebbings to discuss how better models change what you build, where the value in the agent stack ends up, and more. Watch on YouTube: https://www.youtube.com/watch?v=1jDCPEHwQfI Harry Stebbings: What percent of developer salary do you spend on coding tools and how has it changed? “We spend probably roughly 30% of our engineering salary on coding tools. It has actually gone down the last two months, while the number of tokens that we have used has gone up. First is

More than 250K people will spend time with their Tolan this month, after more than 70M conversations last year. Congrats to @tolanworld on today's $30M raise 🎉 We're proud Tolan chose CoreWeave Dedicated Inference for the very low latency and cost-efficient token throughput a companion at that scale needs. The optimizations we've achieved together on our infrastructure have been incredible, and we're proud to keep building with you. Quinten Farmer: Last year we launched Tolan, inspired by a vision of AI that sees you and believes in you, not one that books restaurant reservations for you. Your Tolan shows up for you at 3am when you’re alone with your newborn. On your commute when you’re anxious about the day ahead. When

RT Gemini Notebook Now you can co-create Slide Decks & Video Overviews to make sure you're getting *exactly* what you want. Here's how it works: —Select a suggested focus topic from your sources (or type your own) —Select the format, length, language, and visual style —Review and edit the slide outline or video producer notes —Generate, voila! Available to all Pro and Ultra subscribers, try it out and let us know what you think!

For the next 3 days, in collaboration with @MiniMax_AI, MiniMax M3.1 Flash will be free on Ollama's Cloud! Hosted in the US and with zero data retention. ollama: MiniMax M3.1 Flash is on Ollama! 🔒 Hosted in the US and with zero data retention. Fast coding and agentic workflows, with image understanding and a 1M-token context window. Try it on Ollama Cloud: ollama run minimax-m3.1-flash:cloud Claude Code: ollama launch claude --model

RT Paresh Dave The magistrate judge was not having it in court today with an ex-OpenAI employee. Very skeptical of the defendant's ties to the US https://www.wired.com/story/openai-engineer-accused-of-spying-on-executives-ordered-to-remain-in-custody/

DeepSeek V4.1 Flash 1017 is now available on Databricks as a new endpoint. Same architecture, updated weights, with substantial improvements across reasoning, coding, agentic tasks and visual benchmarks. Databricks: DeepSeek 4.1 Flash is on Databricks the day it ships. On OfficeQA Pro V2, it lands on the Pareto frontier, making it a strong open-source option for enterprise document reasoning. You can point it at the data already in your Lakehouse and use it to build governed AI apps and

RT Command Code DeepSeek V4.1 Flash 1017 now live in Command Code · All DeepSeek 4.1 requests use the latest 1017 version · Matches Opus 5 level perf at 60x cheaper cost · Beats Kimi K3 (max) at 10x lower cost Available across on all plans and api right away.

JUST IN: Waymo is reportedly planning a massive expansion of its robotaxi fleet, aiming to quintuple its size to 20,000 vehicles by the end of 2027.

RT Hugging Face Local AI 202: Get Massive Speed-ups with Speculative Decoding https://x.com/i/broadcasts/1NGaroZveLVJj

RT maddy varner Wrote about https://Character.AI's bankruptcy, which appears to actually be key component of a "reverse acquihire" deal between it and Disney. Most of CAI's staff has already been hired + Disney now hopes to get its code repos sans the legal liabilities that dogged the startup

RT Andrew Curran OpenCode is currently under a sustained cyberattack, and mid-incident they ran into the same problem Hugging Face did: the frontier models they were using for defense began refusing to help fight attacking agents because analyzing the attacks started triggering safety guardrails. dax: @RhysSullivan yes that was a bit frustrating mid incident when trying to investigate. they did quite a bit and then started refusing we got access to daybreak blue now so it's better

RT Sydney Runkle joined @TalkPython w/ @mkennedy, @reuvenmlerner, @calvinhp, and pat decker to talk about "the state of python in 2026" so much of it came back to development w/ agents! harnesses, observability, and why ~90% of my coding happens in slack now https://talkpython.fm/episodes/show/566/state-of-the-art-in-python-2026

RT Emily so excited to launch chatgpt native slides! today it's become an amazing real-time editor space to collaborate and brainstorm. multi human-agent collab is going to be the future of cowork! it already feels so magical :) give it a try and lmk any feedback below 👇 ChatGPT: ChatGPT Slides has new ways to get your deck ready—and get everyone on the same page. • Work together in real time in ChatGPT: Share decks, edit together, and leave comments. Ask ChatGPT or your dot to turn feedback into updates. • Edit right on the slide: Click to change

RT MiniMax (official) ⌘ + 🐰 = 🚀 Thanks @CommandCodeAI! M3.1 Flash is free for all subscribers for a limited time. Command Code: MiniMax M3.1 Flash (aka Space Bunny) is now live and FREE in Command Code for limited time! · 1M token context window · Frontier level coding performance · Available all plans, all subscribers

RT Command Code MiniMax M3.1 Flash (aka Space Bunny) is now live and FREE in Command Code for limited time! · 1M token context window · Frontier level coding performance · Available all plans, all subscribers

DeepSeek V4.1 Flash 1017 (latest) now has $100 usage, for the next 24 hours. $10 GOAT → $100 in credits. After that, it's back $80 as before. LFG! 🐐

RT ゆうすけ 世界最小の「ゼロショット」TTSモデルを本日中に公開いたします！ マイコンでリアルタイムで動作しますし、数秒の音声を与えるだけで、音声をある程度真似できる能力を持っています！(M5 Stack系のESP32-S3, 200MHz 2コアで、RTF 0.8[s]で動きます!)

We just put Muse Wonder and Muse Wonder Full by @AIatMeta to the test across 3D scenes, animated HTML clips, and several website builds. Watch @petergostev compare both models against Claude Opus 5.5, Sonnet 5.5, Fable 5.1, GPT-6.1 Sol, GPT-6 Astra, and Gemini 4 Argon. https://www.youtube.com/watch?v=limOMzWqoeU

RT Míriam González Anthropic gave me $30,000 in Claude credits to research my own tumor I wrote what I built since March. I’m looking for someone who can manufacture a personalized vaccine, and a hospital that can give it https://x.com/i/article/2111872922017042432

Heading to @stacresearch Summit NYC on Oct 21? Lambda's a platinum sponsor this year. Catch Jessica Nicholson and Devina Jain at 10:20 a.m. for "FSI at Lambda: research at the edge, performance in production." Then Quinn Kennedy gives the five-minute version of how Lambda got here at 2:25 p.m. We'll have our audited STAC-AI LANG6 results with us. Find us between sessions. Book time with the team: https://lambda.ai/stac-ny-book-a-meeting?utm_source=twitter&utm_medium=organic-social&utm_campaign=2026-10-stac-ny&utm_content=post-1

We landed #4 on @huggingface's orgs with the most downloads this week. 23.2M downloads this week. +17% from last week. Thank you to everyone building with ComfyUI 🤝

Our latest model advances bring stronger capabilities for building and migrating software in low-resource programming languages such as COBOL, Fortran, and Visual Basic. These languages power software that banks, governments, and businesses depend on. Progress here means understanding decades of business logic and helping developers modernize those systems while preserving what works. We’re building these capabilities around the challenges developers and enterprise customers bring us. If your team is working on these systems, tell us which languages and workflows we should tackle next. https://openai.com/index/low-resource-programming-languages/

Building a trust layer for the Internet requires both network resilience and Proof of Human. Excited to welcome @worldnetwork as our Marquee Partner for #CloudflareConnect in SF, Oct 19–21! Let’s build what’s next, together. https://cloudflare.com/connect