Updated
Updated
Showing low-relevance items too. Hide low-relevance items
Zyphra@ZyphraAIOfficialAI score34AIZyphra says its approach is lossless: the same tokens still reach the same experts, with unchanged architecture, routing decisions, and training objective. By reorganizing where experts and tokens live, it reduces the communication needed to perform the same computation.
Zyphra@ZyphraAIOfficialAI score38AIZyphra reports that its MoE training optimizations speed up token exchange by 1.16x to 2.63x and full training steps by up to 1.41x in Megatron-LM on 8 to 64 GPUs. The gains are largest when each token uses more experts and those experts span several nodes.
Zyphra@ZyphraAIOfficialAI score22AIZyphra reports that patterns in model training emerge early enough to be useful and are inexpensive to measure. Using only 4,096 tokens, its placement and routing predictions come within roughly one percentage point of results obtained with 524,000 tokens.
Zyphra@ZyphraAIOfficialAI score18AIZyphra reports that expert routing in mixture-of-experts models is predictable across layers, since the experts a token uses in one layer indicate which it will need next. Its token shuffling method moves each token to the GPU holding those experts within a transfer that already runs after attention, adding no network traffic.
Zyphra@ZyphraAIOfficialAI score37AIZyphra places experts that are frequently chosen together on the same GPU, so each token is sent to a GPU only once regardless of how many of its experts reside there. On 8 GPUs, this removes up to 58% of token copies sent across the network.
Zyphra@ZyphraAIOfficialAI score32AIIn Zyphra's runs, MoE token exchange between experts consumed 13-24% of step time on one node and 45-60% across four nodes. Because experts are spread across GPUs and nodes, tokens must be sent to their experts and returned, making this communication a major training cost as models scale.
Zyphra@ZyphraAIOfficialAI score23AIIn one layer, just 64 of 8,128 possible expert pairs handled 42% of tokens, far above the 1.6% expected if experts were chosen independently. Zyphra says this structure can be exploited to improve expert placement.
Zyphra@ZyphraAIOfficialAI score22AIZyphra researchers report that exploiting patterns in how tokens are routed to experts makes Mixture of Experts communication up to 2.63x faster on AMD MI300X GPUs. The speedup comes without changing the model itself.
Andrew Curran@AndrewCurran_XAI score9AIAndrew Curran jokes that a newly discovered planet should be named Claude, since he primarily used Opus 5.5 and Fable 5.1. The post is light commentary tied to a quoted thread describing a planet hidden in NASA telescope data for seven years.
Allie K. Miller@alliekmillerXAI score18AIMeta is reportedly placing a new icon on Instagram's profile page, following the same approach it used to promote Reels and Threads. The post predicts millions of additional users within the next week, though it provides no figures or details to support that estimate.
AIGoogle has released AI Edge Foresight, a free experimental note-taking app that transcribes meetings and audio files entirely offline on macOS. It runs on the on-device EmbeddingGemma 2 model and turns shorthand notes into polished notes based on the transcript. Google says files, meeting audio, and notes never leave the computer, and the app is currently optimized only for Macs with Apple Silicon.
Simon Willison@simonwXAI score47AIMicrosoft has released MxC, an open-source sandboxing library that runs on Windows, macOS, and Linux. It builds on processcontainer, bubblewrap, and seatbelt as its underlying mechanisms. The project is aimed at developers struggling with fragmented sandbox tools for AI agents.
Cohere@cohereOfficialAI score20AICohere is hosting a live webinar on the future of search and retrieval, covering Embed 5, Parse 5, and its new retrieval methodology, RCP-nDCG. The post is a livestream announcement and does not include details of the methodology or model performance.
Kilo (acq. by Anaconda)@kilocodeOfficialAI score32AIMistral AI has released Large 4, nicknamed "Le Chonk," and Kilo Code is offering it at 50% off. Kilo Code demonstrated the model by having it one-shot its own video game.
Satya Nadella@satyanadellaXAI score38AIMicrosoft CEO Satya Nadella says Copilot is being positioned as a new operating system for work, paired with a governed headless business layer that gives agents access to CRM, ERP, and other systems of record. He says Microsoft announced over 30 new Copilot skills across Dynamics 365 Sales, Service, and Customer Insights, plus Microsoft Copilot Managed Runtime for IT-governed code hosting. He describes users building custom software or Dataverse extensions through Copilot Code, though the post is an early vision with few concrete specifications.
GitHub@githubOfficialAI score22AICamilla Moraes from GitHub's Tiny Wins team joins the latest GitHub Podcast to discuss the AI slop problem, which has become one of the biggest pain points for open source maintainers. The episode also covers listening to maintainers and deciding what to fix next.
Aravind Srinivas@AravSrinivasXAI score40AIPerplexity's pplx-decider v1.1 scored 643 of 669 on clinical decisions versus 628 for Jev, according to a quoted post. The quoted post says it costs 42% less per decision at about the same speed and that its weights can be downloaded and run inside hospitals.
Waymo@WaymoOfficialAI score14AIWaymo's new research finds that sober drivers have a fatal crash involvement rate 23.3% lower than the overall average. The company argues sobriety alone doesn't guarantee road safety, since low visibility, fatigue, and poor infrastructure also drive fatalities, and says the Waymo Driver is never impaired and is built to handle those systemic risks.
AIAugment Code is selling select assets, including Cosmos, Auggie CLI, and the Code Context Engine, to Harness, and the product team is moving to Harness. The company says Harness's integrated platform delivers these capabilities to customers more effectively than building them independently. Harness describes itself as building the Autonomous SDLC Platform for shipping AI-written code across enterprises.
Why it matters: The announcement shows how a coding AI company is folding its products into a larger software delivery platform, a shift that shapes how enterprise teams will buy these tools.
AILarge-scale AI training is spreading across multiple datacenters, with Google, Microsoft, AWS, Meta, and CoreWeave cited as examples. Because synchronized GPU clusters must exchange data in bursts, inter-site links can become a bottleneck, which Cisco estimates may require aggregate bandwidth about 14x a conventional DCI baseline.
Microsoft Research@MSFTResearchOfficialAI score16AIMicrosoft Partner Research Manager Jennifer Neville tells host Chad Atalla that surprising AI failures can expose where human expectations about intelligence diverge from how AI systems actually work. The discussion frames these less obvious failures as more important than the obvious ones.
meng shao@shao__mengXAI score8AIThe author says that in two years at Megvii they met some of the smartest and most idealistic people, though the company did not achieve what they call a "result" for unnamed reasons. They argue these people can succeed anywhere and will keep thriving after the company's scattering, while a quoted reply reflects on six years at Megvii as "China's AI Fairchild."
Philipp Schmid@_philschmidXAI score46AIGoogle's SynthID Detector is now publicly available, letting users check whether an image, video, or audio file was generated by supported tools. Per the post, it scans for watermarks from Google and partners, including Nano Banana 2.1, OpenAI, NVIDIA, and Kakao, with Apple support coming soon. Uploaded files are deleted right after scanning.
AIAmazon unveiled 8-inch, 11-inch and 12-inch Alexa tablets priced from $230 to $550 and said it is ditching its budget Fire lineup. The devices run Android rather than Fire OS, and preorders open Thursday with shipping starting Oct. 14. Amazon said it will support the Fire lineup for four years after the final shipment but is no longer manufacturing new units.
StepFun@StepFun_aiOfficialAI score38AIStepFun is offering its flagship Step 5 Preview model free in OpenCode for one week, targeting agentic and professional work. The preview supports a 1M context window, multimodal input, and zero data retention.
Gamma@GammaAppOfficialAI score40AIGamma releases Gamma 5, an overhaul it says makes the most visually stunning decks. Its agent plans a complete deck from user context, connectors, and visual references, and a new freeform editing mode lets users drag any slide element anywhere after clicking "convert to freeform."
OpenRouter@OpenRouterOfficialAI score14AIOpenRouter says GPT-6 Luna is currently the fastest Decisions model, at 180ms on global requests. The post also places Jev and Perplexity Decider as following models, but gives no further figures.
AINVIDIA announced commitments valued at $1 billion over the next five years to build U.S. capacity for super intelligence research in fields including quantum computing, healthcare and energy security. The funding will support U.S. higher-education research institutions, American quantum leadership and cloud service providers serving U.S. government mission needs. NVIDIA is also a collaborator on several phase 2 Genesis Mission awards in quantum computing, fusion, accelerator design and microelectronics.
Nous Research@NousResearchOfficialAI score14AIGoogle is launching a "universal" Gemini agent that works across apps and devices in the background, available in private preview to enterprise customers. Users can chat with it and assign tasks from the Gemini Enterprise app, and it works inside Gmail, Drive, Docs, Sheets, and Calendar as well as third-party apps like Slack and Microsoft 365. The source notes it runs in the cloud, keeps the same context across devices, and can use job-specific sub-agents.
Nous Research@NousResearchOfficialAI score31AINous Research says Hermes Agent is now available on the new ASUS ProArt RTX Spark PCs, with MuseTree and ComfyUI integrations and local model support on up to 128 GB of unified memory. The post presents it as a creative stack that runs entirely on the user's own machine.
🚨 AI News | TestingCatalog@testingcatalogXAI score46AIGoogle has announced a single, universal Gemini agent for Gemini Enterprise as part of its Gemini at Work updates. The agent answers questions, handles knowledge work, creates images and media, and writes and runs code. It works inline in Gmail, Drive, Docs, Slides, Sheets, Chat, and Calendar, with new data and analytics skills for plain-language insights and industry-specific tools for financial services and legal teams.
Databricks@databricksOfficialAI score28AIReplit has announced an integration with Databricks that lets teams build business apps and deploy them on live, governed enterprise data. A four-step setup guide covers Databricks admins, Replit org admins, and end users, aiming to take an app from zero to governed production in 15 minutes.
PixVerse@PixVerseOfficialAI score4AIPixVerse shared a registration link hosted on Luma, with the post offering no further details about the event. The post contains only a call to register and a URL, so no product features, dates, or other specifics can be confirmed.
PixVerse@PixVerseOfficialAI score10AIPixVerse will exhibit at NABShowNY, with demonstrations of turning ideas into finished video, faster creative iteration, and variations for different channels. The company invites attendees to visit Booth 565 at the Javits Center on October 21-22 and join a happy hour after Day 1 on October 21.
Guizang@op7418XAI score22AIGuizang (@op7418) says a Grok bot paired with a foldable phone lets him chat on one screen while the bot publishes content on the other. Background post: he set Grok to produce a daily morning AI news video on a schedule, running content collection, code writing, and video rendering entirely on Grok's cloud virtual machine rather than his local computer. He shares the full prompt so others can run the same workflow with their own Grok bot.
AIMeta's Muse personal AI agent, launched on September 8, surpassed 2.5 million downloads in its first two weeks and topped free-app rankings on Apple and Google's US app stores. Its popularity has drawn attention to the emerging market for agent harnesses, where China's biggest internet companies are already competing for position. The excerpt does not provide full details on their specific products.
Orange AI@oran_geXAI score16AIOrange AI (@oran_ge) says Haiku 5.5's pricing is bizarre, with costs rising fivefold once usage exceeds 100k tokens. The post questions why any model would impose such a steep jump at that threshold.
Aidan Gomez@aidangomezXAI score14AICohere is opening an Ottawa office, its third in Canada after Toronto and Montreal. The company says Canada is where it started and that it will keep building there.
OpenRouter@OpenRouterOfficialAI score52AIOpenRouter says StepFun's Step 5 Preview is now live on its platform, priced at $1.00 per million input tokens and $2.70 per million output tokens. Cache hits are 50% off at launch, bringing them to $0.05 per million, and a week of free access is rolling out across partners including opencode, Cline, Nous Research, and Kilo Code.