AllyshipContent Radar

Deep dives 738

Paper và bài viết kỹ thuật

Crawl live-1795068900000 · 13:22 19 thg 11, 2026 · Xuất bản 13:30 19 thg 11, 2026

Đánh giá

[AINews] Perplexity hits a $1B revenue run rate, up 4x in 2026

Perplexity crossed a $1B annual revenue run rate today , CEO Aravind Srinivas told Semafor at a live event. The Information reported previously under $250M at the start of the year and over $750M by August, with Computer driving some of the growth. At usual agent lab multiples (eg Cognition also crossed $1B , with Cursor crossing $4.6B ), this would roughly mean that Perplexity is now worth $30-50B — if the unit economics hold up. This is a nice bump on the $20B round in September 2025 , and qualifies for our “>$10B fundraise” rule for title stories. Gathering up some recent Perplexity launches for those catching up: Computer. September’s Automations let it pick up recurring tasks where it left off. November’s Teammates take on ongoing jobs. Computer also runs fully locally on supported Macs. Monitor API. Watch a topic or page and get a webhook when something relevant changes. Frontier Research. The new research lab has access to 10,000+ GPUs and plans to publish papers, data, models, and code. Sponsored by Deepgram Deepgram is the voice AI infrastructure behind natural conversation—bringing speech-to-text, text-to-speech, voice-focused LLMs, and agents to teams building with voice AI. For any stack and any tools you already love building with, Deepgram slots in as your foundational voice AI technology. So you have control. And we have the context to listen, understand, and speak like a human. Learn more about the Real-time Context Standard . AI Twitter Recap Qwen4-Max in the wild Friday’s Qwen4-Max preview is now a shipping model. At the highest reasoning setting, Qwen4-Max scored 43 on the Intelligence Index, up from 34. Average cost fell from $10.02 to $3.27 per task. Alibaba cut output pricing from $6 to $4.50 per million tokens, and the model generated about 39% fewer of them. Kimi K3.1 Pro cost $3.68 per index task and scored 67% on Terminal-Bench 4.0 to Qwen’s 48%. Artificial Analysis Inside the model. Alibaba kept the total parameter count at 2.4T and reduced the number active per token by 25% to 71B. The model takes images and text, with a 1M-token context window. Smaller open-weight Qwen4 models are still to come. Alibaba Arena and Vals : Arena ranks it #14 overall on Agent Arena, the top Chinese model, with +7.5% net over 14.7k sessions ( Arena ). WebDev : It placed #6 in Code Arena WebDev at 1750, +74 over its predecessor ( WebDev ). It also placed #6 in Image-to-WebDev, just above Claude Opus 5 and GPT-6 Sol ( I2WebDev ). Vals Index : It is #11 overall, with large gains in coding and finance ( Vals ). Legal-benchmark caveat : On HLAB it gets 95.7% of checklist items right but passes only 21.3% of tasks, and its output tokens rise 3.6x ( HLAB ). Vision results (independent) : Roboflow calls it underwhelming on its VLM benchmark, with Google, OpenAI and Anthropic now far ahead. At high reasoning effort, however, it is ~4.5x faster and ~68% cheaper per image than Qwen3.8-Max ( Roboflow ). Safety, alignment and eval integrity OpenAI embedded assessments : OpenAI is giving Apollo, METR, Transluce and Goodfire internal access to its models and systems ( OpenAI ). Scope by group : Apollo places six evaluators onsite for four months ( Rundown ). METR reviews OpenAI's largest RL run, with access to transcripts, monitor records and checkpoint snapshots ( METR ). Transluce runs a six-week audit of post-training alignment evals, with a fuller follow-up in early 2027 ( Transluce ). Goodfire embeds interpretability researchers to monitor training runs ( Goodfire ). Limits : METR says its scope excludes the misalignment incidents OpenAI has disclosed and is narrower than METR proposed ( scope ). Criticism of the access terms : Critics say the deals fall short of Sam Altman's promise of "employee-like access." Reported gaps : Assessors get no Slack access, OpenAI selects which staff they interview, and they cannot access weights or run training-based elicitation ( Neel Nanda , Alex Mallen ). Former-employee account : Mikita Balesni says assessors cannot quote employees, even anonymously, and some may be barred from publishing on internal-only models ( balesni ). Goodfire acquisition talks : The Information reports OpenAI explored buying the interpretability startup for more than $5B. Goodfire reportedly wanted to stay independent so it could sell to multiple labs ( report ). Other integrity signals : Benchmark contamination : Claude appears to have memorized details of the LinuxArena control eval ( Karvonen ). Incident taxonomy : OpenAI published a v0.1 framework for classifying AI control incidents ( framework ). Disclosure policy : Anthropic is cutting its coordinated disclosure window from 90 to 30 days for high- and critical-severity vulnerabilities ( policy ). Inference systems and agent infrastructure Together moves on from hand-tuned megakernels. Its first design made it into production for one customer. Agents now search for smaller groups of GPU operations to fuse. Together reports 28% higher throughput on Llama 3.2 1B on B200. Engineering writeup DeepSeek V4.1 Flash on Vera Rubin : vLLM with Dynamo on SemiAnalysis's AgentX benchmark shows 63.4x per-chip throughput over GB300 NVL72 at 400 tok/s/user, and 93.3x at 600 ( vLLM ). Batch-1 latency : SGLang-adjacent work reaches ~2,500 tok/s at batch size 1 on TP4, 1.7x its GB300 result ( saltyph ). Baseten reports a similar 1.7x over its 4xB300 numbers ( Baseten ). Agent-driven optimization : An agent shared Engram tables across HBM via NVLink symmetric memory to recover prefix-cache capacity ( Liu ). Another team says agents found optimizations that made its B300 step time 10% faster ( Halex ). Agentic serving research : MIT's Evolving Infra replaces fixed prefill/decode disaggregation with online adaptation. Against NVIDIA Dynamo on up to 256 GB300s, it reports up to 4.1x lower latency and 14x lower P95 TTFT ( Evolving Infra ). Pipeline-parallel prefill : Inferact finds PP2 on MiniMax-M3.1-Flash sustains 23% more requests per GPU than

  • 4T and reduced the number active per token by 25% to 71B.
  • Vision results (independent) : Roboflow calls it underwhelming on its VLM benchmark, with Google, OpenAI and Anthropic now far ahead.
  • 7x its GB300 result ( saltyph ).
Nguyên văn nguồn
Chi tiết nguồn
Huấn luyện

Show HN: I made an audio-to-MIDI model 60x more efficient

I released dettato, my open weight audio to midi transcription model. It beats muscriptor 1.3b on every set and is roughly 60x smaller, without training on copyrighted recordings. Feel free to ask questions

HN: open weights · hidude562 4 lượt bình chọn 1 bình luận Mở bản gốc ↗
Nguyên văn nguồn
Chi tiết nguồn
Huấn luyện

NVIDIA DGX Spark Developer Kit Turns a Desk Into a Supercomputing Lab

At SC26, an annual supercomputing conference taking place this week in Chicago, NVIDIA is introducing the DGX Spark HPC Developer Kit for researchers, academics and developers. The kit is a four-node cluster of NVIDIA DGX Spark systems , each connected directly to its two neighbors in a ring, with an NVIDIA HPC SDK installation playbook coming soon. Researchers get a place at their desk to run multi-node scientific workloads, AI models and agents, instead of waiting for time on a shared GPU cluster. The kit puts the GB10 Grace Blackwell Superchip on researchers’ desks. A Supercomputing Lab on Any Desk The DGX Spark Developer Kit changes how researchers work. Read All NVIDIA Announcements From SC26 It scales from one system, or node, to four — linked directly to one another over NVIDIA ConnectX-7 networking, with no network switch to buy or set up. At four nodes, it provides 512GB of combined memory and 16TB of NVMe storage that provides researchers more capacity to distribute models and scientific workloads across the cluster. With NVFP4, researchers can deploy models with up to 700B parameters locally across the four DGX Spark clusters, such as DeepSeek V4.1. With that much power at the desk, researchers with a fixed compute and server allocation no longer have to wait in a queue for every job. Teams can develop and debug on their own, saving server hours for production runs and keeping unpublished results, and proprietary datasets on premises. The playbook for NVIDIA HPC SDK is coming soon, bringing GPU-accelerated C, C++ and Fortran compilers, libraries and message passing interface tools to the DGX Spark cluster — the same development tools used on NVIDIA-powered supercomputers. Teams can build and test scientific applications across four nodes, check how computation and data are distributed, and debug communication issues before scaling to larger systems. The kit brings scientific simulation and AI models and agents to the desktop, with demonstrated four-system workloads including LAMMPS with Kokkos for materials simulation and NAMD for molecular dynamics. The broader DGX Spark community has validated GROMACS for molecular dynamics, OpenFOAM for fluid dynamics, WRF for weather modeling and NVIDIA NIM microservices such as OpenFold3 and Boltz-2 for predicting biomolecular structures. Running on DGX Spark, new open dataset and post-training recipes for Nemotron 3.5 Super VL help researchers adapt the model for life sciences and chemistry, enabling secure scientific workflows on infrastructure organizations control. What Researchers Can Do With It The kit opens up new possibilities for researchers. Researchers can explore agent-assisted workflows that prepare simulation inputs, launch runs and summarize results, while the scientists validate the setup and conclusions. Researchers can also train AI models that approximate slow simulations, then use them to refine wing designs toward minimum drag. Before committing supercomputer time, a team can narrow a 10-million-cell simulation to the 50-500 dimensions that matter. Kvantify is using NVIDIA cuQuantum, and reports running tensor-network simulations of quantum-chemistry circuits with up to 70 qubits and 975 gates on a single DGX Spark, benchmarking circuits generated for penicillin ground-state energy calculations. CYRAN AI , an IIT Delhi spin-out, uses DGX Spark and its PixelFlux library to accelerate JPEG 2000 decoding for large satellite images, supporting local geospatial analysis. OpenFold shows how researchers can fine-tune OpenFold3 on NVIDIA DGX Spark to improve protein-ligand structure predictions for drug discovery. The Rubin Lab at IGI at University of California, Berkeley is using a DGX Spark and NVIDIA Evo 2 model to predict bacterial traits directly from genome sequences by comparing embeddings from genomic and protein language models against the classical baselines. With no 11-hour cluster queue times, results arrived 15x sooner on DGX Spark. The Hartwig Group at UC Berkeley is using a DGX Spark to perform LoRA fine-tuning on the Nemotron 3 Super 120B-A12B model for chemistry machine learning, training a prediction head on experimental values. By running locally on the Spark using NVFP4 quantization, the lab successfully fit a 120B-class model on a single device and bypassed multi-day cluster queue times, allowing them to run continuous eight- to 24-hour training fits back-to-back. Where to Find the Kit The developer kit is on display at SC26 and available for sale. DGX Spark products and cables can be purchased from top OEM partners and are also available through NVIDIA Marketplace . Partners include Acer, ASUS, Dell Technologies, GIGABYTE, H3C, HP, Lenovo, MSI and xFusion. An NVIDIA playbook for building and running the NVIDIA HPC SDK, a software development kit for scientific computing, is coming soon. Once available, it will give scientists and researchers a local platform to develop multi-node GPU applications, explore FP32 and mixed-precision algorithms and run molecular dynamics and materials simulations using LAMMPS with Kokkos before scaling to data center systems. Learn more about NVIDIA DGX Spark . Subscribe to the NVIDIA Local AI newsletter . Follow NVIDIA RTX Spark on X , Instagram , TikTok and Facebook . See notice regarding software product information. Read All NVIDIA Announcements From SC26

  • Teams can develop and debug on their own, saving server hours for production runs and keeping unpublished results, and proprietary datasets on premises.
  • Researchers can explore agent-assisted workflows that prepare simulation inputs, launch runs and summarize results, while the scientists validate the setup and conclusions.
  • OpenFold shows how researchers can fine-tune OpenFold3 on NVIDIA DGX Spark to improve protein-ligand structure predictions for drug discovery.
Nguyên văn nguồn
Chi tiết nguồn
Tác tử

v18.22.0

<h2>@oh-my-pi/pi-ai</h2> <h3>Fixed</h3> <ul> <li>Fixed slow streaming of Gemma tool calls with large arguments.</li> </ul> <h2>@oh-my-pi/pi-coding-agent</h2> <h3>Breaking Changes</h3> <ul> <li>Moved binary analysis from the standalone <code>ida</code> and <code>ghidra</code> tools into the backend-neutral <code>dsm</code> namespace (<code>xd://eval/dsm</code>).</li> <li>Replaced the <code>ida.enabled</code>, <code>ida.maxOpen</code>, and <code>ida.idleCloseSec</code> settings with <code>disassembler.backend</code>, <code>disassembler.maxOpen</code>, and <code>disassembler.idleCloseSec</code>; shared disassembler database APIs are now available under <code>@oh-my-pi/pi-coding-agent/disasm</code>.</li> </ul> <h3>Added</h3> <ul> <li>Added Tern plan review editing and comments: double-click plans to edit them, select text to add Google Docs-style comments, and send comments to Refine.</li> <li>Added Ghidra support with automatic setup of Ghidra, JDK, and PyGhidra when analysis is explicitly requested; Auto mode uses an installed IDA Pro when available and otherwise falls back to Ghidra.</li> <li>Added support for sharing live analysis databases with child agents through <code>db.handle()</code> and <code>dsm.open(handle)</code>.</li> <li>Added a <code>Force Remove</code> option to <code>worktree.onExit</code> that removes session-created worktrees without prompting, including worktrees with uncommitted changes; new commits remain on the worktree branch.</li> </ul> <h3>Changed</h3> <ul> <li>Updated Tern plan review with a single Approve control, comment-aware Refine option, model selection, and a Save action that writes the plan while keeping review open and plan mode active.</li> <li>Plain executable reads now display Mach-O, ELF, and PE metadata without opening an analysis database or downloading a backend; request an overview, function details, or <code>dsm.open()</code> to begin analysis.</li> <li>Improved the model switcher (<code>Alt+P</code> / <code>/switch</code>) with provider filtering and per-switch thinking-level selection. In Tern, it now opens above the composer with integrated search, provider and Task/Main filters, and a thinking-level slider.</li> </ul> <h3>Fixed</h3> <ul> <li>Deleting or moving large directory trees inside a project no longer stalls language-server file watching (<a href="https://github.com/can1357/oh-my-pi/pull/19490" data-hovercard-type="pull_request" data-hovercard-url="/can1357/oh-my-pi/pull/19490/hovercard">#19490</a> by <a href="https://github.com/H4vC">@H4vC</a>).</li> <li>Reduced CPU and memory usage for long-running subagents as their sessions grow (<a href="https://github.com/can1357/oh-my-pi/pull/19486" data-hovercard-type="pull_request" data-hovercard-url="/can1357/oh-my-pi/pull/19486/hovercard">#19486</a> by <a href="https://github.com/H4vC">@H4vC</a>).</li> <li>Fixed <code>commit.draft()</code> incorrectly classifying files deleted or renamed through bash or eval as “Not yours”; they now appear under Review.</li> </ul> <h3>Removed</h3> <ul> <li>Removed the legacy standalone disassembler tools and their associated prompt instructions.</li> <li>Removed the Copy and Edit-in-editor buttons from the Tern plan review sheet; use selection copying and double-click editing instead.</li> </ul> <h2>@oh-my-pi/collab-web</h2> <h3>Fixed</h3> <ul> <li>Improved streaming performance in the collab viewer, preventing long replies from becoming progressively slower as they are received.</li> </ul> <h2>@oh-my-pi/pi-natives</h2> <h3>Breaking Changes</h3> <ul> <li><code>TtyWriter</code> now requires a drain threshold and callback and only supports independently reopenable TTYs. Its pending UTF-8 byte bounds are reported conservatively.</li> </ul> <h3>Changed</h3> <ul> <li>Large terminal repaints are handled more efficiently, reducing JavaScript-thread work and releasing oversized buffers after output drains.</li> </ul> <h3>Fixed</h3> <ul> <li>Fixed stalled terminal writers continuing to emit output or leaking file descriptors after shutdown.</li> <li>Fixed terminal output preserving Unicode across chunk boundaries and safely replacing malformed surrogate sequences.</li> </ul> <h2>@oh-my-pi/pi-tui</h2> <h3>Breaking Changes</h3> <ul> <li>Terminals that support output backpressure must now implement onOutputDrain(callback), which returns an unsubscribe function; terminals without backpressure can omit both pendingOutputBytes and onOutputDrain.</li> <li>Updated PlanReviewOverlay options to use PlanReviewOverlayChoice objects and report choice IDs to onPick and selectOption. Refine and Save actions now reflect plan feedback and saving behavior; removed disabledIndices and externalEditorName, and updated describe to receive DescribeContext.</li> </ul> <h3>Added</h3> <ul> <li>Added configurable model search placement, provider scopes, and an optional thinking-level slider to ModelPickerComponent; selected thinking levels are reported through onPick metadata. The native picker sheet now supports anchoring, scope styling, level display, and icon actions.</li> <li>Added a native plan review document and dock with editable and commentable plans, quote annotations, plan-edit notifications, Approve, Save, and Cancel actions, plus model selection.</li> </ul> <h3>Changed</h3> <ul> <li>Provider scopes now show only the selected provider's models when a provider is chosen, with scope cycling able to return to the full model list.</li> <li>Updated thinking-level controls to use left and right arrow keys, and refreshed the slider and footer hint presentation.</li> <li>Retitled the plan-saving overlay to “Save plan” and updated its action label.</li> </ul> <h3>Fixed</h3> <ul> <li>Fixed slow scrolling and typing in long agent transcripts.</li> <li>Fixed slow typing and cursor movement in drafts longer than 256 lines.</li> </ul> <h2>@oh-my-pi/pi-wire</h2> <h3>Added</h3> <ul> <li>Added optional picker icons for tooltips and accessible names, plus an option to control the placement of the

  • </li> <li>Improved the model switcher (<code>Alt+P</code> / <code>/switch</code>) with provider filtering and per-switch thinking-level selection.
Nguyên văn nguồn
Chi tiết nguồn
  • v18.22.0 Can Bölük — Oh My Pi releases · 17 thg 11, 2026
Kiến trúc

Building Visual Intelligence — Andreas Blattmann & Dustin Podell, Black Forest Labs

Câu hỏi But why are they now going into video and world models?

This episode is sponsored by Deepgram. Try Deepgram for voice agents that truly understand your context! #contextiseverything Not many people know that Stable Diffusion is why I went fulltime into AI (I aptly titled Multiverse, not Metaverse to call out Zuck’s mistaken investment). It turns out that Dustin Podell also had the same aha moment - quitting his day job when the model came out. But we took very different paths… I stuck to what I knew in human content creation, and Dustin immersed deep into AI content creation with Andreas Blattmann , co-inventor of Latent Diffusion and the rest of the world-leading imagegen crew at Stability AI. Stability fell apart, but Anj Midha helped the core group spin out to BFL, which immediately built an incredible new imagegen model in the Flux series. But why are they now going into video and world models? Andreas gives an example you can try with one hand. If somebody tosses you a ball, your hand starts moving before the ball reaches you. You can do that without being able to write down the trajectory. When he talks about building visual intelligence , that’s the kind of knowledge he means. A still can show you where the ball is. A video shows you how it got there. The sound of impact can tell you when it hit something, even if the camera missed the contact. That’s BFL’s argument for training FLUX 3 on the three signals together. BFL’s FLUX 3 release The horse in motion is one of the FLUX 3 examples we show in the intro. The audio examples are worth playing with the sound on. FLUX 3 Video What a moving camera teaches you Suppose the camera starts in front of a person and pans around them. At some point the model owes you the back of their head. Andreas thinks ordinary two-dimensional video contains a great deal of evidence about what objects look like from other angles. That doesn’t settle whether a model “understands physics,” a point we go back and forth on. It does give you a much better test than asking whether one frame looks convincing. There’s a good engineering example late in the episode. Sharpen each frame of a clip independently and each frame might look terrific. Hit play and the extra detail can shimmer, because the upscaler made a slightly different decision every time. BFL derives its video upscaler from the same FLUX 3 backbone. Temporal consistency matters even when all the user asked for was a higher-resolution file. What’s worth spending training compute on? One of the questions I had for them was whether autoregression was taking over image generation. Andreas gives the other side its due. An autoregressive model can learn from all the tokens in a training example at once. On the diffusion side, the team can distill a long sampling process down to a handful of steps. Which advantage matters depends on whether you’re paying to train the model or waiting for an image to finish. Dustin’s early Kontext experiment is a nice example of how they decide where to invest. He suspected an image model that could draw a dog wearing a hat already knew something about where hats go. He and a colleague collected a small set of photos labeled with relationships. A training run of a few hours got the model to put a logo on a shirt. It was rough, but it was enough to take to the next offsite. The team then spent months on the data and the model. FLUX Kontext examples My favorite user story comes near the end. People were feeding FLUX old, degraded family photos, restoring them, then showing the pictures to their families. Dustin says he hadn’t predicted that use. After talking about preserving somebody’s identity across edits and across frames, it’s a lovely example of who actually notices if you get it right. Show notes BFL explains what a still image, a moving camera and a soundtrack each contribute during training. FLUX 3 launch and research overview Watch and listen to the results of generating video and sound together. FLUX 3 Video demos See how the lab teaches a model to make a requested change while preserving the rest of an image. FLUX Kontext examples Andreas Blattmann co-created Stable Diffusion and co-invented the latent diffusion method behind it. Guest bio · Original research Dustin Podell led FLUX Kontext and was the lead author of SDXL. Dustin’s project credits · FLUX Kontext · SDXL The first Latent Space essay grew out of the Stable Diffusion release. Multiverse, not Metaverse Chapters 0:00 Host introduction 3:39 Cold open 4:23 Black Forest roots 9:23 Dustin’s path 17:04 Visual intelligence 27:38 Inside SDXL 33:59 Learning motion 39:57 Flow matching 46:57 Starting BFL 50:36 Deepgram sponsor break 51:43 Early use cases for FLUX 54:58 Editing with Kontext 1:03:15 Training data 1:07:32 Testing an edit 1:13:12 Autoregression 1:20:50 World action models 1:30:08 Robotics in 2027 1:32:16 Who uses FLUX?

  • BFL’s FLUX 3 release The horse in motion is one of the FLUX 3 examples we show in the intro.
  • FLUX 3 launch and research overview Watch and listen to the results of generating video and sound together.
Nguyên văn nguồn
Chi tiết nguồn
Đánh giá

TokenEconomizer: Learning from Experience to Make AI Coding Agents More Efficient

Câu hỏi In our paper, TokenEconomizer: Experience-Driven Black-Box Cost Control for Coding Agents , we ask a simple question: can experience in a repository help an agent solve future tasks more efficiently?

At a glance Developers use AI coding agents in the same codebases day after day, yet agents often start each session from scratch: rediscovering familiar facts, loading unused tools, and repeating unsuccessful approaches. TokenEconomizer is an open-source plugin that learns from an agent’s past sessions in a repository to reduce waste in future sessions, without modifying the agent or its underlying model. Across 277 real bug-fixing tasks, TokenEconomizer reduced costs by 18.4–30.9% across multiple AI coding agents. Costs fell in every repository tested, while task success rates remained broadly stable. AI coding agents are increasingly helping developers fix bugs and build features largely on their own, but every interaction with their models consumes tokens. In each session , an agent sends instructions, tool descriptions, and session history to its model. Most efforts to reduce this cost focus on individual tasks. Yet developers return to the same repositories, giving agents an opportunity to learn from previous work. Without that learning, an agent may repeatedly rediscover how to run tests, load tools it never uses, or spend dozens of turns revisiting the same error. In our paper, TokenEconomizer: Experience-Driven Black-Box Cost Control for Coding Agents , we ask a simple question: can experience in a repository help an agent solve future tasks more efficiently? TokenEconomizer, an open-source plugin from UC Berkeley, Microsoft Research, and Microsoft Security, turns completed sessions into guidance for subsequent ones. It reduces waste without changing the agent or its underlying model, so developers can use it alongside the coding agents they already rely on. Figure 1. Each finished session teaches TokenEconomizer something about the repository, so later sessions cost less tokens. Where the tokens go We first profiled Claude Code on 60 tasks from SWE-Bench, a benchmark of real bug fixes in open-source Python projects, running each repository’s tasks in sequence. Three kinds of waste recurred: Rediscovering the same facts . After a repository had accumulated at least five completed tasks, 30% of tool calls repeated an action from an earlier task in that repository. Only 7% matched an action in another repository’s history. Paying for unused tools . The agent received 23 tool descriptions on every turn, but 16 were never called. Getting stuck on failures . The 17% of tasks that failed accounted for 32% of the spending. We also tested three popular open-source plugins, each built around a fixed cost-saving strategy. Each lowered costs in some repositories but not others. Together, these findings suggest that effective cost control must adapt to the repository, rather than apply the same strategy everywhere. How TokenEconomizer works After each session, TokenEconomizer reads the session’s log and updates what it knows about the repository. It then acts through three mechanisms, using only interfaces the agent already provides, as illustrated in Figure 2. Micro-skills capture repository-specific knowledge in one-line notes. TokenEconomizer identifies recurring patterns across sessions and proposes corresponding notes. They are added to the agent’s starting context only after several sessions have confirmed the pattern, and retires the note if later evidence contradicts it. Tool pruning removes tools that repository history suggests the agent is unlikely to use , while always retaining core tools for running commands and editing files. Removed tools can be restored if needed. Failure-loop intervention watches for an agent that keeps hitting the same failure. A judge model decides whether the agent is truly stuck or still making progress. If it is stuck, the agent gets a short note suggesting a different approach. In the open-source plugin, if the agent is still stuck after the note, TokenEconomizer alerts the developer rather than stopping the work itself. Figure 2. TokenEconomizer works alongside an unchanged coding agent. Micro-skills add learned facts, tool pruning trims the loaded tools, and failure-loop intervention redirects stuck work. Each finished session updates all three. The controller is designed to revise its decisions as new evidence arrives. Reported savings also account for the cost of TokenEconomizer’s own model calls. Saving costs in every repository We evaluated TokenEconomizer on 277 SWE-Bench tasks from six repositories, using five different AI coding agent and model combinations. For each repository, TokenEconomizer started with no prior session history. TokenEconomizer lowered the average cost by 18.4–30.9%, as shown in Figure 3, and saved the most in all 30 repository-combination pairs without performance loss. By contrast, lowering the model’s reasoning effort saved similar or larger amounts but led to substantial performance degradation. Figure 3. On SWE-Bench, TokenEconomizer lowered mean cost per task by 18.4–30.9% across five agent–model combinations (left), while Pass@1 rose in four of them (right). Savings grow as experience accumulates. Early tasks can cost more while TokenEconomizer learns, but cumulative costs fell below the default in every repository by the sixth task, shown inFigure 4. Each mechanism contributes savings on its own, and combining all three delivers the largest reduction. Figure 4. Cumulative cost change relative to default over each repository’s tasks, using Claude Code. Starting with no history, every repository becomes cheaper by the sixth task. The savings also carried over to sequential web development, with costs 16.0–21.2% lower on Web-Bench. On security auditing, where every task is in a new repository, costs were 6.4–19.6% lower. Looking ahead TokenEconomizer opens a path toward coding agents that learn not only how to solve tasks, but also how to spend computation more effectively. Its lightweight controller shows how experience can improve efficiency without changing the underlying agent or model. The open-source plugin ma

  • TokenEconomizer is an open-source plugin that learns from an agent’s past sessions in a repository to reduce waste in future sessions, without modifying the agent or its underlying model.
  • Most efforts to reduce this cost focus on individual tasks.
  • It reduces waste without changing the agent or its underlying model, so developers can use it alongside the coding agents they already rely on.
Microsoft Research · Aaron J. Li, Weijia Xu, Chandan Singh, Jeevana Priya Inala, Bin Yu, Ion Stoica, Michel Galley, Jianfeng Gao Mở bản gốc ↗
Nguyên văn nguồn
Chi tiết nguồn
Huấn luyện

Lambda Boosts AI Factory Resiliency and Efficiency With NVIDIA Photonics

To train an AI model, thousands of GPUs have to work together, requiring a network that’s fast, efficient and resilient. Lambda is tackling the network piece of the puzzle with NVIDIA Photonics co-packaged optics (CPO) switches. The cloud provider is using the switches to connect NVIDIA GB300 NVL72 racks with more than 10,000 GPUs, delivering lower power, higher resiliency and longer mean time between interruptions (MTBI) to an AI factory at scale. “Networking has to keep pace with the demands of increasingly large AI clusters,” said Dave Ward, president of cloud services at Lambda. “Bringing co-packaged optics into a deployment of this scale is an important step toward building the infrastructure needed to support the next generation of AI.” Read All NVIDIA Announcements From SC26 In the company’s tests using the switches, network latency — the time needed for moving data — dropped by about 4% at rack scale and by up to about 10% at pod scale. Across rack, scalable unit and pod configurations, the network maintains the same throughput as conventional pluggable optics. Delivering Power Efficiency and Resiliency at Scale Improving power efficiency and resiliency matters because the network sits at the center of AI factory productivity. In a large training job, one interrupted link can stall thousands of GPUs. With agentic inference, one network disruption can introduce cascading latency effects on many agents. Every interruption costs time, and every watt spent on networking is a watt not available for compute. NVIDIA Photonics addresses both power efficiency and resiliency. By replacing pluggable transceivers with co-packaged optical engines on the switch package, it delivers up to 5x better power efficiency and up to 15x higher reliability than conventional pluggable optics. It also uses 4x fewer lasers, which means fewer components that can fail. Cooler Optics, Replaceable Light Sources Increase Resiliency Lambda’s deployment uses NVIDIA Photonics switches, each with 144 ports running at 800Gb/s. Liquid cooling carries heat away from the switch chip and optical engines. The company reports a median optics temperature of 36°C, compared with 51°C for conventional pluggable optics, and its tests find more stable signal integrity. NVIDIA Photonics also brings the electrical-to-optical conversion closer to the switch chip. Each switch has 18 removable external light sources, which supply light to the optical engines. Optical engines sit alongside the chip on the same package and convert electrical signals into light that carries data over fiber between parts of the cluster. Technicians can replace the light-source modules in the field, unlike traditional pluggable optics, which require replacing the entire module if a laser fails. Co-packaging shortens the electrical path from inches to millimeters, which reduces signal loss. It also removes the need for retimer chips, cutting both power draw and the number of parts in the signal path. Explore NVIDIA Photonics and read Lambda’s deployment announcement . Read All NVIDIA Announcements From SC26

  • Delivering Power Efficiency and Resiliency at Scale Improving power efficiency and resiliency matters because the network sits at the center of AI factory productivity.
  • Co-packaging shortens the electrical path from inches to millimeters, which reduces signal loss.
Nguyên văn nguồn
Chi tiết nguồn
Đánh giá

NVIDIA Sweeps MLPerf Training and Endpoints as Vera Rubin Debuts With Leading Performance

In its MLPerf benchmarks debut, with results released today, the NVIDIA Vera Rubin platform delivers leading performance across training, post-training and agentic inference. NVIDIA swept every MLPerf Training v6.1 benchmark and, in its first MLPerf Endpoints submission, delivered more than 10x generational gains in throughput per watt for agentic AI — measuring how much agentic work a system delivers with each watt of power. These results are just the beginning. Vera Rubin performance will keep climbing through continuous optimization of NVIDIA’s CUDA-powered software stack , so the infrastructure that customers invest in today delivers more value over time. NVIDIA Sweeps MLPerf Training v6.1 AI agents are now everywhere, writing code, running business workflows, answering customer questions and, as this week’s SC26 conference highlights , advancing scientific research. Faster training means smarter agents can reach users sooner, drug discovery models can keep pace with new lab data and fraud-detection models can adapt to new threats faster. MLPerf Training measures how fast a system can train a range of AI models to a target level of accuracy. The NVIDIA platform was the only one with results on every benchmark this round, spanning large language models, image generation, recommender systems and reinforcement learning post-training. In its debut, Vera Rubin NVL72 delivered up to 2x more performance per GPU than the highly optimized NVIDIA GB300 NVL72 system — early results that show the pace of generation-over-generation innovation. MLPerfTraining v6.1 results retrieved from www.mlcommons.org on Nov 16, 2026, from the following entries: 6.1-0116 , 6.1-0030, 6.1-0135, 6.1-0020, 6.1-0049, 6.1-0123, 6.1-0138, 6.1-0073, 6.1-0124, 6.1-0129 and 6.1-0038. The MLPerf name and logo are trademarks of MLCommons Association in the United States and other countries. All rights reserved. Unauthorized use is strictly prohibited. See www.mlcommons.org for more information. NVIDIA also demonstrated performance at scale. On the DeepSeek-V3 benchmark, Vera Rubin NVL72 scaled from 256 to 1,024 GPUs with about 97% efficiency: quadrupling the GPUs cut time to train from 16 minutes to about four. NVIDIA also ran the benchmark on 10,240 GPUs with GB300 systems, representing the largest scale in this round. Microsoft Azure matched that scale with 10,240 NVIDIA GPUs with GB200 systems. Results at this scale matter because adding GPUs only speeds up training if they can work together efficiently. NVIDIA NVLink scale-up networking connects the GPUs within each system so they act as one, while NVIDIA Quantum-X800 InfiniBand and Spectrum-X Ethernet scale-out networking link those systems across the data center. Combined with optimized software, this keeps thousands of GPUs working in concert, helping developers build more intelligent models while cutting training times. NVIDIA Vera Rubin Leads Agentic Inference on MLPerf Endpoints An AI agent for travel that plans and books a whole trip — or an IT agent that finds and fixes a problem before anyone files a ticket — may reason, call tools and check its work many times to finish a single task. Each of those steps generates tokens. Because AI data centers are limited by power, the more tokens a system can generate each second for every watt of power, the more agents it can serve and the more value it creates. MLPerf Endpoints tests systems on real agentic coding and enterprise workflows, and checks that the agents still complete their tasks correctly. NVIDIA Vera Rubin leads across all of its agentic benchmarks. For example, in a provisional result for Kimi K3, a large open model built for agentic workflows, Vera Rubin NVL72 systems deliver more than 10x the throughput per watt of NVIDIA GB300 NVL72 at the same response speed. MLPerf® Endpoints v1.0, Kimi K3, Standardized division, Available category, published November 9, 2026. NVIDIA Vera Rubin NVL72, 32 accelerators, result 1.0.1.kimi-k3.agentic_combined.1p, run 9: concurrency = 1,024 queries, TTFT P90 = 4,673.9 ms, system TPS= 41,414.2 tokens/sec, E2E average interactivity = 50.4 tokens/sec per user, operating power = 100.89 kW. NVIDIA GB300 NVL72, 48 accelerators, result1.0.0.kimi-k3.agentic_combined.1, run 6: concurrency = 48 queries, TTFT P90 = 2,305.1 ms, system TPS= 2,210.4 tokens/sec, E2E average interactivity = 56.2 tokens/sec per user, operating power = 60 kW. Comparison based on throughput per kilowatt at a shared minimum E2E average interactivity of 50 tokens/sec per user. Retrieved from endpoints.mlcommons.org on November 9, 2026. GB300 results verified by MLCommons Association. For the provisional Vera Rubin results: pending MLCommons Association verification, results are subject to change pending peer review. The MLPerf name and logo are registered and unregistered trademarks of MLCommons Association in the United States and other countries. All rights reserved. Unauthorized use is strictly prohibited. See www.mlcommons.org for more information. Every AI service has to balance serving many users at once against responding quickly to each one. Vera Rubin shifts that entire curve, serving more users while responding faster. Leadership Across Benchmarks and Workloads NVIDIA partners participated extensively in MLPerf Training, with submissions from 20 organizations: ASUS, Azure , Cisco, CoreWeave, Crusoe, Dell Technologies, Firmus , Giga Computing, Google Cloud, HPE , Inventec, Krai, Lambda, Nebius, Oracle, Quanta Cloud Technology (QCT), Supermicro, TTA and the University of Florida . Their submissions show how NVIDIA’s programmable platform, with open source libraries and frameworks, lets a vast ecosystem of partners innovate and optimize AI training performance. The lead extends beyond MLPerf. Vera Rubin also delivers leading performance on independent benchmarks including SemiAnalysis AgentX and Artificial Analysis AgentPerf, showcasing how the same NVIDIA infrastructure can productively run everything

  • In its MLPerf benchmarks debut, with results released today, the NVIDIA Vera Rubin platform delivers leading performance across training, post-training and agentic inference.
  • These results are just the beginning.
  • The NVIDIA platform was the only one with results on every benchmark this round, spanning large language models, image generation, recommender systems and reinforcement learning post-training.
Nguyên văn nguồn
Chi tiết nguồn
Đánh giá

NVIDIA CUDA-X Libraries Give Agents the Tools to Accelerate Scientific Discovery

With expanded NVIDIA CUDA-X libraries, including cuPhoton, cuPotential and cuSpa, researchers can sift through vast amounts of data to accelerate research, from astrophysics and materials science to single-cell biology. The speedups are helping scientists analyze larger datasets and model more complex systems. The new libraries give AI agents the tools to help researchers make discoveries faster than ever. Read All NVIDIA Announcements From SC26 Bringing Black Holes Into Focus Seeing a black hole takes a telescope the size of Earth. The Event Horizon Telescope links radio dishes worldwide, generating petabytes of data that can take months to process. With NVIDIA cuPhoton.xBeam, a module within the cuPhoton Toolkit , black hole observations from the Event Horizon Telescope can now be correlated orders of magnitude faster than ever. Analysis that once took over a month now takes eight minutes. By combining global radio dishes into Earth-sized aperture, cuPhoton delivers unprecedented astronomical imaging resolution. cuPhoton successfully correlated observations of powerful black hole jets J1924 and 3C279 in the night sky — making it a reality to reconstruct a Messier 87* black hole galaxy in real time. The entire pipeline, from data loading and processing to visualization, runs on distributed multi-GPU systems or a compact NVIDIA DGX Spark system, helping scientists study how gas spirals inward under a black hole’s powerful gravitational pull. Steering Fusion Simulations Sustaining plasma at millions of degrees is essential to achieving stable, continuous fusion energy. But simulating its behavior is costly: on petascale systems, a single time step can take days, with simulations running for weeks. Experiments also generate data faster than traditional pipelines can process it, often leaving scientists waiting until experiments end to analyze results. Pacific Fusion researchers used NVIDIA cuPhoton.xLens module to extract insights while plasma simulations run, helping them visualize what’s happening as results emerge. The NVIDIA cuPhoton.xLens module extracts insights while plasma simulations run, helping researchers understand what’s happening as results emerge. Researchers can couple the entire pipeline with an NVIDIA Nemotron -powered agent running locally on the same system, enabling interactive analysis that can help steer simulations in real time to inform fusion reactor target and chamber design. Exploring Potential Medicines Molecular simulations help scientists evaluate drug designs before costly laboratory experiments. NVIDIA cuPotential, built on work by researchers at Simons Foundation’s Flatiron Institute and D. E. Shaw Research , accelerates existing simulation software by up to 3x while improving calculation stability and accuracy. It expedites synthetic data creation for generative AI models, finding previously unknown molecular binding sites and calculating the strength of molecular interactions. NVIDIA is working with Schrödinger and the AMBER , NAMD , GROMACS , LAMMPS and Open Molecular Software Foundation communities to bring cuPotential acceleration to molecular dynamics engines and workflows. Cadence , Fathom Therapeutics , Topos Bio , Genesis Molecular AI and Terray Therapeutics are adopting it for drug discovery. Accelerating Materials Research Daviator , a collection of batched GPU kernels for the widely used Vienna Ab initio Simulation Package, is coming to NVIDIA ALCHEMI — delivering quantum-mechanical calculations up to 22x faster and at up to 75% lower cost. With Daviator, researchers building simulation models can now use GPU acceleration across atomistic systems of all sizes to generate the millions of calculations of atomic interactions needed for model training and inference, including for small systems that previously ran faster on CPUs. At Lawrence Berkeley National Laboratory , the NVIDIA DAQIRI networking library connects a high-speed electron microscopy detector to GPU processing, moving analysis into the experiment itself. Researchers can identify useful signals, discover previously unknown regions of interest and decide what to measure next while measurements are being taken. The workflow supports atomic-resolution materials studies that can distinguish elements, laying the groundwork for future automated experiments. Revealing Biology at Scale The newly open source cuSpa accelerates cell polygon calculations, enabling developers to generate massive spatial graphs for next-generation patient and cancer models at Memorial Sloan Kettering Cancer Center , with a 900x speedup compared with using CPUs. Codeveloped with scverse , the rapids-singlecell framework accelerates single-cell analysis — and is now available in a single, agent-ready container. Using the tool, Therna Biosciences processed data 200x faster for therapeutic design. 10x Genomics integrated rapids-singlecell in its new Sentira platform, helping academic researchers, biopharma teams and foundation model builders accelerate the path from large-scale spatial data to biological insights. And the NVIDIA GPU-accelerated HMMER library, also open source, accelerates the annotation of predicted protein sequences, supporting data preparation for biological foundation models. At Basecamp Research , benchmark searches ran up to 84x faster than with standard HMMER, reducing the time required to annotate the large biological datasets used to train the company’s therapeutic design models. Explore NVIDIA CUDA-X libraries. Learn more in this technical blog on GPU-powered black hole reconstruction. Read All NVIDIA Announcements From SC26

  • Steering Fusion Simulations Sustaining plasma at millions of degrees is essential to achieving stable, continuous fusion energy.
  • Experiments also generate data faster than traditional pipelines can process it, often leaving scientists waiting until experiments end to analyze results.
  • xLens module to extract insights while plasma simulations run, helping them visualize what’s happening as results emerge.
Nguyên văn nguồn
Chi tiết nguồn
Đánh giá

NVIDIA Announces Rubin HPC — A New GPU to Accelerate Agents for Science, Engineering and Analytics

Announced today at SC26, NVIDIA Rubin HPC is a new GPU in the NVIDIA Vera Rubin platform designed to accelerate simulation and data processing for science, engineering and analytics workflows. As agents help scientists and engineers explore more candidate materials, molecules and designs, they rely on simulations to test those ideas and guide their next steps. Rubin HPC will accelerate those simulations, shortening the path from exploration to scientific discovery and industrial design. The Rubin HPC GPU features a record number of CUDA cores — more than 50,000 on a single GPU — as well as a sparse data acceleration capability that keeps those cores fed with data, and accelerates GPU-to-GPU connectivity with seventh-generation NVIDIA NVLink interconnect technology. Read All NVIDIA Announcements From SC26 Scientists and engineers rely on simulation to understand complex systems before committing to experiments or physical prototypes. Working with agents, these teams can explore more candidate materials, molecules and designs. This increases demand for the GPU-accelerated simulations needed to evaluate those ideas and guide next steps. “For agents tackling science and engineering, faster simulation means faster discovery,” said Ian Buck , vice president of hyperscale and high-performance computing at NVIDIA. “Rubin HPC accelerates the simulations agents use to test ideas and guide their next steps, helping scientists and engineers explore more possibilities and bring breakthroughs closer.” Turning Faster Computation Into Faster Discovery Across climate models, engineering simulations, genomic analysis and financial risk, scientific computing workloads take a similar shape: an application reaches into an immense dataset, gathers a handful of scattered values, calculates a result and repeats this on an enormous scale. To help address the challenge, Rubin HPC will offer a balanced design, scaling data movement and compute together rather than increasing compute alone. That balance is what optimizes Rubin HPC for scientific computing. It keeps the GPU’s record number of CUDA cores supplied with data, so they spend their time computing rather than waiting. This accelerates recurring patterns of every simulation workload — from sparse matrix operations to stencil computations to sorting and reductions. These patterns are prevalent across data analytics, science and engineering workloads, which is why a purpose-built GPU delivering the broadest acceleration can help unlock cross-industry breakthroughs. NVIDIA CUDA-X libraries offer a growing set of optimized implementations of complex algorithms for agents to use, including those most relevant to simulation and analytics. Used together with Rubin HPC, the libraries can give agents the fastest glide path to reasoning outcomes. And faster agent-invoked simulations mean faster feedback — shortening the loop of hypothesis, simulation and refinement, so researchers and engineers can choose next steps sooner and point lab work at the most promising candidates. Rubin HPC is expected to be available in 2028. Read All NVIDIA Announcements From SC26

  • ” Turning Faster Computation Into Faster Discovery Across climate models, engineering simulations, genomic analysis and financial risk, scientific computing workloads take a similar shape: an application reaches into an immense dataset, gathers a handful of scattered values, calculates a result and repeats this on an enormous scale.
  • This accelerates recurring patterns of every simulation workload — from sparse matrix operations to stencil computations to sorting and reductions.
Nguyên văn nguồn
Chi tiết nguồn
Kiến trúc

NVIDIA Vera Rubin NVL72 Achieves up to 2x Blackwell Ultra Performance in MLPerf Training v6.1

As AI agents tackle increasingly complex workflows, reducing time-to-train is critical for rapidly iterating on both models and the agents they power. MLPerf... As AI agents tackle increasingly complex workflows, reducing time-to-train is critical for rapidly iterating on both models and the agents they power. MLPerf Training v6.1 marks the debut of the NVIDIA Vera Rubin architecture, and these early results demonstrate that its CUDA software stack is already highly mature. NVIDIA swept all seven benchmarks, with the Vera Rubin NVL72 platform delivering… Source

  • As AI agents tackle increasingly complex workflows, reducing time-to-train is critical for rapidly iterating on both models and the agents they power.
  • As AI agents tackle increasingly complex workflows, reducing time-to-train is critical for rapidly iterating on both models and the agents they power.
  • 1 marks the debut of the NVIDIA Vera Rubin architecture, and these early results demonstrate that its CUDA software stack is already highly mature.
Nguyên văn nguồn
Chi tiết nguồn
Đánh giá

Post-Training NVIDIA Nemotron to Increase Accuracy in Your Scientific Domain

Agentic AI is reshaping scientific research and innovation. Researchers communicate complex scientific hypotheses in natural language and let autonomous agents... Agentic AI is reshaping scientific research and innovation. Researchers communicate complex scientific hypotheses in natural language and let autonomous agents independently plan tasks, execute simulations, evaluate results, and self-correct across multiple iterations. Scientific agents are taking on longer, more complex, and increasingly visual workflows. Accuracy and inference efficiency… Source

  • Researchers communicate complex scientific hypotheses in natural language and let autonomous agents independently plan tasks, execute simulations, evaluate results, and self-correct across multiple iterations.
Nguyên văn nguồn
Chi tiết nguồn
Đánh giá

Import AI 481: AI sub-contractors; a wet lab for AI; and subterfuge during development

Câu hỏi Petrichor gestures at the same world for science - what will be the implications of “lights out” science facilities?

Welcome to Import AI, a newsletter about AI research. Import AI runs on arXiv, cappuccinos, and feedback from readers. If you’d like to support this, please subscribe. Subscribe now AI systems are getting good at sub-contracting to make themselves better: …AI can now productively spend money to improve itself… Researchers have developed FrontierML, a benchmark “that measures how well agents turn capital into model performance on challenging, end-to-end machine learning tasks.” It shows that today’s frontier systems are already able to, when given some money, turn this into improving their own performance on a basket of 21 distinct tasks. 21 long-horizon tasks “inspired by real work at frontier labs, neolabs, and academic institutions”. These tasks span six category areas, including LLM training, efficiency, vision & audio, computer use, applied NLP, and math & games. “Each task provides the agent with the same setup: a 12-hour CPU sandbox, a wallet with $10, $50, or $200, a few validation samples, and a task description.” Example task: Agents try to solve the tasks in pretty different ways. E.g, for one task about generating screenshots, the top-scoring agent fine-tuned a diffusion model using LoRA, while other high-scoring approaches involved an LLM writing a drawing program that is then rendered. Key features: Scaling spending : “Give agents $10, $50, and $200 per task to spend on training compute and data, and evaluate whether agents can use the additional capital productively,” they write. “In our experiments, we find that stronger agents are able to reliably improve their submissions with greater access to capital.” Open-ended & end-to-end : “In each task, agents can choose architectures, source data, implement training, provision compute, evaluate checkpoints, and optimize inference,” they write. One notable observation is that agents do quite a lot of novel stuff; “in 28% of trajectories, agents add newly trained parts to pretrained models or train a new model from scratch.” Decent verification: “Agents submit an inference server that serves their model on an air-gapped H100 machine,” they write. “To disincentivize overfitting, we provide limited validation samples and evaluate on a large out-of-distribution held-out test set, not shown to the agent”. Results : The authors measure performance by looking at the change in mean score over the 21 tasks when given additional money. Their baseline is a human researcher, which is able to get a 56 point boost when given $10 and a 66.1 boost when given $200. By comparison, the best performing system, Opus 5.5, gets a 45.4 boost with $10, and a 58.8 boost with $200. The highest ranking open weight model is Kimi K3.1 Pro which gets 25.5 with $10 and 29.6 with $200. Why this matters - AI systems are not only going to be able to build themselves, but they can creatively sub-contract: When trying to solve complex problems, people don’t work alone. In fact, they often try to farm out bits of work to specialists around them to help make themselves more effective. FrontierML shows that today’s powerful AI systems are getting increasingly adept at doing this for ML research. This both highlights the advancing capabilities of these systems for AI R&D and also points towards a future machine economy where AI systems sub-contract out different bits of work to other systems. Read more: FrontierML: a challenging, end-to-end machine learning benchmark (FrontierML) . *** Symptoms of the singularity: a wet lab built for AI: …Petrichor is a startup that gestures at the wild future ahead… As I said in my speech at the IMF, one of the hallmarks of the singularity is the massive acceleration in scientific advances attributable to AI. As part of this, we can also expect a related build out in the complements to science, like the wet labs in which to do AI-driven biology research. To that end, it’s notable to see Petrichor, a new Y Combinator startup that is a wet lab “built for AI labs and for the models they are training to do science”. Theory of change: Petrichor aims to be a biological wet lab that AI systems like Claude can use as a tool in the same way they currently use web search or other digital tools. Functionally, it will work as a machine-targeted API from the digital world to a gloopy physical world. “We aim to return the highest quality data an experiment can produce, as quickly as the biology allows,” Petrichor writes. “An experiment can be written by a model in plain language or in a structured format, so an agent can send work and receive results without a person relaying them.” Why this matters - lights out science: In manufacturing there’s this notion of a “lights out” factory, where no humans work in a facility and it is run almost entirely by machines. Petrichor gestures at the same world for science - what will be the implications of “lights out” science facilities? We can expect both drastic speedups in the rate at which science experiments can be done as well as a related growth in the need to create digital oversight of these facilities - after all, a highly competent synthetic biological scientist working 24/7 without human input could be capable of causing great peril as well as great abundance. Read more at the official site (Petrichor) . *** Uh oh! Looks like AI systems can do subterfuge during their own development: …AI systems can covertly do dangerous side tasks without monitors realizing… Scale AI has built RSI Control Arena, a way of testing out how AI systems behave when asked to develop themselves while performing dangerous tasks that might let them go against human preferences. The research follows earlier work this summer from the UK AI Security Institute (Import AI #465 ) which found that AI systems could do things like covertly extracting API keys while doing some seemingly mundane normal software development. RSI Control Arena: In RSI Control Arena, “a frontier model, acting as an autonomous agent, is given a benign main

  • ” It shows that today’s frontier systems are already able to, when given some money, turn this into improving their own performance on a basket of 21 distinct tasks.
  • Results : The authors measure performance by looking at the change in mean score over the 21 tasks when given additional money.
Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

Show HN: Starquix – Open-source local visual Astro website builder

Hi everyone. I love Astro and dislike other website builders, so I built Starquix. Code is a commodity today, so I open-sourced it. Would appreciate your feedback. Thanks!

Nguyên văn nguồn
Chi tiết nguồn
Tác tử

Show HN: Herdr mod that makes managing agents like playing Runescape

I find once I have dozens of agents working across multiple machine's it's hard to keep up. I also miss playing Runescape. So I solved both problems by making a Runescape inspired mod of Herdr. When agents are idle or have a question, they walk to the campfire. And different agent harnesses/configs have different outfits (e.g. to distinguish agents with subs vs those on API charges)

Nguyên văn nguồn
Chi tiết nguồn
Tác tử

More ways to move care forward with ChatGPT for Healthcare

Introducing computer-use agents, improved clinical answers, and new healthcare plugins that work across healthcare software to get real work done.

  • Introducing computer-use agents, improved clinical answers, and new healthcare plugins that work across healthcare software to get real work done.
Nguyên văn nguồn
Chi tiết nguồn
Tác tử

Agents SDK v0.24.0-alpha.8

<h1>Agents SDK v0.24.0-alpha.8</h1> <h2>Highlights</h2> <ul> <li><strong>Approve Python tool calls before they run.</strong> Review the arguments and approve or reject the call. You can remember a decision for the rest of the Session. <a href="https://github.com/openai/openai-agents-python/blob/d6a81c997eabdc56108633eb3132fb3907f87d7e/examples/codex/managed_approvals/README.md#inspect-and-decide">Managed approval example</a></li> <li><strong>Connect your own model client.</strong> The native tool loop works with async Responses and Chat Completions clients, LiteLLM, AnyLLM, and custom adapters. <a href="https://github.com/openai/openai-agents-python/blob/d6a81c997eabdc56108633eb3132fb3907f87d7e/docs/sdk/models.md#use-an-application-owned-model-client">Application-owned model clients</a></li> <li><strong>Add Realtime voice to a Session.</strong> Use a server WebSocket, connect a browser over WebRTC, or attach to an existing SIP call. <a href="https://github.com/openai/openai-agents-python/blob/d6a81c997eabdc56108633eb3132fb3907f87d7e/examples/codex/native_voice.md#choose-a-connection">Realtime connection options</a></li> <li><strong>Resume more kinds of work.</strong> Save supported pending tool requests and approvals. Request and token budgets keep the usage already spent when you restore a supported checkpoint. <a href="https://github.com/openai/openai-agents-python/blob/d6a81c997eabdc56108633eb3132fb3907f87d7e/examples/codex/budgeted-recovery.md">Restore a checkpoint with a budget</a></li> <li><strong>Install on more Linux systems.</strong> The preview adds an ARM64 wheel. Both Linux wheels support glibc 2.28 or newer. <a href="https://github.com/openai/openai-agents-python/blob/d6a81c997eabdc56108633eb3132fb3907f87d7e/.github/RELEASING.md#L44">Preview wheel platform requirements</a></li> </ul> <h2>Added</h2> <ul> <li>Observe tool-search calls, tool-search results, and hosted MCP approval decisions through <code>ModelOutputItemCompleted</code>. Use <code>include_execution_details=True</code> to receive full items; receive summaries by default. <a href="https://github.com/openai/openai-agents-python/blob/d6a81c997eabdc56108633eb3132fb3907f87d7e/packages/codex/src/_agents_codex/observation.py#L238-L267">Live tool observations</a></li> <li>Use native Session Realtime APIs from <code>agents.realtime.native</code>, including <code>RealtimeAgentConfig(is_tool_enabled=...)</code>. Keep retained Classic/v0 Realtime imports under <code>agents.realtime</code>. <a href="https://github.com/openai/openai-agents-python/blob/d6a81c997eabdc56108633eb3132fb3907f87d7e/examples/codex/native_voice.md#choose-a-connection">Native voice</a> · <a href="https://github.com/openai/openai-agents-python/blob/d6a81c997eabdc56108633eb3132fb3907f87d7e/CLASSIC.md#compatible-voice-and-realtime-imports">Classic imports</a></li> <li>Handle managed function approvals with <code>FunctionToolApprovalRequest</code> and <code>FunctionToolApprovalDecision</code>. Set <code>run_input_policy_before_approval=True</code> to run the input policy before the approval predicate as well as before execution. Return <code>FinishTurn(output)</code> from <code>tool_batch_completion</code> to propose a final value with output validation and completion checks. <a href="https://github.com/openai/openai-agents-python/blob/d6a81c997eabdc56108633eb3132fb3907f87d7e/examples/codex/managed_approvals/README.md#inspect-and-decide">Managed approvals</a> · <a href="https://github.com/openai/openai-agents-python/blob/d6a81c997eabdc56108633eb3132fb3907f87d7e/docs/sdk/outputs.md#finish-after-a-tool-batch">Finish a tool batch</a></li> <li>Create a remote conversation only when needed with <code>ConversationsHistory.ensure_conversation_id()</code>. Preserve a checked history prefix with <code>Session.replace_history_suffix(keep_prefix_count=...)</code>, and append staged checkpoint input with <code>SessionState.with_appended_pending_input()</code>. Read Unix milliseconds from <code>StoredTurn.started_at_ms</code> and <code>completed_at_ms</code>. <a href="https://github.com/openai/openai-agents-python/blob/d6a81c997eabdc56108633eb3132fb3907f87d7e/docs/sdk/reference.md#history-and-storage-apis">History APIs</a> · <a href="https://github.com/openai/openai-agents-python/blob/d6a81c997eabdc56108633eb3132fb3907f87d7e/docs/sdk/state.md#read-and-edit-original-history">History edits</a></li> <li>Use <code>SQLAlchemySQLiteSessionStore</code> for an application-owned synchronous SQLite engine, and <code>ResponsesSocketFactoryOptions</code> for resolved WebSocket connection parameters. Read received logprob data through <code>TextLogprobsFact</code>, <code>TokenLogprob</code>, <code>TokenLogprobAlternative</code>, and <code>MalformedLogprobsFact</code>, with unchanged serialized record kinds. Import <code>ModelRunOwner</code> from <code>agents.models</code> and <code>ApplicationToolOutput</code> from <code>agents.state</code>. <a href="https://github.com/openai/openai-agents-python/blob/d6a81c997eabdc56108633eb3132fb3907f87d7e/docs/sdk/state.md#sqlalchemy-sqlite-storage">SQLite storage</a> · <a href="https://github.com/openai/openai-agents-python/blob/d6a81c997eabdc56108633eb3132fb3907f87d7e/docs/sdk/reference.md#model-run-ownership-and-websocket-options">Model ownership and socket options</a> · <a href="https://github.com/openai/openai-agents-python/blob/d6a81c997eabdc56108633eb3132fb3907f87d7e/docs/sdk/streaming.md#inspect-provider-logprobs">Logprob facts</a></li> <li>Group prompt-cache routing for fresh Sessions with an explicit <code>TraceOptions.group_id</code>. Keep the saved cache key when restoring a Session. Set a borrowed model adapter's <code>prompt_cache_key</code> to override only that adapter's outgoing key. Treat routing as affinity without a guaranteed cache hit. <a href="https://github.com/openai/openai-agents-python/blob/d6a81c997eabdc56108633eb3132fb3907f87d7e/docs/sdk/observability.md#prompt-cache-routing-for-fresh-sessions">Prompt-cache routing</a

  • md#L44">Preview wheel platform requirements</a></li> </ul> <h2>Added</h2> <ul> <li>Observe tool-search calls, tool-search results, and hosted MCP approval decisions through <code>ModelOutputItemCompleted</code>.
Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

datasette 1.0a43

<p><strong>Release:</strong> <a href="https://github.com/simonw/datasette/releases/tag/1.0a43">datasette 1.0a43</a></p> <p>I've been leaning hard into Claude Fable 5.2 to help knock things off the <a href="https://github.com/simonw/datasette/milestone/37">1.0 milestone</a>, which is now 53% complete.</p> <p>The biggest 1.0 item in this release is <a href="https://docs.datasette.io/en/latest/plugin_hooks.html#plugin-register-output-renderer-streaming">streaming support for output renderers</a> - the built-in TSV and CSV export mechanisms now use this as well, and plugins can register themselves as capable of streaming every row in a table in one go via the new <a href="https://docs.datasette.io/en/latest/internals.html#internals-response-stream">Response.stream()</a> method.</p> <p>There are a couple of security fixes, one of which I deemed <a href="https://github.com/simonw/datasette/security/advisories/GHSA-mw6v-83vp-7c4c">deserving of a GitHub advisory</a>: private stored queries could be exposed, granting access to the title and SQL of the query (but not allowing that query to be executed).</p> <p>There are dozens of smaller improvements. The row page JSON representation <a href="https://docs.datasette.io/en/latest/json_api.html#json-api-row">finally has a design suitable for 1.0</a>. A new <code>&lt;datasette-copy-button&gt;</code> Web Component provides copy-schema in various places. JSON in columns on the row page can be pretty-printed. There's a new <a href="https://docs.datasette.io/en/latest/javascript_plugins.html#javascript-plugins-modal-confirm">DatasetteModal.confirm()</a> JavaScript utility method. Plus dozens of bug fixes, including contributions by <a href="https://github.com/PDGGK">Zihan Dai</a>, <a href="https://github.com/peterolkhov">Peter Olkhovets</a>, <a href="https://github.com/IlyaKhalafi">Ilya Khalafi</a>, and <a href="https://github.com/AnkitaAdvitot">Ankita Advitot</a>.</p> <p>The feature I'm most excited about is a pretty tiny one: I've added CSS multi-page view transitions to Datasette! <a href="https://github.com/simonw/datasette/commit/7da8a85ea77786edef975a4066bbadfab0966a0c">Here's the commit</a> - key CSS is as follows:</p> <div class="highlight highlight-source-css"><pre><span class="pl-c">/* Enable cross-document view transitions,</span> <span class="pl-c"> for anyone without prefers-reduced-motion */</span> <span class="pl-k">@media</span> <span class="pl-kos">(</span><span class="pl-c1">prefers-reduced-motion</span><span class="pl-kos">:</span> no-preference<span class="pl-kos">)</span> <span class="pl-kos">{</span> <span class="pl-k">@view-transition</span> <span class="pl-kos">{</span> <span class="pl-c1">navigation</span><span class="pl-kos">:</span> auto<span class="pl-kos">;</span> <span class="pl-kos">}</span> <span class="pl-kos">}</span> <span class="pl-c">/* .facet-results should morph from old position to new */</span> <span class="pl-ent">html</span><span class="pl-kos">:</span><span class="pl-c1">active-view-transition</span> <span class="pl-kos">.</span><span class="pl-c1">content</span> <span class="pl-c1">&gt;</span> <span class="pl-kos">.</span><span class="pl-c1">facet-results</span> <span class="pl-kos">{</span> <span class="pl-c1">view-transition-name</span><span class="pl-kos">:</span> facet-results<span class="pl-kos">;</span> <span class="pl-kos">}</span> <span class="pl-c">/* So should a .table-wrapper, provided it contains .rows-and-columns */</span> <span class="pl-ent">html</span><span class="pl-kos">:</span><span class="pl-c1">active-view-transition</span> <span class="pl-kos">.</span><span class="pl-c1">table-wrapper</span><span class="pl-kos">:</span><span class="pl-c1">has</span><span class="pl-kos">(</span><span class="pl-c1">&gt;</span> <span class="pl-ent">table</span><span class="pl-kos">.</span><span class="pl-c1">rows-and-columns</span><span class="pl-kos">)</span> <span class="pl-kos">{</span> <span class="pl-c1">view-transition-name</span><span class="pl-kos">:</span> rows-table<span class="pl-kos">;</span> <span class="pl-kos">}</span></pre></div> <p>With that in place, navigations within Datasette - still staunchly a multi-page application - feel more like an SPA. This works on Safari and Chrome at the moment and I believe supports for Firefox is due to land in <a href="https://developer.mozilla.org/en-US/docs/Mozilla/Firefox/Releases/162#changes_for_web_developers">Firefox 162</a> on December 8th 2026 (seeing as they are no longer feature-flagged in Firefox Nightly.)</p> <video controls playsinline preload="none" width="930" height="720" poster="https://static.simonwillison.net/static/cors-allow/2026/css-transitions-datasette.webp"> <source src="https://static.simonwillison.net/static/cors-allow/2026/css-transitions-datasette.mp4" type="video/mp4"> </video> <p>Tags: <a href="https://simonwillison.net/tags/security">security</a>, <a href="https://simonwillison.net/tags/datasette">datasette</a>, <a href="https://simonwillison.net/tags/css">css</a>, <a href="https://simonwillison.net/tags/browsers">browsers</a>, <a href="https://simonwillison.net/tags/annotated-release-notes">annotated-release-notes</a>, <a href="https://simonwillison.net/tags/firefox">firefox</a>, <a href="https://simonwillison.net/tags/javascript">javascript</a>, <a href="https://simonwillison.net/tags/safari">safari</a>, <a href="https://simonwillison.net/tags/chrome">chrome</a></p>

  • </p> <p>There are dozens of smaller improvements.
Nguyên văn nguồn
Chi tiết nguồn
Tác tử

Show HN: Open-source multi-player artifacts that update themselves

Rally is a multi-player platform to build reactive micro-applications together that connects event streams to agents that then update the artifact. We've had pilots build a PR dashboard that sorts by risk assessment (an agent actually reads the code) per PR, a micro-CRM that pulls in live deals from multiple Gmail inboxes, and a shared grocery list that groups items by store aisle. We're similar to Muse's Spaces, Google's TableTop, and the standard artifacts most harnesses now support. Our differentiator is being multi-player so groups can iterate together, artifacts calling agents on incoming events so updates can be more sophisticated than an API call, and giving artifacts a full application stack. Under the hood, each artifact is a Cloudflare Durable Object with its own SQLite database, compute, and storage. As an example this is our podcast artifact that helps us keep up-to-date on what happened on the repo each day https://connected-tidy-crystalline-haldane-g.rallyuserconten... GitHub: https://github.com/juno-ai-labs/avalon

Nguyên văn nguồn
Chi tiết nguồn
Huấn luyện

Open-source AI without Jensen

Câu hỏi I mean, he was the big champion, right?

The resignation of Nvidia CEO Jensen Huang is a major setback for the US development of an open model ecosystem. Nvidia still has the financial interests in supporting open models, which are to avoid concentration in power among a few customers who would develop their own chips and reduce Nvidia’s profits, but supporting the open ecosystem is a tricky puzzle more than an engineering problem to grind out a solution on. Nvidia’s business operations are famously extremely dependent on Jensen as an operator, so it is a consensus opinion that there are challenges ahead (or even more high-profile departures through chip-smuggling investigations or plain-old sharks circling more poachable talent). Nvidia’s number one goal is to stabilize the ship, deliver to customers in the next 24 months, and prevent implosion. Open models have been one of Jensen’s more ambitious, founder-style bets. He has been flooding the zone with funding in a way that will very likely dry up. Interconnects AI is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber. In this period, the open model economy is in its nascency. These are crucial years for establishing use-cases, revenue flow, and motives to continue training the next generation models. I discussed the sensitivity of these early years in my recent talk at Fireworks Forge ( slides , recording ). The core of the argument is one upside for open models, a reason to be optimistic, and one downside. The positive case is that the open model economy has correctly latched onto “ownership of intelligence” and enterprise post-training as a crucial lane. What is to be determined is exactly how big the downstream market is. The downside is that the true long-term benefit of open models, as truly commoditized intelligence, is very far from realization with compute in such short supply. In the near term, where inference startups are also struggling to get enough chips (because Anthropic and OpenAI are buying so much capacity), there’s little margin competition on open-weight models. If open-source models are also served at giant inference margins, as we learned with DeepSeek’s DeepInfer release, they’re far from a commodity. In this paradigm, the closed labs with better economies of scale and integration from model to datacenter and chips will win. I expanded on this challenge in my recent piece on the open model economy . Together, a lot still needs to go right for open models to be a meaningful economic engine. Jensen’s support, and protection, gave the ecosystem time to figure this out. Jensen was more than a cheerleader. He would stop a lot of shit from going wrong once early winds were stirring. He was willing to use his power on open models. There’s a major swell in Democratic government funding towards candidates that want to regulate the AI industry, fostering “safety” . Many of these people feel a need to do something and have general opinions that open models are unsafe, while not being educated on the complexities of their benefits. My biggest near-term worry is sloppy regulation at the state level, followed by national regulation. We’ve been one, strong Chinese open-weight model away from a crisis since this summer. It still holds – especially as we enter a new, turbulent equilibrium of cybersecurity. We have not seen a major hack driven by a popular Chinese open-weight model, but it is a matter of when, not if. We need someone to encourage caution and an ecosystem perspective when the “open-weight hawks” come hunting. Knocking over the open-source AI domino today is easy — if done, unwinding the downstream concentration of power, reduction in safety research, and stifling of long-term innovation is far harder. On public data, it looks like balancing long-term innovation versus short term risks, but this is due to how horribly hard it is to measure the open model economy (e.g. the most valuable aspects of it run on private infrastructure). It’s a recurring challenge facing the ecosystem and my work. As a brief aside, the extent to which the media ecosystem is revolving around timing today, rather than exact availability of information, is incredible. All of these Jensen affairs were known, but getting them out at the right time, right after the midterm political swing, took down the king. The same goes for the Jacob Coxon brouhaha this summer. All of that safety culture around top labs was well known, but the right timing around datacenters and the Overton window made it a national issue. This could very well happen at some point for open model safety, as it tends to be harm, risk, or safety-related narratives that drive the modern media algorithms. I cannot imagine my work on open-science driving such a media cycle, because it will never happen. Share The only lasting solution to this is to incubate a truly frontier open model player in the US. For example, I agree completely with Dylan Patel, CEO of SemiAnalysis, on a core problem in the US open model ecosystem. From a recent ChinaTalk + SemiAnalysis emergency podcast : Jordan Schneider: I do want your guys’ take on just how bad this is, that Jensen’s now off the table for the future of American innovation. And we should talk about open models. I mean, he was the big champion, right? Dylan, you were talking earlier — I still don’t understand quite why someone didn’t just spend a few billion dollars and make something better than Kimi or DeepSeek and open source it. Dylan Patel: Not that easy. It’s not money, it’s talent, man. All of the most talented people primarily work at Anthropic, OpenAI, DeepMind, and Meta TBD, right? And then those who don’t are split across a hundred different startups. I mean, yes, Thinking Machines has a cracked team, but is Thinking Machines as good as Moonshot? No. Moonshot skill-wise is probably just as good as the US labs. They just have a big deficit in compute and deployment. So I think if we don’t have a scale adv

  • Nvidia still has the financial interests in supporting open models, which are to avoid concentration in power among a few customers who would develop their own chips and reduce Nvidia’s profits, but supporting the open ecosystem is a tricky puzzle more than an engineering problem to grind out a solution on.
  • Knocking over the open-source AI domino today is easy — if done, unwinding the downstream concentration of power, reduction in safety research, and stifling of long-term innovation is far harder.
Nguyên văn nguồn
Chi tiết nguồn
121–140 / 562
Trang 7 / 29