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

NVIDIA Expands Slurm to Bring Quantum Computing Into Supercomputing

Quantum processor deployments at supercomputing centers have grown nearly sevenfold over the past five years, with nearly 30 supercomputing centers around the world now in the process of deploying over 50 quantum processors. Many of these centers use Slurm to run scientific supercomputing workloads, so expanding its support for quantum processors builds on familiar workflows and tools, enabling computational chemists, nuclear physicists and materials scientists to bring quantum computing into their research domains without friction. In a new release, Slurm builds on previous work with IBM, Pasqal and the wider quantum computing community to allow QPU builders to create a single interface to their hardware for every supercomputing center, through a simple file. This makes it easier than ever for HPC centers and QPU builders to integrate and serve their hardware. The expansion is already being adopted by Alice & Bob , IBM , IQM , Pasqal , Quantinuum and MQSC and supports the widely used QRMI and QDMI interfaces. Read All NVIDIA Announcements From SC26 Beating the Bottleneck Until now, integrating a QPU into a supercomputing center meant a new software challenge for every new QPU or supercomputing center. By building on existing quantum-HPC interfaces such as QRMI and QDMI, the Slurm expansion allows QPU builders to reuse their integrations across all their deployments. It also means that supercomputing centers can add new quantum processors to their existing scheduling and resource-management workflows without developing bespoke QPU-specific solutions. For quantum computing to advance science, it needs to be used by scientists. Now scientists can run quantum-classical applications developed with platforms such as NVIDIA CUDA-Q with Slurm coordinating access to all the processing they need. Slurm 26.11 will be available later this month. Orchestrating Resources in an Open Quantum-GPU Platform The Slurm expansion adds a community-developed orchestration layer to NVIDIA’s open platform for quantum-GPU supercomputing, which centers around the CUDA-Q open platform for hybrid quantum-classical computing. At SC26, new releases expand all aspects of the platform: CUDA-Q QEC allows researchers developing quantum error correction to rapidly prototype and test new ideas, including with a new decoder server . A new TCBP-JOBS decoder is 13,000x faster than the industry standard. CUDA-Q gets an expanded layer of fault-tolerant primitives that application builders can use to realistically compile their ideas. Fermilab has already used CUDA-Q to perform studies in three weeks that would normally have taken five months. CUDA-Q Algorithms provides reusable components for builders of quantum applications, including new capabilities in helping design the loading data into quantum circuits and tools specifically for chemistry and materials applications CUDA-Q Pulse gives quantum hardware builders a layer for shaping the ‘pulses’ needed to control qubits on a QPU, including tools to GPU accelerate their generation. The quantum ecosystem around the world is continuing to build with this platform. Quantum Motion integrated its QUICHE chemistry framework with CUDA-Q. Altera and Oxford Quantum Circuits designed an FPGA-based controller connecting qubit control and readout to GPUs through NVQLink. At supercomputing scale, QunaSys used cuQuantum and 1,024 NVIDIA Hopper GPUs on ABCI-Q to simulate more than 300 million quantum operations, reaching chemical accuracy for a four-hydrogen-atom model. NVIDIA researchers used JUPITER to simulate 104 qubits under realistic noise, testing fault-tolerant designs before they’re built. Learn more about NVIDIA quantum computing technologies . Read All NVIDIA Announcements From SC26

Nguyên văn nguồn
Chi tiết nguồn
Huấn luyện

CUDA at 20: How the Platform Transformed Scientific Computing

This week at SC26, NVIDIA is marking two decades of CUDA by honoring the researchers, developers and engineers who helped make GPU computing a foundation for modern science and AI. CUDA’s 20th anniversary presents an occasion to reflect across computing eras, from the people who wrote the very first CUDA programs to the AI developers building on their work today. At the platform’s core is this developer community whose experiments, tools and persistence expanded what accelerated computing could do. From SC06 to SC26 NVIDIA announced CUDA in November 2006 and brought it to the scientific community at SC06, giving scientists and developers a familiar, high-level programming model for using GPUs. It fundamentally extended accelerated computing beyond graphics use cases to accelerate demanding applications for science, simulation and more. In the two decades since, developers have used CUDA to advance Nobel Prize-winning AI research , accelerate scientific applications on supercomputers and support visual effects in Oscar-winning films . Today, more than 11 million developers build on CUDA across the computing stack. Read All NVIDIA Announcements From SC26 Honoring the Architects of Progress Now, at SC26, NVIDIA is looking back at some of the CUDA luminaries who’ve spearheaded breakthroughs in science and AI: At the University of Illinois Urbana-Champaign, biophysicist Klaus Schulten’s team helped connect GPU computing with molecular biology through Nanoscale Molecular Dynamics and Visual Molecular Dynamics , tools for molecular simulation and visualization. In 2013, researchers at Illinois, the University of Pittsburgh and Vanderbilt University used CUDA-accelerated tools to help resolve the HIV capsid’s atomic structure. They combined experiments with simulations on Blue Waters , one of the world’s most powerful supercomputers at the time. In a critical breakthrough in AI, University of Toronto researchers Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton used CUDA to train AlexNet on just two NVIDIA GPUs. Their image-recognition system won the 2012 ImageNet competition , helping demonstrate the potential of GPU computing for deep learning. That connection between computing and scientific discovery is seen with AlphaFold, which predicts protein structures. Demis Hassabis and John Jumper shared the 2024 Nobel Prize in Chemistry for protein structure prediction. Their team’s publicly released AlphaFold 2 software uses CUDA for GPU execution . These are just a few of many examples that show how one generation’s work creates opportunities for the next. With 1,000+ NVIDIA CUDA-X libraries, developers are now reshaping industries and giving machines the power to see, learn and reason — from discovering new medicines and designing next-generation factories to advancing autonomous vehicles, modeling financial markets and unlocking quantum computing — and more. Start building with the CUDA Toolkit and visit the 20-Year CUDA Celebration Event Wall at NVIDIA’s booth at SC26 . Read All NVIDIA Announcements From SC26

Nguyên văn nguồn
Chi tiết nguồn
Tác tử

NVIDIA Vera Rubin NVL72 Leads Agentic Inference Performance on MLPerf Endpoints v1.0

Agentic workloads are demanding more from inference systems. AI agents carry growing conversation histories across turns and generate a mix of reasoning, tool... Agentic workloads are demanding more from inference systems. AI agents carry growing conversation histories across turns and generate a mix of reasoning, tool calls, and responses. Inference systems must support users at scale while remaining responsive. MLPerf Endpoints v1.0 evaluates this balance between serving capacity and responsiveness using coding and enterprise workflows through API… Source

Nguyên văn nguồn
Chi tiết nguồn
Tác tử

v18.21.0

<h2>@oh-my-pi/pi-ai</h2> <h3>Added</h3> <ul> <li>Added automatic detection of thinking capabilities for previously unknown models, with persistent results and support for alternate effort and token-budget request formats.</li> </ul> <h2>@oh-my-pi/pi-catalog</h2> <h3>Added</h3> <ul> <li>Added Modal as a hosted inference provider.</li> </ul> <h3>Changed</h3> <ul> <li>Changed the local speech model from Kokoro to Phonon 31M.</li> <li>Improved reasoning capability detection by preserving unverified thinking-level support and using endpoint-specific probe results when available.</li> </ul> <h3>Fixed</h3> <ul> <li>Fixed xAI Grok OAuth model grok-composer-2.5-fast being incorrectly identified as non-reasoning; it now supports thinking levels from minimal through xhigh.</li> </ul> <h2>@oh-my-pi/pi-coding-agent</h2> <h3>Breaking Changes</h3> <ul> <li>Updated the <code>AgentSession</code> run lifecycle API: <code>agent_end</code> is emitted once when a run can no longer resume, and <code>agent_pause</code> reports runs that stop while awaiting work. Replaced <code>hasPendingAsyncWork()</code>, <code>settleAsyncWork()</code>, and <code>suppressAutoResumeUntilPrompt()</code> with <code>isRunActive</code>, <code>wakeSources</code>, and <code>waitForRunEnd()</code>. Calling <code>abort()</code> now also ends runs paused on background work; RPC, JSON print, collaboration, and extension streams retain their existing per-loop <code>agent_end</code> behavior.</li> </ul> <h3>Added</h3> <ul> <li>Added configurable automatic subagent prewalk through <code>task.agentPrewalkAuto</code>, allowing per-agent <code>on</code>/<code>off</code> settings or model patterns while preserving unconditional subagent prewalk.</li> <li>Unknown custom and discovered models now automatically detect reasoning support and remember supported reasoning effort levels for future sessions.</li> <li>Added inline image and web-page replies: supported terminals can display local, data, and HTTP(S) images and embedded pages, while other terminals receive linked fallbacks. Added <code>tui.renderEmbeds</code> to disable this rendering.</li> <li>Added interactive embedded pages in Tern replies, including click-to-visit and Escape-to-return navigation; common host, path, and <code>file:</code> URL forms are accepted.</li> <li>Bash redirects can now write to all writable internal URLs, including job cancellation endpoints and handler-backed <code>agent://</code>, <code>cfg://</code>, <code>ssh://</code>, and <code>xd://</code> URLs.</li> <li>Added image pasting to <code>ask</code> form notes and <code>ui.formDialog</code> options via <code>acceptImages</code>; pasted images are returned with <code>[Image #N]</code> markers and described for text-only models. Unsupported prompts now clearly reject image pastes.</li> </ul> <h3>Changed</h3> <ul> <li>Changed the default local text-to-speech model to Phonon 31M for faster synthesis, smaller downloads, and additional voices. Existing Kokoro voice settings migrate automatically; speaking rates now range from 0.8× to 1.3×.</li> <li>Tern <code>/btw</code> responses now appear as folded, interactive notes at the point where they were requested, with follow-up, copy, branch, inject, and stop actions; <code>/btw</code> without a prompt still opens the history view.</li> <li>Overlapping Bash calls now share shell state, including variables, functions, aliases, options, and traps, and background jobs started by those calls continue running independently.</li> <li>Extensions enabled or reconfigured through <code>extensions</code> or <code>disabledExtensions</code> now load into the active session without requiring a restart.</li> <li>Rich-text pastes made during approval steering or in dialogs now preserve Markdown formatting and images when inserted into the chat editor.</li> <li>Moving a task or resuming a session in another project now switches its agents, extensions, custom tools, and MCP servers to match the destination project without a restart.</li> </ul> <h3>Fixed</h3> <ul> <li>Fixed status indicators incorrectly showing the agent as idle while subagents or background work could still resume the run.</li> <li>Fixed subagents and security scans completing while child agents were still running or queued peer messages remained.</li> <li>Fixed conflict-region resolution when a selected side is empty or shares edge lines with surrounding code.</li> <li>MCP stdio server failures before a response now report the process exit code or terminating signal instead of only reporting closed stdout.</li> <li>Fixed prewalk handoffs so model and reasoning-effort changes are preserved, and automatic handoffs return to the planning model when automatic prewalk is changed or disabled.</li> </ul> <h2>@oh-my-pi/pi-mnemopi</h2> <h3>Changed</h3> <ul> <li>Improved file memory bank performance by reducing unnecessary disk synchronization during checkpointing, lowering per-turn I/O overhead for banks shared across multiple processes.</li> </ul> <h2>@oh-my-pi/pi-natives</h2> <h3>Added</h3> <ul> <li>Added the <code>PhononTts</code> binding for on-device Phonon text-to-speech.</li> <li>Added <code>Shell.adopt(fork)</code> to apply variables, functions, aliases, options, and traps from a completed fork back to its parent shell. When both shells set a name, the value from the command that finished last is retained.</li> </ul> <h3>Fixed</h3> <ul> <li>Fixed concurrent <code>commitCreate()</code> calls incorrectly overwriting each other’s commit subject and body.</li> </ul> <h2>@oh-my-pi/pi-tui</h2> <h3>Added</h3> <ul> <li>Added support for pasting images into <code>HtmlFormDialog</code> option notes when opened with <code>acceptImages</code>, including image attachment, removal, and access to retained images through <code>OptionNote.images</code>.</li> <li>Added live web-page and image rendering in assistant replies for supported terminals, including inline images and embedded page previews. Added the <code>/btw</code> transcript no

Nguyên văn nguồn
Chi tiết nguồn
  • v18.21.0 Can Bölük — Oh My Pi releases · 16 thg 11, 2026
Kỹ thuật

New Ask Maps features to help with holiday travel and prep

<img src="https://storage.googleapis.com/gweb-uniblog-publish-prod/images/Ask_Maps_November_herosocial.max-600x600.format-webp.webp">Plan holiday road trips, get route tips, find local appointments, and search your saved lists with new Ask Maps features.

Nguyên văn nguồn
Chi tiết nguồn
Đánh giá

The Huawei-DeepSeek Symbiosis

Right before China’s national day holidays, DeepSeek released Ascend support for TileLang , DeepGEMM , DeepEP , TileKernels , FlashMLA , and DeepSelect . Neither can do this alone. Although they have Pangu, Huawei has no frontier lab to prove their chip can train or infer models, DeepSeek has models but cannot buy the compute it needs. Each supplies exactly what the other lacks: Huawei provides the hardware and native software stack; DeepSeek brings the workloads and engineering needed to make that hardware useful for its models. September 30’s release of Ascend support in DeepSeek repositories is an artifact of this trade, where the exchange runs through the kernel-authoring language. Source: SemiAnalysis - Note that the diagram is a simulation. In 2028 or even 2027, DeepSeek will likely be primarily using Huawei’s stack, especially with the external and internal pressure to build a fully domestic chinese ecosystem. As we show here and in future articles, the Huawei ecosystem is already mature and well developed, with strong foundations from the Ascend time and DeepSeek engineers building further ontop. In this article, we’ll first examine Ascend 950PR’s architecture and benchmark its compute and memory performance. We then explore DeepSeek’s software releases, from TileLang’s compiler capabilities to hardware-specific kernel optimizations and rack-scale communication. Finally, we analyze DeepSeek-V4.1-Flash’s execution profile. Ascend 950 We deep-dived into CANN and the Ascend 950 architecture in our DeepSeek V4 Pro article . Since then, more interesting details have been revealed. Source: SemiAnalysis Starting from the chip, many variants of 950PR/DT have appeared with varying core counts, HBM stack height, and cache sizes. Some are likely either virtualised/simulated, or for testing only. This is because only a fraction of dies and HBM stacks meet the target spec, due to both outright defects and parametric shortfalls in power and performance, so Huawei bins both compute and memory into multiple SKUs and pairs weaker memory with weaker compute, rather than binning only compute as Nvidia does. Source: SemiAnalysis This isn’t new, 910B also has variants such as 910B1, 910B2, 910B3, 910B4. For the following microbenchmarks, we use an Ascend 950PR_9579, with 28 cube cores and 128 GiB HBM. root@cb3bb851f93f:/workspace# npu-smi info +-------------------------------------------------------------------------------------------------+ | npu-smi 25.7.rc1.6 Version: 25.7.rc1.6 | +--------+------------------+---------------+-----------------------------------------------------+ | NPU ID | Name | Health | Power(W) Temp(C) Hugepages-Usage(page) | | | | Bus-Id | NPU Util(%) Memory-Usage(MB) HBM-Usage(MB) | +========+==================+===============+=====================================================+ | 0 | Ascend950PR | OK | 203.3 56 0 / 0 | | | | 0000:12:00.0 | 0 0 / 0 5242 / 131072 | +========+==================+===============+=====================================================+ +---------------------------+---------------+-----------------------------------------------------+ | NPU ID | Process id | Process name | Process memory(MB) | +===========================+===============+=====================================================+ | No running processes found in NPU 0 | +===========================+===============+=====================================================+ Running point chases reading 16 KiB of integer payload per launch, it stays near 134ns through 18 blocks, rises to 182ns at 56, then jumps to 391ns at 64. This is consistent with a second wave beyond 56 Vector cores. Source: SemiAnalysis Then by running a scalar pointer-chase probe separately on every Cube and Vector core, recording load timings at matching addresses throughout a 128 MiB chain, we get the following graph. Each cell’s color represents the mean absolute difference between two cores' address profiles, producing two groups that match the decoded dies. Source: SemiAnalysis Going further, by timing a Vector core’s pointer chase while another core runs either memory traffic or register-only work; the extra slowdown reveals which subsystems share resources, supporting groups such as {0,2,4} and {1,3,5}. Mixed Cube/Vector launches identify each subsystem’s one Cube and two Vectors, which we draw together within those groups, with physical positions remaining schematic. Source: SemiAnalysis CPU within the NPU Each 950PR comes with a variable number of Linx816 AArch64 ELF AICPUs. The MIDR numbers imply these cores are likely the same as the middle cores in the Kirin 9000S/9010. These AICPU cores have instruction set: fp asimd evtstrm aes pmull sha1 sha2 crc32 atomics fphp asimdhp cpuid asimdrdm jscvt fcma lrcpc dcpop sha3 sm3 sm4 asimddp sha512 sve asimdfhm dit uscat ilrcpc flagm ssbs sb paca pacg dcpodp flagm2 frint svei8mm svebf16 i8mm bf16 dgh rng ecv For this SKU, there are 7 configured, 6 visible, and 5 executing AICPU cores. Other than device control it can provide CPU-operator support, such as KV metadata. The AICPU helps out for short work where D2H overhead take up most of the time. For DSA, a small scheduling step reads the current sequence lengths and assigns indexer work to accelerator cores. Cube/Vector cores then compute INT8 indexer scores and select the top 2,048 indices. These planning, index selection, and gather steps are ran on AICPUs in vLLM-ascend implementation, and we compare their performance to running on the host CPU. Source: SemiAnalysis Due to their limited compute performance, AICPU is only good for short work where overhead dominates, and work where data starts and is consumed at the NPU. Microbenchmarks Ascend 950PR is designed for prefill and recommendation. It gets lower memory bandwidth and core than its bigger brother the 950DT. The results below are not from DeepSeek’s contributions but from stock torch.matmul on ascend resolving to aclnnMatmul_MatMulV3Common_MatMulV3 . We compared

  • As we show here and in future articles, the Huawei ecosystem is already mature and well developed, with strong foundations from the Ascend time and DeepSeek engineers building further ontop.
Nguyên văn nguồn
Chi tiết nguồn
Huấn luyện

Solving a 31-year-old grand challenge in mathematics – the KLS conjecture

Microsoft researcher and collaborator combine new mathematical ideas with human-directed AI to resolve a central problem in high-dimensional geometry and probability At a glance Fundamental scientific breakthrough: Researchers resolved the Kannan-Lovász-Simonovits (KLS) conjecture, a 31-year-old problem widely regarded as one of the central open questions in high-dimensional geometry, with implications for statistics, optimization, algorithms, and computer science. A universal geometric principle: The result establishes a universal geometric principle for log-concave distributions, showing that simple flat cuts can capture their narrowest bottlenecks within a fixed factor, regardless of dimension. Human-AI collaboration: The breakthrough emerged from a human-directed agentic research process in which AI systems explored mathematical possibilities while researchers evaluated, pruned, and redirected the search. Independent confirmation: Multiple independent researchers have validated this research, identifying its central criterion as the key breakthrough, followed by a machine-checked formal verification. For decades, a fundamental question about the difficulty of computing in high-dimensional spaces has resisted proof. New research resolves this question, removing a major theoretical obstacle to understanding the efficiency of randomized algorithms used across science and computing. The Kannan–Lovász–Simonovits (KLS) conjecture was proposed in 1995. It asks whether, for a broad class of high-dimensional probability distributions, the best flat or hyperplane cut is always nearly as effective as the best cut of any shape. More precisely, KLS predicts that the two differ by no more than a universal constant, regardless of the number of dimensions. This would establish a dimension-free geometric principle with important consequences for high-dimensional geometry, probability, and theoretical computer science. Because the geometry of these cuts governs how quickly random walks can explore a distribution, KLS also plays a central role in the theory of high-dimensional sampling. Figure 1: An illustration of the KLS Conjecture (simplified) on regular tetrahedron. Since 1995, mathematicians have accumulated evidence that the conjecture was true, but a complete proof remained elusive. Now, two researchers, including Xinzhi Zhang from the Machine Learning and Optimization group at Microsoft Research, have proved the KLS conjecture (opens in new tab) . They have shown that the relationship holds with a fixed bound, regardless of how many dimensions a shape has. Before this work, the best known result (opens in new tab) , established in July 2026, allowed that bound to grow slowly as the number of dimensions increased. This work establishes a fundamental geometric principle: high-dimensional spaces are better behaved than previously known, and the obstacles to navigating them do not grow without bound as dimensions increase. The importance of the KLS conjecture was recognized across the field long before it was resolved. In its 2017-18 Annual Report, the Mathematical Sciences Research Institute (MSRI) (opens in new tab) called it “a central open problem in high-dimensional convex geometry.” The Simons Institute (opens in new tab) described it as “one of the most important open problems in convex geometry.” In the proceedings of the 2022 International Congress of Mathematicians (ICM) (opens in new tab) , Ronen Eldan noted its broad implications for high-dimensional convex and computational geometry. Quanta Magazine (opens in new tab) wrote that KLS “lies at the heart of many questions in statistics and computer science.” Technical details The authors began by improving the July 2026 result in order to make the mathematical bound more precise. Working with AI agents in a structured, human-directed process, they explored more than 100 candidate approaches, many involving lengthy calculations with higher-order tensors. This produced a new mathematical criterion that became the foundation of the final proof. The authors then discovered that the new criterion could be applied recursively. Each improved bound could be fed back into the same argument to produce a still tighter one. In the October 1, 2026 version of the paper, this reduced the dependence on dimension to the iterated logarithm. This function grows extraordinarily slowly, but it still grows with the dimension and therefore did not yet fully resolve the conjecture. The final step was to repeat the entire recursive refinement, rather than applying it only once. This idea was inspired by the authors’ earlier work on fast algorithms for linear programming. The authors proved that the small errors introduced at each stage shrink fast enough that their total remains bounded by a universal constant, independent of the dimension. This eliminated the dependence on the number of dimensions and completed the proof in the October 4, 2026 version of the paper. The recursive method, its repeated application, and the proof that the accumulated errors remain bounded are original contributions of the researchers. AI tools helped explore candidate approaches, carry out detailed derivations, prepare the manuscript, and perform internal proof checks. The researchers determined the mathematical direction, selected the promising ideas, assembled them into the final argument, and verified the proof. Independent confirmation On the same day that the authors posted their complete proof, a separate group of leading experts posted a full proof (opens in new tab) that they developed in parallel. Their work built upon the findings in the first version (opens in new tab) of the Microsoft paper, calling it “the crucial step in the proof.” A third proof (opens in new tab) which appeared separately on the same day relies on a closely related idea. In addition, a Lean 4 formalization (opens in new tab) was subsequently released, combining the key criterion from the first Micros

  • A universal geometric principle: The result establishes a universal geometric principle for log-concave distributions, showing that simple flat cuts can capture their narrowest bottlenecks within a fixed factor, regardless of dimension.
  • Before this work, the best known result (opens in new tab) , established in July 2026, allowed that bound to grow slowly as the number of dimensions increased.
  • ” Technical details The authors began by improving the July 2026 result in order to make the mathematical bound more precise.
Nguyên văn nguồn
Chi tiết nguồn
Tác tử

v18.20.3

<h2>@oh-my-pi/pi-ai</h2> <h3>Breaking Changes</h3> <ul> <li>Removed the public <code>isXdDevicePath</code> export from <code>providers/cursor-pi-args</code>. Ranged reads of <code>xd://</code> documents now use the standard line-selector behavior.</li> </ul> <h3>Fixed</h3> <ul> <li>Fixed rate-limit handling for Mistral and other OpenAI-compatible providers so requests with no available allowance fail immediately with an actionable error suggesting a model or plan change.</li> <li>Fixed Anthropic OAuth requests for pooled Claude accounts to report accurate device and account information, including after credentials are rotated during a request.</li> </ul> <h2>@oh-my-pi/pi-catalog</h2> <h3>Added</h3> <ul> <li>Added publicly available image-sizing helpers for determining model-delivered image dimensions and fitting images to a specified long edge before coordinate mapping.</li> </ul> <h3>Fixed</h3> <ul> <li>Fixed image inputs being dropped for DeepSeek V4 Flash, restoring consistent multimodal image support.</li> <li>Restricted reasoning-effort options for Mistral GLM 5.3, GLM 5.3 Flash, and Mistral Large 4 to the levels supported by those models.</li> </ul> <h2>@oh-my-pi/pi-coding-agent</h2> <h3>Breaking Changes</h3> <ul> <li>Renamed the browser <code>tab.clickAt()</code> option from <code>clickCount</code> to <code>count</code>, matching <code>tab.click()</code>.</li> <li>Browser viewport screenshot coordinates now use the displayed image coordinates when the screenshot is scaled for the model; pass <code>css: true</code> (or <code>{ x, y, css: true }</code> for drag points) to provide viewport CSS pixels explicitly.</li> <li>Added opt-in <code>mnemopi.recallRelevance</code> (<code>off</code>, <code>auto</code>, or <code>on</code>; default <code>off</code>) to filter automatic memory recall for task relevance using the <code>judge</code> role.</li> <li>Added path-scoped entries for <code>bash.patterns</code> and <code>bashInterceptor.patterns</code>, resolved per working directory. Project and <code>--config</code> overlays now prepend these rules to global rules instead of replacing them.</li> <li>Added <code>bash.allowedEnvVars</code> and per-rule <code>env</code> lists to permit explicitly named environment-variable prefixes on matching <code>allow</code> rules.</li> <li>Subagents now preserve <code>@upstream</code> provider pinning on inherited models.</li> <li>Added <code>--no-judge</code>, <code>--no-spotlight</code>, and <code>--offline</code> options to restrict auxiliary services to the active model's provider or disable cross-provider services.</li> <li>Added per-rule TTSR <code>repeatMode</code> frontmatter (<code>once</code>, <code>after-gap</code>, or <code>always</code>) to override the global repeat policy.</li> <li>Compound bash approval prompts now show each command segment separately, distinguishing allowed segments from those requiring approval.</li> <li>Removed the <code>seance</code> agent, <code>/seance</code> command, and the task tool's <code>sourceSession</code> parameter.</li> </ul> <h3>Added</h3> <ul> <li>Added <code>mnemopi.embeddingQueryPrefix</code> and <code>mnemopi.embeddingInputType</code> configuration for asymmetric and instruction-tuned embedding models, including providers that require an <code>input_type</code>.</li> </ul> <h3>Changed</h3> <ul> <li>Browser screenshots are delivered to the model in CSS pixels up to the supported unscaled size, making viewport screenshot coordinates directly usable with browser actions; scaling guidance remains for scaled full-page and element captures.</li> <li>Browser <code>waitForUrl</code>, <code>waitForResponse</code>, and <code>requests({ filter })</code> now support <code>*</code> glob patterns in URL strings, while other strings retain substring matching.</li> <li>The WebKit browser engine now automatically installs and updates to the newest stable ABKit release, with versioned installations that allow updates without waiting for existing browser hosts to exit.</li> <li>Improved browser responsiveness and recovery: unresponsive WebKit pages fail fast with actionable errors, navigation can recover busy pages, and browser operations provide clearer diagnostics for timeouts, closed tabs, unsupported event listeners, invalid routes, cookie data, and viewport changes.</li> <li>Pressing Delete at the end of the Ctrl+R search field removes the selected prompt from history.</li> <li><code>/copy</code> now works while viewing a subagent and copies from that subagent's transcript.</li> <li><code>find</code> now returns unranked literal keyword matches when all judge requests fail.</li> <li><code>take()</code> and <code>drop()</code> now reject nonexistent hunk units with a clear error.</li> <li><code>edit</code> rejection messages now suggest the file's current tag, and edits made before conflict resolution carry over to the resolved file.</li> <li><code>/marketplace</code> and marketplace install/uninstall commands now report picker and catalog-load errors without exiting the session.</li> <li>ACP sessions honor <code>mcp.startupTimeoutMs</code> for MCP servers declared by the client, and ACP form elicitations no longer repeat field labels unnecessarily.</li> <li>Task progress from every <code>task.steps</code> subagent step is now preserved, rather than only the final step's report.</li> <li>Pasting HTML into the composer now preserves spaces from whitespace-only inline elements and indentation inside <code>&lt;pre&gt;</code> blocks.</li> <li>Queued follow-ups edited with Alt+Up are restored as queued messages instead of being sent into the active turn.</li> </ul> <h3>Fixed</h3> <ul> <li>Fixed <code>ida</code> <code>exec</code> output on IDA 9.5+ by suppressing helper deprecation warnings and preserving the updated function metadata API.</li> <li>Expired Perplexity logins and cookies no longer return an error page as a search answer; the provider is reported as expired and the search chain continues.</li> <li>Fixed brow

Nguyên văn nguồn
Chi tiết nguồn
  • v18.20.3 Can Bölük — Oh My Pi releases · 16 thg 11, 2026
Huấn luyện

AI Day Seoul Brings Together Korea’s Industry Leaders, Startups and Developer Communities

Over two days at COEX, more than 2,000 attendees converged for AI Day Seoul, bringing together Korea’s industry leaders, startups and developer communities advancing physical AI and AI agents. At the plenary sessions, Deepu Talla, vice president of robotics and edge AI at NVIDIA, said Korea is “punching above its weight,” with Doosan Robotics, Hyundai Motor Group, LG Electronics and SK Group — as well as startups like Real World and Luda Robotics — using NVIDIA platforms to bring general-purpose robotics into the real world. Jonathan Cohen, vice president of applied research at NVIDIA, showed how enterprises can build agents they own and control. Breakout sessions featured speakers from Coupang Intelligent Cloud, Doosan Robotics, LG AI Research, LG Electronics, NAVER Cloud, SK Telecom and Upstage. “It was a genuinely enjoyable time to have candid conversations,” said Hyunho Yang, research engineer for omnimodal foundation models at SK Telecom. “Seeing the passion of fellow AI developers was a real source of inspiration for me.” Community leaders share how they build with open models and tools at AI Day Seoul. The Agentic AI 10x Bootcamp Community Day brought together Pseudo Lab, Python Korea, PyTorchKR, HAERAE, Hugging Face KREW, Graph User Group, Womxn Who Code Korea, Dnol (Datayanolja) and other developer communities. Five finalists from the Hangul Day hackathon, co-hosted with Fastcampus, built agents trained on the Korean language and culture with the OpenShell runtime. The winning team, FlyGate, received an NVIDIA DGX Spark system. More than 500 developers also took part in full-day NVIDIA Deep Learning Institute workshops and NVIDIA Certification exams in Korean. Workshops covered deploying and optimizing AI inference at scale, building end-to-end robotics workflows with the NVIDIA Isaac platform and developing secure AI agents. Certification included the new AI Model and Application Development exam. The Physical AI Startup Connect event included pitches from 10 NVIDIA Inception startups: AIM Intelligence, Avikus, GoLe Robotics, MOVIN, Neubility, ROAI, ROBROS, TESOLLO, TwelveLabs and XYZ. The AI for Science Forum brought together the Korea Institute of Science and Technology Information, Korean universities and NVIDIA Research. AI Day Seoul sessions will soon be available on NVIDIA On-Demand .

Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

Updating our coordinated vulnerability disclosure policy timelines

We’re shortening the disclosure window in our coordinated vulnerability disclosure (CVD) policy from 90 days to 30 days for high- and critical-severity issues, and to 45 days for all other issues.

Anthropic Research (unofficial RSS mirror) · Chưa ghi tác giả Mở bản gốc ↗
Nguyên văn nguồn
Chi tiết nguồn
301–320 / 680
Trang 16 / 34