Anthropic has formed a 15-person chip design team, utilizing the Claude model to conduct custom chip engineering and design work, said to shorten the time required for chip design.
MarkTechPostSame story · 2 outlets5555NVIDIA boosts the inference throughput of the 124B parameter Ling-3.0-flash model on DGX Spark by 42.5% and speed by 3.5x with a new MoE backend, FP8 output layer, and built-in speculative decoding, without compromising accuracy. This optimization is suitable for single-user desktop systems, improving throughput from 37.81 tokens/s to 53.89 tokens/s and reducing latency by 24.5% on 1,319 GSM8K problem tests.
How the accounts differ:NVIDIA Technical Blog states decoding is the memory bottleneck due to MoE computation and output layer projection, while MarkTechPost highlights the 42.5% throughput increase and 3.5x speedup due to new MoE backend, FP8 output layer, and speculative decoding.
Hugging Face Daily Papers6565CrossSWE benchmark tests the cross-repository consistency of large language models in cross-repository collaborative software development, including 171 instances, covering 165 repositories, 20 technical fields, and 13 programming languages. The evaluation shows that the strongest GPT-6.1 Sol only achieves 58.48% Pass@1, with cross-repository inconsistency being the main failure mode.
Hugging Face Daily Papers3838The IdeaTaste benchmark shows that AI agents are similar to human research communities in citing literature and research directions, but there is a significant gap in specific methods and expected success, with citation matching reaching 26% of human levels and idea matching only 18%.
Databricks Blog3838Claude platform natively exports OpenTelemetry data to Databricks MLflow. Developers can send native tracing data via Anthropic SDKs, Messages API, and Managed Agents, leveraging MLflow to debug agents and evaluate changes.
Hugging Face Daily Papers6565Order-Only On-Policy Distillation (O^{3}PD) trains student models to match the token ordering of the teacher model rather than probabilities, averaging 2.9 points better than OPD on HMMT25, AIME24, and AIME25, with a 5.6 point improvement on AIME25. This technique requires no temperature tuning and is unaffected by the teacher model's temperature.
The Decoder5858Stanford researchers tested 50 widely stated claims about LLMs, finding that 44% have expired, with an average lifespan of only 15.7 months. Claims related to reasoning abilities (such as counting, arithmetic, long context information loss) have all become invalid, while claims related to flattery, deception, and strategic behavior persist or even enhance.
Cohere BlogCohere has integrated third-party open-source models into Model Vault and Model Vault Encrypted, allowing enterprises to run Cohere models and mainstream open-source models like DeepSeek V4.1 Flash and GLM 5.3 Flash on a controlled and encrypted inference platform, with default implementation of zero-shot prompting and output data retention.
NVIDIA claims AI is driving semiconductor foundries into the 'AI-defined factory' phase, accelerating simulation technology to compress engineering simulation from hours and days to just minutes. Applied Materials' Ginestra platform's shape simulation accelerates up to 35 times, reducing engineer decision time from days to minutes.
Meta Engineering3434Meta launches F3 Lightning to address bottlenecks in F3 graph publishing by decoupling feature publishing from execution, enabling independent validation deployment. It adopts optimizations such as stable compute identifiers, a single execution plan, shared memory buffers, and dense index mapping, making the production speed of extraction features approximately 3.6 times faster than F3 without increasing infrastructure compute costs.
OpenAI Agents SDK releases4040OpenAI Agents SDK v0.24.0-alpha.6 released, supporting submission of work through captured Turn control, passive event handling, manual request handling, redirection and cancellation, and extending local storage session state recovery and tool policy features. New features include local storage, native output, experimental external job recovery, and tool selection, while updating the API, such as using Agent.run() instead of Runner, and unified FunctionTool callback and manual tool.
Google Developers Blog5858TypeSafe's Jev classifier can route tasks to the most cost-effective Gemini models, with 50 question tests showing an accuracy of 96%, better than 93.8% of Python regex heuristic methods; setting the inference model MaxOutputTokens limit does not save costs but may interrupt complex task inference.
Hugging Face Blog5050StateBench is a long-context state tracking benchmark that tests how language models maintain tracking changes in entity information within long documents. The benchmark includes four difficulty levels, ranging from simple immutable assignments to complex dependency assignments. It found that cyclic and hybrid models generalize better than Transformers on direct training tasks, but cutting-edge models like Gemini 3.8 Flash, Claude Opus 5.5, and GPT-6 Sol still fail in complex state tracking with long inputs.
AWS Machine Learning Blog2828Amazon Bedrock now supports inference for Kimi K3 within India, meeting strict data localization requirements. Customer Ultrahuman enables their AI health assistant to handle over 14,000 daily conversations through API calls, while keeping health data within AWS, reducing costs by 80% and improving multi-turn conversation speed by 85%.
AWS Machine Learning Blog3838Cartesia uses state space models (SSM) designed for real-time speech AI and AWS infrastructure to achieve sub-100ms text-to-speech and speech-to-text, supporting voice cloning for 44 languages, suitable for scenarios such as virtual assistants and customer service centers. The Sonic model generates the first audio byte in about 90ms, and the Ink model has native turn detection and noise robustness. The API interface supports real-time streaming output.
AWS Machine Learning Blog3434Run Strands Agents on Amazon Bedrock AgentCore using EDB Postgres AI to answer C360 (Customer 360) questions by querying the semantic layer.
AWS Machine Learning Blog2424Gen Digital developed an AI curation news engine for Norton Revamp platform, implementing real-time content grounding through Amazon Bedrock AgentCore Web Search.
AWS Machine Learning Blog3232Amazon Quick integrates structured and unstructured data by connecting Amazon S3 and creating a knowledge base, enabling users to query with a single conversational AI agent. This solution includes a data layer (unstructured content stored in S3 and structured data as semantic datasets), a workspace layer (integrating data types into a single workspace), and an agent layer (custom chat agents linked to the workspace and returning combined responses).
Simon WillisonQwen 3.8 27B can also explain model weights on my laptop. The model successfully identified the Breakout, Pong, and Breakout games, respectively taking 46 minutes 12 seconds, 1 hour 4 minutes 54 seconds, and 52 minutes 25 seconds, with inference tokens of 43,506, 19,493, and 10,884, and confidence levels of 72%, 75%, and 88%.
NVIDIA Newsroom3838AI is changing the way semiconductor wafer factories work, accelerating decision-making for engineers. NVIDIA's cuDSS library accelerates COMSOL's multiphysics simulation 10 times, Applied Materials' Ginestra platform simulation 10 times, ACE+ platform 35 times, and Siemens' Calibre lithography software simulation 17 times. NVIDIA's PhysicsNeMo framework enables Lam Research to predict plasma behavior at the millisecond level, replacing full simulations that took days.
Hugging Face Daily Papers6565Research shows that the latent prediction error in video world models reflects appearance and motion rather than reliable measurements of physical anomalies. Comparisons of latent prediction, masked pixel reconstruction, and simple visual statistics on the IntPhys2 benchmark show that appearance is easy to decode, and V-JEPA2 only slightly reveals whether an event is possible.
Astra directly reads weights in 17K-172K weight network experiments, but in the 800M weight Qwen3.5 mathematical ability improvement experiment, it requires combining Python programming and model execution, not just reading weights; Astra modifying 0.016% weights increased Qwen3.5-0.8B's MATH-500 from 7.9% to 48.1% and tripled IFEval.
The DecoderAnthropic launched the Message Threads feature on the Claude platform as a public beta. This feature only transmits changes since the last message, reducing the transfer size from 500,000 to 1.
Hacker Times is laid out in a newspaper style to present Hacker News stories, linking to the original sources and updating every hour. The project, involving Andrej Karpathy, aims to provide a reading experience different from the traditional browsing experience.
Lenovo Kaitea introduces its first domestic-innovation AI mainframe, equipped with Zhixin KX-7000 processor and HIMO Zhi Neng Qiang LQ50 NPU, providing 160TOPS local AI computing power, with a maximum support of 320 TOPS, capable of running 35B parameter large models. The machine is pre-installed with Tianxi AI Pro intelligent agent, supporting 7×24 digital employee services.
PyTorch BlogPyTorchKR and XCENA co-hosted KVCache Meetup Korea on September 30, 2026, where speakers from XCENA, SK hynix, FuriosaAI, and LMCache shared their achievements in LLM inference and KV cache optimization.
Hugging Face Daily Papers5858CoworkOS enables human-machine collaboration and continuously improves model performance by separating strategies from execution, driving human-machine collaboration towards real-world automation. The framework performs excellently in 12 benchmarks and 6 categories of real-world applications, using GLM-5.3 as the core, and performs RSI-driven automatic post-training on Qwen3.6-35B-A3B, increasing the average score on 8 benchmarks from 33.19 to 42.95.
Ai2 BlogStateBench is a benchmark for testing large language models' ability to track entity state changes in long documents, including four difficulty levels: R-trivial, non-periodic, solvable, and general finite. Research finds that recurrent and hybrid models generalize far better than Transformers when directly trained on these tasks, while cutting-edge models like Gemini 3.8 Flash, Claude Opus 5.5, and GPT-6 Sol still struggle with more complex state tracking issues when input lengths vary.
NVIDIA NewsroomNVIDIA and partners are expanding physical AI and agent AI to global cities, supporting municipal departments, neighboring towns, and local developers to build services without requiring each organization to purchase and manage resources separately.
MarkTechPost6262OpenAI has launched Pando, a Git backend that reduces median push latency by about 80% and processes nearly 1 PB of Git data daily. Pando uses branch-level causal consistency, stores reusable data packages, and achieves high concurrency processing using only a few CPU cores, but it has not yet been open-sourced.
Hugging Face BlogTransformers supports over 750 model types, its test suite includes 88,000+ CPU tests and 200,000+ GPU tests. Hugging Face releases Model Integration Contracts (MIC) standard, including compressed models (such as a 66 MB weight UMT5 fix), runnable examples (such as SmolLM text generation), and human review baselines.
Anthropic ResearchFeaturedSame story · 2 outlets6262Anthropic announced the verifiable reasoning prototype FlashProve, using zkSNARK to prove that the response indeed comes from the committed model weights. For the 400B parameter Llama 4 Maverick (4-bit quantization), a single H100 generates proofs at approximately 110 tokens/second, with proofs for each 1,024 token request being about 19 MB, and verification on a 24-core CPU taking about 17 seconds.
How the accounts differ:MarkTechPost states the proof volume is 18.78 MB; The Decoder states it is about 19 MB.
Why it matters: Anthropic has released the measured data and Lean formal proof of the verifiable reasoning prototype FlashProve, allowing for a cross-comparison with existing zkML solutions.
Xiaomi humanoid robot completes 8-hour work tasks in a factory, achieving exceptional autonomous correction and production line collaboration; the embodied base model Xiaomi-Robotics-1 builds an Agent system, completes long-range task verification for over 30 minutes in Xiaomi Youth Apartments, covering 50+ tasks.
Hugging Face Daily Papers4747The uni-robot-agent framework enables open-source weight robots to surpass proprietary models using agents in zero-shot operation, with its onboard Qwen3.8-Flash-Next scoring 31.39 in the RoboDojo evaluation, exceeding GPT-6 Astra-RoboProbe's 28.97 score and numerous VLA/WAM baselines, and performing excellently in 19 real tasks. Analysis indicates that effective robot use requires combining four general capabilities, suggesting future training should synergistically develop these capabilities.
MIT's experiment allowed GPT-6 Astra to directly read and edit the weights of Qwen3.5-0.8B, changing only 0.016% of the parameters, without training and gradient descent, MATH-500 score improved from 7.9% to 48.1%, and instruction execution ability also increased 3 times.
Why it matters: MIT experiments show that cutting-edge models can directly read and write the weights of smaller models, providing a reproducible path for performance repair of gradient-free training.
Qualcomm AI Hub opens to Chinese developers, integrating Models, Workbench, Apps, and GenieX framework, providing over 230 pre-optimized model libraries and cloud tuning tools. Developers can obtain support through the official website, device farm, and forum to accelerate terminal-side generative AI and intelligent agent AI development.
MarkTechPostZyphra AI has released Zyphra Hierarchical Attention (ZHA), a sparse attention method based on a tree indexer, designed to address the known issue in sparse attention where the indexer still needs to scan the entire context. ZHA can be distilled into a frozen pre-trained model of about 4B tokens, achieving performance close to dense attention in inference, smart encoding, and 256K retrieval benchmarks on Qwen3.6-35B-A3B.
MIT CSAIL researchers propose 'weight space programming', allowing GPT-6 Astra to directly read and modify the neural network weights of Qwen3.5-0.8B, adjusting only about 0.016% (131072 parameters) to increase MATH-500 score from 7.9% to 48.1%, GSM8K from 13.8% to 67.1%, and IFEval from 21.0% to 62.3%.
Aisin and Northeastern University have collaborated to develop the world's first 16nm AI SoC integrated with STT-MRAM, using a near-memory computing design to achieve 50 times the target detection efficiency. This SoC uses an 8/16-bit mixed multiplication architecture and weight sharing technology to complete target detection using Tiny YOLOv3 within 32MB of on-chip STT-MRAM, reducing startup time by 38.59 times and increasing AI processing speed by 10.16 times.
