The DecoderSame story · 4 outlets6565
OpenAI News3838OpenAI has released BioBasicsBench, a benchmark containing 100 tasks, to evaluate the ability of AI agents to interpret raw experimental data and handle daily wet experimental science.
MarkTechPost3838Zyphra has released ZYFA (Zyphra FlashAttention on AMD), a set of FP8 and BF16 pre-filled attention kernels optimized for the AMD Instinct MI355X GPU. On real model activations, its FP8 kernel is 1.16-1.76x faster than AMD's default AITER kernel library, and the FP8 error is 62-78% lower than AITER under 128-dimensional heads.
LangChain BlogFeatured7272LangChain introduces asynchronous sub-agents in Deep Agents, where a resident orchestrating agent interfaces with users, and the sub-agents execute long tasks in the background in parallel and report results upon completion.
Why it matters: The original text provides five management tools and notification mechanisms for asynchronous sub-agents, which can be used to determine how to set up the workflow for parallel long tasks.
Mastra Blog3838Mastra agents can use Ring signals to receive doorbell events, with each event arriving at the agent thread as a notification signal with configurable priority. Agents can retrieve snapshots or video clips from the Ring API and decide on actions when someone presses the doorbell, such as security agents detecting human motion, viewing snapshots, and marking visits requiring urgent attention, or delivery assistants notifying package delivery.
O'Reilly Radar2222Google has released the Gemini 4 Carbon for cybersecurity and multi-step tasks, French Mistral has announced the weights of its 100 billion parameter Large 4 model, and DeepMind has previewed the Genie 4 world model for simulated environments.
Claude Code releases2222Claude Code v2.1.319 released, automatic mode is changed when telemetry or feature flags are disabled, server-side checks are disabled; server-side checks enabled after LLM gateway need to set CLAUDE_CODE_AUTO_MODE_SERVER=1.
AWS Machine Learning Blog3838Taxbit Assist embedded in the Taxbit dashboard, running via Amazon Bedrock AgentCore, requiring user authentication. Intelligently searches for candidate paragraphs in the Amazon Bedrock knowledge base and rearranges them on the server side. Answers must be source attributed and role permission compliant, without providing tax advice. AgentCore manages conversation state in memory, supports regional data residency, and complies with ISO/IEC 42001 standards.
AWS Machine Learning Blog3838Greenlight Financial autonomously processes over 90% of 4,000 security risk cases each month. Its solution, Luna, runs on Amazon Bedrock AgentCore, which includes over 30 specialized agents. Through AgentCore's runtime and memory support for hierarchical multi-agent systems, it achieves second-level security event triage in seconds without human intervention.
Databricks BlogDatabricks Unity Gateway enables public sector organizations to centralize governance of AI, controlling access to and budgets for models, agents, and tools, while meeting compliance requirements such as GDPR and FedRAMP, and reducing costs through intelligent routing.
Google Developers Blog13% of agents are misled by inline TODO(bug/...) or FIXME(bug/...), with 83% of actions completely unguided by developers. 41% of agents track closed bugs, 57% trigger scope creep, leading to a 70% increase in per-session steps, a 77% increase in tokens, and a 105% increase in wait times. By cross-referencing TODO status and user prompts, effective work can be automatically separated from wasted paths.
OpenRouter Blog6262OpenRouter releases the new generation of Auto and Auto Code routers, performance improved 4 times, routing step speed increased 3 times (p90 0.65 seconds vs 1.96 seconds), Auto Code is optimized for coding, comparable to Claude Opus 5.5 in performance but cost reduced by 74%. The new router selects models by evaluating 11 request factors and combining real-time pricing, provider health status, and cache status, supporting any model set and governance strategy.
Modal Blog6565Snorkel uses Modal's GPU and CPU sandboxes to run millions of agent simulations per month, supporting various task evaluations. Through Modal's memory burst feature, Snorkel reduces sandbox costs by 20% and improves runtime reliability. Modal's serverless GPUs and shared/dedicated inference endpoints enable Snorkel to quickly deploy and test custom models.
Hugging Face Blog5858Hugging Face released ToyBlocks, a framework for robot learning and evaluation that provides a controlled simulation environment. The framework supports automatically generating tasks and training demonstrations, and includes metrics to measure changes in robot behavior over time, aiming to study how models improve with more training data and how to combine advanced reasoning with precise physical control.
TechCrunch · AI6565Today is a good day for Harmony OS adaptation progress to be behind, the Huawei development team specifically formed a small group to support; Today is a good day for iOS version to be launched to support sleep data viewing, energy score, personalized sleep suggestions and ecosystem device linkage, and open MCP for various AI to call, while also providing a 7-day no reason return policy.
Latent Space00Inkling 2 Small completes 17A-594B parameter size pre-training within 68 hours, surpassing Muse Spark; supports 12T token context window, users can adjust to 400K via Claude UI, but note that exceeding 300k may trigger workflow issues
Dr. Shang Zhe, General Manager of Midea's Whole-Home Intelligent, pointed out that the industry's bottleneck lies in heterogeneous device data silos, interaction effect gaps, and deployment barriers. Traditional API connection solutions are difficult to popularize due to the exponential increase in costs. Midea's self-developed family intelligence body MevoX achieves cross-brand collaboration by enabling AI Agents to directly communicate. It has completed vehicle-home integration with Chery iCAR, supports fuzzy intent reasoning and long-term memory, and has connected with ecological partners such as Huawei and BYD. As the first home appliance company to achieve A2A collaboration of family intelligence bodies, Midea is driving the industry to shift from hardware scale to Agent reasoning and cross-ecosystem collaboration efficiency competition.
Hugging Face Daily Papers6565SLA model aligns pre-trained VLA strategy with frozen 3D perception model spatially through continuous pre-training and post-training, using feature-dependent gating and probabilistic state dropout to enhance spatial understanding across different robot datasets while retaining the inference architecture.
Hugging Face BlogHugging Face Open ASR leaderboard adds Liva AI private datasets for Egyptian English and Nigerian English, containing 1.34 hours and 1.55 hours of dialogue speech, to enhance the representativeness and reliability of ASR evaluation. The leaderboard now supports more detailed accent classification based on the Kachru three-circle model, covering groups such as US and non-US, inner circle and outer circle/extended circle.
ByteDance Seed · Hugging Face new modelsCola2 is a continuous latent language model that combines dynamic compression VAE, block causal diffusion Transformer (DiT), and cross-segment causal token decoder, dynamically grouping text into semantic segments, representing each segment with a compact continuous latent, generating latent blocks and implementing them as discrete tokens.
Developer releases Mac app Window Sweaters, adds knitted borders to 37 apps, color inspiration from app icons, supports custom styles and modes, installable via Homebrew or manual installation, requires macOS 13 or later.
Hugging Face Daily Papers6565BestRand confirms high-probability strong performance of random sampling in benchmark tests through analysis of error distribution, but exists rare large errors leading to instability. The method is validated in over 70 text, multimodal, and agent benchmark tests, reducing high error risk by 50.2% compared to competing methods, decreasing average estimated error by 16.4%, while requiring only 51.3% of the construction cost, and averaging a 21.8% reduction in OOD estimated error.
Tencent announced on November 5 that due to the surge in Hy4 calls, WorkBuddy has been expanded tenfold and the free experience quota for users has been increased. Hy4 ranks second among Chinese models on the Artificial Analysis Intelligence Index and is ranked 13th on the Agent Arena overall list, ranking first among open-source models.
Why it matters: After the release of Tencent Hy4, calls surged and were expanded tenfold, while also providing a comparison of the rankings in Artificial Analysis and Agent Arena and the cost per single task.
Microsoft's Scott Hanselman, VP of GitHub, open-sources GPU monitoring tool gputaskman, displaying GPU usage rate, memory usage, temperature, and identifying LLM processes to show model details, supporting Windows, Linux, WSL, and Apple Silicon Macs.
Hugging Face Daily PapersWeLM-Image is a unified image base model that integrates multimodal understanding, supporting text-to-image generation, instruction-guided editing, and multi-reference synthesis.
Hugging Face Daily Papers6565TailSteer, by maintaining early denoising dynamics to approach the base model and learning fine deviations in trajectory tails from within the policy through rolling out, significantly reduces accumulated visual drift in multi-round editing while preserving pre-trained generation and editing capabilities. Experiments show that TailSteer can greatly reduce drift on FLUX, SenseNova, and Qwen series models in 15-20 rounds of editing and largely maintains the performance of the base model on large text-to-image and single-image editing benchmarks.
Hugging Face Daily PapersVFlash achieves fast and accurate visual planning through Flash Primitive decoding, with its lightweight renderer generating the entire primitive in a single forward pass, achieving about 25 times speedup in 3D point motion prediction tasks and reaching an average success rate of 70.2% on MolmoSpaces' four task suites.
Hugging Face Daily Papers3535Harness-S achieves general intelligent agent training by learning from its sibling Harness. The model scores 82.5 on the MMLU benchmark, 3.2 times faster than baseline models, with a context length of 128k tokens and a parameter size of 175B. The model is open-source and can be accessed via API calls.
webAI Launches TwIL-8B Inference Model, Optimized for Math and Code Tasks Based on IBM Granite-4.2-8B, Scores Better than Gemma 4 31B in MATH-500 and AIME 2025 Tests, Best in LiveCodeBench Group, IFBench on Par with Gemma 4 12B. Model Inference Capability Reaches 88~93% of Frontier Models, Reduces AI Usage Costs but Has Limited Knowledge Base.
Taobao MNN team open-sources 0.3B mobile model MNN-Pocket-0.3B, achieving 350 tokens/s on real iPhone 18 Pro offline. The model's peak memory is below 300MB, decoding throughput exceeds 200 tokens/s, INT4 deployment file is 173MB, English IFEval score is 73.20, and Chinese instruction following IFEval-zh score is 74.59.
Codex releases00GitHub Copilot rust-v0.169.0-alpha.1 has been released, this version provides code review, collaboration, and other functions.
Ricursive Intelligence releases Ripple-65B Preview model, claiming it has the best agent programming performance in the US open weight model.
Ollama releases6565Ollama now supports Xiaomi MiMo V2.6 model, a 9B parameter code and agent workflow model. The AutoTrust decision model series includes 9b, 35b, and 27b-vl versions, with the 35b and 27b-vl versions capable of handling images. autotrust-jev supports Apple Silicon in this release.
vLLM BlogVeRL-Omni v0.3.0 released, extending asynchronous execution to more multimodal reinforcement learning workloads, setting the V1 trainer as the default option, and expanding post-training support for images, video, audio, and multimodal understanding.
Hugging Face Daily Papers5252Open-HumanScale is the first open-source solution to support the extension of embodied foundation models with egocentric pretraining to 50,000 hours, with its core dataset Open-HumanScale-50K containing 50,000 hours of densely annotated human hand actions and natural language egocentric video data.
