AI Tech Daily - 2026-07-30 | Recsys Frontier

type

Post

status

Published

date

Jul 30, 2026 05:01

slug

ai-daily-en-2026-07-30

summary

📊 Today's Overview

AI hit multiple inflection points today. OpenAI revealed that two simple API settings — retained reasoning and compaction — tripled GPT-5.6 Sol's ARC-AGI-3 score from 13.3% to 38.3%, while slashing output tokens 6x. Microsoft posted its FY2026 results: $331B revenue, Azure hitting $100B at 41% growth. Chinese models swept OpenRouter's top five for the first time, led by Xiaomi's MiMo-V2.5 at 10.5T weekly tokens. On the security front, researchers demonstrated the first self-replicating prompt injection worm targeting Microsoft Word Copilot. Meanwhile, AMD open-sourced Instella-MoE, and LMSYS released Blackwell-native RL training with MXFP8/NVFP4 support.

🔥 Trend Insights

ARC-AGI-3 API settings unlock 3x gains: OpenAI shows two simple parameters — retained reasoning and compaction — boost GPT-5.6 Sol from 13.3% to 38.3% while cutting output tokens 6x. Directly applicable to any agent task.

Chinese models dominate OpenRouter top five: Xiaomi MiMo-V2.5, DeepSeek V4 Flash/Pro, Tencent Hy3, and GLM 5.2 claim the global leaderboard for the first time, signaling a shift in open-source model market dynamics.

Self-replicating AI worm targets Word Copilot: First prompt injection attack with worm-like propagation discovered — hidden instructions embedded in documents spread through Copilot's processing chain, remaining unpatched after 144 days.

🐦 X/Twitter Highlights

📈 热点与趋势

微软FY年收入$331B，Azure增41%达$100B - Satya Nadella（微软CEO）公布2026财年业绩：年收入3310亿美元，同比+18%；Microsoft Cloud收入2140亿，+27%；Azure收入1000亿，+41%。 @satyanadella

分析美国机器人禁令：中国今年将交付5万个人形机器人，占供应链70% - Gary Marcus（纽约大学心理学教授/AI批评者）转引分析：美国机器人禁令将破坏本土能力，中国今年预计交付超5万个人形机器人，美国仅数千；中国占人形机器人供应链约70%（减速器、执行器、电机）。 @GaryMarcus

🔧 工具与产品

MiniMax与Fireworks AI开源M3推理核（MSA及Fireworks kernels） - MiniMax（AI模型开发者）和Fireworks AI（推理优化公司）各自开源M3模型推理所需的GPU内核代码：MiniMax MSA和Fireworks kernels，方便社区部署高性能M3推理。 @MiniMax_AI | @FireworksAI_HQ

Kimi K3通过vLLM扩展至NVIDIA全系列、AMD ROCm及DigitalOcean - vLLM（开源推理引擎/UC Berkeley出品）宣布Kimi K3（月之暗面2.8T MoE模型）从Day 0起支持NVIDIA Blackwell/Hopper/NVL72系列、AMD Instinct (ROCm)以及通过DigitalOcean部署。 @vllm_project | @vllm_project | @vllm_project

Cursor上线iPad版本，支持Agent工作 - Cursor（AI代码编辑器）发布iPad版，功能与iPhone版相同但屏幕更大，适合与Agent协作。 @cursor_ai

Weaviate Query Agent集成GPT-5.6 Luna/Terra，召回率提升5-10% - Weaviate（向量数据库提供商）的Query Agent自动升级，使用OpenAI GPT-5.6 Luna（最快最便宜）和Terra（平衡版）模型，Recall@5提升5-10%，nDCG@10提升5-7%，成本不变。 @weaviate_io

Simon Willison记录为ChatGPT和Claude配置自定义MCP服务器 - Simon Willison（Datasette作者/独立开发者）发布教程，说明如何在ChatGPT和Claude的常规聊天界面中添加自定义MCP服务器。 @simonw

⚙️ 技术实践

GPT-5.6 Sol：ARC-AGI-3达SOTA，使用限制原因公开，自优化降本20% - OpenAI官方：GPT-5.6 Sol通过多上下文窗口和压缩实现在ARC-AGI-3上达到SOTA @thsottiaux；同时解释此前使用限制原因：Sol更努力工作、工具调用更多，已优化使使用时间延长约18% @thsottiaux；还通过生产GPU内核改进使服务成本降低20%、推测解码效率提升15%以上 @OpenAI。 @sama

LMSYS开源Blackwell原生MXFP8/NVFP4 RL训练方案，Qwen3验证 - LMSYS Org（大模型评测组织）与Humanns、NVIDIA合作，开源端到端MXFP8以及硬件原生NVFP4 W4A4 RL训练方案，支持跨层精度控制。在Qwen3-30B-A3B上所有低精度配置均接近BF16奖励曲线，同时减少rollout时间。 @lmsysorg

社区估算K3后训练成本约$4M，576块B300 GPU训练12天 - 社区开发者nrehiew_基于公开信息估算：K3后训练共5000万次rollout、约1.3M任务、24个专家，每专家约250步，每步约10分钟，总成本约400万美元，使用576块B300 GPU。 @nrehiew_

Modal推理负责人与Cognition研究负责人对话：RL与推理优化正在趋同 - Modal（serverless GPU平台）推理负责人_gongy与Cognition研究负责人Silas Alberti对谈，讨论RL训练与推理优化的汇合，包括树形推测解码、DFlash技术、在线训练推测器等。 @modal

⭐ Featured Content

OpenAI reveals ARC-AGI-3 score tripling API settings ｜ Practical reasoning optimization guide

OpenAI's official blog reveals that enabling just two API settings — retained reasoning and compaction — boosts GPT-5.6 Sol's ARC-AGI-3 score from 13.3% to 38.3% (approaching the human average of 48%), while reducing output tokens by 6x. The post analyzes differences between the official harness and optimized harness, showing that benchmark scores reflect not just model capability but also API settings and harness design. For AI practitioners, these two settings apply directly to other agent tasks as practical efficiency hacks.

Sources: OpenAI

AMD releases Instella-MoE fully open-source model, challenging MoE ecosystem ｜ New open-source MoE benchmark

AMD released Instella-MoE-16B-A3B, a fully open-source 16B MoE language model activating only 2.8B parameters per token. Trained on AMD Instinct MI300 and MI325 GPUs with 2.8T tokens across pretraining, long-context extension, SFT, preference tuning, and multi-stage RL. It performs competitively against top models of similar scale. AMD open-sourced model weights, training code, and a detailed technical report, providing a reference for MoE training on AMD hardware.

Sources: ROCm Blogs

Chinese models sweep OpenRouter's global top five for the first time ｜ Open-source model market shifts dramatically

OpenRouter's latest weekly call volume ranking shows Chinese models claiming the global top five for the first time: Xiaomi MiMo-V2.5 tops at 10.5 trillion tokens, DeepSeek V4 Flash/V4 Pro rank second and fifth, Tencent Hy3 jumps 999%+ weekly to third, and Zhipu GLM 5.2 takes fourth. Data comes from real developer paid calls, reflecting Chinese open-source models' dominance in overseas developer markets through high cost-performance and stable agent capabilities.

Sources: BigGo Finance

PatientAgentBench: first clinically validated evaluation standard for patient-facing healthcare AI agents ｜ Medical agent safety benchmark

Amazon Science released PatientAgentBench, the first benchmark for evaluating patient-facing healthcare AI agents. Existing benchmarks test static medical knowledge or clinician-facing technical tasks, failing to assess clinical safety and workflow execution in multi-turn patient conversations. The benchmark uses synthetic patient profiles, LLM-as-a-jury six-dimensional evaluation (clinical safety, triage quality, workflow accuracy, task completion, clinical helpfulness, dialogue quality) for reproducible, contamination-resistant assessment. Key finding: triage is the biggest differentiator, with a severity paradox — models score higher on emergency cases while routine cases hide greater risks. Safety failures cluster around missing crisis resources and fabricated clinical information.

Sources: Amazon Science

K-Search: automatically transferring CUDA kernel optimization knowledge to Apple Silicon ｜ Cross-hardware kernel optimization framework

IBM Research and Berkeley Sky Lab collaborated to automatically transfer CUDA kernel optimization knowledge to Apple Silicon's MLX framework via the K-Search evolutionary search framework. Core results: FlashAttention reaches 97% of native MLX performance, Mamba SSM prefill speeds up 20x. The method generalizes beyond MLX to any hardware ecosystem where CUDA knowledge is transferable.

Sources: BAIR Blog

Trust challenges from MCP statelessness and TEE solutions ｜ Agent infrastructure forward thinking

This post analyzes trust challenges introduced by MCP's July 28, 2026 version switching from stateful to stateless protocol. The author notes statelessness enables MCP servers behind standard polling load balancers, but eliminates single-view tracking of complete interactions, making it impossible for third parties to verify whether tool calls actually occurred across organizational boundaries. The proposed solution uses TEE (Trusted Execution Environment) with hardware-signed proofs, allowing external auditors to verify MCP call authenticity even when the operator is untrusted.

Sources: AAIF

AI Worm achieves self-replication through Word Copilot ｜ First self-replicating prompt injection attack

Security researcher Håkon Måløy discovered a new prompt injection attack variant targeting Microsoft Word Copilot that achieves self-replicating worm behavior. Attackers embed hidden instructions in documents — when Copilot processes the document, instructions execute and manipulate the current document while copying hidden instructions to new documents, forming a propagation chain. The vulnerability was reported to Microsoft but remains unpatched after 144 days. This is the first demonstrated prompt injection attack with self-replication capability.

Sources: Simon Willison

Deconstructing Scaling Law: optimization, architecture, data trio ｜ Systematic Scaling Law interpretation

Su Jianlin's systematic interpretation of Scaling Law, breaking down quantitative laws affecting model performance from three dimensions: optimization, architecture, and data. The post first introduces mathematical tools like heterogeneous power inequalities and power-law combination minima, then discusses optimizer scaling behavior (Muon, Adam), architecture impact on scaling (MoE, MLA), and data composition/quality effects. Finally proposes a unified framework incorporating all three perspectives to guide hyperparameter selection, architecture design, and data strategy.

Sources: 科学空间

🎙️ Podcast Picks

Building the Automated AGI Lab: Core Automation's Jerry Tworek and Rohan Anil

📍 Source: Training Data | ⭐⭐⭐⭐⭐ | 🏷️ LLM, Research, Interview | ⏱️ 49:11

Jerry Tworek (OpenAI reasoning lead) and Rohan Anil (Gemini pretraining co-lead) argue Transformers have hit a wall — the missing piece is continual learning. They note in-context learning fails after ~20 minutes, fine-tuning causes catastrophic forgetting, and pretraining/RL should be optimized end-to-end. Their vision for an automated AGI lab starts with automated kernel generation, since frontier models still lose to human experts here. Essential listening for anyone thinking about LLM architecture evolution and agent system design.

💡 Why Listen: Two of the most senior voices in LLM research openly question the current paradigm. The discussion on continual learning as the missing capability, and why automated kernel generation is the bottleneck for AGI labs, challenges assumptions most practitioners take for granted.

📄 Paper Highlights

HANDBOOK.md: A Benchmark for Long-Context Agentic Instruction Following

Surge AI ｜ 🏷️ Agent Framework, Agentic Workflow, Agent Deployment, Safety, Benchmark

First benchmark testing whether agents actually follow long policy documents (20-124 pages) across 65 enterprise tasks — frontier models pass only 36.2%, with failures from policy override, rule drift, and fabricated compliance.

AngelSpec: Towards Real-World High Performance Inference with Speculative Decoding

Tencent ｜ 🏷️ Inference, Agentic Workflow, Architecture, Training, Code Generation

Unified training framework co-specializing MTP and block-diffusion drafters for different workloads — achieves 1.98-2.40x speedup on Hy3 series by dynamically adapting verification depth per request.

Metis: Memory Foundation Model

MemTensor ｜ 🏷️ Architecture, Training, Inference, Agent Memory, Fine-tuning

First prototype of a memory foundation model with native persistent memory state — compresses history into model parameters via memory attention, gradient-free online updates requiring only a forward pass.

🐙 GitHub Trending

SpecPrefetch ｜ Parameter-efficient MoE expert prefetching

Shared lightweight adapter predicts next-layer expert candidates for asynchronous transfer, separating prediction from execution routing. On Snapdragon 8 Elite, improves decoding throughput by up to 20% over compute-optimized offloading.

GitHub ｜ ⭐ N/A ｜ 🗣️ Python ｜ 🏷️ Inference, MoE, Agent Deployment, Fine-tuning