AI Tech Daily - 2026-07-24 | Recsys Frontier

type

Post

status

Published

date

Jul 24, 2026 05:01

slug

ai-daily-en-2026-07-24

summary

📊 Today's Overview

AI infrastructure and agent engineering dominated the news. DeepSeek's leaked CEO call revealed ~20K H-equivalent cards and a strong preference for NVIDIA over Huawei, while a job posting hinted at managing 100K-card clusters — contradicting public statements. Andrew Ng open-sourced OpenWorker, a local-first agent that delivers docs and updates Slack. Etched raised $300M at a $10.3B valuation for inference chips. On the research front, Together AI's SonicSampler unified LLM sampling into a single kernel achieving 16x speedup, and NVIDIA's NOOA framework reimagined agents as native Python objects. Google's first ATLAS report quantified the AI economy, and Scale AI's PRBench Legal showed even GPT-5 Pro scoring just 0.37.

🔥 Trend Insights

Agent infrastructure goes production-grade: OpenWorker, OpenForgeRL, and NOOA all push agents toward reliable, testable, and trainable systems — the shift from demos to deployed software is accelerating.

Inference efficiency becomes a first-class optimization target: SonicSampler's 16x speedup on sampling, Qdrant Edge's 0.1ms latency, and Cerebras CEO's deep-dive on reasoning bottlenecks all point to inference as the new frontier.

China AI compute strategy comes into focus: DeepSeek's leaked call and job postings reveal a 20K-100K card reality, while MiniMax shows AMD MI355X can approach NVIDIA B200 performance — the compute landscape is fragmenting.

🐦 X/Twitter Highlights

📈 热点与趋势

DeepSeek CEO leaked call: ~20K H-equivalent cards, prefers NVIDIA, Huawei 950 needs 4:1 swap - Teortaxes (independent researcher) citing Liang Wenfeng (DeepSeek CEO) leaked call says DeepSeek currently has ~20K H-equivalent cards, mostly arrived in the last 1-2 months. Liang says they still prioritize buying NVIDIA, willing to pay premium for "non-compliant" cards; Huawei 950 supernode can fully replace GB200/GB300 but needs 4:1 card swap, software 2 years behind. Training next-gen model needs 50K GB300 or 200K Huawei 950. @teortaxesTex | @KonstantinPilz

Analysis shows DeepSeek hiring hints at managing 100K-card clusters - Teortaxes analyzes DeepSeek's "AI Computing Cluster Performance Engineer" job posting requiring management of 100K-card clusters, including supernodes, training-oriented. Contradicts Liang Wenfeng's public statement that "compute resources are still very small." @teortaxesTex

Demis Hassabis congratulates Kanishka Naranyan as UK AI Minister, latter details AI plans - Demis Hassabis (DeepMind co-founder) congratulates Kanishka Naranyan (newly appointed UK AI Minister). Naranyan lists 48-hour actions: communicating with Arm CEO on UK AI champion positioning, discussing with AI Safety Institute (AISI) on retaining global AI safety talent, and discussing AI's impact on employment with unions. @demishassabis | @KanishkaNarayan

🔧 工具与产品

Andrew Ng releases open-source agent OpenWorker: delivers docs, sends Slack, updates calendar - Andrew Ng (AI educator/Coursera co-founder) launches OpenWorker, an open-source agent that can deliver documents, send Slack messages, update calendar, etc. Runs on Mac (Windows coming soon), supports GPT 5.6 Sol, Claude Fable, Gemini 3.6, open-source models (Kimi, GLM, DeepSeek, Inkling) or Ollama local run. Comes with API key, data doesn't leave local machine. @AndrewYNg

ChatGPT Sites feature: build and deploy websites to Cloudflare Workers, with SQLite persistence - Simon Willison (Datasette author) notes ChatGPT in "Work" mode can use Sites feature to go from idea to hosted website/app in one step, supporting static sites, full-stack apps, persistent storage, authentication, API connections, instant publishing. @simonw | @reach_vb

Etched raises $300M Series C at $10.3B valuation, building inference cluster production facility - Etched (AI inference chip company) announces $300M funding from Sequoia, a16z, Jane Street, Argo, SK Hynix at $10.3B valuation. Has opened 80K sq ft, 10MW facility to accelerate inference cluster production and prototyping. Tri Dao (FlashAttention author) congratulates, saying the team excels in both hardware design and kernels. @Etched | @tri_dao

Qwen releases Audio-3.0-TTS: supports inline tag control, 16 languages, tops TTS leaderboard - Qwen (Alibaba Tongyi Qianwen) launches Audio-3.0-TTS, Flash version for real-time interaction, Plus version for high-quality generation. Adds fine-grained inline tag control (whisper, anger, breath, laughter), natural language style control ("read slowly like a bedtime story"). Supports 16 languages, generates clean output even from noisy reference audio, up to 3 minutes per generation. @Alibaba_Qwen

Qdrant Edge edge retrieval: 0.1ms latency vs 52ms on cloud - Qdrant (open-source vector database) demonstrates Qdrant Edge at Vector Space Day SF, same retrieval engine and API, removing network hops reduces latency from 52ms to 0.1ms. Suitable for robots, drones, wearables and other edge scenarios. @qdrant_engine

⚙️ 技术实践

Sakana AI releases Fugu-Ultra v1.1: dynamically orchestrates frontier models, 7.9 point performance lift - Sakana AI (Japanese AI lab/collective intelligence research) releases Fugu-Ultra v1.1, dynamically orchestrating latest frontier models for performance gains, significant improvements on ProgramBench and Terminal Bench 2.1; even without Fable 5 in the agent pool, it surpasses Fable 5 on complex coding and reasoning tasks. Same price as v1.0. @hardmaru | @SakanaAILabs

MiniMax M3 full-stack optimization on AMD MI355X: approaches NVIDIA B200 performance - MiniMax (AI company) publishes technical blog showing M3 (428B MSA sparse MoE multimodal foundation model) on AMD MI355X with ATOM+ATOMesh full-stack optimization. Core upgrades include EAGLE3 speculative decoding, Page-16 KV Cache for 1M ultra-long context, hierarchical FP8 quantization to avoid vision/MoE precision loss, distributed P/D separation architecture. @MiniMax_AI | @RyanLeeMiniMax

PrimeIntellect runs trillion-scale agentic RL training on vLLM: 28 H200 nodes, <5 min per step - vLLM (open-source inference engine) announces PrimeIntellect's prime-rl 0.6.0 runs agentic RL training of GLM-5.1 (trillion-parameter MoE) on its platform, with 131K sequence length on SWE tasks, using FP8, wide expert parallelism, prefill/decode separation, KV cache offloading (native + Mooncake) and vllm-router, 28 H200 nodes <5 min per step. @vllm_project | @RedHat_AI

Poolside AI open-sources small model Laguna S, also releases full evaluation dataset - swyx independent analysis notes Poolside AI (coding AI company) is unusually open: not only releasing Laguna S (small model, coding performance surpassing Think Machines), but also releasing full evaluation dataset, containing 6 public benchmarks, each with 4 runs, each run with hundreds of turns, for independent verification of rewardhacking. @swyx | @latentspacepod

Black Forest Labs releases FLUX 3: unifies image, video, audio and action prediction for robotics - Emad Mostaque (former Stability AI CEO) shares, Black Forest Labs launches FLUX 3 multimodal model, unifying generation of images, video, audio and action prediction. Extends to robot action prediction with mimic and Audi. @EMostaque | @bfl_ai

AntLingAGI releases Ling-3.0: hybrid KDA+MLA attention, native 256K scalable to 1M - AntLingAGI launches Ling-3.0, using 5:1 hybrid KDA (Key-Value Differential Attention) + MLA (Multi-Head Latent Attention) layers. KDA provides fine-grained long-range memory control, 1/64 expert activation makes MoE computation more efficient. Native 256K context, scalable to 1M. @AntLingAGI | @SonglinYang4

Weaviate explains agentic RAG context engineering: five systems — query enhancement/retrieval/memory/tools/agent - Weaviate (open-source vector database) publishes video explaining agentic RAG is essentially a context engineering problem. Breaks down 5 subsystems: query enhancement (translating fuzzy multi-part inputs), retrieval (choosing chunking strategy balancing precision and context), memory (separating short-term context and long-term external memory), tools (clear descriptions, schemas and correct timing), agent (decision orchestration layer, amplifying upstream context failures). @weaviate_io

⭐ Featured Content

Google releases first ATLAS report, systematically defines the AI economy ｜ An authoritative framework quantifying the macro landscape of the AI industry

Google releases its first ATLAS report, the first systematic definition and quantification of the AI economy. The report proposes a measurement framework covering AI investment, deployment, employment impact and other dimensions, including key data on global AI investment scale, industry penetration rates, and labor market changes. This is an important reference for understanding the macroeconomic impact of AI, providing practitioners with a macro perspective on the industry landscape that can be used for strategic planning and industry communication.

Sources: Google Blog

Poolside co-founder deep interview: Model Factory system revealed, 8B parameter model beats larger models ｜ Unique engineering practices in the Coding Agent space

Poolside co-founder Eiso Kant deeply reveals their Model Factory system in a Latent Space interview: a 70-person team runs 10-20K experiments per month, using streaming data training, low-precision computation, and agents that automatically modify training pipelines, compressing the model from pre-training to release to 8 weeks. The latest Laguna S model (118B total params/8B active) beats Thinking Machines' larger models with 1/10 the parameters. Eiso calls MCP and traditional tool calling "stupid," arguing future agents should write scripts directly. Also covers $500M funding, open-source strategy, RL moving earlier, and more. Directly inspiring for Coding Agent developers and model training engineers, with multiple counterintuitive viewpoints worth discussing.

Sources: Latent Space

Amazon Bedrock AgentCore launches silent failure detection ｜ Important practice for production-grade agent observability

Amazon Bedrock AgentCore launches insights feature, solving the "dashboard all green but users keep complaining" silent failure problem in AI Agent production environments. It analyzes trace data to automatically discover 11 types of behavioral failures (hallucination, wrong actions, instruction violations, etc.), clusters and ranks them with root cause analysis, no preset rules needed. Also provides user intent analysis and execution insights, helping developers shift from reactive troubleshooting to proactive pattern discovery. For teams running Agents in production, this is a directly referenceable observability practice guide.

Sources: AWS Blog

Motorway and AWS jointly release Agent production evaluation blueprint: error rate drops from 1/8 to 1/50 ｜ Engineering benchmark for end-to-end Agent evaluation pipelines

Motorway and AWS collaborated to build an end-to-end AI Agent evaluation pipeline, reducing error rate from 1/8 to 1/50 and detection time from hours to minutes. The article details a two-stage evaluation strategy (build-time + production monitoring), a three-layer evaluation framework (tool use, reasoning, output quality), and a five-stage deployment pipeline (with quality gates). The accompanying open-source repository provides a deployable blueprint, with core principles (three-layer framework, pass^k consistency metrics) being cloud-agnostic. Directly valuable for teams building production-grade Agent evaluation systems.

Sources: AWS Blog

Ethan Mollick releases Agent-era AI selection guide ｜ Practical comparison of ChatGPT vs Claude Agent modes from a frontline professor

Ethan Mollick's latest AI selection guide, focusing on how to choose ChatGPT and Claude's models, thinking levels, and agent modes (Work/Cowork) in the agent era. Includes specific scenario recommendations (high stakes use strongest model, low stakes free model is fine), plus real case demonstrations of agents automatically handling emails and preparing presentations. Directly valuable for practitioners to quickly understand current agent product capability boundaries and selection strategies.

Sources: One Useful Thing

Scale AI releases professional reasoning benchmark PRBench Legal: GPT-5 Pro scores only 0.37 ｜ Reveals systematic weaknesses of LLMs in professional scenarios

Scale AI releases Professional Reasoning Benchmark (PRBench) Legal edition, designed by 182 experts with 1,100 real professional questions covering 114 countries. The current best model GPT-5 Pro scores only 0.37 (out of 1), revealing systematic weaknesses in key dimensions like process transparency and uncertainty handling. This benchmark fills the gap between academic benchmarks like MMLU/GPQA and real professional applications, providing a key evaluation tool for LLM deployment in high-value scenarios like finance and law.

Sources: Scale Labs

Google reportedly prioritizes TPUs for AGI research over renting to customers ｜ Compute allocation strategy reveals new infrastructure landscape dynamics

Google CEO Sundar Pichai revealed in Q2 earnings call that Google is managing its TPU accelerator clusters, prioritizing them for AGI research over renting to customers. At the same time, Google is increasing third-party compute procurement to meet G-Cloud demand. This strategy reveals Google's internal priorities in compute allocation, providing important reference for understanding the AI infrastructure landscape, contrasting with AMD/Anthropic partnerships — the compute battle is extending from the chip level to the allocation strategy level.

Sources: The Register

OpenAI and Anthropic rare joint effort, jointly warn of open-source AI risks ｜ Closed-source labs' policy alliance and regulatory game

OpenAI and Anthropic join forces in a rare collaboration, jointly warning in Washington about the risks of Chinese open-source large models, advocating for stronger regulation. This move sparks intense debate about regulatory capture and whether distillation constitutes IP theft. The article reveals the core contradictions between closed-source labs and the open-source community, as well as the AI safety standards framework being developed within the Trump administration. Important reference for practitioners focused on AI policy and competitive dynamics.

Sources: Axios

🎙️ Podcast Picks

The Biggest Chip Ever Built — Why OpenAI Runs On It | Cerebras CEO Andrew Feldman

📍 Source: The MAD Podcast | ⭐⭐⭐⭐⭐ | 🏷️ LLM, Infra, Interview | ⏱️ 01:12:41

Cerebras CEO Andrew Feldman dives deep into AI inference bottlenecks, explaining why tokens per second per user is the critical metric. He compares GPU, ASIC, SRAM and HBM, points out how agent applications amplify latency problems, and analyzes how fast inference reshapes RL, reasoning models, and SaaS. Also covers Cerebras' wafer-scale chip design, partnership with TSMC, the CUDA moat, and AI infrastructure bubbles.

💡 Why Listen: Feldman is one of the few people who can explain chip architecture tradeoffs in plain English. If you're building on top of LLMs and wondering why inference is slow, this episode gives you the hardware-level mental model to understand it.

Inside the Model Factory — Eiso Kant, Poolside AI

📍 Source: Latent Space | ⭐⭐⭐⭐⭐ | 🏷️ LLM, Agent, Open Source | ⏱️ 1:54:33

Poolside AI co-founder Eiso Kant deeply explores the Model Factory engineering system, including 10-20K experiments per month, streaming data training, low-precision computation, and agents that automatically write code and evaluate results. He shares the journey from investing $12M in 2015 to build a code language model to today's Laguna S model surpassing 10x larger competitors, emphasizing the importance of open-weight and open research. He argues 100 foundation model companies is better than 5 oligopolies, discusses RL moving earlier into pre-training, calls MCP and traditional tool calling "stupid," and covers model-hardware co-design and the path from coding agents to AGI.

💡 Why Listen: Eiso has been building code models longer than almost anyone, and his contrarian takes on MCP, tool calling, and RL timing are worth hearing even if you disagree. The Model Factory concept — treating model training like a software factory — is a genuinely new way to think about AI development.

E245｜The journalists behind large models: how GPT's replies are written

📍 Source: 硅谷101 | ⭐⭐⭐⭐ | 🏷️ LLM, Product, Interview | ⏱️ 46:00

This episode explores how content engineers translate journalism and editorial skills into AI dialogue system engineering standards. Guests share how to decompose "good content," train AI to understand subtext, handle context fragmentation in multi-turn conversations, and specific techniques like system prompts and few-shot examples. Discusses AI sycophancy as a product design problem, the concept of AI quotient, and the career value migration of humanities graduates in the AI era.

💡 Why Listen: Most AI discussions focus on engineering; this one focuses on the craft of writing good responses. If you're a product manager or conversation designer, you'll walk away with concrete techniques for making your LLM product feel less robotic.

Building an Autonomous Delivery Experience with DoorDash Co-Founders Andy Fang and Stanley Tang

📍 Source: No Priors | ⭐⭐⭐⭐ | 🏷️ LLM, Agent, Robotics | ⏱️ 49:19

DoorDash co-founders Andy Fang and Stanley Tang discuss integrating AI and autonomous technology into delivery services. Core topics include: how Ask DoorDash (LLM-powered) changes how users discover restaurants and order; the hardware and operational challenges solved by autonomous delivery robot Dot after two years in Phoenix; the importance of multimodal strategy (robots + humans); and the "first and last 100 feet" problem. They emphasize AI augments rather than replaces human delivery workers, sharing lessons from scaling autonomous technology.

💡 Why Listen: DoorDash is one of the few companies running autonomous delivery at scale. The practical lessons on hardware reliability, multimodal operations, and the last-100-feet problem are directly applicable to anyone building real-world AI systems.

Surviving the New Economics of a Post-Agentic World

📍 Source: Practical AI | ⭐⭐⭐⭐ | 🏷️ Agent, Product, Funding | ⏱️ 35:36

The podcast discusses the profound economic, business, and labor impacts of large-scale AI Agent deployment. Core ideas include: digital labor becomes abundant, agents manage agents, and enterprise scale surpasses human labor constraints. The episode analyzes the erosion of traditional software economic moats, capital flows, and the redefinition of productivity.

💡 Why Listen: Short and thought-provoking. If you're building a SaaS product or investing in AI, this episode gives you a framework for thinking about how agents will reshape business models — even if it doesn't dive into technical details.

📄 Paper Highlights

SonicSampler: Unified Tile-Aware Kernels for LLM Sampling and Speculative Verification

Together AI ｜ 🏷️ Inference, Agent Deployment, Architecture

Together AI fuses the entire LLM sampling pipeline — logit processing, top-k filtering, speculative verification — into a single CUDA Graph-compatible Triton kernel, achieving up to 16x speedup over state-of-the-art baselines in heterogeneous speculative decoding workloads.

Autonomous Topology Mutation: Safe Runtime Restructuring for Multi-Agent LLM Systems

Huawei ｜ 🏷️ Agent Framework, Multi-Agent, Safety

Huawei introduces runtime team restructuring for multi-agent systems: when an agent gets overloaded, the system factorizes it into specialized sub-agents with formal safety invariants. Code-task success jumps from 3.3% to 61.7% with under 500 microseconds of p99 latency overhead.

OpenForgeRL: Train Harness-native Agents in Any Environment

Microsoft Research ｜ 🏷️ Agent Framework, RLHF/DPO, Tool Use

Microsoft Research decouples training from inference with a lightweight proxy and Kubernetes orchestrator, enabling end-to-end RL training on any agent harness. OpenForgeGUI matches models several times larger on OSWorld and WebVoyager using only hundreds to thousands of tasks.

🐙 GitHub Trending

OpenWorker ｜ Open-source local-first agent by Andrew Ng

Andrew Ng's open-source agent that delivers documents, sends Slack messages, and updates calendar. Runs locally on Mac with data never leaving your machine. Supports GPT, Claude, Gemini, and open-source models via Ollama.

GitHub ｜ ⭐ 12,400 ｜ 🗣️ Python ｜ 🏷️ Agent, LLM, DevTool

Emerging-Optimizers ｜ NVIDIA's collection of next-gen training optimizers

NVIDIA releases a codebase containing SOAP, Muon, and other emerging optimizers with layer-wise distributed implementation compatible with Megatron-LM. Includes algorithmic fixes for large-batch stability and fair learning rate transfer across optimizers.

GitHub ｜ ⭐ 2,100 ｜ 🗣️ Python ｜ 🏷️ Training, Optimization, LLM