Updates

Agent engineering is rapidly transitioning from proof-of-concept to production deployment: Alook's open-source platform enables role-based orchestration of CLI Agents; Fudan NLP's team offers a fully automated research Agent for academia—complete with free GPU support. Meanwhile, Xiaomi has slashed its domestic large-model API pricing to ¥0.025 per million tokens, signaling the industry's deep dive into an intense 'token price war' [16].

The ultimate form of mobile AI is evolving toward 'imperceptible intelligence'—OPPO's ColorOS 16 and vivo's official website AI shopping assistant both validate a lightweight deployment path combining compact intent-recognition models, Agent workflows, and RAG knowledge bases [0][5]. Meanwhile, AI commercialization faces structural bottlenecks: both advertising and subscription models have hit saturation points, and industry consensus is rapidly shifting toward an 'execution economy' centered on 'task completion' [9].

Tactile embodied AI secures ~$10M angel round; OpenRouter raises $113M in Series B, hitting 25T weekly tokens; SynthID has watermarked 100B AI-generated items and is integrating into Google Search & Chrome.

CUDA 13.3 officially introduces C++ Tile programming and the CompileIQ auto-tuning framework—marking a paradigm shift toward higher-abstraction GPU development. Meanwhile, Stack Overflow—despite a steep decline in user post volume—achieved $115 million in annual revenue through its enterprise AI knowledge base and data licensing services, validating a new commercialization path for developer platforms in the AI era [2][3].

AI engineering is rapidly entering a new era of 'AI building AI': Baichuan Intelligence has launched ForgeTrain—the world's first production-grade pretraining framework fully written by AI—and successfully trained MiniCPM5-1B. Concurrently, foundational architectural innovations—including Domain-Specific Architectures (DSAs), KV cache quantization (e.g., OSCAR's 2-bit scheme), and the newly proposed 'Tao Law'—are being deployed at scale, continuously pushing past compute and energy-efficiency bottlenecks [24][10][22].

Huawei proposes a new chip evolution paradigm—'The Tao Law'—centered on the time constant τ, challenging the traditional Moore's Law trajectory; meanwhile, DeepSeek tops the global large-model API call leaderboard, highlighting the scalable deployment of domestic AI infrastructure [3]; and 'commitment hallucination' in AI conversational products—triggered by anthropomorphic design—is exposing deep gaps between product accountability and legal regulation [0].

AI engineering is rapidly evolving from prompt engineering toward framework engineering and context engineering; standardization of Agent Harness and vertical-domain workflow reengineering have become critical for real-world adoption. DeepSeek has entered the programming-agent market with a 'Mixue Bingcheng'-style low-cost strategy—directly positioning itself against Claude Code [2][6][7][2].

DeepSeek launches a low-price offensive with its permanently discounted V4-Pro API, targeting the Claude Code–level programming agent market; Baishan Intelligence breaks the edge-side bottleneck by achieving 1.58-bit ternary quantization training of a 60-billion-parameter model on Huawei's Ascend platform—reducing GPU memory usage by 6× while retaining 97% of model capability [2]; meanwhile, Gemini quietly overhauls its billing logic, drastically shrinking actual usage quotas for paying users and exposing a trust gap in large-model commercialization [10].

Baobei Intelligence, Tsinghua University, and OpenBMB achieved end-to-end training of a 60B-parameter LLM on Huawei Ascend using 1.58-bit ternary quantization—cutting memory use by ~6× while retaining 97% capability. Meanwhile, continuous-space language modeling emerges as a paradigm shift beyond token-based autoregression, seen as a key step toward AGI.

Agent tech is maturing rapidly—Codex and similar tools are enhancing core workflow capabilities. Meanwhile, Google's CEO acknowledged Gemini's gaps in coding agents and long-horizon tasks, signaling a shift from model benchmarks to real-world task completion. Anthropic's 'should do' > 'can do' framework highlights the growing scarcity of AI judgment.

AI is accelerating into the agent-native era—coding capability is now the key differentiator for agents [5]; AI compute is shifting historically toward inference, expected to consume 70% of total AI compute [17]; Anthropic reveals Claude's new 'dreaming' memory and long-horizon collaboration architecture [9], while Google's CEO admits Gemini lags in coding agents [16].

AI industry focus is shifting structurally: inference compute will rise to 70% of total AI spend; Agent startups now compete for enterprise payroll budgets—not just SaaS budgets—while high-quality medical data and structured security skill repositories emerge as key new barriers.

AI toolchains are evolving from point solutions to persistent, thread-based systems; OpenAI Codex team proposes eight strategies to reshape human-AI collaboration [3]. Meanwhile, Anthropic's valuation nears ¥90 billion, highlighting tension between AI commercialization and investor expectations [1].

AI shifts from model-level competition to system-level innovation: DeepSeek cuts V4-Pro prices permanently and launches the 'Harness' engineering initiative (vs. Claude Code); Google unveils Gemini 3.5 and Antigravity 2.0 agent platform; vertical startups like FlashLabs and invoko.ai focus on business integration and consumer deployment.

AI is accelerating its penetration—from the tool layer down to the foundations of business models and organizational capabilities. Three structural signals are emerging: the collapse of the SaaS subscription model, the mass production and real-world deployment of embodied AI, and the fragmentation of technical stacks driven by sovereign compute infrastructure. Concurrently, the workplace exhibits a paradoxical coexistence of the 'AI overtime paradox' and the '35-year-old premium'—revealing deep tensions stemming from an immature human–AI collaboration paradigm [2][4][5][10].

OpenAI accelerates IPO plans amid pressure from Anthropic's rise, compute funding gaps, and SpaceX's strategic AI talent acquisition; Codex adds macOS-exclusive features for local AI agent workflows; Chinese regulators launch joint investigation into cross-border brokers, signaling stricter AI model compliance.

OpenAI accelerates IPO plans amid pressure from Anthropic's rise, SpaceX's AI investment, and GPU funding gaps. Tencent open-sources Hy-MT2 multilingual translation models (1.8B/7B/30B-A3B) supporting 33 languages. Codex for macOS adds app snapshots and /goal task management.

AI-driven industrialization in film and television is accelerating, with MovieFlow Studio launching an end-to-end AI video Agent; OpenAI's model has achieved a breakthrough discovery in cutting-edge mathematics—refuting a longstanding conjecture in discrete geometry [2]; and the 'small-but-deep' AI hardware startup approach targeting vertical niches has gained validation from Silicon Valley VCs, highlighting the global potential of China's dual-engine advantage: its world-class supply chain + AI tooling capabilities [3].

Anthropic surpasses OpenAI with a $90B valuation and achieves profitability two years ahead of schedule—marking the first major LLM company to enter public-market valuation validation.

Claude Code adds /usage command for granular token tracking of Skills, Agents, MCPs, and Plugins; Moore Threads pushes AI compute to home entertainment devices; diamond emerges as a key thermal management material for AI chips.

AI achieves historic breakthrough in mathematical proof: GPT-5.5 Pro solves the 80-year-old 'unit distance problem'; Tencent's Hy-MT2 supports 33 languages across three model sizes with 1.25-bit extreme quantization; Alibaba Cloud's MaaS revenue surges 15× in 5 months—Qwen 3.7 Max tops domestic models and ranks top 5 globally.

This week, OpenAI, Anthropic, and SpaceX accelerated their IPO preparations—Anthropic has already achieved profitability two years ahead of schedule, signaling that the AI arms race has officially entered the secondary-market validation phase [2]. Meanwhile, embodied intelligence has achieved its first large-scale deployment in real-world logistics environments: StarMotion's Era0 model ranked #1 globally in the RoboChallenge Table30 physical robot evaluation [7].

AI infrastructure is undergoing systemic upgrades—from power supply and wafer capacity to heterogeneous computing—while applications accelerate toward agent-native designs and OS-level integration. Alibaba Cloud launched 32+ new agents; ZhiXiang Future unveiled a 200B-parameter image foundation model; Google redefined search with Gemini 3.5—marking the industry's shift from technical validation to commercial deployment and ecosystem transformation.

Global AI competition has fully entered the Agent-native era and infrastructure reconstruction phase: Google is reshaping the search entry point for 5 billion users with Gemini 3.5; Alibaba Cloud launched Qwen3.7-Max and the Zhenwu M890 chip to advance full-stack Agentification; NVIDIA introduced the Nemotron-Labs-Diffusion model—capable of seamless switching among AR, diffusion, and speculative decoding modes [1][2][8]. Meanwhile, Kubernetes v1.36 strengthens support for AI workloads, signaling that AI engineering is shifting downward—from the model layer to the cloud-native foundation [18].

Google shifts fully to agent-native architecture at I/O 2026, launching Gemini Omni (natively multimodal video model), Gemini 3.5 Flash (high-performance, low-cost), Antigravity 2.0 (agent-first desktop platform), and Gemini Spark (24/7 personal AI agent).