type
Post
status
Published
date
Jun 2, 2026 04:30
slug
ai-daily-en-2026-06-02
summary
AI hit a major capital markets milestone today: Anthropic filed its S-1, kicking off the IPO race with OpenAI. Meanwhile, MiniMax dropped M3 — a model that beats GPT-5.5 and Gemini 3.1 Pro on key benchmarks at just 5-10% the cost, marking the first time a Chinese model has topped US frontier models.
tags
AI
Daily
Tech Trends
category
AI Tech Report
icon
📰
password
priority
1
📊 Today's Overview
AI hit a major capital markets milestone today: Anthropic filed its S-1, kicking off the IPO race with OpenAI. Meanwhile, MiniMax dropped M3 — a model that beats GPT-5.5 and Gemini 3.1 Pro on key benchmarks at just 5-10% the cost, marking the first time a Chinese model has topped US frontier models. NVIDIA released Cosmos 3, the first open-source physical AI foundation model, while xAI's Ethan He made the case that video agents are the next big breakthrough. On the infrastructure side, Google plans an $80B equity raise for AI buildout, and a CNBC report reveals that nearly half of pre-ChatGPT unicorns haven't raised in three years — the AI wave is reshaping the entire startup landscape.
🔥 Trend Insights
- Chinese models leapfrog on cost-performance: MiniMax M3 beats GPT-5.5 and Gemini 3.1 Pro on BrowseComp while costing 5-10% — the first time a Chinese model has topped US frontier models, reshaping the LLM market's competitive dynamics.
- Physical AI goes open source: NVIDIA's Cosmos 3 launches as the first open-source physical AI foundation model, combining vision reasoning, multimodal generation, and action prediction — lowering the barrier for embodied AI development.
- Video agents emerge as the next paradigm: xAI's Ethan He argues video generation will follow AI coding's trajectory — shifting from single-shot output to multi-turn reasoning, planning, and iteration, with video agents replacing better Sora-like models.
🐦 X/Twitter Highlights
📅 2026-06-02 AI/科技信息日报
📈 热点与趋势
- Google计划筹资800亿美元扩建AI基础设施 - amit（投资博主）汇总当日金融市场动态：Google提议800亿美元股权筹资，其中300亿公开发行、100亿私募给Berkshire；Anthropic秘密提交S-1启动IPO流程；AI相关公司年内已融资约3800亿美元，占风投资金87%；数据中心建设支出超500亿美元年化，较2022年增长336%。 @amitisinvesting
- Anthropic秘密提交S-1，启动IPO流程 - 已向SEC提交注册声明草案，等待审核完成后可选择进行IPO。 @AnthropicAI
- OpenAI前沿模型和Codex在AWS Bedrock正式可用 - 企业可通过Amazon Bedrock使用OpenAI模型。OpenAI还宣布Daybreak（网络安全AI能力）未来也将上架AWS。 @OpenAI
- Runway与NVIDIA成立“Cosmos Coalition”，共建开源世界模型 - 新全球倡议将联合多家AI实验室，开放和开源物理AI的前沿世界模型，Runway和NVIDIA为创始成员。 @runwayml
- Runway在伦敦设欧洲总部，投资1亿美元于英国AI生态 - 未来18个月投入1亿美元，至2028年计划翻倍。伦敦将成为其通用世界模型研究的核心新枢纽。 @runwayml
- GTC Taipei：NVIDIA发布AI工厂、自主代理及新一代AI PC - Jensen Huang演讲宣布AI基础设施、Agent、物理AI和计算平台全方向进展。 @nvidia
- OpenAI Foundation拨款1.3亿美元用于“AI韧性” - Sam Altman称基金会正在帮助社会建立对AI的韧性，该拨款覆盖生物、网络、模型安全及青少年影响领域。 @sama | @FoundationOAI
- Andrew Ng分析AI Forward Deployed Engineer角色兴起 - 该角色由Palantir首创，现因Agent定制需求重获热度，但Andrew Ng认为AI Engineer岗位数量将远超FDE。 @AndrewYNg
- Cursor提高Teams用量限制，推出5倍用量Premium席位 - Premium seat费用为3倍，提供5倍使用量。 @cursor_ai
🔧 工具与产品
- Perplexity发布Search as Code：用Python替代函数搜索的Agent架构 - Agent直接编写Python调用搜索栈，取代逐次函数调用，现已在Perplexity Agent API和Computer中默认启用。 @perplexity_ai | @AravSrinivas
- 阿里发布Qwen3.7-Plus多模态Agent模型 - 统一视觉和语言的Agent基座，支持GUI/CLI操作、编码助手、视觉推理，可通过阿里云模型服务平台API调用。 @Alibaba_Qwen
- vLLM day-0支持NVIDIA Cosmos 3和JetBrains Mellum2模型 - Cosmos 3是融合AR推理与扩散的多模态模型，Super (32B)和Nano (8B)已开源；Mellum2为12B MoE编码模型，激活参数2.5B。 @vllm_project | @vllm_project
- Unitree发布H2 Plus人形机器人参考设计，基于NVIDIA Isaac GR00T - 整合Unitree H2身体、Sharp Wave五指手、NVIDIA Jetson Thor机载计算及Isaac GR00T软件栈。 @UnitreeRobotics
- MiniMax M3在Next.js Agent评估中仅次于Opus/GPT5，成本低10-20倍 - Guillermo Rauch（Vercel CEO）发布评测，M3在编码任务上表现出色，当前在AI Gateway上价格再减半。 @rauchg
- Step 3.7 Flash在Kilo代码编辑器中免费提供 - Kilo宣布Step 3.7 Flash专为编码Agent的多步编排和可靠工具调用优化。 @StepFun_ai
- Jerry Liu发布LiteParse v2：用Rust重写的PDF解析器 - 支持50+文档类型，无模型依赖，可输出边界框供编码Agent直接标记来源，已发布为Python和Node原生包。 @jerryjliu0
⚙️ 技术实践
- LMSYS用Intel CPU离线视觉编码加速VLM推理 - 通过SGLang EPD分离和Dynamo加权路由，将视觉编码卸载到Intel Xeon CPU，实现P99 TTFT降低1.2-1.3倍，TPOT降低1.3-30倍。 @lmsysorg
- Modal分享RL规模化训练经验并发布开源库 - 总结帮助团队在Modal上大规模训练的常见模式和学习教训。 @modal
- Pinecone Nexus将推理与知识引擎解耦，减少Agent 90% token消耗 - 将企业原始数据提前转化为任务优化知识，避免Agent盲目的探索性工具调用，实现30倍加速。 @pinecone
- Nous Research与NVIDIA集成Agent Skills目录至Hermes Skills Hub - 教Agent使用CUDA-X、Omniverse、NeMo等NVIDIA平台组件的官方技能。 @NousResearch
⭐ Featured Content
Anthropic Files S-1, Officially Launches IPO Process ｜ AI industry shifts from funding-driven to capital markets maturity
Anthropic has submitted its S-1 filing to the SEC, entering the IPO race with OpenAI. This milestone marks the AI industry's transition from a funding-driven phase to capital markets maturity. For practitioners, it signals profound changes in competitive dynamics, talent flows, and business models. Anthropic's IPO will provide a more stable capital base, accelerating its full-spectrum competition with OpenAI across model capabilities, enterprise customers, and agent products.
Sources: Business Insider
MiniMax-M3 Launches: Surpasses GPT-5.5 and Gemini 3.1 Pro on Key Benchmarks at Just 5-10% Cost ｜ First time a Chinese model outperforms US frontier models
MiniMax released M3, built on the MiniMax Sparse Attention (MSA) architecture with a 1M token context window and native multimodality. It surpasses GPT-5.5 and Gemini 3.1 Pro on key benchmarks like BrowseComp, priced at $0.3/$1.2 per million tokens (limited time), with open-weight release planned within 10 days. This is the first time a Chinese AI model has outperformed US frontier models while maintaining extremely low pricing — potentially reshaping the LLM market landscape. For practitioners, this is a critical signal for evaluating model selection and cost strategies.
NVIDIA Releases Cosmos 3: First Open-Source Physical AI Omnibus Model ｜ Combines vision reasoning, multimodal generation, and action prediction
NVIDIA released Cosmos 3, built on a mixture-of-transformers architecture that supports text, video, image, environmental sound, and action inputs. It generates physically realistic synthetic videos and robot task data, ranking first on VANTAGE-Bench and TAR benchmarks. Deployments with Agile Robots, Linker Vision, and others are already underway. It includes Diffusers integration, post-training scripts, and open-source datasets, lowering the barrier for physical AI development. For practitioners focused on world models and embodied intelligence, this is a key reference for understanding the physical AI technical roadmap.
Sources: NVIDIA Blog ｜ Hugging Face
JetBrains Releases Mellum2: 12B MoE Model with 2x+ Faster Inference Than Peers ｜ Efficient open-source model focused on code and text
JetBrains released Mellum2, a 12B-parameter MoE architecture with only 2.5B active parameters per token, delivering over 2x faster inference than comparable models. Apache 2.0 open-source, suitable for routing, RAG, sub-agents, private deployment, and more. For AI system developers needing efficient, low-latency inference, this is a noteworthy new option — especially for replacing larger models in coding agent and code-related tasks.
Sources: Hugging Face
Video Agents Are the Next Breakthrough: From Single-Shot Generation to Multi-Turn Reasoning and Iteration ｜ xAI Grok Imagine lead shares hands-on insights
xAI Grok Imagine lead Ethan He shared a core thesis on the Latent Space podcast: the next breakthrough in video models isn't a better Sora — it's video agents. Video generation will follow AI coding's evolution from single-shot output to multi-turn reasoning, planning, editing, and iteration within an agent system. He shared hands-on experience building Grok Imagine in 3 months, from NVIDIA Cosmos to xAI, emphasizing that iteration speed and fixing small bugs matter more than grand architecture. He also discussed forward-looking predictions like generative UI (Flipbook) potentially replacing traditional HTML/CSS. Highly insightful for practitioners focused on multimodal agents and video generation technology roadmaps.
Sources: Latent Space
Claude Code vs. Cursor vs. Codex vs. Antigravity: Six-Month Deep Comparison ｜ Practical coding agent selection guide
The New Stack published a deep comparison of four coding agent tools after six months of use. Key findings: Claude Code leads in complex refactoring and cross-file understanding, Cursor excels in rapid iteration and IDE integration, Codex shines in automated testing and documentation, and Antigravity stands out for team collaboration and code review. The article provides use-case scenarios, performance comparisons, and selection recommendations — directly valuable for developers evaluating or using coding agents.
Sources: The New Stack
IBM Research Proposes Agent Logic: Introducing Primitives Like Knowledge Graphs at the Agent Layer, Reducing Token Consumption by 30x ｜ New approach to enterprise agent scalability
IBM Research introduced the concept of agent logic — introducing software primitives like knowledge graphs and program analysis at the agent layer to actively guide LLMs toward enterprise workflows, reducing context space. Validated across four real-world scenarios (legacy code understanding, test generation, incident response, compliance modernization), it reduced token consumption by ~30x compared to pure LLM approaches while maintaining or improving performance. The article provides specific architecture and quantitative results — directly valuable for building scalable enterprise agents.
Sources: Hugging Face
AI Wave Destroys Pre-ChatGPT Startups: Nearly Half of 857 Unicorns Haven't Raised in Three Years ｜ Startup ecosystem reshaped
CNBC exclusive: The AI wave is destroying pre-ChatGPT-era startups. PitchBook data shows nearly half of 857 US unicorns haven't raised funding in three years. Companies that last raised in 2021 have seen valuations drop an average of 68%, while 2022 rounds are down 52%. Over 220 companies that once reached billion-dollar valuations have become 'fallen unicorns.' The AI boom has sucked up over $250 billion in funding, completely resetting startup valuation systems. For practitioners, this is key data for understanding the current startup funding environment and market landscape shifts.
Sources: CNBC
🎙️ Podcast Picks
Why Video Agent models are next — Ethan He, xAI Grok Imagine
📍 Source: Latent Space | ⭐⭐⭐⭐⭐ | 🏷️ LLM, Agent, MultiModal | ⏱️ 1:43:26
Ethan He shares hands-on experience from NVIDIA Cosmos world models to building xAI's Grok Imagine in 3 months. His core thesis: video model intelligence comes primarily from LLMs, not video data, and the future direction is video agents. Deep dive into engineering details — data, VAEs, diffusion transformers, inference acceleration — with emphasis on iteration speed and fixing data/training bugs over grand architecture. Predicts video generation will follow AI coding's trajectory from single-shot output to multi-turn reasoning and planning, eventually replacing traditional UI.
💡 Why Listen: Ethan He built Grok Imagine from scratch in 3 months — this isn't theory, it's battle-tested. If you care about where multimodal agents are heading, this is the most concrete, technically detailed take you'll hear this week.
📄 Paper Highlights
Mellum2 Technical Report
JetBrains ｜ 🏷️ Architecture, Training, Inference, Code Generation, Reasoning, MoE
JetBrains' 12B MoE model (2.5B active params) combines 64 experts, sliding window attention, and a multi-token prediction head for speculative decoding — competitive with 4B-14B models at 2.5B compute cost.
Learning Agent-Compatible Context Management for Long-Horizon Tasks
Tongyi Lab, Alibaba Group ｜ 🏷️ Agent Framework, Agentic Workflow, Reinforcement Learning, Long Context, Web Search
AdaCoM trains an external LLM to manage a frozen agent's context via RL, cutting token waste while preserving task constraints — and reveals a Fidelity-Reliability trade-off across agent capabilities.
Configurable Reward Model for Balanced Safety Alignment
Meta ｜ 🏷️ Fine-tuning, Safety, RLHF/DPO, Reward Model
Meta's CSRM achieves SOTA on configurable safety benchmarks (94.6% F1 on CoSApien) without extra human annotation — a practical path for LLMs to adapt to evolving safety requirements.
Probing the Prompt KV Cache: Where It Becomes Dispensable
AWS AI Labs ｜ 🏷️ Inference, Architecture, Transformer, KV Cache
Systematic analysis shows prompt KV cache redundancy is about chat template scaffolding, not content — replacing upper-layer cache with neutral filler recovers near-perfect accuracy across Qwen3, Gemma 3, and Llama 3.
MosaicLeaks: Privacy Risks in Querying-in-the-Open for Deep Research Agents
ServiceNow AI Research ｜ 🏷️ Agent Framework, Safety, Agentic Workflow, RLHF/DPO, Fine-tuning
Deep research agents leak private info through external queries — MosaicLeaks benchmark shows PA-DR framework cuts leakage from 34% to 9.9% while improving accuracy from 48.7% to 58.7%.