📊 Today's Overview

🔥 Trend Insights

• Inference infrastructure becomes the new battleground: Fireworks ($15B) and Baseten ($11B) hit decacorn status as capital pivots from training to scaled reasoning — AWS, NVIDIA, and xAI all shipped major inference-facing products today.

• Agent safety and cost scrutiny intensify: Anthropic published sandboxing guidelines, Microsoft Copilot Cowork leaked data via prompt injection, and Uber's COO publicly questioned AI ROI — the industry is waking up to agent operational risks.

• Device-cloud coordination matures: Hera (Alibaba) introduced step-level device-cloud routing, Arize AI showed 3B models matching Claude Sonnet quality, and PrismML released 1-bit image models for local hardware — the "small model + smart routing" paradigm is gaining traction.

🐦 X/Twitter Highlights

📈 热点与趋势

• Greg Isenberg (Pi Day founder) shares SF insights: MCP becomes the new SEO, "agent debt" concept emerges – After 5 days visiting 3 billionaires and frontier model teams: billionaires are acquiring SaaS companies at 40-70% discounts and rebuilding them agent-first; frontier model companies desperately want to use data to fill API blind spots; consumer AI is undervalued — Cal AI hit $50M ARR in 18 months; MCP endpoints are being passively pulled into deals — if an agent can't find you, you don't exist; seed round valuations at $25-50M, Series A hitting $450M; open-source models (Gemma, DeepSeek) are sufficient for 80% of use cases — "which model" is being replaced by "which task needs which model"; "agent debt" refers to loose agent workflows that after 6 months develop system prompt conflicts, memory pollution, and tool overlap @gregisenberg

• IREN signs $1.6B Blackwell system procurement deal with Dell – Supporting the previously announced $3.4B AI cloud hosting contract, deployed at Texas Childress data center, starting early 2027, expected annualized revenue increasing from $3.7B to $4.4B @IREN_Ltd

• Figure signs with JCPenney operator Catalyst Brands for large-scale humanoid robot deployment – Initial pilot at Nevada Reno warehouse; Catalyst also owns Aéropostale and Brooks Brothers @Figure_robot

• SoftBank insiders fear OpenAI could become WeWork 2.0, Altman says they need to become an "inference company" – SoftBank executives banned from discussing failure risk; Son liquidated Nvidia/T-Mobile positions to concentrate bets on OpenAI with zero board seats. Meanwhile Sam Altman says "we must become an AI inference company"; analysts note Anthropic's inference compute already at 300MW (SpaceX Colossus), approaching 1GW via Amazon by year-end — inference capacity is being priced as delivery commitment, not a research project @GaryMarcus (MIT Professor Emeritus) | @demian_ai (Independent Analyst)

• Uber COO directly states AI cost and capability improvements are increasingly hard to justify – First time a major company executive has publicly said AI spending is "getting harder to rationalize" @edzitron (Tech Writer / Better Offline Host)

• China restricts Alibaba, DeepSeek and other companies' AI experts from traveling abroad – Polymarket cites reports of new travel restrictions targeting top AI researchers @Polymarket

🔧 工具与产品

• vLLM officially merges Rust frontend, preprocessing throughput 5x Python – 837 req/s single process vs 162 req/s; enable with `VLLM_USE_RUST_FRONTEND=1`, built on stable Rust @vllm_project

• EAGLE 3.1 released: long-context acceptance length doubles, NVIDIA involved in training verification – FC normalization + post-norm hidden state feedback architecture solving attention drift bottleneck; native vLLM support, open-source Kimi K2.6 draft model @vllm_project

• SenseTime open-sources SenseNova-U1 full training codebase: 8B dense + A3B MoE, Apache-2.0 – Single training stack covering text-to-image, editing, interleaved generation, text and vision understanding; hybrid WP/TP/PP + ISP parallelism, scalable from 1×8 GPU to multi-node @SenseTime_AI

• Qwen3.7-Max ranks 4th on Code Arena frontend, Hermes Agent already supports it – Ties with Claude Opus 4.6, highest ranking for a Chinese lab on agentic web dev tasks @Alibaba_Qwen | @NousResearch

• NVIDIA releases Vera CPU: built for agentic AI, 1.5x performance lead over x86 – Linux kernel compile 2x faster, STREAM TRIAD memory bandwidth 4x @nvidia

• PrismML releases 1-bit/Ternary Bonsai Image 4B image generation model – Designed for local hardware, running high-quality diffusion inference from laptop to phone @PrismML

⚙️ 技术实践

• Theo (prominent YouTuber/indie dev) and Greg Brockman (OpenAI co-founder) praise GPT-5.5 as excellent coding model – Theo says it took 2 months to adjust prompting style and agents.md, now can't use anything else for coding @theo | @gdb

• Anthropic publishes Engineering Blog: agent permissions should evolve with capability, limit destructive operations via sandboxing – In-product sandboxing parameters to scope any potentially destructive actions @AnthropicAI

• PyTorch and NVIDIA publish blog: FP8 PTQ quantization of CLIP using Model Optimizer – Complete workflow from PyTorch checkpoint export to quantization, reducing VRAM usage and inference latency @PyTorch

• Coinbase Base launches MCP, allowing AI agents (ChatGPT/Claude) to manage crypto wallets and DeFi apps – Direct interaction via chat interface @CoinMarketCap

⭐ Featured Content

📄 Paper Highlights

Hera: Learning Long-Horizon Coordination for Device-Cloud Collaborative LLM Agents

CUA-Gym: Scaling Verifiable Training Environments and Tasks for Computer-Use Agents

ECHO: Terminal Agents Learn World Models for Free

🐙 GitHub Trending