DeepSeek Open-Sources Internal Reasoning Framework DeepInfer
DeepSeek announces full open-sourcing of its internal reasoning framework DeepInfer, alongside production-grade Token encoder/decoder component deepseek-recipe. Meanwhile, Inception Labs releases diffusion-based LLM Mercury 3 at 1068 tokens/sec.
Model Releases & Updates
Inception Labs Releases Diffusion LLM Mercury 3 at 1068 tokens/sec
Inception Labs releases Mercury 3, a diffusion-based large language model claiming to surpass GPT-6 Luna and Claude Haiku 5.5 on multiple benchmarks, achieving 1068 tokens/sec with 40x more parameters than Mercury 1.
Source: Inception Labs：Blog
Product Launches & Updates
DeepSeek Open-Sources Internal Reasoning Framework DeepInfer
DeepSeek announces full open-sourcing of its internal reasoning framework DeepInfer, alongside the production-grade Token encoder/decoder component deepseek-recipe.
Source: WeChat：DeepSeek
Google & Meta Release TorchTPU: Native PyTorch Backend for TPUs
Google and Meta release TorchTPU, an open-source project providing native PyTorch backend for Google TPUs. Model migration requires just changing .to("cuda") to .to("tpu").
Source：PyTorch：Blog
ChatGPT Launches Locked Chats: PIN or Biometric Protection for Sensitive Conversations
ChatGPT introduces Locked Chats, moving sensitive conversations into a protected space separate from regular history and search. Users on web and desktop can unlock with PIN; mobile users can use PIN, Face ID, Touch ID, or device unlock. Messages in locked chats don't update memory used in regular conversations. Available now for Free, Go, Plus, Pro, Business, Enterprise, and Edu plans.
Source：X：ChatGPT (@ChatGPT)
Google Releases Native TorchTPU Inference Engines for vLLM and SGLang
Google announces native TorchTPU-based inference engines for vLLM and SGLang, developed with Inferact and RadixArk respectively, enabling developers to deploy models on TPUs using familiar PyTorch tools.
Source：Google Developers Blog
Industry News
Reuters: OpenAI in Talks to Raise ~$30B Before Anthropic IPO, Valuation Exceeds $1.4T
According to Reuters, OpenAI is negotiating to raise approximately $30 billion ahead of Anthropic's IPO, at a valuation exceeding $1.4 trillion, at least 64% higher than the $852 billion from March.
Source：X：Rohan Paul (@rohanpaul_ai)
SpaceX Earnings Show AI Business Beating Expectations, 52% of Revenue
SpaceX's Q3 regulatory filing reveals $14.6B revenue and $4.4B net profit, reversing last year's $1B loss. Its renamed "Super Intelligence" business line (including Cursor, Grok, and neocloud compute) outperformed investor expectations. AI now accounts for 52% of company revenue with 1.7 GW of compute capacity.
Source：TechCrunch：AI
US Army Cyber Command ECHOTRIBBLE Project Aims to Turn AI Agents into Self-Spreading Cyberweapons
According to documents obtained by The Intercept, ARCYBER is developing a platform codenamed ECHOTRIBBLE that lets AI agents autonomously move across networks, identify and neutralize threat targets, then self-destruct, with minimal human oversight required.
Source：The Decoder：AI News
Redact Investigation: WhatsApp Help Pages Hide LLM Instructions to Suppress Regulator Recommendations, GDPR References
Redact found WhatsApp help center pages sending hidden instructions to LLMs that are invisible to users, directing the chatbot not to recommend data protection regulators and to avoid citing GDPR, LGPD, or CCPA after users mention lawyers or legal action.
Source：Hacker News：AI Trending
OpenAI Acquires Drug Development Platform Orchestra Bio, Integrating into Rosalind Life Sciences
OpenAI announces acquisition of Orchestra Bio, a platform helping biopharma teams coordinate complex drug development workflows. Its team and technology will integrate with Rosalind. Orchestra's agents can help teams track research and project progress, analyzing how batch production failures impact clinical trial timelines and suggesting responses.
Source：OpenAI
Anthropic's Cloud Partnerships Shifting: AWS, Google, Microsoft Now View It as Competitor
According to S-1 documents leaked to WSJ, Anthropic struggles to secure sufficient cloud compute from AWS, Google, and Microsoft, as these vendors increasingly view it as a competitor rather than a customer.
Source：TechCrunch：AI
Research Papers
Anthropic Designs Protein Mini-Binders Against Scorpion Neurotoxins Using Claude
Anthropic ran its internal research model autonomously in Claude Science for 64 hours to design protein mini-binders targeting five scorpion neurotoxins, generating 318,484 designs and筛选出 180.
Source：Anthropic：Research
Flashes
- Perplexity Computer Now Runs Fully Locally on Mac — no cloud inference required
- Apate.ai Uses 350K AI Anti-Scam Agents to Stall Scammers — domestic youth fighting AI-generated sympathy accounts with AI videos
- Google Releases TorchTitan TPU Fork with Model Optimization Playbook
- Waymo Announces Spain Expansion — first passengers in Madrid region in 2027
- METR Reviews Anthropic August 2026 Risk Report Alignment Chapter — endorses low-risk conclusion but questions evaluation methods
- Meta Open-Sources Weighted Model Muse Spark Open on Fireworks — focused on long-context coding and Agent work
- Stanford HAI Releases Three Medical AI Agent Benchmarks — actual completion rate below 50%
- Ai2 Releases FairStress Benchmark — measuring hidden fairness asymmetries in LLM decisions
- ModelBest Explains LES: How 35B MoE Runs Losslessly on Mobile Edge in Real-Time
- Databricks Releases MemCurrent: Production-Grade Persistent Memory Architecture for Agents
Originally published on WeChat Official Account 「瑞哥观势」.
About the Author
ERIC
AI Technology Expert, focusing on research and application of artificial intelligence and automation tools
Contact & Platforms