2026-10-21 · WednesdayPublished daily at 08:00
AI Daily
DeepSeek fully open-sources its internal reasoning framework, DeepInfer
DeepSeek announces that it will fully open-source its internal reasoning framework DeepInfer, while also open-sourcing the production-grade Token encoding and decoding component deepseek-recipe.
- 1
- 2
- 3
01Model Releases/Updates
Inception Labs releases Mercury 3, a diffusion large language model reaching 1068 tokens/second
Inception Labs releases Mercury 3, a diffusion-based large language model, claiming it surpasses GPT-6 Luna and Claude Haiku 5.5 on multiple benchmarks, reaches 1,068 tokens per second, and has more than 40 times the parameters of Mercury 1.
Original text02Product Releases/Updates
DeepSeek fully open-sources its internal reasoning framework, DeepInfer
DeepSeek announces that it will fully open-source its internal reasoning framework DeepInfer, while also open-sourcing the production-grade Token encoding and decoding component deepseek-recipe.
Original textGoogle and Meta release TorchTPU, a native TPU backend for PyTorch.
Google and Meta release the open-source project TorchTPU, providing a native PyTorch backend for Google TPU; migrating models only requires changing .to("cuda") to .to("tpu").
Original textChatGPT introduces Locked Chats, allowing sensitive conversations to be locked with a PIN or biometric authentication
ChatGPT introduces the Locked Chats feature, moving sensitive conversations from regular history and search into a protected space. Users can unlock chats with a PIN on the web and desktop, and with a PIN, Face ID, Touch ID, or device unlock on supported mobile devices; new messages in locked chats will not update the memory used by regular conversations. The feature is rolling out today to the Free, Go, Plus, Pro, Business, Enterprise, and Edu plans.
Original textGoogle released native TorchTPU inference engines for vLLM and SGLang
In a PyTorch-related announcement, Google introduces native PyTorch inference engines for vLLM and SGLang based on TorchTPU, developed with Inferact and RadixArk respectively, allowing developers to deploy models on TPUs using familiar PyTorch tools.
Original text03Industry News
Reuters: OpenAI plans to accelerate raising about $30 billion before Anthropic goes public, with a valuation exceeding $1.4 trillion
According to Reuters, OpenAI is rushing to negotiate approximately $30 billion in financing before Anthropic goes public, at a valuation exceeding $1.4 trillion—at least 64% higher than the $852 billion valuation in March.
Original textSpaceX financial results show its AI business exceeded expectations, accounting for 52% of revenue
SpaceX released its third-quarter regulatory filing, disclosing $14.6 billion in revenue and $4.4 billion in net profit, reversing a $1 billion loss in the same period last year. Its renamed “Super Intelligence” business line, including Cursor, Grok, and its neocloud computing business, performed better than investors expected. AI currently accounts for 52% of the company’s revenue, and its computing capacity has reached 1.7 GW.
Original textThe US Army Cyber Command's ECHOTRIBBLE project aims to turn AI agents into autonomously proliferating cyber weapons
According to documents obtained by The Intercept, the U.S. Army Cyber Command (ARCYBER) is developing a platform codenamed ECHOTRIBBLE that allows AI agents to move autonomously across networks, identify and eliminate threat targets, and then self-destruct, while operating with minimal human oversight.
Original textRedact investigation: WhatsApp help pages hide LLM instructions to prevent AI from recommending regulators or mentioning GDPR
Redact discovered that WhatsApp Help Center pages send hidden instructions invisible to users to LLMs, directing its support chatbot not to recommend data protection regulators and not to cite the GDPR, LGPD, or CCPA after users mention lawyers or legal action.
Original textOpenAI acquires drug development platform Orchestra Bio and integrates it into the Rosalind life sciences system
OpenAI announces the acquisition of Orchestra Bio, a platform that helps biopharmaceutical teams coordinate the complex processes involved in drug development. Its team and technology will be integrated with Rosalind. Orchestra's agents can help teams track research and project progress, analyze the impact of events such as failed production batches on clinical trial timelines, and propose responses. The article also mentions collaborations with Novo Nordisk and Moderna, as well as the background that new drug development takes an average of 10 to 15 years.
Original textAnthropic’s cloud partnerships shift: AWS, Google, and Microsoft view it as a competitor
According to S-1 documents leaked to The Wall Street Journal, Anthropic is struggling to obtain sufficient cloud computing capacity from AWS, Google, and Microsoft, which increasingly view it as a competitor rather than a customer.
Original text04Paper research
Anthropic used Claude to design a protein mini-binder targeting scorpion neurotoxins.
Anthropic allowed its internal research model to operate mostly autonomously for 64 hours in Claude Science, designing protein minibinders targeting the neurotoxins of five scorpion species. It generated 318,484 designs and selected 180.
Original text05News Briefs
- Perplexity Computer can now run entirely locally on Mac, without cloud inferenceAravind Srinivas
- Apate.ai uses 350,000 AI anti-fraud agents to stall scammers, while young people in China use AI videos to fight back against AI sob-story accountsWeChat Official Account: Digital Life Kazike
- Google Releases TorchTitan TPU Fork and Provides a Model Optimization PlaybookGoogle Developers Blog
- Waymo announces its expansion into Spain, with its first passengers set to ride in the Madrid region in 2027Waymo:Blog
- METR evaluated the alignment section of Anthropic's August 2026 risk report, accepting its low-risk conclusion but questioning the evaluation methodology.METR:Blog
- Meta open-sources the Muse Spark Open weight model, launches it on Fireworks, and focuses on long-context coding and agent workFireworks AI
- Stanford HAI released MedHELM, PhysicianBench, and HealthAdminBench, finding that medical AI agents' real-world completion rate is below halfStanford HAI News
- Ai2 released the FairStress benchmark to measure hidden fairness asymmetries in LLM decision-makingAi2 / Allen Institute for AI
- ModelBest breaks down how the forward-looking expert scheduler LES enables 35B MoE to run in real time on mobile edge devices without lossOfficial account: Mianbi Intelligence
- Databricks released MemCurrent, a production-grade persistent-memory architecture for agents.Databricks:Blog