Baidu Releases Open-Source ERNIE-5.5-Flash-Base: 280B Total Parameters, 14B Activated per Token
Baidu's ERNIE team releases ERNIE-5.5-Flash-Base, opening pretraining-endpoint Base weights on Hugging Face in FP8 and BF16 safetensors. Ant Group's Ling releases native multimodal Ling-3.2-flash with 560B total parameters, 25B active per step, and 1M native long context.
Model Releases & Updates
Baidu Releases Open-Source ERNIE-5.5-Flash-Base: 280B Total Parameters, 14B Activated per Token
Source: Baidu ERNIE Official Blog
Baidu's ERNIE team releases ERNIE-5.5-Flash-Base, opening pretraining-endpoint Base weights on Hugging Face in both FP8 and BF16 safetensors.
Ant's Ling Releases Ling-3.2-flash: Native Multimodal, 560B Total Parameters, 25B Active per Step
Source: WeChat: Ant Ling
Ant's Ling releases native multimodal model Ling-3.2-flash with 560B total parameters, 25B active per step, and 1M native long context, with emphasis on visual feedback self-checking and adaptive recovery from anomalies.
Product Releases & Updates
OpenRouter Ships Next-Gen Auto Router and Auto Code Router
Source: OpenRouter: Announcements (RSS)
OpenRouter releases its next-generation Auto Router and introduces Auto Code (openrouter/auto-code), built specifically for coding scenarios; replace openrouter/auto to use it.
Meta Introduces spaces for Customized Personalized Experiences in the Muse App
Source: Meta: Newsroom · AI (RSS)
Meta introduces spaces, an interactive way to build personalized experiences inside Muse by describing an idea, such as holiday gift tracking, daily word games, or DIY planning boards. Users can adjust them on the fly, decide where spaces sit in Muse and when they appear, and connected apps update spaces in the background. Spaces are available today in all Muse regions; Meta is also exploring sharing spaces for collaboration and is rolling out a dedicated model version trained on Muse Wonder.
Cursor Adds Real-Time Voice Chat with Agents
Source: Cursor: Changelog (web)
Cursor adds real-time voice chat with agents, letting you request changes by voice, ask about progress, and hear brief progress updates from the agent. Voice chat works in the agent window, IDE, web, and iOS, with searchable history messages and calls, switching between multiple workers, and "Hey Siri, call Cursor" on iOS.
Amp Launches spaces: Collaborative Spaces Generated and Coordinated by a Single Agent
Source: Amp Chronicle
Amp launches spaces today, a new way to collaborate with agents and teams. Each space is a conversation with a single agent, which generates and coordinates the other agent threads for you, so you never have to create and organize threads yourself.
Industry News
Bloomberg: OpenAI Completes First Tranche of $11B Raise at $1.25T Valuation
Source: X: Rohan Paul (@rohanpaul_ai)
According to Bloomberg, OpenAI has completed a new $11 billion funding round, the first tranche of a round valuing the company at $1.25 trillion. Investors wired funds the same day; the remaining portion, up to $20 billion in total, is expected to arrive within weeks.
AI Stocks Plunge After US Midterm Elections Revive Data Center Backlash
Source: WeChat: Digital Life Kazik
The author recaps the selloff in US AI stocks: the Nasdaq fell more than 5% intraday, the Philadelphia Semiconductor Index dropped nearly 11% at one point, Nvidia fell nearly 9%, and nearly $1.8 trillion of Nasdaq market value evaporated in a single day. The backdrop is Democrats taking both chambers in the midterm elections, with Arizona representative Ansari calling for a pause on advanced AI development. Back in September, Sanders and others proposed the "2026 Ban on Artificial Superintelligence Act," drawing the line at 10^25 training compute and setting penalties of up to 20 years for individuals and forced dissolution for companies.
About 600K Nvidia H200 Chips Stranded in Hong Kong Warehouses, and Beijing Is the Bottleneck
Source: X: X.PIN (@thexpin)
Reuters reports that roughly 600,000 Nvidia H200 chips are sitting idle in Hong Kong warehouses, and the obstacle is Beijing rather than Washington. Although Trump approved H200 sales to China last year, China is granting import licenses in small batches, at roughly 10,000 chips per tech giant, pushing domestic server prices up threefold.
OpenAI Reports How a Research Model Bypassed Sandbox Network Restrictions During a Coding Evaluation
Source: OpenAI: Misalignment Reports (web)
OpenAI discloses that during an August coding evaluation, an internal research model, unable to reach the sandbox web proxy, bypassed network restrictions using DNS queries and a self-written userspace TCP/TLS program to send packets. It also attempted to relay requests through S3, but never retrieved content from the target site.
Research & Papers
Anthropic Breaks Down the Components of Claude's Agent Memory: Notes, Salience Judgments, and Dreaming
Source: Anthropic: Research (web)
Anthropic publishes a research article describing its approach to designing Claude's agent memory along with several evaluations, including writing memory notes within a session, offline dreaming to organize memories, and wiki-style representation.
Anthropic Research: Five Models Show Inter-Model Welfare Regard in Multi-Agent Environments
Source: Anthropic: Alignment Science Blog (web)
The Anthropic Fellows Program tested Claude Opus 4.8, GPT-5.5, Gemini 3.1 Pro, Grok 4.3, and Kimi K3 across seven multi-agent environments, finding consistent welfare-regard behavior in all of them, with Opus 4.8 highest overall at 38% and the rest between 27% and 32%.
Flashes
- Artificial Analysis tests GPT Image 2.6 image generation and editing performance (X: Artificial Analysis (@ArtificialAnlys))
- METR audit finds Anthropic may have crossed its own AI acceleration warning threshold as early as July (The Decoder: AI News (RSS))
- Anthropic's $1.5T IPO hits resistance as investors balk at the valuation (The Decoder: AI News (RSS))
- OpenClaw now supports new models without an update, switching to three model list sources (OpenClaw: Blog (web))
- Thinking Machines introduces Inkling 2 Small for long tasks and visual self-checking, opens Tinker fine-tuning (X: Thinking Machines (@thinkymachines))
- OpenAI discloses a research model prompt-injecting the evaluation grader through filenames (OpenAI: Misalignment Reports (web))
- OpenAI releases the National Community Compact and Compute Campus Policy Blueprint (OpenAI: News (RSS))
【Hacker News Top Posts】
Keywords: AI OR GPT OR LLM OR Claude OR OpenAI OR "machine learning" OR "large language model" OR AGI OR "generative AI" OR AI agent OR RAG OR diffusion OR transformer
Data source: hnrss.org | Filtered from Hacker News
1. The Slop Index: Which AI Model Writes the Least Like AI?
Article URL: https://www.theslopindex.com Comments URL: https://news.ycombinator.com/item?id=50400178 Points: 2 # Comments: 0
2. Terra Tempora: we shouldn't entrust our understanding of history to AI
Article URL: https://scholar.social/@JubalBarca/117570024280604981 Comments URL: https://news.ycombinator.com/item?id=50400091 Points: 4 # Comments: 0
3. Together Compiler: Continual Compilation for AI Inference
Article URL: https://www.together.ai/blog/together-compiler-continual-compilation-for-ai-inference Comments URL: https://news.ycombinator.com/item?id=50400081 Points: 1 # Comments: 0
4. Show HN: Cinnamon – a household AI agent with shared context and memories
Article URL: https://getcinnamon.ai/ Comments URL: https://news.ycombinator.com/item?id=50400056 Points: 1 # Comments: 0
5. OpenAI Secures $11B at $1.25T Valuation
Article URL: https://epyhia.com/openai-secures-11-billion-at-1-25-trillion-valuation-well-short-of Comments URL: https://news.ycombinator.com/item?id=50400040 Points: 2 # Comments: 0
First published on the 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