**Baidu’s ERNIE 5.0 Outperforms GPT-5.1 as AlphaProof Cracks Olympiad Math**
New AI Models & Benchmarks
Baidu Releases ERNIE 5.0, Claiming Frontier Performance Across Multimodal Domains
Baidu unveiled ERNIE 5.0, a multimodal AI model outperforming competitors like GPT-5 (High) and Gemini-2.5-Pro in text, visual, and audio understanding tasks. The model showcases advancements in benchmark performance, positioning Baidu at the forefront of multimodal AI capabilities.
-
Ernie 5.0 released, achieving frontier performance across multimodal domains
-
- Note: Refers to ERNIE-4.5-VL-28B-A3B-Thinking, a separate but related model with MoE architecture (28B total/3B active params), open-sourced under Apache 2.0.
OpenAI Launches GPT-5.1 with Enhanced Conversational Abilities and Customization
OpenAI rolled out GPT-5.1, an upgrade focused on smarter, more natural conversations and new personalization features (e.g., chat style customization). The model is being deployed first to paid subscribers, with broader availability planned.
-
We’re rolling out GPT-5.1 and new customization features. Ask us Anything.
-
- Note: Includes UI screenshots of the new model selector.
AI Research Breakthroughs
Google DeepMind’s AlphaProof Achieves Olympiad-Level Mathematical Reasoning
DeepMind’s AlphaProof combines reinforcement learning with formal languages (e.g., Lean) to solve complex math problems at Olympiad levels. The system is trained on auto-formalized problems, marking a leap in AI’s ability to verify and generate rigorous mathematical proofs.
Open-Source Tools & Developer Resources
NVIDIA Open-Sources Live VLM WebUI for Real-Time Vision Model Testing
NVIDIA released Live VLM WebUI, a tool enabling real-time video streaming to Ollama vision models with AI analysis overlays. The open-source project supports multiple backends and platforms, simplifying local VLM development and testing.
Speculative Decoding: Accelerating LLM Inference Over Networks
A developer proposed speculative decoding—a technique using a draft model to speed up LLM inference via network-side verification. The method is being integrated into the open-source ArchGW project (a sidecar proxy for agents), with community feedback sought.
Open Science & Accessibility Initiatives
Project AELLA: 100M+ Research Papers Summarized by LLMs for Open Access
Project AELLA aims to democratize scientific knowledge by using LLMs to generate structured summaries of over 100 million research papers. The open-science initiative seeks to improve accessibility and discoverability of academic literature.
Product Launches & Mobile AI
Perplexity AI Opens Early Access for Comet Android App
Perplexity AI began beta testing Comet Android, a new mobile app offering AI-powered browsing and interaction. Selected power users received early access to the beta version.
Industry & Business News
OpenAI’s Inference Costs Outpace Revenues by Billions in 2025
Leaked documents reveal OpenAI spent over $8 billion on inference costs in the first nine months of 2025—more than double its revenue for the period. The financial strain raises questions about the sustainability of its high-cost, high-scale AI operations.