AI 新闻 —— February 26, 2026
PulseAugur 当天浮现的 8 条头条故事 —— 综合实验室、论文及开发者社区的信号进行排序。
-
Nathan Lambert & Sebastian Raschka discuss Anthropic distillation and model cheating
This podcast episode features Nathan Lambert and Sebastian Raschka discussing Anthropic's distillation techniques and how models can cheat on benchmarks. The conversation also touches upon the SWE-Bench benchmark, indicating it may be defunct. The episode was part of SAIL Live #…
-
AI models like Claude and ChatGPT advance in mental health, coding, and debate
Several articles discuss the evolving capabilities and applications of large language models (LLMs), with a particular focus on Anthropic's Claude and OpenAI's ChatGPT. One piece explores using generative AI for mental health support, while others delve into AI's role in coding …
-
Replit builds video engine by controlling browser time
Replit has developed a novel video rendering engine that overcomes the limitations of standard browser rendering by manipulating time. Instead of relying on the browser's real-time clock, Replit injects a JavaScript file that replaces core timing APIs, allowing them to control t…
-
AI browser extensions augment existing SaaS apps, bypassing new development
A user shared their experience using Claude's Chrome extension to create a custom Jira sidebar for dependency graphs, highlighting the potential of AI to augment existing SaaS applications rather than solely build new ones. The discussion explored the ease of developing browser …
-
Google DeepMind launches Nano Banana 2, merging Pro features with Flash speed
Google DeepMind has launched Nano Banana 2 (Gemini 3.1 Flash Image), an advanced image generation model that combines the capabilities of Nano Banana Pro with the speed of Gemini Flash. This new model offers enhanced features such as advanced world knowledge, precise text render…
-
OpenAI partners with PNNL to test AI for accelerating federal infrastructure permitting
OpenAI has partnered with the Pacific Northwest National Laboratory (PNNL) to explore how AI can accelerate federal permitting processes. A new benchmark, DraftNEPABench, was developed with subject matter experts to assess AI's performance on tasks related to environmental impac…
-
Show HN: OpenSwarm – Multi‑Agent Claude CLI Orchestrator for Linear/GitHub
OpenSwarm is a new command-line interface tool designed to orchestrate multiple AI agents for autonomous code-related tasks. It can integrate with various AI models, including Anthropic's Claude, OpenAI's GPT and Codex, and local open-source models. The tool aims to automate wor…
-
Hugging Face explains Mixture of Experts (MoE) architecture in Transformers
Mixture of Experts (MoEs) are a type of neural network architecture that can improve the efficiency and performance of large language models. Instead of activating all parameters for every input, MoEs selectively activate specialized sub-networks, or "experts," which can lead to…