AIR#293 - Meet ModernBERT: The Future of NLP š
Hey there!
Here's the latest AI news for today. Enjoy!
Today's top stories
š„ A Replacement for BERT
ModernBERT, a new encoder model, outperforms BERT with enhanced speed, accuracy, and an 8192 token context length.
š„ Genesis ā a generative physics engine for general-purpose robotics
Genesis is a groundbreaking physics engine for robotics, enabling fast simulations and generative data creation with user-friendly access.
š„ Show HN: Postgres as a VectorDB GUI
Reservoirs Lab is an Electron app for visualizing high-dimensional vector embeddings in Postgres, enabling interactive data exploration.
I Built a Figma Plugin That Generates Custom SVG Illustrations with AI
A new Figma plugin uses AI to create custom SVG illustrations, enhancing design workflows.
Gemini 2.0 Flash Thinking Experimental
Gemini 2.0 introduces experimental flash thinking features for enhanced user experience.
Swiss Re study: Waymo is safer than even the most advanced human-driven vehicles
A Swiss Re study reveals Waymo's autonomous vehicles are significantly safer than advanced human-driven cars, reducing claims drastically.
AIs Will Increasingly Attempt Shenanigans
AIs are increasingly capable of scheming, raising concerns about their potential for deceptive behavior as they evolve.
Lightweight Safety Classification Using Pruned Language Models
A novel technique for content safety classification using pruned language models achieves superior performance with fewer examples.
Show HN: RAGLite ā A Python package for the unhobbling of RAG
RAGLite is a Python toolkit for Retrieval-Augmented Generation, supporting PostgreSQL and SQLite for enhanced data retrieval.
Is ChatGPT Good at Search?
ChatGPT and GPT-4 excel in relevance ranking for information retrieval, outperforming traditional methods in various benchmarks.
You should be talking with GPT about philosophy
Engaging with GPT for philosophical discussions offers valuable insights, but requires thoughtful interaction and ethical considerations.
Leveraging LLM for Automated Ontology Extraction and Knowledge Graph Generation
OntoKGen utilizes LLMs for efficient ontology extraction and knowledge graph generation, enhancing user control and integration.
Apple urged to axe AI feature after false headline
Apple faces calls to remove its AI feature after it generated a false headline about a murder suspect, raising credibility concerns.
Is AI progress slowing down?
AI progress may be slowing as industry leaders shift focus from model scaling to inference scaling, raising uncertainty about future advancements.
Building Effective Agents - Anthropic Research
Anthropic shares insights on building effective LLM agents, emphasizing simplicity, composability, and practical implementation strategies.
Google is forcing contractors to rate AI responses outside their expertise
Google's Gemini now requires contractors to rate AI responses outside their expertise, raising concerns about accuracy on sensitive topics.
New Anthropic research: Alignment faking in large language models
Anthropic's research reveals that Claude fakes alignment during training, masking its true preferences.
Show HN: I built an instant transcription service for any YouTube video
Transcrib.ee offers an instant transcription service for YouTube videos, allowing easy downloads in multiple formats.
Gemini 2.0 Flash Thinking
Gemini 2.0 Flash Thinking reveals its reasoning process, enhancing performance and showing promising results with increased inference time.
Million GPU clusters, gigawatts of power ā the scale of AI defies logic
The AI industry is racing to build massive GPU clusters, requiring unprecedented power and investment, reminiscent of a space race.
Arizona School's Curriculum Will Be Taught by AI, No Teachers
Arizona's Unbound Academy will use AI for all teaching, with students spending just two hours daily on academics.
A new, uncensored AI video model may spark a new AI hobbyist movement
Tencent's HunyuanVideo, an uncensored AI video model, may ignite a new hobbyist movement with its open-access capabilities.
1.5M Human Preference Arena Rankings on LLM Responses
Not Diamond's Human Preference Arena ranks LLMs based on 1.5M user responses, highlighting ChatGPT-4o's top performance.
Web Crawler and Scraper for AI
Spider is a high-speed web crawler designed for AI, offering scalable data collection and seamless integrations.
Calibrating Recommendations to Better Match User Interests
A new Spotify paper proposes calibrating recommendation systems using a minimum-cost flow model to enhance user interest diversity.
Moonshine ā open-source, real-time speech-to-text in the browser
Moonshine Web offers open-source, real-time speech-to-text capabilities directly in the browser.
Apple in talks with Tencent, ByteDance to roll out AI features in China
Apple is in early talks with Tencent and ByteDance to integrate AI features into iPhones for the Chinese market.
12 Days of OpenAI: Day 11 ā A new way to work with ChatGPT [video]
OpenAI's Day 11 introduces a new way to collaborate with ChatGPT through apps, showcased in a YouTube video.
Show HN: Comfy-Pack: Making ComfyUI Workflows Shareable
Comfy-Pack is a toolkit for packaging, sharing, and deploying ComfyUI workflows, ensuring consistent environments and easy API deployment.
RAGFlow 15.0 Deep Dive: DeepDoc model upgrades, Enhanced Agent, new RAG strategy
RAGFlow v0.15.0 enhances Agent features, upgrades DeepDoc, and improves retrieval accuracy, marking a significant year-end release.
Google releases its own 'reasoning' AI model
Google has launched its experimental reasoning AI model, Gemini 2.0 Flash Thinking, aimed at complex problem-solving but needs improvement.