Weekly AI and NLP News — June 20th 2023

ChatGPT became cheaper and GPT-4 can now use third-party functions

Here are your weekly articles, guides, and news about NLP and AI chosen for you by NLPlanet!

😎 News From The Web

  • The first AI model based on Yann LeCun’s vision for more human-like AI. The I-JEPA model employs self-supervised learning to capture common sense knowledge about the world and avoid limitations of generative approaches. Yann LeCun's vision for human-like AI is the foundation of this model.

  • Google Lens can now search for skin conditions. Google Lens now has a feature that can identify skin conditions and other physical maladies by analyzing uploaded images. The feature can be integrated with chatbots to provide accurate answers about objects in photos.

  • Europe moves ahead on AI regulation, challenging tech giants’ power. The European Parliament has approved the EU AI Act, which regulates AI to protect consumers from risks such as discrimination and misinformation. The act bans certain tools and sets limits on high-risk technologies, promoting responsible AI development through transparency and a risk-based approach. It may also impact international policies and urge companies towards a unified regulatory landscape.

  • Paul McCartney records Beatles song with help of AI. Paul McCartney used AI technology to restore John Lennon's voice and record a Beatles song from a 1978 cassette tape labeled "For Paul." This collaboration showcases the potential of AI in music preservation and opens up new opportunities for artists and fans in the future.

  • Function calling and other API updates. OpenAI has released several new updates, including a 16k context version of GPT-3.5-turbo, a new API for calling user-defined functions, cost reductions for models and input tokens, and updated versions of GPT-4 and GPT-3.5-turbo. The updates aim to improve GPT's capabilities and make it easier to connect with external tools and APIs.

  • People Are Pirating GPT-4 By Scraping Exposed API Keys. OpenAI warns of stolen API keys being advertised for unauthorized access to GPT-4, resulting in unexpected charges for the account holder. Users are advised to safeguard their keys and rotate them immediately if exposed. Automated scans are conducted to revoke identified exposed keys.

📚 Guides From The Web

  • The Secret Sauce behind 100K context window in LLMs: all tricks in one place. Researchers have developed optimization techniques to enable Large Language Models to train on longer context lengths, including replacing Positional Sinusoidal Encoding with ALiBi, using Sparse Attention, employing FlashAttention, implementing Multi-Query attention, and using Conditional Computation to optimize speed and accuracy. These tricks enable training on up to 100K tokens, making it easier to generate longer texts.

  • Harnessing the Falcon 40B Model, the Most Powerful Open-Source LLM. The Falcon 40B is a powerful open-source LLM that is making waves in the AI community. This guide using Hugging Face can help enthusiasts and professionals start exploring its potential without barriers.

  • How GPT works: A Metaphoric Explanation of Key, Value, Query in Attention, using a Tale of Potion. The article explains how GPT models use Key, Value, and Query in Attention to understand the context of words in a text. The author uses a potion-making metaphor to describe how these elements work together. Understanding this interplay helps in predicting the next word with coherence.

  • How Me and My Internet Friends Are Using AI. The article discusses the real-world applications of AI to boost productivity and creativity, with examples such as ChatGPT and DALL-E 2. AI is useful for generating ideas and visuals but should be combined with personal touch to create exceptional content.

  • GPTeam: A multi-agent simulation. GPTeam is an open-source multi-agent simulation inspired by Stanford's "Generative Agents" paper. It allows creating a world with agents having unique personality, memories, and directives resulting in fascinating emergent behavior. The simulation setup is easy with JSON files, and it can be easily run with one command while observing agents' decision-making in real-time via a web interface.

🔬 Interesting Papers and Repositories

  • Build, Manage & Run Autonomous AI Agents. SuperAGI is an open-source framework for building and running autonomous AI agents, aimed at giving developers a powerful and quick tool to create useful autonomous agents without constant human intervention.

  • FinGPT: Open-Source Financial Large Language Models. FinGPT is a new open-source language model designed specifically for the finance sector, emphasizing high-quality financial data, and has various potential uses in finance such as robo-advising and algorithmic trading.

  • INSTRUCT EVAL: Towards Holistic Evaluation of. INSTRUCT EVAL is a new evaluation suite for large language models that assesses problem-solving, writing skills, and alignment with human values, improving practical usability and model performance scalability. The study can be found in the arXiv repository.

  • On the Reliability of Watermarks for Large Language. A study shows that watermarks can effectively detect machine-generated text despite human paraphrasing and non-watermarked language models, with detection accuracy remaining high even with GPT-3.5 and purpose-built models. The findings suggest the potential of watermarks in identifying and combating spam, social media bots, and valueless content on the internet.

  • MIND2W EB: Towards a Generalist Agent for the Web. MIND2W EB is an open-source dataset of over 2,000 open-ended tasks from 137 real websites across 31 domains, designed to develop agents that can follow language instructions and complete tasks on any website. The dataset also addresses the challenges of using LLMs to create web agents by filtering HTML using a small language model, helping GPT-like models to acquire information and perform actions on HTML websites without predefined APIs or retriever tools.

Thank you for reading! If you want to learn more about NLP, remember to follow NLPlanet. You can find us on LinkedIn, Twitter, Medium, and our Discord server!