Weekly AI and NLP News — June 12th 2023

Discovering faster sorting algorithms with RL, and LlamaIndex getting fundings

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

😎 News From The Web

  • AlphaDev discovers faster sorting algorithms. AlphaDev, an AI system using reinforcement learning, has developed faster sorting algorithms for organizing data by starting from scratch and selecting computer assembly instructions using reinforcement learning. The new algorithms are up to 70% faster for shorter sequences and integrated into the LLVM libc++ standard library.

  • Building the data framework for LLMs. LlamaIndex has raised $8.5M in seed funding and developed a toolkit to integrate user data with LLMs, enabling the creation of knowledge-intensive LLM apps such as search engines, chatbots, and analytics helpers. The project has impressive traction with 16K stars on Github, 20K Twitter followers and 200K monthly downloads, and 6K active Discord users.

  • Bard is getting better at logic and reasoning. Google has combined the capabilities of advanced language models and traditional code to enhance Bard's reasoning and math abilities. This new method of implicit code execution has boosted its accuracy by 30%.

  • RedPajama 7B now available, instruct model outperforms all open 7B models on HELM benchmarks. Introducing the new RedPajama-INCITE models, optimized for few-shot tasks, which outperform similar models on HELM benchmarks. The project analyzed differences with previous models and incorporated community feedback. The models are available under Apache 2.0 license for AI professionals.

  • ChatGPT comes to iPad, adds support for Siri and Shortcuts. OpenAI's latest app update allows iPad users to use Siri and Shortcuts with ChatGPT, one of the most popular chatbots available, offering custom shortcuts for hands-free operation. Apple has updated its App Store rules to prevent ChatGPT clones from being submitted.

  • generative AI learning path, by google. Google's new learning path on Generative AI covers everything from Large Language Models to deploying solutions on Google Cloud, with courses on Responsible AI, Image Generation, advanced topics, and a Quest to explore Generative AI Explorer in Vertex AI.

  • Japan Goes All In: Copyright Doesn’t Apply To AI Training. Japan has announced that it will no longer enforce copyrights on data used for AI training, allowing for unhindered AI research and competition with the West.

📚 Guides From The Web

  • 🤗 Open LLM Leaderboard. The Open LLM Leaderboard allows researchers to track advancements in large language models and chatbots by submitting Transformers models for automated evaluation on a GPU cluster. The leaderboard tests these models on various tasks including science questions, inference, multitasking accuracy, and truthful answers. It's a helpful resource for staying up-to-date and comparing LLM and chatbot models.

  • Why AI Will Save the World. The article discusses how AI has the potential to revolutionize various fields, including coding, medicine, law, and arts. While there are concerns about its negative impact, its benefits outweigh the risks if developed ethically and safely. AI can augment human intelligence and help us achieve better outcomes in every domain of activity while becoming an AI tutor for every person maximizing their potential.

  • Are AI startups too easy to copy? AI startups face tough competition and investors worry about their ability to differentiate themselves in a highly saturated market. VCs look for network effects and proprietary datasets to invest in successful AI startups that gain quick traction.

  • Is AI Killing the Stock Industry? A Data Perspective. AI-generated images are transforming the stock industry, resulting in boosted revenue for some agencies like Adobe Stock, but not all agencies accept them. It's still too early to determine AI-generated content's impact fully, as there are still unknown variables, and professional AI engineers are rare.

  • GPT best practices. This guide on GPT best practices discusses strategies and tactics for using GPTs effectively. It emphasizes the importance of providing context and details to help GPTs produce better results, and suggests tactics such as breaking down complex tasks and measuring performance.

🔬 Interesting Papers and Repositories

  • Video-LLaMA: An Instruction-tuned Audio-Visual Language Model for Video Understanding. Video-LLaMA is a new language model for video understanding, built on top of BLIP-2 and MiniGPT-4. It has two components: the Vision-Language component and the Audio-Language component, and can improve video accessibility by assisting in automated captioning, search, and navigation.

  • Fine-Grained Human Feedback Gives Better Rewards for Language Model Training. The article introduces fine-grained RLHF as a solution to improve language models' output quality, offering detailed rewards for explicit training signals and tailoring the language model for specific needs. This method achieves better performance than traditional methods.

  • Orca: Progressive Learning from Complex Explanation Traces of GPT-4. Orca, a 13-billion model, enhances AI models' capabilities through imitation learning and outperforms other models in complex reasoning benchmarks. It also performs well in professional and academic exams like LSAT, GMAT, SAT, and GRE.

  • Simple and Controllable Music Generation. MUSIC GEN is a single Language Model that generates high-quality music conditioned on textual descriptions or melodic features, allowing control over the output. It achieved superior quality compared to previous models and uses an unsupervised melody conditioning technique to follow specific harmonic and melodic structures.

  • Tracking Everything Everywhere All at Once. OmniMotion is a new motion estimation method that surpasses traditional optical flow and particle video tracking methods. It utilizes a globally consistent motion representation to ensure accurate tracking, estimate full-length motion trajectories for each pixel, and model camera and object motion. It outperforms previous approaches both quantitatively and qualitatively on TAP-Vid benchmark and real-world footage.

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!