ML twitter is in a flurry over ToolFormer, a model that is getting LLMs to use tools in a self-supervised manner. New implementations of ChatGPT continue to enter the market and an AI agent recently piloted Lockheed Martin’s VISTA X-62A for more than 17 hours. Let’s dive in!
Research Highlights:
- Researchers from Meta are claiming that language models may learn how to use external tools through straightforward APIs. They presented Toolformer, a model trained to select which APIs to contact, when to call them, what arguments to send, and how to best incorporate the outcomes into future token prediction. Without compromising its basic language modeling capabilities, Toolformer claims to achieve much enhanced zero-shot performance across a range of downstream applications, frequently competitive with much bigger models.
- A 22B-parameter ViT (ViT-22B) was trained using a recipe provided by Google researchers, and the resulting model was subjected to a wide range of trials. It is stated that ViT-22B exhibits scaling up performance when tested on downstream activities. Other advantages of scale that the researchers noted included a better trade-off between fairness and performance, state-of-the-art alignment to human visual perception in terms of shape/texture bias, and greater robustness.
- Researchers from Copenhagen introduced MarioGPT, a fine-tuned GPT2 model trained to generate tile-based game levels– in their case Super Mario Bros levels. According to the authors, MarioGPT is the first text-to-level model and can not only create a wide variety of levels but can also be text-prompted to produce controllable levels, resolving one of the main issues with current PCG methods.
ML Engineering Highlights
- You.com, a search engine startup with headquarters in San Francisco announced the release of YouChat 2.0, a new “multimodal conversational AI” system. According to YouChat 2.0, it is the first web search to bring together cutting-edge conversational AI with user-created apps. YouChat 2.0 may respond to user queries with charts, photos, videos, tables, graphs, text, or code inserted thanks to its C-A-L (Chat, Applications, and Links) blended big language paradigm.
- Yext, an online brand management platform, unveiled Yext Chat, an AI-powered chatbot for enterprise use cases. This chatbot is claimed to be differentiated by a partially proprietary back end. Yext Chat aims to be able to interface with current systems when it launches to the public, such as ticketing systems and Slack workspaces.
- At the U.S. Air Force Test Pilot School (USAF TPS) at Edwards Air Force Base in California, an AI agent recently piloted Lockheed Martin’s VISTA X-62A for more than 17 hours. This was the first time AI had been used to a tactical aircraft. It is anticipated that the experimental training aircraft will provide the framework for a future wave of planes totally controlled by computers.
Tutorial of the Week
Lots of talk lately about ML models and the applications that they power. To demystify the process of actually building an AI-powered application, we put together this tutorial that covers how we deployed OpenAI’s Whisper into a simple but powerful web UI.
Community Spotlight
Want your work featured? Contact us on Discord or email us at [email protected]
- Another week, another merged docs PR from the community. We love to see it. Cleaning up documentation is a great way to land a PR in a high-profile repo, get used to GitHub’s workflow, and work your way up to becoming a pro. 😎
- Open MatSci ML Toolkit is a single framework by Intel Labs for prototyping and scaling out deep learning models for materials discovery, built on top of OpenCatalyst, PyTorch Lightning, and the Deep Graph Library. The OpenCatalyst Project, jointly developed by Fundamental AI Research (FAIR) at Meta AI and Carnegie Mellon University’s (CMU) Department of Chemical Engineering, encompasses one of the first large-scale datasets to enable the application of machine learning techniques containing over 1.3 million molecular relaxations of 82 adsorbates on 55 different catalytic surfaces.
- This blog post from our community member Justin Goheen covers hyperparameter optimization by creating a custom Lightning Logger to use with Lightning. It uses the Weights & Biases experiment manager and demonstrates an efficient way to perform hyperparameter optimization. Nice!
Lightning AI Highlights
- The AI Buzz, Lightning’s very own podcast hosted by StatQuest’s Josh Starmer and our CTO Luca Antiga, is now available on Spotify and Apple Podcasts. The first three episodes are currently available, with more to follow!
- Rumor has it that Lightning has some cool things in store for the next few months. 👀 If you want to be the first to hear about our latest upgrades, products, and features, make sure to join us on Discord where we hang out, build together, and recruit active users for beta and early access testing!
Don’t Miss the Submission Deadline
- UAI 2023: International conferences on research related to knowledge representation, learning, and reasoning in the presence of uncertainty. (Pittsburgh, USA). Submission deadline: Sat Feb 18 2023 03:59:59 GMT-0800
- IROS 2023 : International Conference on Intelligent Robots and Systems. Oct 1 – 5, 2023 (Detroit, Michigan). Submission deadline: March 1, 2023
- InterSpeech 2023: International conference on the science and technology of spoken language processing. (Dublin, Ireland). Submission deadline: Thu Mar 02 2023 03:59:59 GMT-0800
- CoLLAs 2023: 2nd conference on lifelong learning agents. (Montreal, Canada). Submission Deadline: Tue Mar 07 2023 03:59:59 GMT-0800
- ICCV 2023: International Conference on Computer Vision. Oct 2 – 6, 2023. (Paris, France). 1. Submission deadline: March 8, 2023 23:59 GMT
