All About LLMs

Researchers are exploring if emergent abilities of LLMs are a mirage and Microsoft’s AI-powered Bing is now in open review. Sebastian Raschka’s latest unit all about LLMs just dropped and we’ve partnered with Stability AI on a meetup to keep AI open source on May 19th in NYC. Let’s dive in!

Research Highlights:

• Stanford researchers proposed an alternative explanation for emergent abilities in large language models, which are claimed to be abilities not present in smaller-scale models. The researchers argue that emergent abilities are not fundamental changes in model behavior, but rather creations of researcher's analyses. They confirmed their findings by testing predictions using a mathematical model and through three complementary ways. The researchers found strong evidence to support their alternative explanation that emergent abilities may not be a fundamental property of scaling AI models.
• Berkeley researchers claim that OpenAI models such as ChatGPT and GPT-4 have memorized a diverse range of copyrighted materials through a name cloze membership inference query. The extent of this memorization is linked to the frequency of passages from these materials appearing on the web. The authors argue that this finding has implications for cultural analytics as it may compromise the validity of assessments and contaminate test data, with models performing significantly better on memorized books than on non-memorized ones for downstream tasks. They suggest that open models with transparent training data can help address this issue.
• Researchers explored the challenge of managing ambiguity in natural language processing. Ambiguity is crucial for human language understanding, but it remains challenging for pre-trained language models (LMs). To evaluate the LMs' ability to recognize and disentangle possible meanings of ambiguous sentences, the researchers designed a benchmark called AmbiEnt. The study claims that the latest GPT-4 also struggles to recognize and disentangle ambiguity, and the researchers demonstrate how ambiguity-sensitive tools can identify misleading political claims. The authors encourage the NLP field to acknowledge the significance of ambiguity.

ML Engineering Highlights:

• Microsoft moved its AI-powered Bing and Edge to open preview, meaning users no longer have to wait. It is adding more visual search capabilities to Bing and is expanding the number of languages supported by its Image Creator tool to more than 100. Chat sessions are claimed to be more productive with new chat history and share functionalities, while Edge mobile will soon include page context. Microsoft is also turning the product into a platform, adding third-party plug-ins that will help people to book restaurant reservations, for instance.
• Slack announced new AI integrations to enable customers to link directly to enterprise applications without having to switch tasks. The company plans to incorporate AI natively into the user experience with SlackGPT and EinsteinGPT, and developers can also build AI steps into workflows. The new SlackGPT will let users and developers tap into AI-driven experiences, using content from Slack as a starting point for building models related to the platform.
• The UK's Competition and Markets Authority (CMA) is conducting a review of AI foundational models, including large language models (LLMs) like OpenAI's ChatGPT and Microsoft's New Bing. The review aims to examine competition and consumer protection considerations in the development and use of these models. The CMA is expected to publish the review in early September and is accepting responses from stakeholders until June 2.

Upcoming NYC Event: Keep AI Open Source!

Calling the AI community of NYC! Come celebrate the power of open-source AI in our meetup with Stability AI!

When: May 19th | 6 PM

Where: NYC | Secret Location

What: Show off your AI demos, be inspired and join the movement to democratize AI!

Tutorial of the Week

Dive into natural language processing and large language models with Sebastian Raschka’s latest unit! Start Unit 8 of his Deep Learning Fundamentals course.

Get ready to:

?? Master text data encoding for ML models

?? Explore deep neural networks for sequence data

?? Grasp attention mechanisms & finetune transformers for predictions

Don’t Miss the Submission Deadline

• BMVC 2023: 34th annual conference on machine vision, image processing, and pattern recognition. Nov 20 - 24, 2023. (Aberdeen, United Kingdom). Submission deadline: Fri May 12 2023 16:59:59 GMT-0700
• NeurIPS 2023: 37th conference on Neural Information Processing Systems. Dec 10 - 16, 2023. (New Orleans, Louisiana). Submission deadline: Wed May 17 2023 13:00:00 GMT-0700
• ICCVS 2023: The 14th International Conference on Computer Vision Systems. Sep 27-29, 2023. (Vienna, Austria). Submission Deadline: Mon May 29 2023
• ICMLA 2023L: The 22nd International Conference on Machine Learning and Applications. Dec 15-17, 2023. (Jacksonville, Florida). Submission Deadline: Sat Jul 15 2023

Want to learn more from Lightning AI? “Subscribe” to make sure you don’t miss the latest flashes of inspiration, news, tutorials, educational courses, and other AI-driven resources from around the industry. Thanks for reading!