Should Open-Source AI Prioritize Developing Foundation Models?

Berkeley researchers argue that there are more viable solutions than model imitation. Google announces Search Labs at Google I/O, the EU resorts existing laws to regulate AI, and Sebastian Raschka released the last unit of his popular Deep Learning Fundamentals course. Let’s dive in!

Research Highlights

• UC Berkeley researchers conducted a critical analysis of a method to enhance weaker language models by fine-tuning them on outputs from stronger proprietary systems like ChatGPT. The researchers finetuned several language models to imitate ChatGPT using various base model sizes, data sources, and amounts of imitation data. The authors claim that the imitation models initially showed promising results, further evaluations revealed that they did not significantly close the performance gap with ChatGPT on tasks that lacked strong support in the imitation data. The researchers conclude that model imitation is not a viable solution and suggest that improving open-source models should focus on developing better base language models rather than imitating proprietary systems.
• Researchers from the University of Edinburgh and Heriot-Watt University claim that Large Language Models (LLMs) struggle to generate accurate Python code when default function names are swapped, perhaps indicating a lack of understanding of programming semantics. Some LLMs in the study exhibited increased confidence in their incorrect predictions as their model size increased, contrary to the expected trend of improved performance with larger models. Their study claims that despite relatively good performance in typical scenarios, LLMs still lack a deep, abstract comprehension of the content they handle, rendering them inadequate for tasks that deviate from their training data, emphasizing the need for more than mere scaling to enhance their capabilities.
• Researchers from Stanford developed an approach called Direct Preference Optimization (DPO) that they say achieves precise control of large-scale unsupervised language models (LMs) without the complexities of reinforcement learning from human feedback (RLHF). By leveraging a mapping between reward functions and optimal policies, DPO claims to solve the constrained reward maximization problem through a single stage of policy training, eliminating the need for reward model fitting, LM sampling, and extensive hyperparameter tuning, while being significantly simpler to implement and train.

ML Engineering Highlights

• Google opened up access to its new generative AI capabilities in Search through its Search Labs program, allowing users to sign up for experiments and test the Search Generative Experience before its wider release. The AI-powered Search experience aims to help users understand complex topics faster by providing a snapshot of key factors to consider when entering a query. It also offers quick tips, shopping integrations, and the ability to ask follow-up questions.
• OpenAI is pursuing a new way to fight A.I. ‘hallucinations’ wherein models generate false information. The approach, called "process supervision," focuses on training AI models to reward correct steps of reasoning during problem-solving, aiming for more explainable AI behavior. While the research has been met with some skepticism, OpenAI's efforts to address logical mistakes and misinformation in AI systems are seen as steps toward building more capable and accountable models.
• Character.AI, the a16z-backed chatbot startup, tops 1.7M installs in first week . The app was developed by Noam Shazeer and Daniel De Freitas, who previously worked at Google on the LaMDA language model. Character.AI's growth is said to be attributed to the popularity of its web experience.

Evolving AI Policies

• Europe turns to existing laws to regulate cutting-edge AI like ChatGPT due to the absence of specific AI legislation. The EU is addressing AI concerns, with the AI Act set to provide comprehensive guidance on AI tools, facial recognition, and biometric surveillance. However, in the absence of dedicated AI regulations, national data protection authorities are leveraging existing laws such as the General Data Protection Regulation (GDPR) to address privacy and data breaches related to AI applications.
• OpenAI reconsidering European presence after EU's AI Act . The divergence between the US and the EU in their approaches to AI regulation and responsible innovation poses challenges and potential inefficiencies in the multi-trillion-dollar industry. The EU's regulatory influence, known as the "Brussels Effect" may be weakened if major companies like OpenAI choose to leave, which could impact global tech norms and innovation.
• Top executives say AI is risky. On Tuesday, CEOs of leading AI companies, including OpenAI, DeepMind, and Anthropic, along with hundreds of other AI scientists and experts, released a unified statement highlighting the risks posed by AI to humanity. The letter calls for mitigating the risk of extinction from AI and emphasizes the need for global prioritization of AI safety. The letter's signatories occupy influential positions within AI labs and tech companies, and it comes at a time when governments and organizations are increasingly recognizing the urgency of regulating AI.

Tutorial of the Week

Sebastian Raschka just wrapped his Deep Learning Fundamentals course!🚀

Start Unit 10 now

Key takeaways:

🔍 Considerations for trustworthy & reliable prediction pipelines.

💪 Constructing confidence intervals for ML models.

🔬 Exploring Fabric: an open-source library to boost PyTorch models with multi-GPU & mixed-precision training.

Don’t Miss the Submission Deadline

• **AI World Barcelona 2023 :** International Conference dedicated to the field of generative AI and autonomous agents. September 7 - 8, 2023. (Barcelona, Spain). Submission Deadline: Wed Jun 07 2023 16:59:59 GMT-0700
• CoRL 2023: International conference focusing on the intersection of robotics and machine learning. Nov 6 - 9, 2023. (Atlanta, Georgia). Submission Deadline: Fri Jun 09 2023 04:59:00 GMT-0700
• ACML 2023: The 15th Asian Conference on Machine Learning. Nov 11 - 14, 2023. (Istanbul, Turkey). Submission Deadline: Sat Jun 24 2023 04:59:00 GMT-0700
• WACV 2024: IEEE/CVF Winter Conference on Applications of Computer Vision. January 3-7, 2024. (Waikoloa, Hawaii). Submission Deadline: Sat Jul 15 2023: Wed Jun 28 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!