The Future is Fine Tuned (with Dev Rishi, Predibase)
May 24, 2024
auto_awesome
Dev Rishi, CEO of Predibase, discusses the future of AI fine-tuning, challenges in model hosting, and competition with OpenAI. They explore the benefits of multi-modal language models, optimization methods like Lorax, and the advancements in audio modeling. The conversation also touches on balancing work and personal life in the AI industry.
Fine-tuning language models is the future of AI optimization.
Prior experience with language models facilitates swift use case implementation.
Innovations in audio model development overcome complexities in speech data.
Deep dives
Prettbase's Evolution Towards Focusing on LLMs
Prettbase, founded in 2021 with a focus on Ludwig, a low code framework for building deep learning models, shifted its focus to Language Model Models (LLMs) in 2023. The company's unique approach revolves around fine-tuning and deploying these task-specific LLMs. This strategic move was influenced by the demand and relevance of deep learning, specifically in the realm of large language models, leading Prettbase to refine its offerings towards LLM fine-tuning and deployment.
The Significance of Language Models in Early Industry Adoption
The early application focus on language models stems from their flexibility and prior adoption in various use cases. Language models like the transformer architecture have excelled in language and computer vision tasks. The established foundation in language modeling has facilitated swift use case implementation, demonstrating the advantage of prior experience and existing datasets in driving industry adoption.
Challenges and Advancements in Audio Model Development
The launch of advanced audio models, such as GPT-4, showcases a significant advancement in low-latency audio modeling, surprising many with its real-time response capabilities. The complexities in audio modeling, including cleaner text data, present challenges compared to the inherent complications in speech data. Innovations in training data usage, synthetic data creation, and magnetic architectures contribute to overcoming these hurdles in audio model development.
The Evolution of Fine-Tuning AI Models
The podcast delves into the evolution of fine-tuning AI models, highlighting how optimizing hyperparameters used to be challenging, but has become more accessible. By partnering with companies like Gretel and leveraging synthetic data, the process of fine-tuning models for specific tasks has become more efficient. The discussion emphasizes the significance of tackling data-related challenges in machine learning, especially in scenarios where no specific data exists.
Advancing Towards AGI and Future AI Functions
The conversation also explores the path towards achieving Artificial General Intelligence (AGI) and the potential future roles of AI in various industries. The speakers discuss the importance of addressing narrow, repetitive tasks through automation before moving on to more complex AI applications. They envision a future where AI can handle a multitude of specialized tasks efficiently by utilizing specialized adapters and routing approaches for different use cases, leading to a more creative and impactful AI landscape.
Dev Rishi is the founder and CEO of Predibase, the company behind Ludwig and LoRAX. Predibase just released LoRA Land, a technical report showing 310 models that can outcompete GPT-4 on specific tasks through fine-tuning. In this episode, Dev tries (pretty successfully) to convince me that fine-tuning is the future, while answering a bunch of interesting questions, like:
Is fine-tuning hard?
If LoRAX is a competitive advantage for you, why open-source it?
Is model hosting becoming commoditized? If so, how can anyone compete?
What are people actually fine-tuning language models for?
How worried are you about OpenAI eating your lunch?
I had a ton of fun with Dev on this one. Also, check out Predibase’s newsletter called fine-tuned (great name!) and LoRA Land.
Remember Everything You Learn from Podcasts
Save insights instantly, chat with episodes, and build lasting knowledge - all powered by AI.
