How will enterprise AI strategies evolve over the next decade?
Here are 3 crazy insights from Naveen Rao (a PM leader at Databricks) in a recent a16z podcast.
1/ New chips may be built more often to keep up with latest AI model innovation.
AI models have meaningful updates as often as every ~6 months. Even GPT-4 saw major updates 6 months into being launched. With the funding companies get these days it isn’t unrealistic to build new chips optimized for the latest models every ~6 months. Crazy when you think about it!
2/ We may start seeing custom chips for AI model architectures.
Right now the industry as a whole (including silicon innovation) is standardizing on the transformer architecture of AI models like GPT-4. As AI models evolve, we may see customized chips that are optimized based on the needs of the AI model architecture and use-case
3/ It may be cheaper and, wait for it… BETTER to run a small fine-tuned model than a large-scale LLM for enterprise use-cases.
Enterprises are considering training smaller models with domain-specific data (down to 1/100th the cost and size of a GPT-4 like model). These smaller models provide more contextual responses unlike a larger, more generic model that requires a lot of data and $$$ to train and serve.
— —
Big takeaway:
As enterprise AI use-cases and patterns start to emerge, we’ll likely see more specialized end-to-end solutions — all the way from the Silicon to model training and inference.
— —
Link to the full 45min a16z episode: https://a16z.com/podcast/scoping-the-enterprise-llm-market/