Enterprise VC-engineer-company builder. Early investor in @databricks, @tubi and 6 other unicorns - @cohesity, @eightfold, @turing, @anyscale, @alation, @amperity, | GP@Foundation Capital
Nobody calls in a SWAT team to change a light bulb. This overkill approach doesn't work in everyday life, and it won't be the way forward for AI either. On one side, you have players like OpenAI who believe in going all-in on gigantic models. Their explicit goal is AGI, and they're willing to spend seemingly without limit to get there. But Snowflake, Databricks, and others in the "small-but-mighty" camp envision a different future. They believe the key to unlocking AI's potential isn't solely about scale. Instead, it's about combining multiple, specialized models into compound AI systems that can handle complex tasks from start to finish. Rather than forcing one huge model to learn everything, it's about teaching a team of smaller models to work together efficiently. More in my latest newsletter:
Ashu Garg - Compound AI Systems, designed for specific use case is the way to go. Some additional notes 1. narrow AI/fine-tuned LLMs 2. Domain-specific SLMs 3. Language Models with Embedded Code 4. Deterministic validators and quality filters 5. Multi-agent systems will be critical parts of the compound AI systems
What are the odds we might enter a different world where our time tested ideas and principles no longer work?
Agentic nature will force the AI in the direction of highly specialised models that are safe and reliable.
Ashu, how does eightfold.ai approach the balance between large models and specialized systems in its AI strategy?
Love reading your blogs man, great insight every time 🧠
Couldn’t agree more
Very interesting factoid Ashu Garg on "JPMC possessing 150 petabytes of data, >150x the size of the dataset used to train GPT-4." At SentinelOne we're innovating along the "small-but-mighty" trend of enabling customers to use their own data for automating their security posture. There is potential to integrate S1's Data Lake with Databricks and Snowflake to extend the AI capabilities we can offer to customers.