--:-- Host: The AI revolution is here, but adoption is not always evenly distributed. While consumers are quick to try the latest AI apps, we haven't seen much traction in the enterprise. Awareness isn't the issue. The Financial Times reported that nearly 40% of S&P companies mentioned AI in their earnings last quarter. So, why is it so hard for enterprises to adopt generative AI?
--:-- As companies realize the value of their proprietary data, new questions emerge. Are enterprises right to fear giving away their data? Can they build better models themselves? Do they need the highest accuracy? With OpenAI's recent release of ChatGPT Enterprise, will this change everything?
--:-- Today, you'll hear from a16z co-founder Ben Horowitz and Databricks co-founder and CEO Ali Ghodsi as they answer these questions and more, including their perspectives on open source, whether benchmarks are BS, and the role of universities in this wave they helped start decades ago. You'll also hear how Databricks' acquisition of Mosaic ML fits into this.
--:-- This episode continues our coverage from a16z's exclusive AI Revolution event, featuring influential builders across the ecosystem, including OpenAI, Anthropic, and Roblox founders. Check out the full package at a16z.com/airevolution.
--:-- As a reminder, this content is for informational purposes only and should not be taken as legal, business, tax, or investment advice. A16z and its affiliates may have investments in the companies discussed. For more details, visit a16c.com/disclosures.
--:-- Host: Let's dive into generative AI. As a VC, we see companies with amazing traction, but none in the enterprise. Why is it so hard for enterprises to adopt generative AI?
--:-- Ben Horowitz: Enterprises move slowly. That's one reason. They're also super concerned about data privacy and security. They've realized how valuable their data is and are cautious about sharing it. There's also the fear of data leakage, with LLMs potentially exposing sensitive information.
--:-- Another challenge is the need for accuracy in enterprise use cases. They're being cautious, which is typical for large enterprises. Additionally, there's internal politics over who owns generative AI within the company. It's unclear if it's IT, product, or business lines, leading to slowdowns. The prize is huge for whoever cracks the code.
--:-- Ali Ghodsi: Are enterprises right to fear giving their data to OpenAI or others? They can gain value by putting their data into large models, but leaders now realize their data's strategic importance. They want to own the IP and build it themselves. This is why Databricks acquired Mosaic ML—to help enterprises build their own LLMs.
--:-- Ben Horowitz: For specific use cases, smaller, tailored models can be more effective. However, larger models have their place if you can afford the cost and complexity. The key is finding the right balance between model size and specific needs.
--:-- Ali Ghodsi: Training large models is expensive and resource-intensive. Mosaic ML helps enterprises train models from scratch, which can be better for specific tasks. However, it's not for the faint of heart. The goal is to make models more efficient and accurate for particular use cases.
--:-- Host: How do you think about fine-tuning larger models for specific tasks?
--:-- Ali Ghodsi: Fine-tuning is the holy grail. Modifying existing models to excel at specific tasks without rebuilding them is challenging. Techniques like prefix tuning and LoRA show promise but aren't yet perfected. The ideal future model would be a large foundation model with specialized, efficient add-ons for specific tasks.
--:-- Ben Horowitz: The industry is like the early internet, where everyone focuses on infrastructure. But the real value will come from applications like Uber or Facebook. Specialized models will dominate, similar to how Cisco dominated routers but didn't monopolize the internet.
--:-- Ali Ghodsi: I think there will be lots of specialization. Base models will be common, but each application will have its own tuned version. This is more efficient and accurate for specific tasks.
--:-- Host: How do you think about open source in this context?
--:-- Ben Horowitz: Open source has been crucial, like Llama, but proprietary models will always be ahead. Open source lags but can sometimes leapfrog, like Linux. However, training giant models requires significant resources, which limits open source progress.
--:-- Ali Ghodsi: Universities are struggling to keep up due to resource constraints. They're looking for ways to make AI more accessible and affordable. Open source will continue to improve, especially as techniques like efficient transformers emerge.
--:-- Host: What about the ethics of large models versus open source?
--:-- Ben Horowitz: The real risk is malicious use, similar to how hammers can be misused. AGI is a hypothetical risk, but we're far from machines with free will. Current models lack consciousness and are just sophisticated tools.
--:-- Ali Ghodsi: Benchmarks are flawed. They measure test-taking skills, not real-world application. True medical benchmarks, for example, would require secretive, unbiased tests. Current benchmarks don't reflect practical abilities.
--:-- Ben Horowitz: AGI is a concern, but we're far from it. The asymmetry in training and using models is a barrier. Reproduction and self-improvement in machines are still distant. For now, the risk is manageable.
--:-- Host: Thanks to Ali Ghodsi for joining us. If you liked this episode, share it with a friend or leave a review at ratethispodcast.com/a16z. Producing a podcast can feel like talking into a void, so your feedback matters. Thanks for tuning in, and we'll see you next time.
