Great article by Ashu Garg and Jaya G. from Foundation Capital on the large-scale adoption of LLMs and what lies ahead in this field. They delve into why we still don’t see widespread use of LLMs in production, pointing out: "There are also a host of adoption challenges, including the fact that most company data needs significant preprocessing before it can be used by a model, and use cases that worked in demos quickly break down when scaled up to production." Solving for this needs no magic, but first principal problem solving. You might have smartest people available to work for you (read SOTA LLMs), but you do need to onboard them to your task to expect quality work. Similarly, to productionalize LLMs, focus on personalizing them for your use case. Here are some straightforward steps: 1. Record and Refine Data: Start by refining your production datasets or MVP GPT-4 calls to create robust training data. Curate a golden dataset for effective evaluation. 2. Fine-tune: Fine-tune popular LLMs that perform well in your specific tasks, and practice them to perfection. 3. Evaluate: Assess performance using your golden dataset or conduct A/B testing. 4. Deploy: Deploy the model on infrastructure that meets your traffic RPS (requests per second) requirements. 5. Repeat and Upgrade: Regularly align your model with the latest research and upgrade accordingly. This is a full-time endeavor for an Applied Researcher, but emerging tools promise to automate this process effectively. The future lies not in delegating intelligence but in owning and maximizing its effectiveness for your use case.
Enterprise VC-engineer-company builder. Early investor in @databricks, @tubi and 6 other unicorns - @cohesity, @eightfold, @turing, @anyscale, @alation, @amperity, | GP@Foundation Capital
What’s next after LLMs? Every CEO, whether they make computer chips or potato chips, has announced an AI strategy. But deployment of LLM-based applications remains nascent, and LLMs have plenty of shortcomings. At Foundation Capital, we think 3 innovations will be huge for building in AI’s next era. 1. Multimodal models are moving beyond text. Inputs / outputs as images, audio, and video opens up a world of new use cases. Customer service chatbots get faster + more accurate voice agents. Multimodal unlocks healthcare, where there’s a mixture of data types. GPTs get deployed for cancer treatment 2. Multi-agent systems transform automating complex tasks. Hands down, autonomous agents calling multiple interacting systems is one of the biggest breakthroughs since ChatGPT. Multi-agents transform AI from passive tool to active player 3. New model architectures will address some limitations of transformers. State-space models (Cartesia), large graphical models (Ikigai), RWKV. Post-transformer architectures have the potential to be equally (or more) performant than LLMs, especially for specialized tasks, while also being less computationally intensive, exhibiting lower latency, being easier to control Founders may feel they’re building on quicksand, with every layer of the AI stack moving fast. But the flywheel also presents a once-in-a-generation chance to build magic. ✨ Latest blog in the comments 👇
Virtual CMO and Go-to-Market Builder for Tech Startups
2wAshu and Jaya have highlighted a critical challenge in the adoption of LLMs, and their strategies for overcoming it are spot on. I’ve seen how critical it is to tailor LLMs to specific use cases. When an LLM is implemented for customer service automation, the key to success was curating highly representative training data directly from our historical customer interactions. For marketing content writing, giving content based on a library of inputs from previous materials greatly improves the quality...