AI is learning fast. Comparing GPT-3, GPT-3.5, and GPT-4, you can see the apparent growth, and it happens within one year or even less. When we applauded for Stable diffusion to paint a picture from words, Midjourney has nearly conquered the challenge of drawing fingers. The heroes behind them are scientists and engineers who feed the whole internet information to hundreds or thousands of the most powerful GPUs and tens of millions of users to provide human feedback. They are like human kids to learn, practice, and iterate. But never go to sleep or play. Within certain domains, it won’t take a long time that AI will be more precise than humans.
LLM model is more generic to remove bias. During a particular time, learning everything means mastering nothing (may not be right in AI). To prevent biasing to one topic over another one. LLM model is trending to build with general capability and avoid an apparently wrong answer. That leaves room for further fine-tuning to reduce the scope and focus on specific domains. Alpaca and LoRa are two interesting projects to fine-tune LLM models and miniaturize them for inference deployment.
The open playground is a low-cost, crowdsourcing, human-feedback community, and it also fosters customers for future momentize. Nearly all the LLM communities have opened the model for a “playground” where customers could play with different prompts and thumb up/down the result. Taking ChatGPT as an example, it takes a fast speed to reach 100M customers and at least billions of training/feedback data, just like there are 100M teachers to teach one kid. That’s why it could grow fast.
The fine-tuned model is smaller enough to be deployed economically. The LLM usually has 10B level parameters (aka Gb size) which generally requires A100 to infer, and the operation cost is too high to commercialize. However, the community never stops, and several projects could miniaturize the model to Mb size and inference with CPU. That makes the future business model reasonable.
What’s next? AI requires a flagship business company like Microsoft for office, Google for search, Amazon for e-commerce and cloud and computing, and Facebook for social media. That will prove the value of AI. Microsoft and Google have explored to integrate it into productivity software like Office and Workspace. But it may not be enough. I see two potential directions –
a) be an engine of metaverse. Yes, that was the hottest name and later thrown into Ariane Trench. However, if AI could generate dialogs — verbal communication, AI could paint pictures — visual interaction. Then it could build a metaverse for you and your friends
b) model fine-tunes and host MaaS solution. As I said, the LLM model is very generic, and data is the critical differentiated value for a business. How the company leverages LLM and easily fine-tune them for their specific business purpose will be next opportunities