???: The CEO keeps asking me to do something with artificial intelligence...
Haebom
1
Recently, Foundation LLM has shown excellent performance in various tasks. As a result, many industries are drooling over the introduction of AI, but I personally think that we need to think about it a little at this point. Right now, there is no ROI in terms of cost, and in a time like now when new technologies and models are continuously emerging, it is inevitable to end up in a chicken game. If we blindly introduce AI, the number of users may increase, but the deficit in retention and sales will increase. (Personally, I think that a cost-effective alternative will come out within 1~2 years.)
However, if you think you need to introduce AI, I recommend introducing a small and lightweight model. It's like saying, "Don't kill a chicken with a slaughter knife." Sometimes, you need a model for a specific task. There are many concerns about the role and importance of a model for a specific task (task-specific), and about jumping into it at this point. There are three main reasons.
When handling personal data
If your personal data, such as legal, medical, business, etc., is not publicly available on the web, you need a model for a specific task where the model needs to know about this data. And this data must be a significant amount. Otherwise, it can be easily solved by a model with a long context or search.
When performing domain-specific tasks
Even when you need to do something fast and cheap in a specific domain, models for specific tasks are useful. For example, Codex requires only coding skills and can be used quickly and cheaply at scale. Similarly, rather than using a large and burdensome LLM to solve a domain problem, it is better to develop a small, domain-specific model. (Of course, the limitations are clear in areas outside the domain.)
If you are looking for the best performance
When you want to get the best performance on a specific task, you need a model for that specific task, if you can easily keep up with the development of the next GPT-N+1 model. Even if you train a model for that specific task and beat GPT-4, it is only a short-term satisfaction. As the GPT-n+1 model grows in scale and gains access to better inference and knowledge, there is a risk that the model for that specific task will fall behind. (In the end, it is not a battle of money, but a battle of learning methods and speed.)
Conclusion
Models for specific tasks are needed when dealing with personal data, performing domain-specific tasks, or pursuing best-in-class performance. However, given the pace at which general models are evolving, focusing too much on a specific task can be dangerous.
Subscribe to 'haebom'
📚 Welcome to Haebom's archives. --- I post articles related to IT 💻, economy 💰, and humanities 🎭. If you are curious about my thoughts, perspectives or interests, please subscribe. Would you like to be notified when new articles are posted? 🔔 Yes, that means subscribe. haebom@kakao.com
