In the previous article, I showed a very simple way to build a custom machine vision AI model online. However, you maybe aware that OpenAI recently released ChatGPT4 that can “look at” images.
Do you still need to know how to build your own AI models if ChatGPT can do everything?
The answer of course depends on what your jobs and goals are. I want to provide some insights and viewpoints to consider so you can come to your own conclusions.
To address the topic, I present a talk given by Dr. Andrew Ng, one of the foremost AI researchers and educators. You are encouraged to listen to the whole presentation, but I will discuss some relevant points he made.
He talks about the profound changes LLMs (Large Language Models) brought to the relationship between AI and the business world. We certainly had many forms of AI before the introduction of ChatGPT. But the introduction of LLMs made adoption of AI into business much easier.
Let’s say we want to have AI analyze Yelp reviews and categorize them into either positive or negative reviews. Before LLMs, we still could have trained a natural language processing AI model to do the job. But we would have needed a team of developers collecting immense amounts of data then building and training the AI model.
With ChatGPT, the task is as simple as asking ChatGPT: “Is this review positive or negative?” Okay, if you want to process thousands of reviews you still need a software engineer.
But you don’t need to build a language model from ground up. This cuts down the development time drastically. Many niche business that could not use AI because of the development costs can now use AI. Dr. Andrew Ng himself makes an example where even your local pizza shop can make use of AI.
This sounds all rosy and great. Until you come to the following points.
ChatGPT doesn’t actually do everything. Not yet anyway.
ChatGPT and other powerful LLMs have deceptively high cost.
Dr. Ng doesn’t cover the #2 point much in the above video. Therefore, I will get back to this point using other sources.
To illustrate #1, let me use an example cited by Dr. Ng himself in the video. He talks about an actual real-life startup that he helped create. He and other researchers discussed the possibility of using AI to help maritime shipping companies save fuel with their ships, thus not only save money but also help the environments. He discussed the process of brainstorming, finding the right people, and launching the business.
A very fascinating talk, and I wish he went deeper into more technical aspects of this system. Nevertheless, one thing is clear. They didn’t build the system by just asking ChatGPT which ways to steer the ship. It must have been a custom AI based on meticulous research, possibly built from ground up.
Dr. Ng Talks about developing AI for the fields that are increasingly more niche like the maritime shipping example he mentioned. For those fields, some level of custom developments would be inevitable, not to mention accurate understanding of how AI works.
But wait, isn’t Artificial General Intelligence around the corner? Wouldn’t we have general super-duper AI that can do everything soon? At least, Andrew Ng doesn’t think so. In the video above, he opines that AGI is still decades away. His view contrasts some other experts who believe AGI is very close. Google Deepmind CEO Demis Hassabis, for example, says we are only a few years away from AGI. I won’t judge which camp is correct in this article.
Second, ChatGPT and other LLMs are not as cheap to run as many people think. According to the following article, it costs $700,000 a day to run ChatGPT.
OpenAI is consuming billions in investments to maintain its operations. Given the current rate, OpenAI could face bankruptcy by 2024. While it's likely that additional capital will be provided to prevent this, one can anticipate that, unless OpenAI achieves a 100x efficiency gain, it may need to significantly increase its fees in the coming years.
https://technext24.com/2023/08/14/chatgpt-costs-700000-daily-openai/
I am not here to worry whether Sam Altman can afford his mortgage next year. But the important point is that running powerful LLMs costs resources.
If OpenAI cannot become profitable soon through other means, it may shift the cost to anyone using their API to build apps.
I believe Zoom is a good corporate example where they built their own AI to handle transcripts—smaller and simpler, but cheaper to operate than an OpenAI API.
Another reason to go back to simpler and customized AI models involves edge computing. Edge computing refers to the practice of processing data closer to the data's source, rather than in a centralized cloud-based system. This approach can reduce latency, as data doesn't have to travel long distances to a data center; it's processed right where it's generated, making operations more efficient. If we were to apply AI to edge computing, we would need AI models that can run on your laptop, your phone, if not your toaster oven.
I believe if you were to build a serious, huge-scale, time-sensitive AI app (finance, transportation, manufacturing), you need to implement edge computing.
More advantages of edge computing
Reduced Latency: By processing data closer to the source, edge computing minimizes the delay (latency) of sending data to a centralized cloud server and back. This is crucial for real-time applications such as autonomous vehicles and industrial robots.
Bandwidth Efficiency: Transmitting vast amounts of raw data over networks can consume significant bandwidth. Edge computing allows for initial processing and filtering at the source, ensuring only necessary data is transmitted.
Improved Privacy and Security: By processing data locally, there's a reduced risk of data interception during transmission. It can also allow sensitive data to be processed without leaving the device, bolstering privacy.
Reliability: In scenarios where constant connectivity to a central server isn't guaranteed, like remote locations or during network outages, edge computing ensures that devices can still operate effectively.
Cost Savings: Transferring less data to the cloud can lead to savings on data transmission and cloud processing costs.
ChatGPT and other cheap LLMs available online today surely provide excellent means for research and development. But I believe the importance of custom AI engineering will rise over time.
That means it would still be very important to understand how AI works at the fundamental level. Again, I believe the best way to learn is by doing. I will discuss more fundamentals of AI in future articles.