Artificial Intelligence Trends That are Here to Stay
There is a well-known quote from Jeff Bezos about how he plans for the future.
“I very frequently get the question: ‘What’s going to change in the next 10 years?’ And that is a very interesting question; it’s a very common one. I almost never get the question: ‘What’s not going to change in the next 10 years?’ And I submit to you that that second question is actually the more important of the two…”
Understanding what is not going to change in AI and AI-adjacent fields over the next 10 years is a challenging task, but likely a more reasonable one than predicting what will change. If I had a chance to wager on which AI trends are here to stay (psst Kalshi), I’d put money on the following.
Unstructured Data Continues to Grow in Importance
Unstructured data (e.g., voice, text, image) accounts for ~90 percent of the overall digital data universe. The amount of unstructured data that exists has grown incredibly fast thanks to the internet and will continue to grow thanks to GenAI tools like ChatGPT that make the generation of data incredibly easy. This feeds the incentive for software engineers and data science professionals to continue to develop novel methods for extracting insights from unstructured data. Some of today’s mainstream methods include optical character recognition (OCR), sentiment analysis, and computer vision. In the next ten years, unstructured data will continue to grow* in volume, importance, and insights.
*Unstructured data has already been growing in importance. This is a statement that the trend will continue.
Compute Continues to get Cheaper
“The observation that the number of transistors on computer chips doubles approximately every two years is known as Moore’s Law.” This law has accurately modeled the improvements in compute from around 1965 to 2020. Whether or not you believe Moore’s Law will be a good model moving forward, the trend of compute getting cheaper over time will likely continue. Not every AI use case necessitates using a State-of-the-Art (SOTA) model. Software engineering is littered with approximate algorithms that closely mirror the accuracy of SOTA algorithms. One common example is the approximate nearest neighbor (ANN) algorithm that mimics the more compute-intensive, but accurate k-nearest neighbors algorithm (KNN). Hugging Face has already developed several SMOL models which are smaller and less compute intensive. Compute is virtually a guarantee to get cheaper from a hardware and software perspective.
Python and SQL Maintain Popularity
Arguably the most successful use case for GenAI/LLMs has been its ability to generate code. A key detail of this success is that they are only successful at generating code for some programming languages. Many of these LLMs were trained on the internet, websites like StackOverflow, and Github. Therefore, the models often struggle with languages that were “outside the original distribution” of the LLM training data set. Because python is open source and has been popular for 10+ years, there are a ton of python script examples that feed these LLMs. SQL has been around for decades and PostgreSQL is open source so there is a ton of SQL code available on the web as well. Because LLMs are proficient at these languages, it will create barriers to entry for newer languages. Why should a developer write in a new language from scratch when it can generate python or SQL in seconds with an LLM? Ever heard of network effects?
Software Eats the World
It doesn’t matter whether all of the new AI use cases pan out. Either way, software will continue to eat the world. Waymo uses computer vision to help develop the software for driverless cars, mobile apps power marketplaces everywhere (e.g., Uber, AirBnB, etc.), and self-checkout machines are in thousands of stores. Developing and managing software products is a core competency of at least 7 of the top 10 largest U.S. companies. Betting that software will continue to automate more of our world is a guarantee.
Systems and Data Grow in Complexity
Software decays over time. Maintaining that software comes with a cost. The more systems that exist in an organization, the more complicated it will become to manage and maintain those systems. Layering AI on top of systems will only increase the complexity (and cost) of managing the software because machine learning models decay over time, too. On top of that, AI models introduce new vulnerabilities, making network and information security even more difficult to manage. The most common example might be CAPTCHA that is no longer impervious to AI.
Every system produces data. More systems equates an increase in the volume and variety of data. This necessitates larger investments in managing and wrangling the data if leaders want reliable information to make decisions.
Conclusion
If there’s one thing we can predict, it’s that technology will continue to defy our expectations. To help plan for the future, organizations should utilize Jeff Bezos’ mental model of focusing on what is likely not to change over the next 10 years. The AI trends that are most likely to continue are the increase in unstructured data, the deflationary advancement of compute, a continued popularity in python and SQL, the continuation of software eating the world, and the increasing complexity of systems and data in an organization. Organizations would be wise to account for these trends in their AI strategy.
~ The Data Generalist
Data Science Career Advisor