10 Tips About AI That Will Never Change
Because of how ingrained AI tools, like ChatGPT, are in our life it is essential to understand the basic fundamentals of how these models work. While many articles talk about “prompt engineering” tricks or “the best AI tools”, the goal of this article is to explain AI and AI-adjacent principles that are timeless*. I want readers to open up this article in 10 years and still find value. The more you understand about how these AI tools work, the more you will understand how to use them well– which is an incredibly powerful skill. Below are 10 timeless principles on AI that I believe will last your lifetime:
If You Have a Hammer, Everything Looks Like a Nail
If you have an amazing set of AI tools, everything looks like a problem solvable by AI. This is a mirage. Not everything is solvable by AI nor is AI the best tool for many problems. You need to choose the right tool for the problem you want to solve.
AI Tools are Pattern Matching Machines, Not Magic
AI Tools are built on top of sophisticated algorithms that are incredible at finding patterns (i.e. neural networks maximizing an objective function). Humans are incredible pattern matching machines, too. Any pattern matching process; whether AI-driven or human-driven, is prone to mistakes.
Data is AI’s Fuel and All Data is Biased
All AI tools are fueled by data. The AI model might have been pretrained, trained, finetuned by data and/or it could retrieve data as part of the AI system.
All data is biased no matter which method you use for data collection. This implies that all AI tools are biased. While all data and all AI tools are biased, some are more so than others.
Garbage in, Garbage Out
“Garbage in, garbage out” is a relevant concept for ANYTHING that relies on data. That includes every AI tool, AI model, dataset, information repository, etc. Most, if not all, large language models (LLMs) that AI tools are often built upon utilized internet data as part of their training dataset. If there is garbage on the internet, there could be garbage coming out of the AI tool.
AI is Limited by the Data it Has Access to
AI does not inherently have access to all information nor is AI born with the same 5 senses that humans have. AI needs help from humans and/or other tools and instruments to access information. Any information that can be collected and transformed into a machine-readable format could be used to train/finetune an AI model or be used by an AI tool.
No AI Tool Has “Human Memory“
Any AI tool that appears to have memory (or remembers what you said) is simply a set of sophisticated engineering behind the scenes to mimic “human memory”. Some methods (and tools) that help mimic memory include prompt engineering, caching results, databases, combining prompts before feeding the AI model, and information retrieval algorithms. None of these methods (and tools) will perfectly replicate the equivalent of “human memory” but they might try to.
AI Predicts What’s Most Likely, Not What is Certain
The most common AI tools (e.g., ChatGPT, Gemini, Claude, Grok) are powered by LLMs which are inherently probabilistic. They output responses that are most likely correct and will produce different responses despite being given the same input. Alternatively, deterministic processes produce the same output every time you give the same input. Deterministic processes are consistent. Some common deterministic features include sorting by best sellers on your clothing website, navigation buttons on a browser window, calculator mobile applications, or data preprocessing scripts (not powered by LLMs). Probabilistic features include Instagram’s social media feed, Spotify’s shuffle mode for playlists, or Netflix’s “Recommended for you” section. Some implications of this include:
- Mistakes from a probabilistic process will be difficult, if not impossible, to identify in many situations
- It might never be possible to detect AI-generated output**
- If you don’t like the output from an AI tool, you can repeat the same input and get a new result.
- If you see an AI tool that automates data collection, preprocessing, and analysis, it might have made some mistakes if it is powered by a probabilistic algorithm for any of the steps.
You Only See the Tip of the Iceberg
While you might see a simple website/app, there could be MANY deterministic and probabilistic algorithms powering the features on that website/app. Netflixs’ applications are likely powered by a combination of sophisticated reranking models, A/B testing to optimize UI/UX, natural language processing to categorize movie titles, and more. The best software tools will appear simplistic and intuitive, but will utilize a combination of deterministic and probabilistic algorithms to power their features. ChatGPT is NOT good at math so if you ask it to do math, ChatGPT will likely attempt to feed the numbers into a deterministic process (e.g., calculator tool) to perform the calculation and then output an answer to you. The AI tool user might only see what they input into the AI tool and what the AI tool outputs to them.
Many AI Tools Cannot Explain Why
For many complex AI systems, it is difficult to trace back exactly how they arrived to a particular answer. The lack of transparency and the lack of human understanding of neural networks, which underpin large language models, makes them less explainable. This is why you will often hear them referred to as black box models. Humans nor machines are able to explain how and why neural networks produce specific outputs. AI tools will attempt to provide traceability and to explain their reasoning, but it will always be limited by our understanding of the neural network algorithms that underpin them.
It’s Not the Tool—It’s Who Uses It
A carpenter is better at diagnosing construction problems that require a hammer, better at understanding how to use the hammer to solve the problem, and better at understanding whether the hammer fixed the problem. The same principles apply to AI tools. Ideally, you want a person who understands the subject matter of the problem and understands how to wield the AI tool appropriately to solve the problem.
~ The Data Generalist
Data Science Career Advisor
*None of these statements are likely to change [for the mainstream] anytime in the next 10+ years. Technology advancements that could change any of these statements would require huge progress such as if/when AI reaches the singularity, brain computer interfaces (BCI), nuclear fission, quantum, or using new AI architecture that is significantly better than the Transformer for processing data.
**Assuming the current (or similar) architecture of the large language models that rely on “black box” neural networks