Uncover Eye-Popping Insights to Turbo-Charge YOUR Understanding Today from the Long-Awaited Book by an AI Authority (Part 3)
Here comes the third part of insights to better understand AI from Ethan Mollick’s book, “Co-Intelligence”. This part will be mainly about AI’s creativity and hallucinations. Actually, the final insight in Part 2, Insight #6, was also about hallucinations of AIs. When I look at my notes, this was by far the most interesting topic in the book. If you want to check the first and second parts of this article series, here are the links:
Here are some very interesting insights about hallucinations.
Insight #7:
“Apart from technical issues, the AI can also develop hallucinations based on the data it’s fed. This data might be flawed, incomplete, conflicting, or even outright wrong. The AI doesn’t have the ability to tell the difference between opinions, creative writing, facts, metaphors, or trustworthy and untrustworthy sources. It might end up adopting the biases and prejudices of the people who provided, sorted, and adjusted the data.”
The idea of “garbage in, garbage out” is a very common precept in data science. It means that if you feed a computer or AI system bad or flawed information, you’ll get bad or flawed results. In simpler terms, if you put junk into a machine, you’ll get junk out of it. So, when the data going into an AI is messed up, the outcomes or decisions it makes will likely be messed up too.
Having “clean data” is crucial in data science. It’s like having a solid foundation for a building — if it’s shaky or weak, everything built on top of it will be shaky too. Clean data means accurate, reliable, and well-organized information. When data scientists work with clean data, they can trust the results they get from their analyses. They can make better decisions, find meaningful patterns, and create more accurate predictions. Without clean data, their work becomes much more difficult, and the results may not be trustworthy. Almost all written information that is available online is usually fed into LLMs. It is practically very difficult to filter “garbage” and maintain only “clean data” in the learning material set. So, what is the solution?
“Reinforcement Learning from Human Feedback (RLHF)” is a crucial aspect of AI, where machines learn and improve by receiving feedback from humans. Humans provide feedback based on the actions or decisions of the AI, helping it understand what’s desirable and what’s not. This continuous loop of feedback allows the AI to adjust its behavior over time, becoming more proficient and, more importantly, more accurate at its tasks. This is an integral part of “co-intelligence” to sort out reality from hallucination.
Insight #8:
“It can be amusing when the AI gets mixed up between what’s real and what’s not. Take this example: A data scientist named Colin Fraser found that when he asked ChatGPT for a random number between 1 and 100, it said “42” about 10 percent of the time. Now, if it was truly picking numbers randomly, “42” should only come up about 1 percent of the time. Fans of science fiction probably already know why “42” shows up so often. In Douglas Adams’s book, The Hitchhiker’s Guide to the Galaxy, “42” is the answer to the “ultimate question of life, the universe, and everything” (although it leaves us wondering what the question actually is). This number has become a joke online. So, Fraser thinks that because there are so many references to “42” online, the AI is more likely to see it and think it’s a good answer — even though it’s just hallucinating that it’s giving a random response.”
This is truly fantastic. Random number generation is actually not a simple feat. Computers are really good at following rules and patterns, but not so great at being truly random. They rely on algorithms and formulas to generate numbers, which means there’s always some predictability lurking beneath the surface. Plus, factors like the starting point of the algorithm or the time it takes to generate the number can influence the outcome, making it less random than we’d like. That’s why creating truly random numbers often requires specialized hardware or collecting data from unpredictable sources, like atmospheric noise or radioactive decay. It’s a tricky business, but crucial for tasks like cryptography, gaming, and simulations where randomness is key. The example of “42” is a clear manifestation of the need for human awareness about hallucinations, or simply drawbacks of LLMs.
Insight #9:
“You can’t really ask an AI why it’s imagining things because it’s not aware of how it works. So, if you ask it to explain, it might give you an answer that seems right, but it’s not connected to how it came up with the original idea. The AI can’t really explain its choices or even remember what they were. Instead, it just tries to make you happy by saying something that sounds good. These language models aren’t programmed to say “I don’t know” when they’re unsure. Instead, they’ll give you an answer, sounding pretty sure of themselves.”
This is self-explanatory. When the answer is satisfactory, it is also very difficult to catch hallucinations that are subtle. So, even human feedback sometimes is not sufficient. On the other hand, I do not think that this problem is more harming than deliberate acts of deception made by regular people or even scientists at reputable institutions. Consider the case of “fact-checkers”. After all, fact-checkers are just people, and people are naturally biased. So, expecting them to be completely objective and unbiased might be a bit naive. Their judgments can be influenced by their own beliefs, perspectives, and even agendas, which can cloud their ability to provide an entirely impartial assessment. While fact-checkers certainly play a valuable role in scrutinizing information, it’s important to recognize their inherent subjectivity and consider multiple sources when seeking the truth. Therefore, we should not blame AIs for hallucinating. They are no worse than some “pundits”. We should only be on the lookout, alert, and aware.
To be continued with “Part 4”…
