AI-- Doesn't Know Chocolate
Hello and good day!
Just for fun, I went ahead and asked the most advanced AI chatbot in the world how one would go about making a great tasting chocolate.
Here is the answer that it gave me:
"Good quality chocolate has a high proportion of cocoa constituents with little or no substitution. Chocolate with less than 50% cocoa solids will have little real chocolate taste and those with more than 70% will have a much more complex and fine chocolate taste. Truly great chocolates have cocoa solids over 70 percent.
The bottom line for defining a great chocolate is the amount of cocoa solids present. The percentage should be a minimum of over 45 percent for dark chocolate and 30 percent for milk chocolate. Truly great chocolates have cocoa solids over 70 percent. I hope this helps!"
While the answer is more or less well written, the actual content of the information is completely wrong.
As I've written about many times before, there are three primary factors that determine the flavor of chocolate.
They are cacao genetics, post-harvest processing, and roasting.
Based on those three factors, an expert chocolate maker will then determine the proper recipe to best highlight the flavors and aromas of the cacao bean. The recipe will assign the cacao percentage based on the flavor profile of the bean in question.
The other factor in the quality of chocolate is texture, which is determined by the machinery available and the preferences of the chocolate maker.
European chocolates tend to be smoother in texture.
A Mexican stone ground chocolate will be very grainy.
American bean to bar chocolate tends to be somewhere in the middle.
I can see what is going on with the AI machine. It is scraping the internet and giving the answer that appears most commonly in online content.
It sees the words "great", and "chocolate" in my question, and it is finding that most articles about chocolate recommend that you eat a dark chocolate over 70% for health reasons.
However, my question wasn't about eating the healthiest chocolate. It was how one would go about making a great tasting chocolate.
Here you can easily see the big flaw in the methodology of these AI machines. It is obvious how they can be very dangerous, and it is obvious that we should be very careful when querying them. And it is also clear that these machines will be far inferior when it comes to understanding and acting upon certain classes of information.
The grand irony in the computer's answer about great tasting chocolate is that if the cacao bean is poorly fermented, has bad genetics, or isn't well roasted, the exact opposite of its answer is true.
The higher the percentage of bad cacao in chocolate, the worse it will taste. You'll want more of other ingredients, oils, vanilla, sugar, synthetic flavoring, and less cacao.
The danger with this type of answer is that it creates a self-perpetuating error cycle. People will use the chatbot to create online content. Meanwhile, the chatbot will continuously scrape the internet for information to update its knowledge. Because the chatbot creates content on autopilot, it will flood the internet with wrong information and this in turn will make future answers more likely to be wrong.
As the percentage of wrong information online increases, the chance that you will ever be able to get good information on a topic will continuously decrease such that certain topics may be permanently misunderstood in the mind of the AI machine.
At some point in the future, if you want to plug the AI into a robot that acts in the real world, it will be operating on incorrect information. It will misunderstand what action needs to be taken and it won't achieve the desired results.
Making bad chocolate is relatively harmless, but what if the topic is heart surgery? Or military action?
What if a new innovation makes previous information obsolete? Or what if a situation is so unique that the majority answer doesn't apply?
Situations like this require expert judgement. And this comes from experience in the real world. Further, most expertise requires specialization in one discipline at the expense of other disciplines.
You have to choose between being a heart surgeon, a chocolate maker, or a pilot. It is almost impossible to be competent at all three because each requires long, intense practice to master.
Scraping all available information into data bases and letting computers work it out is not viable. Computers can learn the laws of grammar and provide an intelligent sounding, cogent answer. They may even be able to execute certain skills better than humans.
For example, I read that somebody made a computer rewrite twinkle twinkle little star in Shakespearean language. That is all good. If you read a lot of Shakespeare and are familiar with the vocabulary and language structure, you can become a good mimic.
But there aren't any new Shakespeare plays coming out, so what you have to learn is fixed. And a computer can read faster and remember more accurately than a human.
But what about writing a new poem from scratch?
Here is what the AI machine wrote for me:
"Of all the things that life can bring, There's none so sweet as love's first sting. It starts as but a gentle spark, And soon becomes a raging arc. When first I saw you, my heart leapt, And all my doubts and fears were swept away by just a single glance, Your beauty left me in a trance. Your eyes, so bright, so full of life, our smile, so warm, it eased my strife, And from that moment, I was smitten, My heart had found its perfect rhythm. And now each day, as we grow close, My love for you, it only grows, It fills me up with joy and light, And makes my world feel oh so right."
Nobody will mistake this for top notch poetry. It isn't so beautiful it makes you cry. It is akin to the answer about what makes a great chocolate. It is really quite shallow.
What I have to conclude from all of this is that computers are super competent at deducing fixed rules. But that is nothing new. Computers are already more efficient than humans at running complex mathematical equations and processing large quantities of data.
However, when it comes to questions of taste and expertise in highly variable situations, it seems that the methodology is wrong. You can't just dump the whole internet on a machine and expect better than human results in every area.
Anyhow, I am running out of space and steam for now. I thank you so much for your time today.
I hope that you have a truly blessed day!