AI: What is it good for?

The term “Artificial Intelligence” (AI) was first used back in 1956.

From the beginning, most Computer Scientists hated the term Artificial Intelligence.

Thanks for reading! Subscribe for free to receive new posts and support my work.

AI is certainly not intelligent in the human sense. It cannot “understand” concepts without extensive human training. The logic it uses is completely different from that used by humans. Regardless of how human-like a computer program may behave, as we sit here today, computers cannot yet in any sense have human understanding or consciousness.

It is also not artificial because it is currently constructed using vast amounts of natural and computing resources, and human labor.

But, the train has left the station. AI is a broadly accepted term now, and it is fruitless to continue pointing out it is not accurate in the strict sense.

To gain a perspective on this topic, you need to start with its history.

In the mid 1960s, Joseph Weizenbaum at MIT created what we now call an AI Chatbot (a Chatbot is a computer program that you can type questions, and receive an answer from the computer) called “ELIZA”. Its most famous incarnation was DOCTOR, which simulated a psychotherapist of the Rogerian school (in which the therapist often reflects back the patient's words to the patient). You can chat with DOCTOR here: https://web.njit.edu/~ronkowit/eliza.html

To Weizenbaum’s surprise and shock many people, including his own secretary, believed they were actually chatting with a human while using DOCTOR.

In 1972 a program called PARRY modeled the behavior of paranoid schizophrenics. Psychiatrists only successfully identified about 52 percent of the schizophrenics. (i.e. they were no better than a lay person guessing).

So AI has a long history, and Chatbots are not a recent invention.

What has captured our attention in the past few years is an AI innovation called Large Language Models (LLMs). Most lay persons now are using the terms AI and LLM interchangeably, but AI is the broader concept.

LLMs are a significant breakthrough in the development of AI. From Wikipedia: They are a type of machine learning model designed for natural language processing tasks such as language generation. LLMs are language models with many parameters, and are trained with self-supervised learning on a vast amount of text.

I realize people outside Computer Science may not fully understand the preceding paragraph. You can think of LLMs as “compressing” human text information in a form that computers can use to quickly answer questions from humans. Oversimplifying, LLMs take a sequence of words and predict the subsequent word or phrase.

In the past couple of years, as a good example, LLMs have gotten very, very good at writing computer code. I will go out on a limb and say that writing code is one of the things LLMs are most suited for. This isn’t a surprise if you have ever written a computer program. The grammar rules of programming languages are much, much less complicated than natural language (e.g. English, Chinese).

There is an important caveat though. You have to learn how to ask code-generating LLMs what you want. A person with experience can get much better results. i.e. it takes practice to get good results.

The models that can manipulate and extract information from images and videos are also impressive, but, again, they take practice.

Right now LLMs are for “power users”. They seem simple to use, but you need quite a bit of understanding and experience to realize their potential, and avoid the “gotcha’s”.

Let’s talk about the “dark” side of LLMs.

The environmental impact of LLMs is a problem. For example, the burden on the electrical grid is enormous and growing. Big companies are spending billions building new data centers. The incoming president is talking about subsidizing this buildout.

Are these data centers going to be needed? Who knows?

The problem is that the LLM (AI) hype is like a run away train now. Would you want to be the CEO who decided not to build, and be proved wrong in a few years?

Some people have pointed out this is like the way railways were over-constructed in the 1800s. Enormous investment and massive environmental impact, which resulted in a huge numbers of duplicate lines from different companies. (Not to mention bankruptcies).

The aforementioned image and video models, though impressive overall, have displayed racial bias.

You have probably already seen examples of LLMs “hallucinating”, i.e. giving obviously false or worthless results. The problem here is how do you know if a model is “hallucinating” when it is not obvious?

In the last year we have seen the word “Slop” associated with LLMs. Slop describes AI-generated content that is both unrequested and unreviewed. My favorite example is the LLM-generated internet article listing a food bank as a tourist attraction. There is a LLM-generated book about mushroom foraging sold on Amazon that contains false (and dangerous) information.

From the NY Times (quoting Simon Willison): “Society needs concise ways to talk about modern A.I. — both the positives and the negatives. ‘Ignore that email, it’s spam,’ and ‘Ignore that article, it’s slop,’ are both useful lessons.”

All you have to do to start a fight on social media is to say “LLMs are useful”. There are plenty of reasons — the environmental impact, the lack of reliability (e.g. “hallucinating”, “slop”), ethics (e.g. racial bias), and the potential impacts on people’s jobs.

It is wise to be skeptical about AI (LLMs in particular), but we need to acknowledge there are good applications, and the models are continuously improving.

Thanks for reading! Subscribe for free to receive new posts and support my work.