ChatGPT

Lacking words to describe how utterly impressive it is.

A very impressive version 1.0

The first version of a new piece of technology is often unimpressive. Here are the specs for the first version of the iPhone:

Just look at that. 128MB of RAM, what can you run on that? 4GB of storage, not RAM, storage. That is 4GB of storage for the entire OS, apps, and data.

15 years later it's clear that despite what the first version may have looked like the iPhone is highly useful. Today's iPhones are much better. They have orders of magnitude better specs. This is the nature of emerging things, initially fairly terrible but they improve, sometimes rapidly.

ChatGPT is a new technology. For a version 1.0, it is an unbelievable piece of technology.

The impact on my workflows

It has completely changed how I operate in my day-to-day job. Here are examples of questions I have asked it:

• How do I parallelize the apply function in Pandas?
• how can I see how much data is in an s3 bucket under a certain key?
• on Linux can I list swap usage per process?
• Please critique the outline of this presentation?

The answers that come back are mind-blowing. Most of the time they are correct, and not only that. They are well formatted and always explain the answer in a way that is easy to understand. For example, all shell commands will explain all flags and options. Furthermore, the answer is aimed at my specific question, not a related question as would be the case on stack-overflow.

This is revolutionary for my workflows in many cases. For example, I am not familiar with the pandas API. I find the documentation dreadful and generally unproductive to interact with.

As a consequence, I have stayed away from pandas and notebooks for a very long time. I certainly wouldn't work with them in my spare time. This has changed completely now.

I can just plainly ask ChatGPT how to do something and it will give me a working example.

Tips and tricks

I only really have one, but it is very important.

You can significantly cut down on the hallucinations you get by rephrasing questions in terms of "can I do X" instead of "how do I do X". ChatGPT will respond correctly to "can I do X", but will sometimes hallucinate when prompted to tell "how do I do X". It will almost always also explain how something can be done if it can be done.

After reading "The Product Manager's Prompt Book"

The premise here is pretty simple. He gives a list of prompts, mostly to ChatGPT, that can automate a lot of the textual work that a product manager does.

This includes

• Meeting agendas
• Product roadmaps, requirements and acceptance criteria
• Marketing material and release notes

In general, this looks rather impressive. Certainly, a human expert could probably improve a little bit and there are lots of things ChatGPT can't do (like moderate a meeting). However, it looks like it takes a very good Product Manager to be able to compete. Or more interestingly: A very bad product manager will level up significantly by using this.

This is a pattern I have seen before.

When the Internet first came out we had a similar situation. We would often run into various situations that were simply blockers:

• The printer stopped working
• The computer would stop working
• etc.

For many of these things, one would often need to call a technician and wait. You would be stuck, sometimes for hours or days or weeks until a human could fix the problem. What the internet did was give people a low-effort way to solve (sometimes trivial) problems themselves. For a little while this turned some people into superhumans simply because they were willing to type a few words into a search engine and follow the lead from there. The discipline of IT support shows that this is still a useful skill! :D

Another area where the internet had a similar effect was home cooking. Before the internet, you would be limited to whatever recipes you had at hand at home. The internet gave people access to a huge amount of recipes and cooking tips. Before the internet, you would have to freestyle or stick to the tried and true. Freestyling without a lot of experience is usually difficult and the results were often questionable. The internet made it possible for people to cook much more interesting meals at home. Again the new technology significantly leveled up people who had low skills but were willing to put in the effort to learn.

One of the relatively obvious consequences of this is that the bar for expert advice rises. Since fairly competent advice was now at hand for free for many tasks that meant that expert advice now had a higher bar. One place where this is on prominent display is in physical stores. Many places have salespeople who are mostly there to facilitate discussions with people who are too lazy to search online for readily available information. The internet meant many of these in-store salespeople became obsolete. This happened slowly, but it is still going on. When this started clothing was one of the areas that were expected to never be hit. How can online provide the kind of advice you need to buy well-fitting clothes? This turned out to not be an obstacle. Quick and cheap shipping meant that you could try on clothes at home and return them if they didn't fit. I haven't bought clothes in a physical store in years. I simply have no reason to go there, the salespeople in the stores know nothing that I can't figure out myself. They provide zero value. You still have physical clothing stores today, but they usually come in a couple of flavors

• Stores where the staff are just there to facilitate the sale and don't know anything about the products: H&M, Zara, etc.
• Stores where the staff will help: Most of these stores cater to the older crowd, and this breed of store is dying.
• High end stores where the staff are experts: These stores are doing fine, but they are a niche market (tailored clothing).

The internet has been great for troubleshooting specific problems. One area where with room for improvement is broader problems. For example, I can paste in a specific error code and usually get fairly high-quality answers. I get a highly useful answer to a very specific question. However, if I have a more general problem like: I would like to buy a high-quality low price toaster that is easy to clean. This is more difficult. Usually, google just lists a bunch of "best toasters" lists. All of these lists are designed to be slightly useful to a wide range of people rather than highly useful to a small number of people. This is what ChatGPT solves. It can ingest all of these "slightly useful to many people" lists and a specific problem and provide an answer that is highly useful to a specific person.

That is the lasting impact I expect generative models to have. They will level up base-ability of people who are willing to learn how to use the new tools similar to the internet. This will for a time allow some people to be superhumans, because their base ability in a wide variety of tasks is now competently high.

The effect on education

Traditional schooling has a similar problem to "best toasters" lists. They also provide a service that is designed for breadth rather than depth. I predict that with ChatGPT we will see parts of the education system being completely revamped. Instead of relying on textbooks, students will have access to a highly personalized learning experience through chatbot-like interactions with ChatGPT. The main function of human input in education will shift from teaching to testing and certification. While some base material will still be necessary for ChatGPT to provide useful answers, the role of publishers will likely change. The question of whether publishers will still be necessary in a chatbot-driven education system remains to be seen.

The impact of ChatGPT on education will likely resemble the impact of the internet on physical salespeople. Teachers who are not significantly better than ChatGPT in providing personalized and tailored advice may become obsolete, while teachers who have strong connections with students and provide individualized support will remain in high demand. This shift could lead to a more efficient and effective education system, where students receive the guidance and support they need to achieve their goals.