Given the current fuss about OpenAI's newest release, the ChatGPT3 model, I wanted to share my thoughts about it. I played with it for a little bit, followed what people used it for or say it can be used for, and tried to find some more practical uses that will enhance my current workflows.

Long story short, I'm just going to give my honest opinion right out the gate, and then I'll go into the details of why I think that. In my opinion ChatGPT3 at this moment is good for three things: entertainmet, some idea generation (but very rudimentary, and if you try to iterate on the same area of idea generation, it will very fast become repetitive, and at best you will get the same overall idea but with changed words) and accessing some simple high-level information, fast.

Unfortunately I didn't find ChatGPT useful in other situations, and it will not make anyone's job redundant anytime soon. More than often, it offers wrong information (confidently). A lot of people are impressed by the code it writes, but usually it produces very simple code snippets that have no real practical use, and it contains subtle bugs that would not be introduced by a real competent programmer (especially with the error handling).

What can ChatGPT be used for?

For entertainment purposes, you can spend some time giving it prompts to see what it produces. This activity can keep you engaged for some minutes, but loses fun pretty fast. At least for me, after 5-6 prompts I run out of ideas of things to give it do, so I close it down. But after a break of either a few hours or days, you get new ideas to resume your "fun".

Regarding the code it writes, it is not producing good enough code for any direct purpose. Maybe just learning, for beginners, but that's the best it can do. And the code that it produces often is overly verbose (we don't need comments for each line!), and contains subtle bugs.

But with some refinement it can provide some value for developers, and it is really useful as a way to access the documentation of the code and get some snippets to get you started. I can definitely see the value in that, as guidance and as a quick replacement for some StackOverflow usage.

But what it misses the most in this aspect (and StackOverflow clearly has the edge) is the extra context and comments on the subject. More than often, on StackOverflow you also get some side-comments about on what version of some library it works for, some links to their documentation, some alternative ways to do it, etc. Usually ChatGPT just gives you comments about what each line of code does (even for the simplest assignments).

For quick information access, it is good enough, as long as the information you request is pretty basic and not too complicated. I found it very good for accessing science related information (because for exact sciences there are fewer ways to be subjective about a question), and not that good for getting recommendations. It is programmed to avoid tricky topics such as politics, sport events and other things that could be used for ulterior motifs, but overall, the information you can access with it is pretty decent. I will definitely use it to get fast information I need for some projects, and get some starting points for some ideation processes (for example, getting some spell names for a game I am working on).

What about Dall-e?

Dall-e is another interesting AI project, that can give you some generated images based on a text prompt. It is pretty good at generating basic objects, but I found it having very big difficulties generating anything with a face.

It can definitely be useful to generate some placeholder art for some projects, but for more "industrial" use-cases, I found it to be not good enough. Sure, the prompts you give also play a big role in the quality of the generated images, but without some clear guidelines or information about what produces better results, most people will just brush it off.

A cool combination I tried was using ChatGPT to give me some better prompts to use with Dall-e, and ... it didn't really raise to my expectations. Usually, the prompt ChatGPT gives back is basically a more detailed version of your original prompt, but with more details that I feel don't really add much value (ex. adding non-visual words such as "powerful" (what does powerful look like, it is certainly up to the reader how they interpret this term) or the prompt it gives is actually less specific that my original prompt (ex. for "give me a prompt to use for dall-e to generate an image with an alien planet seen from space, which has some lower orbit objects flying around" I got back "Generate an image of an alien planet seen from space, with several smaller objects orbiting around it.", which as far as I know, the Dall-e prompt should contain just the description of what you want generated instead of a phrase directed to Dall-e directly.

Conclusion

The advancement of generative AI is definitely awesome, but we are still a long way to getting the real-life benefits and automatize our work even more. Right now, the scopes of the work they can produce are pretty limited, and can at best be used at a hobby level. Professionals will still have to rely on humans for a while.

The biggest use-case for ChatGPT right now, that directly impacts me, is getting a direction for some code problems I encounter. For example, for the prompt "show me how to save the image from an url as pdf" gave me a good starting point. I still had to do all the error handling myself and tweak the code a little bit, but it was definitely faster than searching for the specific pieces of knowledge independently and then combining them myself (so it definitely helped me save some time with that). It would serve a good basis for quick answers to some questions, but I don't really know if it is in the state to provide a sustainable business. At the moment, all its usage is free, and it consumes a lot of resouces to train and use (the GPT-3 model it uses has billions of inputs), and in the moment it will become a paid service, 99% if the current users will back off, as its value given isn't really worth anything right now (you still have decent free alternatives, and it's a nice to have thing rather than a must have).

But one this is for sure: I am looking forward to seeing the advancements, as I feel in the past years we've seen bigger and more frequent breakthroughts in this field.