My personal experience with AI chatbots - Part 1

Artificial Intelligence tools I have used and pros and cons of each in my use case

The coffee chain Starbucks is famous for making highly customized (some even label them “customerized”) orders. But if they were to make a beverage right now based on the flavour in town it would be “Artificially Intelligently Delicious Frappuccino”

Quite simply put AI and text to text and text to image AI’s have exploded in the last one year, prompting both awe and horror. This article is not a social commentary or tips and tricks. Rather its my experience with AI based tools.

Its important to give a short background of myself first. I have worked in corporate sector as a tech researcher. Iam currently pursuing academic research. One of the key things in academic research is you are always looking for a way to get a leg up. So, we are looking for new and useful things all the time. Which got me into AI based tools for both images and text.

First let us go over text to text AI bots starting with

1.??????ChatGPT

ChatGPT has taken over the world by storm, ChatGPT itself is a bot built on top of GPT. GPT is a large language model currently in its 4th iteration. Before it went paid ChatGPT was my go-to for generating titles, giving me snippets of code in R and C#. In terms of fluency of conversation ChatGPT is hands down the best there is. The success of such bots is determined by how well they understand contextual cues. ChatGPT is the industrial leader by a wide margin at this point. It’s my second favorite text to text AI bot.

Pros: Super easy to use

Cons: The latest version is paid

2.??????Bing Chat

Bing Chat is essentially a chatbot within Bing search, which is Microsoft’s competitor (if you can call it that to Google search). Bing Chat is based on GPT4 and after some initial hiccups where the AI was too free (do read about ReviewGeek article on how they made it go nuts). Now the model I feel is too restrained. While I love that it has three modes (Creative, Balanced, Precise) its no go list can be pretty unforgiving at times. Still presently its my go to bot. Why? It’s integrated with Edge (which has been my default browser for 3 years because Chrome eats RAM faster than a hungry hippo) the ability to switch between search and text to text data without Alt Tabbing is an underrated bonus feature.

Pros: Perfect integration with Bing search and Edge browser makes MS Cloud workflow super smooth, three modes and also its FREE

Cons: Sometimes restricts some bizarre prompts

3.??????Google Bard

The newest kid on the block, Bard went public yesterday (which would be 10th May 2023) and is based on LaMDA language model. And its pretty rough compared to the competition. Tell you what you can test this your self

I asked Bard a rather simple query

“Write me a code in R to make a Arellano bond GMM model on heterogenous panel data”

Arellano-Bond is a GMM model often used for heterogenous panel data and is very popular in academic circles (aka me) its reply is pasted below

“I do not have enough information about that person to help with your request. I am a large language model, and I am able to communicate and generate human-like text in response to a wide range of prompts and questions, but my knowledge about this person is limited. Is there anything else I can do to help you with this request?”

So, despite Google claiming its trained on R, Julia and 20+ something programming languages. Its still not as good as ChatGPT in either reading cues or generating responses.

I contrast, here’s what Bing search or Bing chat generated in the precise mode

# Load the plm package

library(plm)

# Load your data

data <- read.csv("your_data.csv")

# Convert your data to pdata.frame format

pdata <- pdata.frame(data, index = c("id", "time"))

# Specify your model formula

formula <- y ~ lag(y, 1) + x1 + x2

# Estimate the model using the pgmm function

model <- pgmm(formula, data = pdata, effect = "twoways", model = "twosteps")

# View the model summary

summary(model)

Which is a better response. So yes, Google Bard has a long way to go

Pros: Its free

Cons: Does not understand context very well as of now

As I also write stories, I often visualize on a chart or through images (stock or otherwise) my story. Something commonly referred to as storyboarding. These are text to image bots and there’s so many of them, that I have lost count. But I will review primarily four. And yes, all of them are not good at human features especially hands or in my very specific case, weapons.

1.??????Midjourney

Currently in its 5th version, when it comes to photorealism. Nothing matches midjourney as of yet. The ability to over specify what you want, upscale and even look at a large vibrant community’s creations make it the go to for hardcore professionals. I have at times created images on it which have fooled people into thinking these were photographs. The problem, pricing. Midjourney was free in the past, not anymore. As the models have improved, processing times have also increased, leading to you getting less for more. Any text to image bots offloads their processing to a server and basically charge you for minutes or time used per generation. Midjourney costs the most.

Pros: In terms of photorealism the best bot there is, great vibrant community gives you many ideas on prompts etc.

Cons: The price and struggle with certain art styles

2.??????Dall E or Bing Image creator

?Have clubbed these two together as they are one and the same thing. This was the text to image bot which started the whole craze. But with time I feel it has fallen behind Midjourney in all but one category. Which is surreal and horror-based imagery. Dall E understands art styles slightly better but the outputs leave a lot to be desired if you are going for photorealism. Still, it is very generous in terms of free credits (the Bing Image Creator gives me 100 credits a day, that is between 100 to 400 prompt-based generations a day. But again, there is a tradeoff, in Bing image creator (based on Dall E) there is not creative upscaling, Dall E charges credits for upscaling.

Pros: Great at surrealist, horror images, better application of artstyle

Cons: Not as photorealistic, lots of prompt restrictions, some of which are bizarre (like try entering any weapon name into Bing image creator like say the AR 15)

3.??????Stable Diffusion

Presently on version 2.1 this bot is absolutely free, you can create as many images as you want. But that obviously comes with the catch that the images are not that detailed. One very good use of this is to create backdrops, like say forests, outer space etc. If you want very high customizability, Stable Diffusion is the go-to. But installing it and using it offline is bit of a process and the user interface leaves a lot to be desired (it runs on python currently and performs better if you have Nvidia GPUs)

Pros: Free, highly tweakable

Cons: GUI could be better, not as precise

4.??????Leonardo.ai

A new but quickly rising kind on the block Leonardo.ai allows you to use multiple models (including Stable Diffusion 2.1) to generate images. I have used it for about a month and an impressed with its abilities to do facial features and human faces. But besides this, a lot remains to be done with respect to more abstract imagery (which to me alongside “does it look like a photo”) are the tests a text to image AI bot simply has to pass for me personally. Very generous with credits. For every day you get 150 credits, each generation of 4 images costs 2 credits. So, you can do theoretically 300 per day.

Pros: Good at facial features

Cons: Not good with abstract ideas yet

In a future version of this article, I shall look at Office 365 which has been enhanced with a lot of AI features, follow up on Google Bard and hopefully get access to Adobe’s Firefly (its text to image AI). Adobe Firefly from what I’ve seen has a lot of potential. Have a nice day, hope this article helps you ??. I will write a follow up article in the next 2 months.