Preamble: Due to cost, Inthe510 did not maintain the website and it is currently not longer available. You can still visit a old, non functioning site here
Maybe someone will be inspired to create a similar project. But enough extra text, onto the actual article:
This project hit me personally
Usually we report about other people needing assistance, having problems or showcasing great causes. This is the first article about how I needed help and went out looking for it. I found it at the strangest place possible: In the arms of a robot. First let’s meet the man that created that robot.
The developer
Inthe510 is a python developer specialising in the creation and implementation of large language models. A current project he is working on is the combination of LLM’s, non-violent communication and cognitive behavioural therapy. He’s the perfect candidate for such a project:
“I’ve been fucking around with computers for a really long time. And I’ve never had so many normal people, random people on Twitter be like, hey, that helped me. Like, hey, that was cool. Like I was just going to these people, like random Americans who are having social difficulty, right? And I’d be saying, hey, try it.“
Inthe510
Grieving is something that might not be reserved entirely human, but definitely out of reach for computers. Showing a person that you are there for them but also giving them the opportunity for them to cope with their feelings is a very fine line to walk.
An introduction to LLMs
This article is for everyone, even those who think language models are a new Gucci design. If you are already familiar with the topic or concept, fast forward to the next headline.
A brief introduction to language models for all the others: In a traditional machine learning model, the model is trained on a set of labelled data, and the model’s performance is evaluated on another set of labelled data. The model can then answer questions about the topic the data has answers for. A simple example might be the same as homework. A teacher might ask the class how to assigns the class a book they have not read before.
You can picture labelled data as a sort of “ground truth”. No matter how you look at it, it is always correct. For language models, they are questions and answers. How does an orange taste? Sweet, sour with a hint of bitterness. How does sugar taste? Sweet. The model’s able to make predictions, but is limited to the patterns it has learned from the training data. In this case, it’ll conclude that the orange must contain sugar. This is of course very simple and only helps the people that are a big fan of oranges and weight loss. The answer “It might not contain sugar” to the question “What can I do after my boyfriend breaks up with me?” is not very helpful.
How LLMs work
In contrast, large language models (LLMs) are trained on a much larger amount of data, and the model’s training process includes the ability to learn from unlabelled data. This allows LLMs to learn patterns that may not have been present in the labelled training data, making them more flexible and adaptable to new data. Additionally, LLMs are often pre-trained on a task that is related to the task the model will be used for, such as language translation or text generation.
This pre-training allows the model to be fine-tuned on the specific task for which it will be used, further improving its performance. For example: Spotting fake credit card charges, writing captivating headlines or writing a new, 5 part crime series.
In our case: Writing condolences and helping yourself and grieving loved ones. Inthe510 explained it the following way:
“So it’s easy for the LLM to come up with some stuff. So then I just said, look, translate this into nonviolent communication.”
Inthe510
About the project
https://feelings.ltd/ is a website run by Inthe510, a developer from California. A passion project which aims to show off a way LLMs might be used to assist and teach us. Since a big part of the interaction between you and the site is also the changing of your own understanding. The project had quite an impact on his life as well:
“I’ve never had so many normal people, random people on Twitter be like, hey, that helped me. Like, hey, that was cool. […] People would come back to me and go like, wow, dude, that was so great. And I’ve just never had that experience where just super normal people had a real positive outcome with something I’ve done with a computer.”
Inthe510
I’ve had the pleasure of not having to write a lot of condolence letters in my life. But a few weeks ago, the wife of a family member passed away. While good with words, I am not perfect with expressing my feelings. Then I stumbled upon his website.
How I use the website
The premise is quite basic. You enter the recipient, the “problem” and maybe some information or special occasion you remember. This is not only for when people are passing away, also diseases, accidents and unfortunate events, like your favourite football team not making it into the next season.
So I started playing around a bit. The first prompt I gave the site was:
Write a WhatsApp message to my uncle. His wife just passed away.
After a few seconds and rotation of the loading bar, the following message appeared:
I am so sorry to hear about your wife’s passing. Please know that I am here for you and thinking of you.
The goal of non-violent communication is to express a specific feeling in the person you are trying to reach: I care about your pain and sadness and am here to support you. Please let me know if there is anything I can do to help.
I tried to make sure to convey that I was here to help him. My next prompt was:
Write a WhatsApp message to my uncle. His wife just passed away. Mention that I’m always here if he wants to reach out.
Dear Uncle, I am so sorry to hear about the passing of your wife. My heart aches for you, and I want you to know that I am here for you and thinking of you. Please don’t hesitate to reach out to me if there is anything I can do to help. With care,
That was already pretty close, maybe a bit too formal, so I tried:
Please write a short WhatsApp message for my uncle. His wife died from cancer. Make it a bit more familiar and less formal
Bar spins, answer:
Dear Uncle, I’m so sorry for your loss. I’m sending you love and strength during this difficult time. You and your wife’s memory will remain in my heart forever. With care and compassion,
I want it to make it more personal. Since the LLM of course does not know my uncle, I need to give it some more personal information:
Please write a short WhatsApp message for my uncle. His wife died from cancer. Make it a bit more familiar and less formal. The wife was a really good basketball player and coach
Bar, wait, answer:
Dear Uncle,
I am so sorry for your loss. I know words cannot express the grief you are feeling right now. I want you to know that I am here for you and thinking of you.
I remember how your wife was a great basketball player and coach. She was an inspiration to many, and her legacy will continue to live on.
You are not alone in this difficult time. I am here to listen and support you as best as I can.
With love,
This is already pretty close to what I want. I manually picked out the best of all outputs. The message I sent my uncle in the end was:
Hey Stefan,
I am so sorry to hear about the passing of your wife. My heart aches for you and I want you to know that I am here for you and thinking of you. Please don’t hesitate to reach out to me if there is anything I can do to help.
I remember how your wife was a great basketball player and coach. She was an inspiration to many, and her legacy will continue to live on.
You are not alone in this difficult time. I am here to listen and support you as best as I can.
Let’s go on a walk sometime!
Flo
This was a thousand times better than the stuff I was able to come up with.
Final Thoughts
The project was free. It was not only a useful tool in extreme situations, but makes all of us a better, non-violent communicator. Let us hope someone will pick it up again !
Even if you are in the fortunate position to not write a condolence letter, it is still worth learning from it. Maybe next time, you don’t have the opportunity to consult a LLM.
There is still a problem: Robots can not solve everything. They might be able to make us better at talking but they cannot carry our groceries. If you’re up for that or need help check out wirhelfen.eu and write your needs or haves there.
