This is a short post about how I am using Language Model (LM) based tools in my day-to-day life.
There are two subscriptions which I heavily rely on:
- A Chat Bot service: So far, I have used ChatGPT Plus, Claude.ai or Gemini (whichever is performing the best at that point). ChatGPT and Claude’s free subscription allow usage of their most advanced models (GPT-4o, Claude 3.5 Sonnet), but the number of allowed requests are very limited, often not sufficient for my need. There is a drastic performance difference between the most advanced model and the default/cheaper model offered by these chatbots (e.g. GPT-4o vs GPT-4o-mini or Claude-3.5-Sonnet vs Claude-3-Haiku). The cheaper models often make naive mistakes creating an impression that language models are not yet up to the mark.
- Github copilot: This is a super helpful tool specifically geared towards software development activities like coding. It integrates well with editors like VSCode and helps to maintain the state of flow during development. I use Claude as well for the same purpose. Claude’s artifact feature is mind blowing (Example: Claude cloning an app). But, very recently (August 2024), I am observing performance degradation of Claude, while performance improvement of GPT-4o (Used by Github Copilot). Have I used Cursor? Not yet, but planning to use it pretty soon.
- Perplexity: Consider this as a search engine with whom you can chat while looking for relevant information. It’s already performing much better compared to Google. However, I have not yet bought their subscription. My limited need is mostly fulfilled by their Pro-Mode. I have heard lot of good things about Kagi as well, but used it till now.
There are several occasions where these tools have helped me complete my work in few hours which could have easily took me weeks just a year before. With these tools, I have dealt with domain specific problems, which I had no clue about, in a matter of hours/days.
I use the Chat bots regularly at home as well. I use those while helping my daughters’ with their study. LMs elaborate most of concepts in a very intuitive way compared to the text books. We use those:
- while introducing to new concepts (What is group theory? Explain for a high school student)
- to look for answers of adhoc questions (How sand paper is harder than metal?) arising in their mind while studying
- for understanding images from their text books which lack detailed description (take a photo, upload and ask LM to describe).
- to generate images accompanying their writings
I ask LMs to suggest me books on my topics of interest. I’ve been using Amazon’s recommendation system intensely for last several years. It has helped me to find out many interesting books for myself, my wife, daughters and our parents. Initially I was skeptical if LM would be able to do justice to my need. Over time, I am realizing that GPT does it really well. For a topic, it returns probably the best 4/5 books. Many times it has introduced me to the books which are not widely popular, but critically acclaimed.
Chatbots helps me with boring house hold tasks as well:
- During income tax filing, I had to deal with a some numbers from a not so friendly pdf document. I quickly took a photo and asked claude to create a CSV out of it saving an hour on a Saturday morning.
- Here is another interesting example. I sent my 12 year old car for servicing. The agency sent a bunch of photos of parts which are broken. I asked GPT to describe those parts and how those are related. For a person like me whose knowledge about car is limited to clutch, break and accelerator, that was a great help.
Few cases where I don’t use LMs:
- First one is while solving (advanced) math problems with my daughters. I prefer to sit on a problem for hours than to look for a solution. I believe, in coming years, our mind and brain are going to be the only thing which would differentiate us from the machines. So, we better increase their usage.
- During our morning walk, my wife often becomes curious about the trees we see around our neighbouring lake. For Indian origin trees, GPT has not worked well so far. Google’s image search still performs better.
- Finally, I don’t use LMs to write posts or emails, but I ask it to check for spelling and grammar (“You are my editor, but don’t modify the content”). I strictly prefer to stick to my “not so great” style of writing. After spending hundreds of hours with LMs, most often I am able to distinguish AI generated texts from human composed texts.
During my college days, one of my math teachers (probably, in his early 50s) told us that he doesn’t want to learn to use computer in his lifetime. I am sure he regretted his decision pretty quickly. So, don’t be like my math teacher. If you are a knowledge worker, start using LM based tools for day-to-day life. This will not only increase your productivity, but will give you a glimpse of the “irreversible artificial future” where humanity is heading towards sooner than we all can imagine.
Thanks to my wife, Indrani for reviewing and editing this post. The feature image is generated by Gemini Imagen 3.
arnab's blog 