Are LLMs worth the hype? Where do they add value? — POV of a daily user
In a world where AI is getting chatty (Llama2 to blame :P), we are getting lost in a linguistic jungle but its mostly our fault like how we sometimes take the wrong directions while following Maps on mobile unconsciously. I work at a renowned organization and have been primarily working in NLP since 5 years. From Nov’22 there has been lot of hype around ChatGPT that took AI industry by storm. Fortunately I got the chance to download, test, finetune lot of famous LLMs released ever since. Whatever I have written is from both my professional and personal experience on using LLMs.
I have joined numerous discord groups, followed twitter threads(now X! — Surely Musk hasn’t taken help from any LLM to rename the microblogging app), read and highlighted many medium articles and most importantly I’ve noticed that each individual has a distinct approach to prompting an LLM for the same task, much like their unique fingerprint. It’s intriguing to think that in the future, identifying cybercriminals might involve analyzing how they prompt LLM models — a sort of digital fingerprinting in itself! :P
The numerous use cases LLMs are able to solve is unimaginable. More than general tasks, solving domain specific tasks made LLMs unique in selling itself to target audience. If you join any discord channel of a new model that gets released, you can see people asking how to finetune this model? how to use parametres? how to make quantised versions(for occupying and running it on less memory machines)?
The aim of a business professional looking to harness LLMs goes beyond tackling intricate math problems or asking LLM to plan budget-friendly vacations. It involves utilizing these models to address industry-specific challenges through fine-tuning with tailored datasets. This strategic approach adds a touch of ingenuity and professionalism to their endeavors. For example, a team working in supply chain industry knows lots of specific problem where there is lot of manual work that goes in, to reach a solution, instead if they had proper data to finetune or make LLM learn their specific task, it could bring down their manaul costs/resources to lesser extent saving more worth of time.
Forget solving, AI Agents powered by LLMs can alone create, plan and execute tasks with out any extra pair of eyes (Damn! I need my parents, friends, neighbours, unknown person walking beside me — pairs of eyes to even check my travel tickets to confirm if I have entered correct date and time of my travel plan!). Of course now I don’t need another person to double check my tasks, I only have to rerun the same prompt twice to be sure if LLM returned the same response or it hallucinated. LLMs analyze mountains of text like Sherlock analyzing clues. They’re great at guessing, but even Sherlock gets it wrong sometimes!
Let me address the reality, LLMs aren’t perfect. They’re like that friend who’s super smart but occasionally messes up the punchline in a joke. But that friend utters each and every word not from conscience but from a statistical model thats gives highest probable next word to utter and that word might not be repeated at all if we ask it again(Can produce hallucinated or few mishap blocks of text). Hmmm! So do you trust your new friend? or give some time for it to settle in? No we can’t wait! The world doesn’t wait, the 1M downloads chatGPT got in first 1 week of its release didn’t wait.
This rush towards understanding LLMs, trying out the knowledge of it from millions of users over the world has opened up a stupendous wave of technological advancement of executing LLMs only in Multi Nodal GPUs(A100, A10, H100 etc.,) on cloud to single GPU enabled laptop then to CPU then on mobile phone and may be tomorrow in your wrist band too. All because of the big elephant in the room — GPUs. Those who know pytorch library (kinda back bone in inferencing/finetuning) LLMs, they have built it amazingly that work on CUDA (Compute Unified Device Architecture) which is a parallel computing platform and API developed by NVIDIA Corp.
Its like iOS works great for Apple Inc products. People like iOS, Apple is a trillion dollar company. People like LLMs, Nvidia’s GPUs are in high demand. There are number of articles on internet that mention very specific details on how many orders are there from the Big Clouds( AWS, Azure, GCP) for the GPUs for future and its mind blowing. But lots of tech enthusiasts are figuring out on how to make an LLM work with less resource and compute with decent performance.
Despite their quirks, LLMs are the ultimate research companions. They’re like librarians on turbo mode, helping you weed out internet nonsense. If I want to know the summary of the famous novel Ayn Rand’s — The fountain head, google/safari/bing etc., does give me lots of links that I can navigate to and scratch my head. But to an LLM, I can prompt it to explain like a friend/ explain like how a 5 year old can understand/ explain only the crux of the story with important events of the protagonists journey etc., and all of them in a single webpage/UI. I don’t need to switch tabs, I don’t need to forget which link I was referring to before.
All these are okay but what if you are using it to code your test cases/ asking it to do your homework/ asking it to write your email and All from scratch? Does this help you in anyway? Are LLMs killing human creativity or Do they fuel it? Are LLMs worth the hype — Absolutely YESSSSS! — In this era of Time is Money — Knowledge is Money, only a responsible user understands the value of LLM to exponentially increase knowledge in short time to increase worth.
They add value to help you become a better coder, better writer, better planner, better manager, better learner, better therapist to advice your friends relationship problems :P(Google’s Deepmind is about to release a personal life coach, so no surprises later!) and overall a better human being when used wisely and having a filter to take only what you need from the output. LLM saga continues, they are powerful, they’re here to stay — Makings the outputs and life unpredictable!
Reach out to me on linkedin https://www.linkedin.com/in/dharani-jonnalagadda-828737120/ if you want to know more about LLMs.
