My life with ChatGPT or The Times They Are A-Changin'
I have been super bullish on Generative AI ever since ChatGPT released. In fact, some of us at Glance were exploring Generative AI even before it became a well-known concept or term. We experimented with image inpainting and summarization tasks several years ago - when they were barely any good. One of the tools we developed was built on the idea that AI could generate content, leaving humans to simply moderate or edit it. That was 2.5 years ago—a topic worthy of a separate blog post. But let's fast-forward to the present.
My first time
Around October 2022, my team at Glance had begun utilizing the GPT-3 Davinci model for tasks such as summarization. Me and my other AI enthusiast friends tried to make use of ChatGPT (whenever it was available & and not down). One of the smartest people I know (not you Rags) started using ChatGPT for his day to day coding activities. Inspired, I finally gave in and got Copilot, which, to put it simply, is auto-complete for coding. It wasn't long before our creative minds (which seem rather short-sighted in retrospect) realized that we could leverage ChatGPT to write blogs. I mean how else are we getting those LinkedIn likes and clout? And indeed, many people jumped on the opportunity to create more professional and beautifully written content with the help of this powerful AI tool.
In the months to come, I have gone on to make use of these offerings for the obvious usecases & some super non obvious usecases. Getting inspired from the countless exercises people did with ChatGPT, few of them were good enough to get it integrated at work.
Every time I experiment (actually more like play) with these technologies (GPT, ChatGPT, Copilot), I find myself completely amazed by their capabilities. During my first week using Copilot, I was so impressed that I felt compelled to share my excitement in the company's channel with all Tech folks, sharing just how incredibly magical this tool was. When I was experimenting a bit too hard with ChatGPT, I found myself brimming with ideas that I had to bounce across people around me. One of those ideas was - how easy it is now to create authentic fact based news content using Generative AI and simple seed content. This was before Buzzfeed had committed to it & saw its fortunes improving (briefly).
Both Copilot and ChatGPT are accessible through paid subscriptions. Copilot costs 10 USD per month, while ChatGPT Plus is priced at 20 USD per month. Me being a frugal person, I didn't buy these premium plans until months after they were launched. It really didn't feel to me before that my life would change significantly having these 30 USD odd monthly recurring payments. After all its just predicting the next token right?
Cue in - the sceptics
There are many sceptics of Generative AI. Some critics argue that it is just a big transformer model that was trained on big amounts of data and lacks true knowledge (looks at Shank), some express concerns around the biases we introduce in these models. Some are unhappy about how it will potentially impact jobs for all - tech, copy writing, any and everything. Some dismiss it as a hype train or bubble, while others worry about the rise of AI girlfriends or friends and the potential for a fake news crisis. Hallucinations don't help either. People don't like that an organization that started out as Non Profit is now a Big Corporate.
The most recent thing people have taken up is an open letter calling for a six-month pause on all Generative AI experiments. People including Elon Musk, Yuval Harari & Steve Wozniak. While Musk always had his doubts and expressed concerns about potential threats posed by AI / AGI. Yuval Harari in his 2015 book "Homo Deus" alluded to the disruptive influence of data and AI on our lives, as well as the need for ethical considerations. If I am being super blunt, you wouldn't see someone from Google, Microsoft, Meta or Open AI sign this. You can't help but notice how much of an existential crisis Generative AI has put Big Tech in. And can't help ponder why anybody would want to pause "AI experiments" just for ethical considerations & policies.
A Computer Scientist on Lex's podcast just called for air strikes (and nukes) against GPUs.
As Roy Keane would say "Gimme a break"
My favourite take on LLM is from Geoffery Hinton
But honestly - in some what selfish way, how much should I care ?
Its not like Open AI is casually releasing these models. They are taking measures (and have been for years) on making sure what their models don't get out of control. These technologies have made me better, efficient.
Ok. So stopping the general commentary about Generative AI & focussing more on how my life is changing with this.
Aah man, the progress..
Every day I go on Twitter and see thousands of people create extensions and wrappers around ChatGPT. I can't help but feel jealous, amazed and mad at myself for not thinking about it before and executing it. Indie developers build cool tech and Big Tech eventually planning integrations in their systems. Say extensions for writing good docs or emails. And then Microsoft announces Copilot for Web. Pichai shares that Generative AI is coming to Google Workspace So definitely, new ideas come, get traction, the idea gets integrated into our day to day tools.
it makes me a bit jealous & much more optimisitc
Ambitious ideas of having a "copilot" for your laptop, or running an LLM on a macbook locally are casually done by developers over a couple of days 🙂 It took what, a month ? for people to create a comparable open source (?) equivalent of ChatGPT. Things are changing too fast. The super ambitious moonshots are weekend projects for some. And here I am, waiting for the perfect problem statement to commit to.
But .. my personal ambitions aside, ChatGPT is an integral part of my day to day living.
be it my main work or random chores
If I am solving an ML classification problem statement at work - classifying content into buckets and more buckets - its just prompt engineering away. Gone are the days where you had to create a dataset, train a model and deploy it. Many such problem statements that were non trivial before can be just prompt engineered now.
There's a anti bribery course work & tests that I have to submit at work? you know what, let's see if flan-ul2 can solve it for me. Shout out to Huggingface for making ML inference so easy.
a smart ass who knows most of the things
3 of us ML Engineers at work have been maintaining & creating React Web Apps with MSFT Authentication for our internal tools. For the longest time (2.5 years exact) I couldn't get the MSFT Auth to seamlessly work. All it took is me conversing with GPT 4 for 20 mins. It provides a solution, you add it in your code and see if it works. You ask it how you would do it with these new constraints, it modifies its solution to cater to your asks. You question the solution and check if edge cases are taken care of. GPT 4 will help with that too. It explains the code yaa.
having dedicated juniors aiding you with your work..
GPT solutions are like having dedicated juniors for you to boss around - give them assignments (coding, research, etc) and then hours or days later when they do it, you review it and take the points you want out of it and integrate it with whatever you were doing. You ask the juniors for edge cases, ask them their reasoning, get their help in understanding something. They know more than you after all. All of this is literally now possible in seconds!! ChatGPT is the guy that is just there for you to ask help from. And yes, it makes mistakes and hallucinates (confidently shares wrong things) - but over time you can learn on how to make most of it and fact check via Google.
personal guide to the new pastures..
If I am learning a new topic - I ask ChatGPT to give me pointers on how to approach it. When I cooked chicken for the first time and had a headache at night, I asked it if its related. How is it different than you looking at 10 articles and forming your own opinion based on somebody else's? I mean obviously its very different but the gist is most of your tasks are taken care of by these new technologies.
I am guessing search was probably a big deal when it first came. Mobile phones with internet access, essentially putting computers in our pockets, were next. ChatGPT like technologies is the next big thing. Don't take my word for it - experience it yourself.