Five things I have learnt about AI so far
Learning AI is partly an exercise in curbing enthusiasm. There is too much to learn and it is easy to get carried away or overwhelmed by the sheer amount of knowledge out there. What helps is cutting out the noise and letting go of some preconceived notions. When I need a change of flavour from calculus, probabilities, numpy and pandas, I try to get a high level view of things from some of the brightest minds in the AI space. Here’s what I have learnt so far.
Automation is not AI
For someone not familiar with what really constitutes an AI based system, it is common to mistake a reasonably well-programmed chatbot or any other rule based automated workflow as AI. Even some companies are guilty of propagating that their solution is intelligent. Rule based systems do not learn. They operate with static knowledge that was programmed at creation. Rules may be tweaked by humans from time to time, but the system itself does not become an improved version of itself based on interaction with users or emergence of new data.
What is obvious may not be the solution
A robot that vacuum cleans the house does not have to look like an android wielding a broom. That is why you have disc shaped cleaners that are so popular. They are compact and can reach nooks and crannies easily. Popular culture often shows robots with arms, legs and a head serving humans or doing human like work. This image is very cinematic and works well in movies. AI practitioners are not fixated with a robot resembling a human. They prioritise function over form. Consider autonomous cars. Although currently they resemble conventional cars in the sense that they still have steering, accelerator, brake pedals, and all other input devices, these are not really needed. Perhaps they are incorporated so that the passenger can take control during an emergency. An autonomous car can be designed to eliminate all these such that the inputs are given directly to the moving parts.
Data collection at scale is more important than how you collect it
This is a valuable learning from the Lex Fridman podcast featuring Andrej Karpathy. During the podcast Andrej was answering the question about whether cameras are better than Lidar as sensors in autonomous vehicles. In the overall scheme of things, prioritization is important. In an AI based system, it is more and better data that is the priority, not how it is collected. He also made a point about granularity of data using the example of maps. Excessive details in maps may not be desirable because humans are able to drive without maps (yes, despite being spoiled by Google Maps) and hence AI should be able to do that too. This is not to say that maps are completely useless for autonomous vehicles. Of course they are important for route planning, but such vehicles need not know if a passing shop on the right side is a famous hair cutting saloon rated 4.5/5 on Google. For us folks from business, this rings true in many ways. We obsess over irrelevant details in a marketing campaign, or a strategy or an app/website, while we see a competitor running away with customers with much less sophisticated versions of the same thing.
Incremental gains are better than Big Bang solutions that have a high risk of failure
Another learning from the same podcast. For businesses, the AI journey should be a series of steps and not a giant leap. The technology is evolving very rapidly, with new thinking and new frameworks arriving on the landscape every day. Companies should be vary of throwing all their eggs in a very big, expensive basket. I recall that somewhere around Covid times, no-code platforms had become all the rage. I had myself tried to use one, and despite being familiar with programming and databases, I had really struggled to use one. Now I don’t encounter anyone talking about them (I may be wrong). Vibe coding is the new norm with generative AI having become very powerful in clean code generation.
There are other reasons why smaller steps is the way to go—AI talent is not very widespread (although this is changing); organizational acceptance requires retraining employees who will be dealing with AI tools and you start seeing results sooner, which is a motivating factor.
The most sought after AI outputs are pretty mundane stuff
As per McKinsey’s State of Global AI survey, the most common outputs of AI deployment are (in order) – Text, Image, Computer Code, Video and Voice & Music. In terms of functional areas, AI is most prevalent in – Marketing & Sales, Product & Service Development, Service Operations, Corporate IT and Software engineering. Putting the two together, it is obvious that AI from a business perspective is primarily Generative AI as of now. This is the area where CXOs see immediate returns on their investments. So no, AI is not flying planes, operating large factories or launching robots into space. It is currently busy writing ad copies, automating repetitive processes, generating code and replacing low level staff. It will eventually do bigger things and come after white collar jobs too.