AGI: A Marathon, Not a Sprint

Artificial General Intelligence(AGI) has been explained as the ability for machines to match or surpass the cognitive ability of humans.?

It has been a divisive discussion topic where proponents of AGI talk of it as the magic that would solve all (if not most) problems that humanity faces today making us super efficient and leaving us time to pursue what we all love while living under a utopian universal basic income model. Skeptics of AGI on the other hand talk about how AI learns from humans and our deepest secrets lie in some form of domination, destruction and greed and this would make super-intelligent machines to be trained on these instincts leading to chaos and potential destruction of the human race.

I personally have a different opinion on AGI - I believe AGI would be phased like self-driving is and achieving the full range of cognitive abilities of a human would be very difficult for a machine. I foresee political, religious, ideological? and legal hurdles more than technological challenges. Going back to just one of the many tasks that humans perform on a daily basis - driving - we have probably 10 years and millions of miles of data but we still would not trust machines with edge cases or the problem solving ability or the presence of mind of a human with 10 years of driving experience.

Why is AGI so difficult

AGI is interpreted as a “Digital God” , something that knows everything & can do anything. While this sounds easy in theory - let’s put all the world's knowledge into a large computer and train it to act like any human would. The trouble with this is there is also a lot of tacit knowledge gained by experience not translated into any medium at the moment that makes us unique - the problem with the current approach is that the G in GPT is focused on General instead of Generative.

The best way to reach AGI

Designing systems that are focused on being everything for everyone sounds great in theory but would be difficult to achieve practically. Imagine a software having the ability to do EVERYTHING a human can do, would be like asking someone - tell me everything you know about every topic. Sounds bizarre.

Instead the best way to think of this solution is to divide the problem into as small parts as possible i.e. take large problem sets or tasks or skills and divide them to the most granular steps and teach machines on being a super specialist in that one specific task to the point that if there’s an error or issue or a fire or an accident - how does it act, who does it call, etc. The idea is then to do something similar to a Lego block where each block is a super specialist with controllers and super controllers that dictate what action is to be taken and when.

You can already imagine how this might be very difficult with a subsystem performing the task and one monitoring for progress and checking for errors or what else to call as a remediation step should a particular task not go as planned.

This scaled up to thousands of tasks over millions of chips in a data center where a chip could melt due to the heat and one of these systems down could mean the entire set up fails. The only solution to this is exponentially better technology in terms of both processing power and temperature management.

This is probably why each of the mega tech companies need to invest billions and more to even be considered a strong player in this space which explains their massive investments in the space with each of them trying to get into chip design and data center architecture.

An alternative solution

While achieving a comprehensive "know it all" solution appears impractical in the foreseeable future, continued exploration remains crucial.. In my opinion the best way to get this is to have “Digital Super Specialists” -? a system that knows everything about one topic - which could be for example be one that knows everything about treating a specific disease. These digital super specialists are created by companies that have expertise in that space i.e. subject matter experts.

Everyone else would have digital assistants or Co-pilots that would just have metadata i.e. have an idea of the problems and which digital super specialist to call based on the nature of the problem, budget, turnaround time, etc. I expect every organization to have their own Digital Super Specialist which would be created by using all the knowledge within the organization. So basically every document you’ve written could be worth a lot.

In such a model organizations that take a lot of notes i.e. doctors, healthcare workers, etc. become very valuable to making humanity better.

The creativity conundrum

One aspect I haven’t touched upon so far is creativity - can we rely on AGI to help us be more creative? Or put another way - have we figured out where creativity comes from? How did a chef think of a dish? How did an artist think of a painting? Is creativity a spark that is inexplicable or is there something that triggers that spark? How is the solution thought about? I still don’t know.

My personal opinion is that creativity is something humans will continue to practice while AGI makes us more productive and maybe even gives us more time to be creative.

How do we use AI going forward?

The way I look at AI being very useful is as a companion or co-pilot in most cases. It doesn’t “perform” the task but assists you with the options, writes a first draft, brainstorms with you and presents expected outcomes of specific decisions. A productivity booster of sorts - helping get the mundane and routine out of the way so one could focus on true value add.

When will we get there?

I expect AGI to me more evolution and less Big Bang - just like every technology so far, be it automobiles, airplanes, phones - there might be a pivotal moment but getting to true AGI may be a long process or even a mirage given human ability also has been improving with the dawn of new technologies and AI based on the way it’s designed, may be playing catch-up. Of course there are areas where I expect AI to be much better than humans - anything involving the process or recall of a known solution but novel solutions may be difficult.?

That said - the ability for machines to come up with permutations and combinations of compounds, proteins and potentially even potential variants of a virus, gives us the ability to plan and react faster to make the world a better place.

Which goes back to AI is a productivity booster but getting to AGI will be in phases with specific problems being solved in siloes then the ability to bring them together, measure dependencies and tradeoffs and chart a path forward.