Let’s bring it back to the good old days.
Not Taylor swift old. Or Pre-Internet old. Or pre-electricity old. Way, before the pyramids of Giza old. Before we had language — heck before humans existed. Wayyyyy beyond the meteoroid dinosaur extinction.
Forget it this is going to take forever. We’re throwing it back 3.5 BILLION years ago when the first organisms formed.
13.5 B years ago Universe went bang. Matter, space, time and energy were created. A quick 300K years later matter banged together to form atoms and those atoms formed molecules. Billions of years later those particles banged and turned into complex organisms we call life. They were pretty ugly, to begin with:
Billions of years later, Humans are born 🤩. 70,000 years later, they’re destroying all the matter, energy, molecules and organisms that were created before them! Go us!
Trending AI Articles:
1. Paper repro: “Learning to Learn by Gradient Descent by Gradient Descent”
2. Back-Propagation is very simple. Who made it Complicated ?
3. How to train a neural network to code by itself ?
4. Basics of Neural Network
I could get going on the Anthropocene, but that’s a topic for another time.
Tons of stuff has happened these past few billion years.
Throwing it back to single-celled organisms days at the beginning of life, they were damn simple. These organisms just slithered around. Imagine that! No taxes, homework or complicated relationships.
Today, life is more than just single-celled bacteria. How did we get from simple organisms to complex human beings??
? = Evolution
Evolution has created intelligence… humans! Almost hard to believe, but yes, we’re the most “advanced” form of life…
Stupidity is part of our nature, but overall we’re decently smart.
Sooooo, if evolution has created intelligence (let’s call them humans) to create AGI (artificial general intelligence; equal to humans) or even ASI (artificial superintelligence; surpassing human intelligence), we’ll likely need to follow copy the process of evolution.
If evolution has literally created intelligence from matter, it’s probably a trustworthy technique. To create artificial forms of intelligence, we need to follow the biological process to make organic intelligence.
That’s where genetic algorithms are all the hype.
Genetic algorithms 🤯😍
Genetic algorithms are kind of like the designers of mickey mouse. Over many years the designers tweaked and iterated the design of the mouse and voila:
Walt Disney artists probably started with a bunch of sketches of potential mice. Once they’ve curated 100 or so designs, they went to the board of critiques so they can get feedback and choose the best-looking mice sketches.
The artists would take notes on what the critiques say about the mice, and which sketches are the best. They’d then go back to the drawing board and sketch new mice based on the top-voted old mice, implement the criticism and add in new aspects of design.
Then they’d go back to get feedback; over time the design would get better and better.
Then, in 1928 after several renditions and iterations, they’d release.
Over the next 5 years, they would go about improving the design. In 1928 all the mickey mouse sketches were black and white.
But they’re making a 2.0 version of Mickey and the artists want to experiment with color. They’ll start by making 100 or so sketches, similar to the 1928 Mickey, but with slight variation. Some would have color, others not. Over time and feedback, they’d come up with the 1933 design:
Mickey has had so many upgrades in his life. We can see he has evolved over time, and it was a simple process:
- Create a bunch of Mickeys
- Get feedback on them
- Create a new set of mickies Mickeys based on old feedback
- repeat 2&3 until satisfied with the design
Walt Disney evolved the idea of a mouse named Mickey; it likely took a ton of time and effort, but he’s their most popular character so the process was worth it.
I hope you’re convinced that evolution is the best process for creating amazing things. (Humans & Mickey Mouse) 😇.
The process that created humans and Mickey mouse is very effective. How can we do it with computers?
The steps to evolve stuff in computers is almost identical to evolving mickey mouse:
- Create samples of whatever you’re trying to evolve (e.g. if you’re trying to evolve the weights of a neural network, you’ll generate many sets of combinations of weights for a neural network)
- Get “feedback” / test the fitness function of the samples
- Choose the top scoring/ functioning samples and make them have babies. Add some mutations to the babies (change things up a Lil’).
- Repeat step 2&3
….And this process is similar to nature’s
- Create a population with a genetic variety
- Individuals of the population blessed with the best genetics for survival and sexual attractiveness will be “chosen” to reproduce (because they survive for longer and find it easier to get mating partners)
- The “fittest” individuals will make a new population
- Steps 2&3 will repeat for the new population.
And this is what happened:
The process of evolution can be applied to aNyTHinG. → Whatever we want to “evolve” we must generate a population(step 1) and set a fitness function(step 2). [Note: Evolution is where the “fittest survive” ]
Evolution in computers example 🤖
If we wanted to evolve an object (population)to get to the bottom of a hill (fitness), we would start by generating a bunch of random objects(Step 1).
Then we would test the “fitness” of the objects (step 2). In this case, the task is to roll down the entire hill, so we would simulate the object rolling down the hill on a computer, then take the objects with the highest fitness scores to create a new population and eliminate the rest.
We generate the new population by making babies with the top scoring objects and throwing in some random mutations to add variety — just like nature does it.
Genetic algorithm is like a sped-up version of nature’s evolution. Except, unlike the universe, we don’t have billions of years to do most of our optimization tasks.
We can see that after 1 iteration, our objects have already started to look more round, and will score higher than the previous generation. [red number are fitness scores; 0 is the object didn’t roll any distance down the hill, 1 is object got all the way to the bottom.
We can do this for ANY task using genetic algorithms (GA)B . From optimizing the layout of a kitchen to developing high-functioning robotic arms to combining molecules in the best structure to create drugs!
Can we talk about that AGI thing again??
One of the most far-fetched, but interesting applications of GA is creating computers at the AGI or even ASI level of intelligence.
Right now, we have ANI (artificial narrow intelligence); AI that is good at specific tasks. Ex. A self-driving car won’t be good at predicting stock, but it’s intelligent at driving. In the future, we’re striving for AGI which can drive cars and predict stocks + everything in between.
In order to reach AGI, we’re going to need to get to work.
Designing AGI is basically creating the human brain, on a computer. Easier said than done.
For context, we’re only able to simulate the 1 thousandth of human brain activity on a computer. We still have a long way to go.
Scientists are getting lazy. It would be ideal if AGI designs itself. I empathize with these scientists, it would be ideal if my essays could write themselves! For the meantime, I’m stuck writing my essays, but scientists might not be stuck programming AGI…
AI designing AI… how?
You guessed it. Genetic algorithms.
The brain powers on 20 Megawatts and we’ve already created computers that run on more. So in terms of computation, we’re set. What we’re lacking is architecture for our powerful computers to run intelligently. So in theory, we can use our super powerful, 24 MW computers and have them use the genetic algorithm to find the best structure for their code.
They would evolve their code by:
- Generating different sets of code to run themselves on
- Fitness Function: Run the codes on various tasks to test intelligence. (We want the AI with the most generalized intelligence, so this function would choose the models with most spread out knowledge; rather than the models that are excellent at certain tasks but terrible at others).
- Take the top scoring general intelligence models and make babies out of them, by combing sections of code and adding random mutations just for giggles.
- Repeat steps 2&3
The step 2 function will be very lengthy and complex. General intelligence has tons of components, we’d be testing the code on many functions from image recognition, to the ability to write an essay, to solving Inverse kinematic problems, to making chocolate chip cookies.
This will take many years, even decades to develop.
The probability that AGI will appear 30 years from now is around 80%
– Sam Altman, Open AI
Louis Rosenberg believes we’ll reach singularity (ASI) by 2030.
Patrick Winston and Ray Kurzweil predict between 2040–2045.
In a study asking researchers when they think there will be a 50% chance of creating AGI, the median was 2099
Other studies give a 90% chance by 2075.
In other words, there is 0 consensus when the heck this will happen. It seems like AGI is a comfortable way away 🤔.
But… it might come even sooner?
“The pace of progress in artificial intelligence (I’m not referring to narrow AI) is incredibly fast. Unless you have direct exposure to groups like Deepmind, you have no idea how fast — it is growing at a pace close to exponential. The risk of something seriously dangerous happening is in the five-year timeframe. 10 years at most.” — Elon Musk wrote in a comment on Edge.org
Uh oh… 🥺
“I’m increasingly inclined to think that there should be some regulatory oversight, maybe at the national and international level, just to make sure that we don’t do something very foolish. I mean with artificial intelligence we’re summoning the demon.” — Elon Musk warned at MIT’s AeroAstro Centennial Symposium
WAIT! There’s no government regulation?? Human stupidity 😍.
I truly believe that AI can be harnessed for good: however, if we don’t become careful it might not be. Not having regulations is probably a very dangerous decision: maybe we deserve to be taken over by self-coding machines.
How smart would ASI even be; is it even worth it?
Well, I’ll give you some perspective. Cancer has been around since as long as we can remember, the first cancer patient dates back more than 5 thousand years to 3000 BC. We’ve invested billions and billions — if not trillions of dollars into research and we haven’t cracked the cute.
We have tons of people working on this problem it’s been too complex to solve thus far. Humans are just complex. Goddamn evolution why you gotta do that to us??? Just let us cure cancer!
Anyway, it would probably take ASI minutes to cure cancer.
Climate change, Alzheimer’s, Drug discovery, longevity? Poof: gone.
For all of human history, we have been tackling these problems and ASI will solve them no problem.
It’s like the world is a textbook with so many questions that humans have been tackling for centuries and ASI will create the answer page.
Personally, when I do my math homework I’m a big fan of the answer pages at the back, so maybe ASI won’t be a bad thing??
But, to be honest, as a machine learning developer and researcher, I still have no clue what ASI will do to humanity. What happens when we’re “worthless”? No idea. But I’m just fascinated how AI is transforming lives right now in its early stages of intelligence: from disease detection to manufacturing.
And… there’s potential to advance our species. Everyone can be intelligent!!
There is a real danger that computers will develop intelligence and take over. We urgently need to develop direct connections to the brain so that computers can add to human intelligence rather than be in opposition.
– Stephan Hawking
Some researchers even believe that thanks to ASI we’ll be immortal and free to explore the universe forever. I mean… maybe instead of going on Vacation to Japan or Australia, you’d go visit other planets or even galaxies… maybe universes??. ASI will be smart enough to solve all of our problems… humans would be left to do whatever the heck they want… potentially? Will it take away from the meaning of life thought??
No one knows.
I’m just a 15-year-old kid excited about the fact that we can create intelligence that solves the world’s biggest problems. We will get there with the help of genetic algorithms (or other strategies similar to evolution). Nevertheless, the largest concern is how we’ll live person+ machine, not person vs machine.
Dang, too many ethics questions for my little brain to think about… maybe ASI can solve this problem too… oh, wait?? Good thing we have until 2030/2045/2050/2075/2099/oh nevermind.
Sometimes I worry about the fate of humanity. But then I remember 2 things: we are destroying the planet super duper fast so if ASI doesn’t kill us we’ll kill ourselves, and there are so many assholes that exist maybe being immortal would be a harmful situation.