ai

There are two types of artificial intelligence (AI) though most of the time, you won't see the distinguishing prefix: general AI and specific AI. Now, with no further information, one might default to assuming these two types are equally valuable and their usefulness might be 50% of the market, each:

2019-2020 has seen large marketing campaigns and huge venture capital interest in AI:

However, these are all attempts to develop specific AI. You recognize specific AI when you hear about it beating people at Jeopardy! and chess. But the chess AI program doesn't do anything except calculate chess moves. It's not widely-adaptable. At best, these are specific AI programs that will perform a niche analysis or task as well as a top human could. At worst, they are dressed-up statistical analysis firms utilizing large data sources to draw out new insights (having virtually nothing to do with actual artificial intelligence). There will be money made, but none of these have the enormous upside potential of general AI.

General AI, on the other hand, will learn and do literally anything. It is functionally a digital human brain. Consequently, the value comparison between these two actually looks like this:

This will not be apparent to most at first, least of all investors. There will be some disappointing outcomes before people learn to better-distinguish between low-value and high-value AI.

I don't use memes as data points, but they are useful in revealing the zeitgeist around a particular topic.

My views on AI are principally from three books. Between them, I recommend reading On Intelligence, but skip directly to chapter 2 (chapter 1 mostly bashes contemporary competitive firms and legacy approaches to designing AI).

Written in 1980

Written in 1994

Written in 2004

on intelligence summary:

  • Your brain sits in a dark, silent box. All of the pleasure, pain, and other sensory input is just small electrical pulses.

  • The cerebral cortex is a few credit cards thick and about the area of a large dinner napkin, crumpled over the surface of the outside of the brain. This is where "you" reside. We can take any part of you away (arm, leg, gall bladder) and you are still you. But if we take any part of the cerebral cortex away, you will change personality/functionality/memory.

    • In a way, the human body is a space suit designed to allow your cerebral cortex to survive on the space rock "Earth."

  • The mind is not a computer, it doesn't take input and calculate what the input means. Rather, it is a giant memory bank with a short-cut system. It pattern-matches previous stimuli and produces the "same actions as last time."

  • The inputs from our five senses arrive at the lowest level. Depending on the particular blend of inputs, a cascade of electrical impulses head up through the layers, progressing from a broad network of 'roots' to a more narrow, focused signal at the top.

    • Interestingly, we may get input from an eyeball that is quickly racing back and forth across a face--saccading--but the top-level signal in the brain just gets a report from the lower levels as "face," not: eye...0.04 seconds...cheek...0.10 seconds...teeth...0.09 seconds...etc...

  • It is the holistic report that our brain entertains, not the individual, minute inputs.

  • When we are touching or seeing or hearing a thing we have touched or seen or heard before, the recalled symbol in our mind causes the reverse cascade: top to bottom, branching out toward the individual input-level neurons.

BIG DEAL #1:

  • Our world makes sense and we know we are experiencing 'reality' when the inputs flow of signals from bottom-to-top are consistent with the memorized signals flowing top-to-bottom. The world is consistent with our memory.

    • We obtain input from all nerves in our bodies, but we only pay attention to a small set at once.

    • The slightest deviations register as anomalous, especially doorknobs, height of your staircase steps, or the sound your car makes when it starts. We interact with these objects frequently and if your doorknob were 2" lower on the door the next time you went to use it, there would be the DJ 'record scratch' in your mind and all of your attention would focus in.

    • If you are reading this in your home, your brain is monitoring (but not paying attention to) the smell. If smoke or fresh bread odors were to enter your olfactory system, you would notice immediately.

BIG DEAL #2:

  • There are two types of artificial intelligence: general and specific.

    • The hundreds of companies working AI into their products are pursuing specific AI

    • What the world really needs is general AI.

  • Specific AI programs like Alpha Go and CV6 can play chess or write paragraphs of text better than people. These highly-customized programs excel in just one domain. The world is not revolutionized as a result of their existence.

  • General AI programs would revolutionize the world. It would work like this:

    • A company has 1000 trucks with motorized steering wheels, brakes, and throttles, but not enough drivers.

    • A GenAI unit is brought in, but it has never seen or driven a truck before. So, it is placed inside a truck driving simulator and given an objective function: Move A to B without receiving a fail (crash into a digital fence, digital pedestrian, etc).

    • GenAI unit achieves only 0-10% of the journey on its first 24 hour day of driving. Progressively, it twists the wheel, adjusts its speed to do the one and only thing it wants: achieve the objective function.

    • After a few weeks, the GenAI unit can twist a steering wheel and adjust the throttle and achieve the objective function even better than humans. The software function is copied to a thousands units and a thousand trucks hit the road. They all immidately drive as well as the trained unit and continue to get better with time.

Now here is the cool part:

  • The exact same unit can be placed in a hospital and told to identify cancerous tumors on X-rays as opposed to harmless cysts. A library of past images and their associated test results are fed through the imager and the unit starts by making guesses, but quickly picks out details and learns.

  • The exact same unit can be placed in a grade school with audio input from teachers and students and fed the grades of the class, with the objective function of determining how each student learns (semantic memorization, procedural repetition, or episodic recall of story/song).

  • The exact same unit can be hooked up to a Da Vinci robotic surgeon and trained to peel open grapes & stitch them closed again. The AI doesn't get bored, so it can peel a million different grapes, bananas, kiwi fruits, etc until it is exceptionally accurate and fast. Then it trains on animal and human corpses doing actual surgical tasks. Then one day it graduates to actual appendectomies and lumpectomies.

  • The key difference is the GenAI objective function is NOT FIXED. It is a general AI in that any objective function can be established and it will learn to achieve the objective on its own given a suitable training environment. This is why general AI has almost unlimited potential to revolutionize the world.

the art of doing science and engineering summary

  • There is a checkers program written by Art Samuel that had a few parameters to help decide between a few legal moves: exert control the center, capture a piece, pass an opponent's piece, block a piece. There were weighting factors to decide between them, set to arbitrary values. The program played against itself ten times, adjusted settings, played another ten games, repeat until the parameters asymptotically approached their optimized values.

    • Samuel's program could then beat the state chess champion. Was this machine learning? Almost certainly yes.

    • Google's AlphaZero was essentially the same thing, parameters to decide goodness-of-moves, completely tuned by playing itself.

  • Now how is this different than human learning? In the human mind, we expand the number of parameters until we are satisfied we have suitable means for achieving the objective function. If our existing strategy involves turning ten "knobs" but we suffer 5 losses in a row, we might stop and ask "what is in common here? What did the moment look like right before the games went south? What was the best move in each case?" From this, a new "knob" (parameter or heuristic) is brought into existence to provide proper move guidance in those situations.

  • A machine can optimize a given set of parameters. A human can optimize AND expand the number of parameters. A string of misses/losses/failures prompts examination to determine the common thread. This is where the human's "giant memory bank" helps because we automatically overlay/pattern-match our losses/misses/failures. We automatically start to notice the common features.

gödel, escher, bach summary:

  1. Self-referential systems can produce contradictions. Weaker systems don't share this flaw, but they also don't share the same abilities to describe mathematical theorems.

  2. There are layers between the input layer and the top-most layer where a 'symbol' is generated.

  3. The upper layers can think about lower layers, but lower layers cannot fathom or describe (or even think about) higher layers. There is a confinement of ideas to the layer where they manifest, or lower.

BIG DEAL #1:

Everyone has the same lower levels. We all have cerebral neurons receiving input from our nervous system. Your accent, your values, your world view all arise out of how that lowest level is organized. The same exact set of neurons can be organized to run totally different programs. The analogy is to that of an ant colony where groups of ants dig tunnels, other groups forage, and still others care for larvae. They are all identical ants, but the colony structure arises as a result of the specialization.

The Ant Fugue illustrates this via the most interesting story in the book. There is an intelligent person who converses with an ant colony by writing symbols into the sand, then observing the patterns the ants adopt when crawling across them. The colony has a personality and sense of humor, even though literally none of the ants is aware the conversation is taking place. The personality has a name: Aunt Hillary. One day, Aunt Hillary dies when a heavy rainfall floods the anthill, causing panic and the disoriented ants frantically run about, but none of them die. Though every ant survived, the new emergent structure has a completely new personality, and calls itself JSF. The sense of humor is gone, but a keen mathematical awareness has emerged. Reading this, I had the epiphany hit me and a recollection of Taleb's words from Skin in the Game "Don't use the behavior of a person as a guide for how people will behave. They are a different as a toaster is from a rabbit." The neurons in our brains are all pretty much the same and we have pretty much the same number of them. How they become integrated and structured "neurons that fire together, wire together" determines our individual personalities.

In one of the final chapters of On Intelligence, Hawkins wrestles with a definition of consciousness, ultimately calling it undefinable. In fact, Hawkins challenges anyone to prove they are, in fact, conscious. In contrast, Hofstadter seemingly effortlessly gives a compelling description-definition of consciousness:

  • to respond to situations very flexibly

  • to take advantage of fortuitous circumstances

  • to make sense out of ambiguous or contradictory messages

  • to recognize the relative importance of different elements of a situation

  • to find similarities between situations despite differences which may separate them

  • to find distinctions between situations despite similarities which may link them

  • to synthesize new concepts by taking old concepts and putting them together in new ways

  • to come up with ideas which are novel

conclusion

As long as we chase specific AI, any discussion about artificial intelligence becoming self-aware and taking over the world is preposterous. Don't even worry for one second about it.

General AI will be exciting when (if?) it arrives. But the challenge is so much greater, so be very skeptical when reading anything about breakthrough AI. For now, it is very safe to assume anyone talking about "AI" means the far less revolutionary specific AI.

Bonus

For an entertaining look at how specific AI programs found ways to achieve their objective functions, but not in the manner intended by the original programmer, have a read through this list:

https://docs.google.com/spreadsheets/d/e/2PACX-1vRPiprOaC3HsCf5Tuum8bRfzYUiKLRqJmbOoC-32JorNdfyTiRRsR7Ea5eWtvsWzuxo8bjOxCG84dAg/pubhtml

The list owner's site is here: https://vkrakovna.wordpress.com/2018/04/02/specification-gaming-examples-in-ai/