Machine intelligence is in a restart. Do you remember your first attempts at using Google? You probably entered a whole sentence and never expected a dump truck of those “search results”! You never wanted search results in the first place. You wanted answers, but were instead handed a list of educated guesses.
Missed Expectations
There is something very special about innocence. It can dream and go places where no engineer can go. Innocently, you and I both likely had expectations that were not met.
Missed expectations applied to our experiences with Google. Even in its earliest days, though, Google was a wonderful tool. You simply needed to alter your expectations.
A funny thing about expectations: After a reset, they can always be met. Should we expect Google or Bing to only find and aggregate articles – never analyzing them? That expectation will always be met.
What’s the problem, then? Well, we’re obviously lowering the bar. Those tools are now not so satisfying as our original expectations would project they could be. Lets then examine the idea of machine intelligence.
Machine Intelligence and Artificial Intelligence
Machine intelligence is a machine’s demonstration of human-like intelligence. (This is another area of missed expectations.)
In my innocence, I imposed my expectations of machine intelligence on Google, expecting human-level analysis of information. I had wanted an analysis on my topic, and the back-stories that justified that analysis. Google should do my research for me. Silly me!
Current methods of machine intelligence are at best artificial. The name artificial intelligence (AI) is so apt, deserved and appropriate.
Each decade since the 1950s introduced a new facet of machine intelligence, but none produced human-level intelligence. The list of human-level wannabes is pretty long. They all come under the moniker of “artificial intelligence.” Don’t set your expectations by them though:
- Massively parallel operations have been stated to be the driver of AI’s salvation. Honestly, the Illiac IV of the 1970s was to have done that. It did improve weather forecasting, though.
- So was CUDA, the use of massively-parallel graphics processors in a SIMD mode. It sure made animations better, though.
- Then there was the hope for gate arrays, and for massive silicon neural networks. If your problems are very narrow, they may be able to solve them. I would continue breathing, though.
- The genetic algorithm (GA) was to arrive at human-like intelligence with the GA’s most optimal algorithms. So not. The academics and Darpa coddled them, though the market rightly yawned.
- Then there was machine learning, a current foible of bio-mimetic theory. It applies a correlation-like brain stem activity to recognize patterns in sounds, images and “data”. It’s a great I/O process for a brain-like machine, but is no brain. The market raises a hopeful eye, only one, but it looks weary.
- The current ChatGPT craze is a text retrieval system – much like a database query. It looks cool because it interacts in English, but uses heuristics rather than human-level intelligence. The hyped crowd will again become disappointed in the end game.
What’s up with all these? They are all tool, artificial and not implementations of human-level intelligence..
The Human Brain Model
What is missing? A cohesive brain model, for one. It must tie together emotion, learning, experience backed decision-making, and a strong accounting for personal and social culture.
People repeatedly observe that people could make much better decisions without emotion that gets in the way. How’s that working for us? In another article, we show its fallacy.
The only machine emotion I have heard is the snarky “Re-calculating!” when you don’t follow a GPS navigator’s route plan. The annoyed tone of voice is only a guess at emotion, not the real thing. The navigator always complies, giving in in the end, and we expect that.
One big difference between a computer program and a human is that you teach a human facts and give him a set of tasks. Ideally, the tasks are in his realm of experience, so you don’t have to tell him how to do the task, but what to do. (Of course, there are variations on that.)
By contrast, the computer program is rigidly scripted. If customer’s needs are well defined, and the programmer implements all possible cases, you have a happy customer. What if either the customer/client or the programmer did not think of some situation that could arise? (By the way, the customer might be NASA and the program is riding a mission to Pluto.) Failure, maybe loss or death.
There is always the script Revision 2. New decision points are added, no new features. Not a good scenario, but pretty common today. Not very intelligent.
If your machine intelligence was human-like, it could have analyzed the current situation (never before encountered) and propose an alternative decision path. It could have also gone ahead and executed the decision, if its phone link home was dead. It’s better than the mission being dead.
Bio-Mimetic or Psycho-Mimetic?
The mere suggestion of bio-mimetic usage is a tacit acknowledgement is that God (for which the euphemism is “nature”) did it right. It is that some intelligence greater than ours defined a system (the brain) using carbon technology, and did it very right. The latter premise is implicit in that the biological system was done right enough to be deemed worthy of a copy.
There is a fundamental problem with bio-mimetic, though. God implemented the brain function in a carbon system. Some processes were implemented specifically because their associated systems were implemented in carbon. That process part may be very sub-optimal when implemented in software, or in silicon. I’m just sayin.
So said, the bio-mimetic option is really moot. It’s only appropriate if you’re implementing some carbon-specific item such as DNA programming. That’s not common. You can do a wonderful implementation of a bio-mimetic process in software, not helpful if it is a purely biological function of a carbon brain.
A Psycho-Mimetic Implementation
A psycho-mimetic implementation is needed. To do that, be prepared to understand its underpinnings. Be prepared to deal with the breadth of real-world issues such as emotion, the spiritual person, personal culture, motivation, finance, the political realm and complex humor. That’s reasonable, isn’t it? To most people this is an unbounded problem set.
I will challenge any person of science to develop a system of instruction that optimally trains and broadly educates a generation of K-12 human minds. If a person of science cannot do this, they are even less likely to understand how to create a mind-like emulation of the human. If they were capable of such creation, then they would well know how to educate a generation of K-12 minds.
Self-Imposed Restrictions
The barrier to entry for the psycho-mimetic implementation is exceedingly high. A rather confounding requirement to understand brain function is… humility. Most academic or DoD agency scientists are self-limited by their lack of that, in the name of pure science.
For example, within the brain, matters of conscience appear to almost purely relate to external communications with the non-physical realm. That is is hard to justify for pure science! Therefore, to not go there restricts that realm from options open to science. After all, a related scientific validation plan is impossible to either propose, create or accept.
The barrier to entry for genuine machine intelligence is therefore not intrinsic to the nature of the problem. Rather, it is a self-limiting, self imposed barrier.
Remember the earlier reference to innocence? Their innocence is indeed lost, so there are some places the DoD academic simply cannot go, dream or explore. This is a major reason why the psycho-mimetic implementation has been traded out for the bio-mimetic model, a serious charge.
Many external observers of these two institutions are fully aware of the above institutional mind set… and simply avoid engagement with them.
Conclusions
In this paper, I have examined the expectations of machine intelligence, exploring the different forms of machine intelligence. Most of these are artificial in nature. I have examined what is needed to implement it and why the most capable people fail to produce real machine intelligence.
In a subsequent paper, I examine the more recent area of linguistics-based Human-Like Intelligence (HLI). It replaces the artificial aspects of machine intelligence.
Thomas A. Visel