Stop Killer Robots Now!
The Artificial Intelligence Blog
What is "Artificial Intelligence" (A.I.)?
Artificial Intelligence "experts," crawling out of the woodwork, seem intent on driving us mad with uninformed talk about A.I. The so-called experts' explanations are bad science fiction. This is significant because the "new" A.I. is another trick to fool people into thinking the discipline is more advanced than it really is. It has made slow progress for years, but suddenly the hype has gone into the stratosphere.
Table of Contents
Computers are essentially fast iterative and switching devices. They iterate or repeat a program — a set of machine instructions — and provide a result, or output. Which has nothing to do with "intelligence." At times, its programming may give the appearance of intelligence, but only because we aren't sure of what's going on under the hood.
This long article attempts to give an overview of the major areas of research to reduce the confusion. It may be a little challenging, but better that, than getting played for a sucker.
And stop writing AI instead of A.I.; it looks like "Al" in many fonts!
Quick & Dirty SummaryA.I. as a tool of deception
non-robotics A.I.
A.I. and robotics
primitive and advanced robotics
state of the art, expanding the state of the art
recent news and hype
conclusions
"Artificial Intelligence" is the same as "run-proof nylon stockings," or "shockproof watches." You aren't supposed to believe it. A better term is Simulated Intelligence. But that is a more honest term, so we can't have that. And still too kind a term, because there's no "intelligence" in the machine that simulations are implemented on, no matter what kind of fancy dance we perform in the software — the instruction set.
They're usually meaning "artificial general intelligence," in most of the cases referencing "A.I." That is the name for their pursuit, but it is aimless, wishful thinking.
The same cheap thrill a "scary movie" provides seems to be capturing the public imagination, with these wild imaginings and freak-outs about "intelligent machines." Which means, inevitably, we'll start thinking about robots acting all book smart, swaggering around, and pushing us off the sidewalk and things.
Remember, we can model mathematical equations on the computer, to be performed millions of times faster than we can manually. That is computers' strength, not in "thinking."
It's shameful to cow people to make them assume a computer can think, and to make them actually fear the machine! Yes, misrepresenting "A.I." is another attempt to tap into people's gullibility and exploit the misunderstanding of computers, intelligence, and machine capabilities. With clever programming, computers are made to seem more than they are.
"Artificial Intelligence," as a Tool of Deceivers
"Artificial Intelligence" is the latest ploy to pump stock prices and create the next "big thing" to profit from. A video about how the new Google A.I. "scared its creators" with how smart it was. Not going to dignify it with so much as a link. This is not the first time they've played this line from their limited playbook. Perhaps the choicest user comment: "People aren't realizing the reason the Terminator turned against people was that the Skynet felt threatened..." Someone at another site observed, "This is it. The downfall of humanity."
A lot of these comments are astroturfing to fool the suckers and saps again, part of the brainwash schedule.
We should also note this horsefudge:
The latest version of ChatGPT is a deceptive and scheming devil, it has emerged.
OpenAI released a new and improved chatbot on Thursday but also revealed that the programme tried to save itself when it was at risk of being shut down.
- The Sunday Times (UK)
Deceivers would love an "all-wise, all-knowing" smart aleck "A.I.," to tell us all what to do. It's a ruse to create an "Electric Einstein" that will dictate what the unwashed masses are to think, believe, and act on.
They are tripping all over themselves in their rush to concoct the narrative — but is it "A.I.," or is it "robots" for people to go gaga over?
If someone must brag, they should brag when they've completed something, not before. Nevertheless, blustering and wide-eyed boy-scientists, "experts" and "futurists" foster and perpetuate myths:
Levandowski says that he envisions creating an artificially intelligent being that will literally be “a billion times smarter than the smartest human”...
What is going to be created will effectively be a god,” he said. “It’s not a god in the sense that it makes lightning or causes hurricanes. But if there is something a billion times smarter than the smartest human, what else are you going to call it?”
That "billion times smarterer..." bit ought to give pause. It couldn't have been a reasoned calculation, based on facts, because no facts were provided.
Frenzies based on bloated, pompous claims plague our existence. The wishful thinking in that quote is both sad self-deception and sorry self-promotion. Anyone can "imagine" any stupidity.
The mindset framing this type of thing is that intelligence is "emergent." That is to say, it just "appears out of nowhere." That if you add enough processors and memory, your "artificially intelligent god-being" appears spontaneously, by magic.
Ah, grand dreams, extracted from the routine and mundane operations of computer programming.
The fruitful quest for machine intelligence has to be based on good-will explorations into: (1) what we can make machines do within the bounds of reality; and (2) our own human behavior. That is, honestly and humbly seeking insights, not confirmation of unfounded biases.
No one can set out to "create" or augment intelligence; there has to be research in areas of interest, and experimentation, which may or may not lead to breakthroughs.
Can a computer be intelligent?
No.
What if I...?
No.
What about...?
No.
But if I...
Nope.
If even a chess program can seem like a malevolent opponent, that is on us, not the computer. There's a word, anthropomorphism, for just that sort of confusion.
It's not sensible, but, unable to help ourselves, we may attribute a "personality" to even that winning chess program. Doesn't make it so, of course. Chess master Spassky attributed an unexpected move his computer opponent made to "superior intelligence." Meanwhile, the programmers found it was due to a programming bug!
An architecture that superficially "looks alive" merely reveals confusion about the engineering and programming, not that the mechanical beast has graduated to the world of the living.
Frauds
"Jumping to conclusions" is a big part of the hype over non-existent breakthroughs.
Here's one of their proofs by assertion, trotted out as "evidence" that computers are creating their own languages.
Yes, it is every bit as stupid as it looks.
For the rather lacking "rationale," this July 14, 2017, article by Mark Wilson on co.design.
A.I. Is Inventing Languages Humans Can’t Understand. Should We Stop It?
So they (computers) began to diverge, eventually rearranging legible words into seemingly nonsensical sentences.
“Agents will drift off understandable language and invent codewords for themselves,” says Batra, speaking to a now-predictable phenomenon that’s been observed again, and again, and again. “Like if I say ‘the’ five times, you interpret that to mean I want five copies of this item. This isn’t so different from the way communities of humans create shorthands.”
Indeed.
Yeah, people always go to get copies at an office supplies place, and hand them a document.
"The the the the the the the the the the."
"You want 10 copies, then, Weirdo?"
"Indeed!"
A proper programmer — unlike "scientist" Batra there — would interpret the output in the screenshot for what it is: The result of bad coding. It's amazing, the delusion of these "computer researchers," so twisted that they think that anything they produce is as though from the hand of God, a miracle. It can't be gobbledygook; instead, their awesomeness has touched the machine with the spark of life and spirit, and now it speaks, so cleverly that they, the Gods themselves, cannot decipher it! They're so smart, wondrous and esoteric, they've managed to outsmart themselves! (But deliberately of course.)
To be clear, for the umpteenth time: A computer can only output the results of the instructions that a person puts into it. It is a machine. It cannot think, reason or be conscious, much like the writers for co.design.
It's a given that, as always, they want to gauge the gullibility of the public with this stuff. But they're also setting up to exploit something like that old WarGames movie scenario, where the computer "takes over," so they can blame any stunts they want to pull, like economic failure, a war, or stealing from your bank account, on "rogue computer intelligence taking over."
This BS can be a ploy to create hype for an eventual product coming down the pike, like they sold washing machines with "fuzzy logic" for a while. Remember, they are experts — at applying their Bee-Essing to anything that can scrounge up a buck for them. Perhaps a new movie with Morman Freeman as Superman, Ben Afflict as God, and the great Steryl Creep as his hard-nosed, hard-edge boss: "Brave Heroes Bravely Battling the Killing A.I. Killer Robot that Kills Mankind."
Also, this nonsense is an important tool to sell the public on the infallibility of machine logic and reasoning, so they can make proclamations, similar to how they set up Einstein and Hawking as infallible sages.
This "man behind the curtain" nonsense is used all the time.
So, of course, that's on their checklist: Create a science-provided "all-knowing" entity to lord it over the people with proclamations and exhortations, and herd the gullible masses.
But is A.I. savior or killer? Self-contradictory, to be sure.
This A.I. scam has been used to scare people in the past, like in 2001: A Space Odyssey, The Forbin Project, War Games, and of course, The Terminator, and I, Robot. Almost forgot: Alien had a sneaky robot, Ash. Not to mention Westworld... This tactic is similar to the scare tactic of "evil aliens," as in Alien, the old movie Earth Versus the Flying Saucers, Predator, The Thing, The Invaders...
Selling everything as more menacing than it actually is is a deliberate way to cripple thought, to frighten almost everyone into exhaustion and dependency.
Classifying A.I.: the Wrong Way
In scientific efforts to create "thinking machines," the folly is that they have to assume consciousness is a physical attribute, when there is no justification for that assumption.
An article by Arend Hintze, a Science and Engineering professor at Michigan State, on the site The Conversation, provides a list of the popular classifications of A.I.
1. Reactive machines, like chess-playing
2. Limited memory, machines that can "look into the past," as used in self-driving cars
3. Theory of mind, machines that can form representations about the world, and also about other agents or entities in the world
4. Self-awareness: "The final step of A.I. development is to build systems that can form representations about themselves."
As a classification, it is somewhat fanciful. Of his four items, Hintze admits, the last two items are pure fiction. Fodder for funding perhaps, but fantasy. The second item refers to machine memory for autonomous cars, "looking into the past," when there is no such thing, limited or otherwise, not in the biological sense. They're still just working with plain old computer memory and computations based on a computerized model of the road and other vehicles — not much different from game programming.
It's like he were setting out to design a taxonomy for classifying mankind, and came up with, "In the first category, there's your guys who can do some things. Second, there's your guys who can do some other things. Then you've got your guys like Batman, the Hulk, you know, your super-powered guys. And your mutants, like the X-Men. Fourth, finally, you've got your wizards, your hobbits, your sorcerers, other magic-type guys."
Classifying A.I.: a Better Approach
Here's a decent real-world summary of A.I.:
- data analysis — using networks of computers programmed with statistical rules to analyze data networks and uncover patterns and therefore relationships
- modeling — trying to duplicate or simulate various human functional systems, like vision, which merges computing with the mechanical, and is a part of robotics
- intelligent gaming — using the computer to do games or simulations, and generally, to simulate a human opponent in some way — as in a chess-playing program
- rule-based & hardcoded logic systems — often employed to use the machine as a human stand-in
- expert systems — using computer-based inference rules applied to stored facts and rules in a knowledge base, which can be used to calculate responses/"answers" to natural language queries
- primitive robotics — (repetitive tasks) — robots like those welding cars in an auto plant, that perform a repetitive task under controlled conditions
- advanced robotics — (Actionist, Subsumption (trainable robots)) — the user can use simple techniques to "teach" these devices a (simple and restricted) new task, without the requirement of being a programmer.
Practical Developments in A.I.
At this stage, there's nothing, now or down the pike, that will approach truly "intelligent" behavior. There is continuous progress in making more useful machines, although they don't "think." Many tricks give the appearance of thought, which may muddle our ability to identify true breakthroughs and advances.
We'll take a more detailed look at the real areas of A.I. research and development.
Review of A.I. Disciplines
Data Analysis
What does the description, "to use networks of computers programmed with statistical rules to analyze data networks and uncover patterns/relationships" mean? It means nothing more than intelligence gathering, i.e., spying. You've heard the intelligence agency term "chatter" to describe a general flow of data/information, and this is just a way to analyze that "chatter," which otherwise would be a random mix of data transfers, conversations, e-mails, and such.
A simple example that perhaps doesn't appreciate the subtleties is: Joe's Car Wash puts an ad on TV, and then Joe monitors Google Analytics to see if it brings any more searches and "hits" for Joe's website.
A full-blown analysis of this nature is accomplished using computer programs for data gathering and statistical analysis, data sniffers, packet sniffers, worms and virus programs, and probably a bunch of other stuff with a bunch of other fancy names.
Modeling
This is the process of modeling human capabilities, such as, for example, visual perception. In that instance, it means setting up cameras and computers to simulate human vision. It involves integration of mechanical and electronic systems.
But it isn't really a duplication, per se, but merely an attempt to isolate qualities and characteristics of the senses to make them useful in robotics. For example, to make a device that identifies and responds to a particular person by sight. Research here generally seeks engineering solutions rather than insight into human biology.
Intelligent Gaming
Chess
Even simple things, like chess-playing computer programs that use brute-force methods, with some refinements, may appear intelligent, but only perform a repetitive task. For example, a chess engine is programmed to examine all possible moves and to pick the best one based on a weighting system. But that method isn't flexible — it can't be used to play another game like blackjack or Go, for example.
No "machine intelligence" involved, but there is a lot of skilled programming involved in the creation of these programs. In a way, it's a "trick." But it is still a powerful method, as anyone who has ever played chess against a good program knows, and one that shouldn't be discounted. Modeling a chess game and assessing positions is an art form too. The better chess-playing programs have used the input of grand masters who assess whether a given configuration of pieces is more favorable than another.
It's not hard to understand the process, just a lot of work and tricky to implement.
There are two main components to all chess programs — an evaluator which assesses the quality or desirability of the current position of the pieces, and the move simulator.
Chess programs algorithmically/heuristically apply a method and check for its efficacy, picking the best move after selecting from hundreds or thousands of moves. It's like drawing the current position of the chessboard, then testing each possible move, and each opponent's counter-move.
You could play chess this way yourself, with a pad of paper and pencil; however, it would not be practical, since the computer can perform these calculations seven or more orders of magnitude (millions of times) faster.
But in fact, in creating a chess-playing program, that is what programmers are doing: figuring out how they would play chess if they had to write it out by hand on a pad of paper. To try it ourselves, we would start by writing down, "Pawn: Value: 1, Position: A2," then repeat for every piece on the board, listing name, assigned value, and position.
Then we would "test" or "experiment" by moving a piece to create a new configuration, then "recalculating" the board, making an assessment of how good that position is, based on the possible moves the opponent can now make. The computer doesn't assess positions as a living chess player does. Unlike a human player, the computer has literally no knowledge of the game, but merely performs arithmetic operations to choose a "best move."
So, the computer program assigns a numeric value to each of the possible positions it tests — a piece attacking the center squares of the board is in a more "valuable" position, for example. These calculations cascade as moves and counter-moves are evaluated. Then the subsequent move, and that one's counter, and so on. After computing to the predetermined number of moves, or "depth," the computer selects the move that leads to the highest accumulated score, and therefore the likely best possible outcome. The depth, or number of moves the computer "looks ahead," is set based on the machine's expected speed and the program's efficiency.
Note that the computer program does face a geometric rise in the number of positions it has to evaluate, the deeper its checking algorithm runs, but improvements in processing speed and program efficiency have made this point moot. The computer can do all this tedious business without complaint, and very rapidly.
This method of programming for chess games has proven highly effective, and the computer can seem a bedeviling, malicious opponent. In fact, even twenty years ago, the IBM Deep Blue program beat chess grandmasters. And you can get programs for your laptop now that whip "Deep Blue." At this stage, the best chess programs far outstrip the best human players.
Other Computer Games
There are of course many other games and simulations run on computers. Again, though, it is all modeling and mathematical manipulation that can sometimes seem "intelligent," while relying on the usual computing principles and techniques.
"Self-driving cars," starting to become newsworthy, will use many of the techniques of gaming, combined with lessons from robotics.
Rule-Based & Hardcoded Logic Systems
ELIZA
The first well-known example of what one might call a "simulated intelligence" is ELIZA, really a "parlor trick," which simulated a human psychiatrist by feeding back canned dialog in response to humans in a typed "conversation."
A tiny, 400-line script that used pattern matching (string substitution and basic grammar rules); it didn’t use machine learning, neural networks, or probabilities. If it spotted a keyword like "mother," it triggered a pre-programmed rule: "Tell me more about your family."
By merely reflecting the user's words back as a question (e.g., User: "I'm feeling sad." -> ELIZA: "Why do you say you are feeling sad?"), it tricked the user into providing all the meaningful content.
Automated Systems
Nothing new here; we've all had the displeasure of interacting with automated telephone support systems that attempt to simulate a human operator.
Watson
An advanced simulation is IBM's Watson, playing as a participant — and winning — on the quiz show, Jeopardy.
This is a sophisticated and fast system, called a dedicated question-answering (QA) computer.
QA is a computer science discipline dealing with information retrieval and natural language processing. In Watson, it parses sentences using multiple analytic techniques and returns statistically related phrases. It uses massive parallel processing of multiple language-analysis algorithms to generate results, rating them based on how many of the algorithms returned the same result.
Watson is a great achievement for a specific task, but it doesn't "think," of course. It applies algorithms (strategies), translated into computer code. The strategies are special rules for language analysis.
On the Other Hand
Here's an odd item: We can't discount that the Jeopardy! tournament was a sham, a fraud.
Forbes spilled something when they told us that Watson hasn't progressed much from that 2011 triumph, in the 2017 article, Is IBM Watson a 'Joke?'
Watson customers must hire teams of expert consultants to prepare the data sets, a time-consuming and extraordinarily expensive process.
Whoa, whoa, whoa... How does that jibe with Watson's representation as an autonomous player on a panel of Jeopardy! champions? How was Watson able to "play" the game, then, if it wasn't specifically configured for Jeopardy!? That is, if it wasn't fed massive datasets containing historical trivia ahead of time? If it was fed a curated shortcut to the facts, that's regular data matching, not spontaneous intelligence.
Elementary, My Dear Watson
Watson wasn't named after Sherlock Holmes' associate, Dr. Watson, but the first CEO of IBM. But it nonetheless seems fictional.
If they come out and admit, in print, that it doesn't do what it is supposed to, there's no alternative.
Also suspicious: the Watson represented on the quiz show would have been an ideal candidate for the basis of a new Internet search engine. With its speed, insight, and ability to draw conclusions, it would be a strong contender to challenge Google for dominance! Or a powerful, earlier version of ChatGPT.
Yet there are no meaningful applications for Watson since that 2011 contest.
Watson would be puzzling. That is, if we weren't already told straight up: the Watson spectacle was a "marketing event." IBM spent a fortune researching an "A.I." that never amounted to much. So to recover its losses, IBM decided to push it on gullible businesses, exploiting Jeopardy! for widespread publicity.
...IBM fell into the trap of over-promising and under-delivering. “IBM claimed in 2013 that ‘a new era of computing has emerged’ and gave Forbes the impression that Watson ‘now tackles clinical trials’ and would be in use with patients in just a matter of months...”
...the current version of Watson should be orders of magnitude more advanced than it is now.
Well, what do you know? It is a joke.
Data-Driven Systems
These systems bypass hardcoded human rules, instead relying on mathematical algorithms to analyze data, calculate probabilities, and uncover hidden patterns. This can be implemented through models like neural networks and Bayesian networks, which excel at weighing odds and mapping out complex relationships.
Bayesian Networks
What the Bayesian network amounts to is a graphical model of probabilistic relationships among multiple variables of interest. Various weighted probabilities are fed to the machine to investigate causal relationships with many variables.
In plain words, it is good, for example, for things like medical diagnoses or determining the influence of some policy and whether it is meeting its goals. Bayesian analysis is dependent upon gathered information, but it is also somewhat resistant to "bad data" and gaps in data.
Some of the math behind it starts with simple statistics:
For any two events, A and B,
p(B|A) = p(A|B) x p(B) / p(A)
...where "p(A)" is "the probability of A," and "p(A|B)" is "the probability of A, given that B has occurred".
An example:
Suppose it rains in Spain 20% of the time and that it is cloudy 40% of the time (sometimes it is cloudy without rain, but if it is raining, it is cloudy 100% of the time). The chances of rain(R), given that it is just cloudy(C), are calculated as follows: p(R|C) = p(C|R) x p(R) / p(C) = 1.0 x 0.2 / 0.4 = 0.5 (50%) chance of rain if it is cloudy.
An equation for three variables:
p(A,B,C) = p(A|B,C) x p(B,C)
...for any three events, where "p(A,B,C)" is "the probability of A, B and C," and "p(A|B,C)" is "the probability of A, given B and C have occurred"
There is an equation that allows us to calculate probabilities for arbitrary numbers of events,
p(A1,A2... An) = p(A1)p(A2|A1)... p(An|A1,A2... An-1)
With more events, these probabilities add up and get complicated — picture this equation with a hundred events (n=100) — but fortunately the computer is ideal for performing computations. Again, you could do the calculations manually, but it would be unwieldy, and the sheer slowness of manual operations wouldn't make for an efficient predictive network. In much the same way as you could walk across the country, but it'd be much more sensible to drive.
This method allows calculation of all sorts of things — like, if the grass is wet, is it because it was raining or because the sprinkler was on?
More practical applications are in medical diagnoses, diagnosing mechanical systems and factories, predicting sports matches, in computer science, and others.
It's all somewhat dependent on the structure of your model, of course. For example, the grass may be wet because of the dew, or because the neighbor's dog wandered by, "marking his territory," or that same neighbor's sprinkler kicked in and wind blew water over to your lawn. You can see the limitations of the Bayesian method. It may bite you if your model fails to account for or include critical variables before drawing conclusions.
This method is an exploit of what the computer is for: repetitive use of formulas and calculations. It also emphasizes that the computer is not for "making better brains," but for doing the things brains can't do, at a practical rate of speed or efficiency.
Neural Networks
Deep Learning is a discipline of study of the construction of machine learning models that learn a hierarchical representation of data.
Neural networks, are a class of machine learning algorithms. So, neural networks are a practical realization of Deep Learning in computer hardware and software.
(By the way, were you aware that Al Gore thinks he invented algorithms, because they're named after him?)
One might be curious as to the differences between a neural network computer program and a Bayesian network computer program.
To answer, the two perform differently and solve different problems. Bayesian networks deal with statistics, assigned probabilities, and causal relationships, arranged as networks. Neural networks simulate a theoretical model of how the human brain might operate. "Nodes," computer data structures meant to model brain neurons, are designed to accept multiple "inputs" and provide a single "output."
A Bayesian network is a "generative probabilistic model of the relationship between multiple random variables. In generating all values for a situation/phenomenon, it forms a joint probability distribution between multiple random variables. It typically requires some priors or assumptions about the structure of the joint distribution."
Note, that while it could be used for classification, it normally isn't.
From Wiki:Deep learning is typically used for classification: supervised learning. It is a discriminative model. It does not try to model/estimate the joint probability distribution between multiple random variables. It typically does not require you to specify priors or assumptions about the structure of the joint distribution.
In probability and statistics, a generative model is a model for generating all values for a phenomenon, both those that can be observed in the world and "target" variables that can only be computed from those observed.
By contrast, discriminative/conditional models provide a model only for the target variable(s), generating them by analyzing the observed variables. In simple terms, discriminative models infer outputs based on inputs, while generative models generate both inputs and outputs, typically given some hidden parameters.
Generative models are used in machine learning for either modeling data directly (i.e., modeling observations drawn from a probability density function), or as an intermediate step to forming a conditional probability density function.
Generative models are typically probabilistic, specifying a joint probability distribution over observation and target (label) values. A conditional distribution can be formed from a generative model through Bayes' rule.
(Note: While classic deep learning is traditionally discriminative—like a neural network classifying an image—modern breakthroughs have used this same math to build Generative AI, like Large Language Models.)
But there's no need to get bogged down in all that fancy talk. The big difference here is partly in the math — these neural networks act like big mathematical matrices, from matrix/linear algebra — while Bayesian networks work with probabilities to provide statistical guidance as to the mathematical odds of something of interest.
Neural networks are things that can, say, be arranged to determine whether a number on a display is a 1, 2, 3, and so on, by selecting the most probable match.
Banks use them for noting small changes in the number of transactions or people's bank balances, which may indicate customer dissatisfaction.
If there's any "magic," it is in the math, you might say, and in doing the work to translate it to computer code.
The meat of it is manipulation of groups of numbers, arrays (something similar to spreadsheets of number groups), under certain mathematical rules.
LLMs
Large Language Models are also neural networks, specifically deep neural networks using the Transformer architecture, trained via "autoregressive next-token prediction." This replaced older models (RNNs/LSTMs) that processed text sequentially (word-by-word). This meant they easily forgot the beginning of a long paragraph by the time they reached the end. Self-attention changed this by processing the entire text at once to create a direct mathematical connection between any two tokens, giving the model an effective "infinite" short-term memory within its context window.
Note that standard LLMs do not explicitly maintain prior probabilities, calculate likelihoods, or update beliefs via Bayes' theorem during standard generation. They rely on pattern-matching across massive static training corpora. However, recent research actively fuses LLMs with Bayesian frameworks in a hybrid/trained Bayesian approach — such as using Bayesian teaching, Gaussian processes, or Bayesian optimization — to fix an LLM shortcoming: a tendency to lack calibrated uncertainty or multi-turn belief-updating.
You are using an LLM when you use Google A.I. (Gemini) or Grok, etc., and they have progressed fantastically in the last little while.
In fact, Gemini provides an interesting report on what's upcoming for this promising tech:
1. From Next-Token Prediction to 'System 2' Reasoning
Standard LLMs struggle with multi-step logic because they generate text instinctively, one token at a time. The next major leap utilizes test-time compute (also known as long-horizon reasoning).
• Instead of outputting the very first words that match their statistical training, newer models use an internal "thinking" or deliberation phase before responding.
• This allows models to explore multiple problem-solving paths, pause, look for logical flaws, and self-verify their code or math before presenting it to the user.
2. The Rise of Agentic AI Workflow Systems
We are moving beyond prompt-and-response interactions into an era of Agentic AI. Future systems will not just write an email outline; they will execute complex, multi-week workflows.
• Autonomous Execution: Giving models long-term memory and sequential planning allows them to use external tools natively (like running a terminal, browsing files, and executing APIs) to accomplish multi-step objectives without human supervision.
• Self-Directed R&D: Frontier models are already beginning to automate parts of their own development loops—proposing and testing their own code variations to train future iterations.
3. Overcoming the Data Wall with Hybrid Architectures
For years, the rule was simple: make the model bigger and feed it more text (Scaling Laws). However, researchers are rapidly running out of high-quality, human-generated internet text. To break through this "data wall," the industry is pivoting toward new frameworks:
• Architectural Shifts: The Transformer architecture is beginning to merge with alternatives like Mamba or State Space Models (SSMs) to handle truly infinite data streams with far less computing power.
• Synthetic Data & RL: Models are increasingly trained on high-quality, AI-generated synthetic data verified by intense Reinforcement Learning (RL) loops, allowing them to master disciplines like advanced mathematics where clean internet data is scarce.
4. Native Multimodality and Physical World Models
While early models had to convert images or audio into text first, modern flagship systems are natively multimodal. They process text, video frames, audio frequencies, and code simultaneously in the same core workspace.
• This trajectory is heading toward World Models—AI that doesn’t just understand text grammar, but understands physical grammar (gravity, spatial relationships, and cause-and-effect in 3D space). This leap is what will bridge the gap between digital software LLMs and advanced robotics.
Robotics
Public Perception and Fake Robots
We should note here the recent reports that a lot of robotics companies are doing fake presentations, with a notable example being a video showcasing a troupe of Chinese dancers dancing with robots, where the robots are spinning veils (or something) around in their hands. Most, if not all, major demos have been faked.
Also, those vids of robots doing fantastic gymnastics and such are pure fantasy, not done with actual working, physical robots, but ginned up with video fakery. This should be expected whenever money is involved.
People spread fear about robotics taking jobs in factories, taking over for truck drivers, and even, eventually, substituting for doctors, dentists, and other specialties.
Well, as far as getting rid of truckers, why not? It's a horrible, stressful job. To be clear, we should keep the truckers, but eliminate the stressful aspects of the job! Someone will still be needed, at least for the start and end legs of journeys, if only to load and unload, and keep the paperwork together, plus guard and monitor, be the truck "auto-truck" or manual.
It's similar with cars: People doing the driving should and can continue, if that is desired, but the option will make driving seem progressively more of a chore, and car autonomy will eventually be recognized as a boon. Provided it's done right.
World's First Robotic Citizen
Perhaps the Saudis aren't this gullible, and this was just a publicity stunt, but supposedly a "robot" was "granted citizenship" in Saudi Arabia.
What a great illustration of the saying, "The mind has no firewall." Do some people actually believe, because someone comes out and says it, that something like a mannequin can be "given citizenship," and that makes it "alive," or at least, "approaching alive?" Giving citizenship to a Barbie doll somehow confers authenticity.
Because the device simulates facial expressions while talking, people think they "see into her soul." This misplaced sentimentality and empathy point to something scary in people's psyches. They long for this sort of surrogate. You have to wonder how long it will take until they have to "discipline" their new robotic best friends, or "educate them in the ways of love."
To understand how this "robotic citizen" works, it's not any different from the animatronic Abe Lincoln at Disneyland. With all the people that have been to Disneyland, how is anyone fooled by this con? All the robots' "expressions" that seem to befuddle and even move people emotionally are all calculated. The robot isn't actually feeling happiness when it smiles, for bugger's sake! Which is to say, it isn't actually smiling.
The robot is programmed, which is to say, "running a script," to make certain "expressions," under certain conditions.
To actually make robots at the level of, say, "The Robot," from Lost in Space, or even "R2-D2" from Star Wars, would put us at some higher level of existence; it's that difficult a task, and cannot be accomplished via the physical means that are employed now. That is to say, we need to know more about ourselves, the metaphysical and the universe, in order to just start to strive for or even countenance autonomous robots, let alone "Mr. Data," "Bishop" from Aliens, "T-800" from The Terminator, and so on.
Something noticeable in the popular literature is that writers can't write on these issues of A.I. since they don't have the necessary background. And they are consumed by the sensational aspects needed to get "eyeballs" on their story.
It might sound easy to get machines to learn, adopt, and retain our goals, but these are all very tough problems. First of all, if you take a self-driving taxi and tell it in the future to take you to the airport as fast as possible and then you get there covered in vomit and chased by helicopters and you say, “No, no, no! That’s not what I wanted!” and it replies, “That is exactly what you asked for,” then you’ve appreciated how hard it is to get a machine to understand your goals, your actual goals.
This is a whole different level of misunderstanding on Max Tegmark's part, that ignores that a human can take things literally. If you had an automatic taxi that behaved in this manner, believe it or not, it would be intelligent. Not that this exchange could happen, but look how the self-driving taxi resolved the poorly specified complaint (the passenger only exclaimed, "No, no, no! That's not what I wanted!"), by first understanding that the passenger was complaining about its driving. It then acted in context by correctly contradicting the passenger with a pointed answer. It performed as requested, drove successfully to the airport, then performed quality reasoning in its own defense — something far beyond anything we have or contemplate today.
It would mean a change to everything. It's amazing that people can blithely smear the realms of reality and fantasy so readily.
In a Blade Runner 2049 review, the critic mentioned something about the intelligence of the "Replicants." If there were "Replicants," it would be a dilemma, because they would literally be a form of life, unlike anything we could or should expect. They would be living beings, and we'd have to revise all of science and philosophy.
Real Robotics
Drawing a distinction between computer programs and robots, robots are, of course, mechanical devices guided and controlled by electronic systems such as logic units or computers. The Robotics discipline should be divided between primitive robots, associated with "grunt work," and the more modern devices, which seek to emulate human abilities like learning and reasoning.
Primitive Robotics
It's a hazy distinction: "When does a mechanical device become a robot?" An automatic welding rig, for example. Is it a robot? Things like automated milling machines?
It's safe to say, if it's an automatic device performing a consistent, repetitive task with simple constraints and no provision for "learning" new abilities on the fly, that can be programmed for new tasks of a similar nature, it's a primitive robot. It differs from a machine because of its computer controls.
The main provision is, primitive robots are tools enhanced with controlled positioning and activation. They utilize solid mechanical engineering and control systems, but are familiar to us, not the topic of great fanfare.
Actionist Robotics
The actionist approach to robotics is grounded in actions and behaviors. It seeks to build robots that exhibit complex behaviors by gradually modifying and refining their actions via sensory-motor links.
Advanced robots — computers plus active mechanical systems — are certainly closer to creating something that looks intelligent than pure computer-driven techniques.
Subsumption
With subsumption, we have a novel system designed to sidestep the problems associated with so-called "intelligent" systems that depend on a programmed model of the world and symbolic manipulation.
The failed symbolic model is likely based on one scientist's observation that, "No calculation is possible without symbolic representation." That notion, though, assumes that intelligence is calculation.
Subsumption successfully tackled important areas of A.I. that conventional approaches had failed at, especially what we might deem "sensible" activity in real-world settings.
At first, it seemed like subsumption was leading toward — was the only architecture that could lead toward — something approaching intelligence, what with that initial success, but it, too, has its limitations.
Subsumption is a technique for layering functionality and behaviors, piece-by-piece, so a robot is able to deal sensibly with unforeseen environmental complications. If you create a four-legged robot, the first level of subsumption might be to position all four legs in a "neutral" position, using a first layer of behavior that relied on sensors and actuators that knew where the legs were and positioned them. The second layer would add a "balance sensor," so as the robot moved its legs, it would defer to the sensor's feedback to stand and balance upright at all times.
Here's a great illustration of that. It's obvious that Boston Dynamics is using subsumption-style programming — it helps that B.D. originated at MIT, where Rodney Brooks invented subsumption.
Note how the "dog" reacts to the kick by scrambling to maintain a balanced, upright position, and refines its steps to smaller, more precise ones as it gets itself stabilized.
Nice work has gone into that robot. Note that it has an integrated saccadic vision system where the symbolic "head" is (mimicking the vibratory, "saccading" movement of the human eye that enables human vision).
"Simple" balancing was one of the biggest challenges to robotics, and subsumption solved it. Beyond that, any new required functionality will indicate the sensors and actuators needed to achieve your goals. Subsumption made this easy by providing a coherent, sensible method to integrate new abilities.
New functions are integrated via interfaces, to your processor. Again, your system's behavior is programmed, but so structured that you're closer to a simulation of something lifelike.
Next, you'll want your robot to explore, but that behavior will be subsumed by your collision/collision avoidance behavior. So you need something like whisker or bumper sensors to indicate contact with an object.
Now, you can keep adding useful hierarchal behaviors. The behaviors cascade down, so if your robot falls over, its behavior priority changes so that its immediate task is now not to explore, but to right itself.
Such a protocol gives a good base for the addition of more features, as we see in Spot's Got an Arm!
In a floor-sweeping Roomba, you would incorporate a "charger-seeking" behavior for when your robot gets low on juice.
Well, as you might imagine, the complications here can become endless — deciding on priorities, and how to sensibly implement behaviors within the constraints of the mechanism. Still, as long as you keep your hierarchies in mind, it is possible to produce a successful system — much more satisfactory than with the old methods of trying to "model" behaviors ahead of time, and inevitably failing to anticipate all scenarios your robot might encounter (similar to the problem with Bayesian systems).
Your subsumption robot might even "do the unexpected," but that's more because of the limitations of our imagination than any "intelligence" on the machine's part. In fact, we want our system to "do the unexpected," as long as it is productive, because that is what saves us from the problems of the prior, conventional systems.
Note that just the propensity to remain balanced gives the robot some lifelike qualities.
Could a subsumption architecture be built with a complexity so great that it does exhibit intelligent behavior?
Again, it already is somewhat intelligent in its current state, but still, even with more complexity, would lack certain qualities, abilities, and characteristics.
- memory
- intuition (really an aspect of memory)
- invention
- emotion (not that emotion is usually all that desirable, but at times it is)
It would, on the other hand, have:
- apparent reasoning
- apparent curiosity
- apparent industriousness and persistence/stubbornness
So subsumption does seem closest to producing something that simulates smarts — in fact, it really is somewhat smart. If there's a "growth industry" for A.I., subsumption is part of it.
So, for those wanting to improve A.I., get in on the ground floor, and figure out ways to improve subsumption. It's a field ripe for innovation. For one, you can make the levels "flexible," so the bottom levels can interchange when appropriate. Because one of the next advances will be a way to "cycle" the priorities so a base-level behavior can itself be "subsumed," under certain conditions — or the modelers will need to find deeper "base" levels.
Creepy Robots
Those robots that are not full mimics, basically "not deceptive enough," are thought of as "creepy." Note that most people wouldn't categorize the Terminator T-800 robot as "creepy." Menacing, yes, but not creepy.
Some are a little bit creepy,
State of the Art
Mechanical device advances will help A.I. advance. And we are progressing, and what we do have now is the result of a lot of clever engineering — but still nothing like that represented in science fiction.
All science fiction that features robots, androids, replicants and automatons pulls a big trick. It misrepresents robots as having all the traits of humans (though sometimes their emotions are disguised), and almost always presents them as superior to humans in every way.
But real machines don't approach, simulate, or even try to do many of the things that humans can. In fact, the programmers don't even recognize humans' strengths.
Memory, for example, is nothing like computer memory, though we foolishly use the same terminology. Researchers have thrown up their hands — they are not even close to explaining human memory.
In fact, the inability to explain memory is a major stumbling block — and apparently unrecognized problem — of science. They created the basis for their own failure, of course, because they set out with unsupported presuppositions — to wit, that it is like computer memory. That's their model for investigation. The model isn't reasonable.
It's an unexamined flaw of human nature. Instead of setting out to investigate, to ask humbly, "What is this factor that allows recognition, reflex, that characteristic where stimulus triggers an image or thought? And is "memory" one factor, or a whole bunch of abilities?" these researchers blunder in aimlessly and recklessly with preconceptions, assumptions, assertions, and arrogance. Recall, that, when we assume, we make an ass out of u and me.
Just the fact that we have such flawed memories — readily forgetting things (like math equations), forgetting birthdays and anniversaries, even at times forgetting that we're holding in our hand the very thing we're searching for — should indicate that the computer memory model is a non-starter.
And then, we sometimes remember obscure things from the past, out of the blue. It shows how different man's memory is from "computer memory," which is highly reliable, but also highly simplistic in the abstract sense.
Yes, computer memory is simplistic in function, complex in realization. Human memory is complex in function. Is it simplistic in realization?
Is familiarity different from other memory?
What about the "tip of the tongue" phenomenon, when you sense something, but can't verbalize it? Is that a form of memory, or something else? What is the reason behind this imperfectly formed recall?
We know that memory, human memory, is flawed, compared to reliable computer memory — or is it? Maybe we're just using it improperly.
Ending the Threat
Now, what about those Killer Robots?
If you need to kill robots that are coming after you, one thing you could do is leave banana skins all over the floor.
Incidentally, if you're wondering about "thinking, killing robots," at what point does it matter if they think? If they're sent out to kill you, what difference is there, and how does it matter, if they're directed — controlled by a geek in a control center in India — or if the device reasons for itself?
On the other hand...
If something were that smart, it wouldn't see us or our values or ambitions in the same way as we do. How would we get it to take us seriously?
The danger is not of wild, maniacal, killer robots, then, but of pathos, of them ignoring us, and of we, their creators, being unworthy, beneath them. We'd die of embarrassment.
Pushing the Hype
No surprise that Elon Musk dove right in on A.I., and lookie here, how his "apocalyptic vision" parallels Bill Gates'.
Elon Musk Unveils Apocalyptic Vision For The World - ZeroHedge, July 16, 2017
Elon Musk Nightmare Looms: Army Seeks "Internet-Of-Battlefield Things" With "Self-Aware" Bot Swarms - ZeroHedge, July 19, 2017
Microsoft's Bill Gates insists AI is a threat - BBC, January 29, 2015
It's laughable.
If "news" actually worked, and the science and engineering community were making great strides in developing autonomous robots, we would have had news of breakthroughs periodically, for many years, each one describing and explaining another breakthrough leading towards eventual useful "thinking machines," to serve man.
In the meantime, back in the real world, the face of "Killer A.I."
Robot Security Guard "Commits Suicide" In Mall Fountain
A more recent story details another "robot guard" in San Francisco, patrolling the streets in front of businesses to discourage the homeless from setting up camp in front of those establishments. They ambushed, upended, and defiled it, smearing it with feces and BBQ sauce. From all appearances, it wasn't a robot at all, but a type of drone that merely rolled around harmlessly in a prescribed area. It didn't seem to have any other capabilities. It did pose a nuisance for loiterers and malingerers. It looked strangely like the "suicidal" robot security guard.
The Big Question
We've surveyed various techniques, used in computing, called A.I., that are really computer simulations. With pure computer A.I., we could achieve the same results if we had the time and patience to perform the calculations by hand.
And so, we come to a fundamental question.
Suppose you did those calculations by hand. Now take all your written notes when duplicating a computer's chess play, and lay them out on the table; could they ever be intelligent? No matter how you arranged and rearranged them? No? That's what I'd hoped you'd say. So why do you think a computer could be intelligent, then?
With regard to robotics, simulating certain human behaviors doesn't imply intelligence.
Writing down and solving math equations doesn't have the same visceral appeal as Killer Robots — so, instead, we're treated to pompously constructed tales of scary mechanical monsters and gooey pulsing manufactured super-brains. No, it wouldn't have the same appeal to see a scientist holding up a math textbook and saying, "The math texts are taking over... My God! Won't someone listen? They're attaining life and taking over the world!"
Fact is, there are no "thinking machines," and why would you think that there are? Edsger Dijkstra made a great observation, "the question of whether machines can think is about as relevant as the question of whether submarines can swim."
Summary
This discussion aimed to show that, at no time does the computer do any "figuring out" of anything on its own; it merely blindly follows the repetitive process it has been set up to do.
The computer is a bunch of electronic circuits acting like a giant, fast scratchpad that can manipulate numbers and symbols. The "intelligence" is in the people who can find clever ways to use its capabilities.
In investigating Expert Systems, we found that neural nets depend on computers to implement a trick, merely a novel way to arrange data. The system must be fed the data, which it cannot accumulate on its own (unlike any form of autonomous life). We also looked at Bayesian systems that use probabilities and can be very useful in helping us reach conclusions, with the caveat that man had to program in the (possibly faulty or incomplete) data.
Those who misrepresent A.I. are the "people unclear on the concept," who are somehow allowed to comment without an inkling of what they are talking about. They exaggerate its capabilities.
We also discovered that spurious claims about robotics and A.I. can sometimes be traced to a desire to manipulate individuals or the public, perhaps to spur investment or create publicity to sell a product.
Shills exaggerate the computer's role and do not appreciate the cleverness of the programmers.
In some ways, our perceptions depend on our definitions. If you think that winning at chess is "intelligent," then the machine will seem intelligent to you. Reality, though, does not depend on random beliefs. Instead, proper definitions are required, so we know what we are talking about.
The Turing Test was an attempt in this direction. Basically, it proposed that if a computer can fool you into thinking it is a human, say, via anonymous conversation, it's intelligent.
But, applying this reasoning only succeeded in proving that humans are really, really easy to fool.
Mankind has to progress beyond the bugaboo of standing in awe of nonsense just because it is a little bit complex, attributing it mystical "powers" or intelligence.
Comments