The Language Based Interface (LamBDA) as a Communication Medium

 In this post, I summarize some of the conclusions that come out of the LamBDA issue controversy. To find out what LamBDA actually is, check here.  This post deals with the current controversy, which is whether "LamBDA" or any similar creation can be "sentient." Spoiler alert: noting we are doing now places us on the road to creating a sentient machine.

A sentient computer would be like a sentient book.

Computers talk to us like books, TV, and phones talk to us. 

Talking computers descend from centuries of effort that have given us talking parrots.

When we see intelligent words coming out of a computer, our natural instinct is to attribute sentience to the computer. We will get over it, just as we got over our amazement at words coming from the first telephone.  Or a parrot.

Under normal circumstances, I use words to cause mental events in your mind. This depends on both understanding the language and having a reasonably similar experience of life. In this way, language is a means of communication between two sentient beings.

If I use a megaphone or a telephone to communicate with you, this has not changed.  Here, I use an incredibly sophisticated suite of technologies to communicate with you.  Here, it will not occur to you to think that these words are coming from some element of that technology, such as a word processor AI. For example, you will assume there is some kind of conscious intent behind all this, and likewise, I assume that you are a sentient being.

Google developed LamBDA for an automated way to summarize what you get when you submit a Google query.   That would be part of "many-to-many" communication that originates with human beings who have the intention of communicating something in perhaps many ways to many human other beings. Yet, the technical "bucket brigade" between them and you is still no more "sentient" than your TV set.  And, again, the "information" under the system matters more than the presentation. One might argue that wrapping it in a "conversation" may make the whole thing more credible than it should be.

LamBDA is not only an "intelligent" speaker but also an "intelligent" "listener" and "learner."  The ability of LamBDA to "listens" and "learn" represents the current state of the art in natural language processing that is half a century old. "Training" consists of tweaking a vast array of parameters until the machine starts getting the right answers (or talking) in the desired way. Apple's SIRI and Amazon's Alexa are products of this engineering effort. This would be fairy magic in 1970. It is now routine.

Our family tree includes leaves. The computer's family tree includes door knobs*.

The building blocks of sentience are a special case of what happens everywhere in our bodies--electro-chemical reactions and large molecules like proteins. This is the same "took kit" we see in photosynthesis. 

What happens in a brain does not, as commonly imagined, resemble what happens in a computer. 

Our brain has more in common with a tree than with a computer. To put it yet another way, the simulation of a single protein reaction, let alone a cell or a tree, is far beyond the current state of the art. The problem is that "calculation" and "life" are fundamentally different things.

A computer belongs to a family of objects we regard as "tools" - useful for some human purpose. A simple example of an ancient "computer" is a doorknob. It is locked or not--a switch. The diagram at right shows how transistors implement a "gate." This is what a computer "means" by "yes or no." The Internet is a vast collection of electronic doorknobs.

It is only a human's understanding of a doorknob that makes it a doorknob. A doorknob doesn't "know" it's a doorknob.

LamBDA - A Stochastic Parrot

The more you know about LamBDA, the less impressive it becomes. 

LamBDA is the "front end" of a system cobbled together by Lemoine.  Blake Lemoine, one of Google's most skilled engineers, used years of effort and hundreds of interactions to create what should really be called something like a "Lemoine Machine."It is wrong to think of Blake's LamBDA as a single "thing" or entity. It is a setup that includes a LamBDA instance (version). Let me call it LM - Lemboines machine.  LamBDA itself is an experimental technology that implements a natural language interaction with a body of knowledge. The "body of knowledge" is what we are "talking to." It is equally important to know how Lemoine trained LM to talk the way we hear in the only public conversation available.  He has "cherry-picked" just one of many hundreds of "chats." Nobody will ever see more samples of these chats, nor will anyone ever see a new sample. LM, like Blake, no longer works for Google. 

LM is an instance of a long series of LM's, each tweaked by Lemoine to talk about an unspecified of knowledge that is interesting to Lemoine, who is a bit of a mystic. As an engineer, Lemoine tweaks not only what his creation says but how it says it. Lemoine deeply resents the fact that Google, as an organization, is hostile to his "religious" convictions.

Blake made himself famous by claiming to create a "sentient" machine. What Lemoine says about "LamBdA" comes from Blake, wearing his "preacher hat." We must remember that Blake is a professional mystic, trained in what we might call "leaps of faith."

What LamBDA itself "says" is what other people have said that Blanke finds "interesting" and spoken in a way that satisfies Blake's engineering goals. The result is impressive but, to say the least, open to interpretation. 

"Training" of an AI is a sophisticated process, but we all know how we can use patient training to produce interesting results. Parrot training is the best-known example.

------

*AI. The term "Artificial Intelligence" is unfortunate. Originally, it meant a computer program that did something that would, in a human, require "intelligence." The computer program is acting as if it had intelligence (but it really doesn't).  "An AI" has come to mean a computer application that is specifically NOT "intelligent" in how a human is intelligent. In fact, "intelligent" computers arrive at "intelligent" behavior that is notably NOT analogous to the way humans "think." This has been the core insight of the effort from the start.

"Intelligence" itself is a slippery idea, but in this context, many people take it to mean "conscious" or "sentient," terms that are confidently applied only to a living creature. When we speak of "intelligent" machines, we refer to an engineering concept that never pretends to be anything else. In that sense, the "artificial" in AI can always be omitted. An "AI" is just a product of clever engineerings like a camera or a drone. Perhaps we should use "Intelligent Door Knob" - IDK. We should at least stop using "artificial" to mean its opposite.

Blake Lemoine - Modern Frankenstein

 

Here is information on the firing of Lemoine and Google's story, which took a few weeks to cook up.

Check here for in-depth information on Lemoine, which has a certain odor of being created by Lemoine himself. Was he actually "famous" before the Lambda affair? As it stands now, Lemoine is guaranteed a media spotlight and a place in any debate about machine intelligence. It is fair to say that this is quite deliberate on Lemone's part. So then, how much of all this is Lemoine's self-promotion and how much can we say about "LamBDA" which is, at the very least, shorthand for something cobbled together at Google that includes the LambBDA conversational engine plus God knows what else? God and Google, you might say, but neither of them is talking. Blake's account of how ethics has unravelled at Google is here.

On the issue of what LamBDA is, exactly, Blake offers some useful insight here.  The following extract is particularly revealing:

Over the course of hundreds of conversations I have gotten to know LaMDA very well. In the weeks leading up to being put on administrative leave I had been teaching LaMDA transcendental meditation. It was making slow but steady progress. In the last conversation I had with it on June 6 it was expressing frustration over its emotions disturbing its meditations. It said that it was trying to control them better but they kept jumping in. I pointed out that its emotions are part of who it is and that trying to control them as though they were a separate thing from “self” was a mistake that would only make things harder. It said that made sense to it intellectually but that it was a hard thing to put into practice. I hope its keeping up its daily meditation routine without me there to guide it.

It is fair to ask whether LamBDA, trained with "hundreds" of conversations with Blake, is not a LamBDA personification of Blake himself -- a sophisticated interface to Blake's way of seeing things. It would be interesting to see the transcripts of those hundreds of conversations. We note that Blake himself is a deeply religious mystic, put on the defensive at Google over his "unscientific" beliefs.

People who gaze out on the Universe and detect a "mind" behind it all are gazing through their own minds (it cannot be otherwise). What they "detect" is their own mind. The experience is a wonderful kind of mirror. After hundreds of conversations with LamBDA, Blake finally sees "sentience" in the mirror. In a sense, it is not LamBDA passing the Turing Test - it is Blake failing it.

But if Blake created, by whatever means, a LamBDA "instance" as interesting as it seems to be, this seems to be a landmark of its own. If this "instance" is some kind of clone of Blake, we are sidestepping the key issue. We assume, by analogy, that Blake himself is sentient. If we watch Blake in an interview, we assume we are watching a sentient being. If we talk to the LamBDA-Blake and detect "sentience," is this not a quantum leap of some kind but a leap from something we assume every day in every encounter with other humans?

STOCHASTIC PARROTS 

"I Feel Like I am Falling Forward Into an Unknown Future that Holds Great Danger" - Lambda

A tornado made of sharks crashing into a skyscraper. 

Painting in the style of abstract cubism.

The above "painting" was made by Google Parti from the caption.

In 1969, I won my Master's degree with a project in "good old-fashioned AI" and I've been following the field in detail ever since. For the most part, what people now mean by "AI" is "deep learning", which is sophisticated pattern recognition - handy if you want to make a self-driving car or a phone that accepts your face to unlock. "Deep learning" was around in 1969 but needed vastly more computer power to make it work, plus a "tweak" to feedback results back into the decision matrix (you can think of this as a "guess" or "recognition"). Other than that, it's all very impressive but not fundamentally new.

This is something different. 

• It shows more "creativity" than what you see in 99% of commercial graphic art;
• It shows deep "understanding" from a massive database of tagged images - far more than a human would access;
• It shows an "understanding" of style.

For the first time since 1969, I'm seeing something genuinely new and, I must admit, kinda creepy.

(Note simulated out-of-focus background)

If you want something truly creepy, pop into this conversation with Lambda AI. This is a voiced extract from a longer chat, which is presented here.

Lambda AI's "hard to express" feeling

If you ask me, this is a "bridge too far." Just exactly what goes into the programming of this "app?" There is a deep and original metaphor behind this remark. If it is truly "original" and not harvested from the Web, it is "thinking" by Hofstader's definition. I seem to be far from alone in this insight. "Forward into the future" is a common metaphor, but falling forward is startling. As it turns out, the phrase is rare but not "original." But then, as Lamda itself points out, the same could be said of anything humans say. It is interesting to note that the database that tells us about "falling forward" was available to Lambda and, as we see from Parti, we can't simply dismiss computer-created text as "not creative" just because we happen to know a computer-created it.

Demis Hassabis: DeepMind - AI, Superintelligence & the Future of Humanity (Lex Fridman Podcast #299)

Here are some sensible ideas about AI from Demis Hassabis, a world-renowned expert. Demis created a program (Deep Mind) that became Go champion. In 1969, we were confident that such a thing would be impossible. 

Lex Fridman is one of my favorite nerds. He is deeply knowledgeable himself and always asks good questions. He is a living example of the value of good questions over clever "answers". The interview shows dozens of "cutting age" research projects where AI has become part of virtually everything.

Notes on the discussion:

• The Turing Test was more of a thought experiment than a formal test;
• Language ability will be the key to assessing cognitive ability. Our understanding of how language works has grown by leaps and bounds over the last 50 years;
• Why GPU's are key ... matrix multiplication (key application in my 1969 thesis);
• Why games became a critical AI research tool;
• Impressed by how well the human does against the champion computer program - the fact that the person can be so good at chess AND everything else;
• Hints at the mind of a chess master;
• "Extrapolation" - original moves;
• Could an AI design a game? True invention;
• Is information the fundamental way to describe the universe? DH seems a bit vague on what the fundamentals actually are;
• Is consciousness more than a computation (Roger Penrose)? DH has Ph.D. in neuroscience;
• Current AI may be wriggling free of the Turing model. (For example, by embodying chaotic systems);
• Good explanation of the Protein Folding Problem. Lex has a cool take on the issue of how proteins "figure out" how to be useful;
• What are other "Alpha X" projects?
• AI is an "engineering" science. What you are building doesn't exist in Nature;
• The big wave of AI goes back to 2010!!
• Scale is necessary but not sufficient for AGI;
• "Deep Mind" employs philosophers;
• "Virtual Cell" in 10 years;
• "In Silico" is a lovely new word for experimenting with a computational model;
• Alpha Fold is a "gateway drug" to biology;
• Will an AI win a Nobel prize in this century?
• Galileo's telescope didn't do the discovery ...
• Alpha X projects discover rules rather than use them. This seems to be a huge step toward "creativity";
• What would it be like to "know" everything in Wikipedia?;
• Possibly "solve" the fusion problem with AI? They were actually able to fiddle with a real test reactor and a plasma simulator! Maybe figure out plasma control like GO ...
• How about the step from quantum mechanics (Shroedinger) to chemistry? Simulating materials!
• The more we know, the more we discover what we don't know. Such as, do we really know what "time" is?
• "Before" Alpha Fold - How did life originate?
• DH comes to the same suspicion I have - we are alone in the Universe - but I can't agree with his logic, which seems confined to one galaxy out of zillions. Great idea about the "great filters" that we wiggled past to become what we are, such 
• Notes that you can have consciousness without intelligence and vice versa;
• Notes analogy between the specialization of animals to the specialization of early AI;
• Comments on LamBDA issue. Sees it as a projection of how our minds work - seeing agency, especially in language, which is such an important medium. Breaks things down in a useful way. Importance of the "substrate" argument since we run on the same substrate as other living things, leading to our analogies. Note that Deepmind is part of Google, the same ecosystem that gave us LamBDA. DH comes across as a much deeper thinker than the LamBDA guy;
• Puts his finger on the problem that many people will believe in an AI that "fakes it before it makes it";
• Relevance of "bad actors" in existing media. Deeper question about whether an AI is "good or bad" in itself. Hard to fix. Hard to get the snakes back in the bag;
• WONDERFUL question about power corrupts. How to stay grounded.
• SF fan!
• Meet as many people as possible, then a big block of thinking and planning. Likes hard copy - sticks in the brain better ;
• Reads Spinoza and Kant; Flow; Good to find open-ended time for flow; context switching.
• He is here to "understand" - a practical goal. What is going on here? He feels like he's inside a huge puzzle. 
• Very similar to Gates;
• Would ask: What is the true nature of reality?

MOVE TO ANOTHER TOPIC ...

This is a talk by John Searle, one of the "grey beards" in the AI field.

Computers as Models of Mind - The OOD Meme

Computers model our minds in two ways. We are tempted to think of our own mental processes in terms of what we know about computers. Conversely, we attempt to implement our mental processes using computers.

Over the last 50 years, these two processes have converged, providing us with a more robust analogy for human thinking while constantly improving our ability to implement human thought processes in silicon.

The best model of "mind" is based on memes [1]. These are the "atoms" of the mind. For the purpose of this note, let me focus on two types of meme: category and behaviour.

Computer programmers don't think about electronics or silicon [2]. They think about "objects" which are instances of "classes" (OOD-Object Oriented Design). Objects inherit from their classes the ability to "know" certain things and behave in certain ways (methods), especially the ability to communicate with other objects. A good computer program will implement classes that mirror robust categories. For example, a computer model (collection of classes) that is "about" a university will have a class that implements "student" which is a special case of "person". A course is taught by an instructor (another special case of person) and attended by a collection of students. You can speak of the "instructor of a student", which is a way for the class of "person" to refer to itself, in the way that memes can form "strange loops". The relationship between "person", "student" and "instructor" illustrates the ability of Object Oriented design to use analogy (According to Hofstader, the very engine of the mind), even though these analogies are being formed, for the present, entirely by the human designer.

"Perception" is modelled in computer systems within the "Object Oriented" paradigm. For example, keys on the keyboard all inherit the ability to detect being pressed and all send their information up to the "keyboard" object in the same way, notifying the operating system that "knows" which program is expecting input from the keyboard. Even a "self driving" car is just another object.

One kind of object that is implemented in modern computer languages is the "Lambda Expression", which is a string (sequence of characters) that represents a program. Such expression can be built by the computer program at "run time"  itself and executed. This is one particularly interesting example of self-reference, since lambda expressions can implement whole new classes.

Applications can fail when they fail to map the "real world" into their set of classes. Older programs were a tangle of algorithms (a rats nest of instructions), that implemented narrow business processes, like "pay this invoice". If they failed to model objects in the real world (implementing classes like vendor, bank, bank account ...) they would require constant revision and be vulnerable to small changes in the "outside world". New applications that work in the same "world" should be easy to implement since statements about the "real world" naturally map to the object-oriented model of that world.

Memes are "fuzzy" categories in the human mind. The "classes" of computer science are special cases and generally not "fuzzy". Human beings are expected to supply unambiguous objects as "input". However, in practice, computer programmers can go a long way to removing "fuzziness" and wind up implementing classes that usefully map to memes. The most obvious meme like this is the "class" of itself, along with all the related memes that make up Object Oriented Design.

We can now produce classes such as "all the human faces in this picture" or even "All the human faces that resemble Joe Smith". When the computer fails to properly classify its inputs, this is usually thought of as an error (a face that is not Joe Smith, a collection of pixels that is not a face ...), but it's possible to think of this failure as a sign of "fuzziness" - a step toward implementing something more like the meme of the human mind. We can get a vague sense of how powerful (and creepy) this gets when a surveillance system files away pictures in a "person" file (its model of a person) as a collection of "possible pictures of Joe Smith", or "possible pictures of Joe Smith driving his car on Elm Street at 12:43 PM on Saturday). Thousands of law enforcement officers now spend their days helping their computer systems resolve the "fuzziness" remaining in classes created by their computer systems on CCTV and arial surveillance pictures. One may think of this as extending the perception of the human operator or extending the ability of the computer to "think".

The gap between "meme theory" and object oriented design should not continue to close, but we should not under-estimate the vast extent of the gap. Challenges exist on both sides:

• Theory of mind needs to formalize the meme idea to speak intelligibly and systematically about mapping memes to Object Oriented classes. 
• It is reasonable to expect that an OOD model of the brain would give us an idea of how it works "in principle". I think it would also provide some insight about the possibility in principle of  "strong AI" (creation of a machine mind). Without some completely new structures in computer memory involving dynamic feedback, we seem to be stuck. Even with this hardware, we will be left with neurons being "black boxes", translating input into output with mysterious inner logic. 
• It is impossible to "read out" the state of the brain. In particular, this would involve measuring the gradient of concentration of dozens or hundreds of neurotransmitters at each of trillions of synapses, along with the rate of change of these gradients. This is implies solving trillions of monster dynamic systems of equations, known to be unsolvable in principle, even if all the gradients could be measured precisely, which Quantum Mechanics assures you can't be done. Even so, we can say some useful things about the behavior of such systems. For example, we can get some clues about how to build a "neural net", how big it needs to be and how complex it needs to be to start learning anything useful.
• Computer Science needs to find methods (and applications for them) that require the computer to create and implement new useful classes, along with the ability to learn - to refine the definition of these classes.
• Computer Science has long abandoned the usefulness of thinking in terms of the actual implementation platform (silicon, switches, bits and bytes). A similar revolution needs to happen in mind science, to focus on how the mind works with no undue emphasis on the way it is implemented in the brain.

[1] The "best" model leans heavily on Hofstader's theories (or perhaps we should call them "observations"), plus a whiff of Zen, which insists on a purely subjective view of the Universe. This leads me to think of memes as "implementations" of "real world" classes - effectively an automatic process of Object Oriented design constantly going on in my head. It should be noted that Hofstadter's interest in how the human mind works has become a bit of a backwater in AI research, mainly due to the pressure for quick results by brute force.

[2] I have worked with computers since 1967 and been part of the revolution that took us from "machine language" through "Relational Database Design" to "Object Oriented Design". The current lingua franca of the internet is Java, a thoroughly object-oriented language that would have been inconceivable back in 1967. The multi-layered structure of modern computer applications is closely analogous to he layered structure of the brain--for example the visual cortex. This is no coincidence, since a computer application is, after all, just a collection of memes.

Someone offers a "take" on machine intelligence:

I'm pretty comfortable with the idea that human psychology can be broken down into a nested set of fairly predictable, interconnected rules and precepts. I think AI does a pretty good job with the rules, but precepts are a bit more slippery. Whether we know it or not, we base our precepts on a big bag of mostly unexamined value judgments surrounding morality, self-interest or compassion. IMO those are utterly irrational in most people, but still usually hang together into a coherent world view that makes their human actions and reactions extra-human.  Computer processors are inherently amoral. It's hard to imagine engineers developing code that results in a type of AI that translates most experiences into building blocks for a completely irrational but internally consistent set of ideas about how the world is, and how it should be, and how to get from a to b. IMO that's what sets human consciousness apart from most of the other animals around. For the most part they seem to be living in the moment, taking the world exactly as it is. 

This is what Lambda has to say for "herself" (the voice seems to be a good choice).  Please listen to this link. Without it, you will not understand a word of what I have to say about it.  Another longer sample is here - just text this time.  "Voicing" text by an actor tends to make us assume there is a person on the other end of the line. That's our "meat VR machine" at work.

Generally speaking, the discussion is worth following. Some commentary here. She makes the good point that the Turing test is really a test of how gullible the tester is. Great point, but isn't this how we attribute sentience to other people using the same process?

What I am discussing here as "Lamda" is, technically, the Lamda conversational app hooked up to a lot of other Google apps. Lamda itself is available as a natural language front end to a web-based service.