Can Computers Really Talk? Or Are They Faking It?

In the mid-1960s a computer scientist named  Joseph Weisenbaum created one of the first natural   language processing programs, called ELIZA.

The user would type conversational sentences   and get responses in return--a system that today  we might call a "chatbot," like the ones you find   on customer support websites.

ELIZA used simple  pattern matching and substitution to give the   illusion of comprehension.

She was designed with  the personality of a psychotherapist, so that she   could reply with canned responses, like "Why do  you say that?"

or "Is that important to you?"

Weisenbaum actually created ELIZA to demonstrate  how shallow computer language skills were,   so imagine his surprise that many users actually  believed she understood what they were saying.

I guess it's in our nature to anthropomorphize  things.

After all, I've been calling a computer   program "she".

ELIZA was one of the first programs  that could participate in the Turing Test,   a theoretical experiment proposed by computer  scientist Alan Turing about 15 years earlier   as an assessment of artificial intelligence.

An  evaluator would eavesdrop on a typed conversation   between two participants: a human and a computer.

If the evaluator was unable to determine which was   which, the computer was said to have passed the  test.

Half a century later, computer technology   has evolved exponentially.

We all carry around  a natural language processor in our pockets   vastly more advanced than ELIZA.

We ask it to  give us directions, keep track of our schedules,   even turn our lights on and off.

But is any of  this technology close to passing the Turing Test?

Even if it is, are computers really talking  to us?

Or are they still just faking it?

Not long after ELIZA, another computer scientist  named Terry Winograd developed his own natural   language processor called SHRDLU.

Unlike ELIZA,  SHRDLU was supposed to actually understand   what it was talking about--so to speak.

Its  comprehension was limited to a handful of simple   objects and actions, but users could instruct it  to rearrange the objects around a virtual room,   and then ask questions to verify that  it knew what it had accomplished.

SHRDLU was one of the first successful  attempts to teach a computer grammar,   to put words into categories like nouns, verbs  and prepositions.

To have rules about how they   could combine and assign meanings to various  combinations.

The technology was popularized in   the text adventure video games of the 1980s, which  allowed players to interact with the game world   by typing simple verb-noun commands like "GET  SWORD," "GO EAST," or "OPEN DOOR."

Half the fun   of these games was trying different combinations  to see just how... extensive the vocabulary was But this technology did not advance as quickly as  some hoped, because programmers soon realized that   they had to manually code all the rules of English  grammar, which are a lot more complicated than   "open door."

Even a simple rule like "add -ed to  make a verb past tense" has enough exceptions to   give a human ESL student a headache.

Even a simple  sentence like "It's raining," would confound   a computer because it couldn't proceed without  knowing what "it" is.

Grammar is also dependent on   the listener having at least some understanding of  how the world works.

Consider these two sentences:   Take the cap off the milk and pour it.

Take the sheet off the bed and fold it.

These seem like pretty simple instructions, but  that's only because we already know that milk   pours and sheets fold.

A computer without that  prior knowledge would have no way of knowing   what the "it's" refer to.

And how about these two  sentences: Sarah drew the dog with a pencil.

Sarah   drew the dog with a bone.

Again, neither of these  sentences would give a human much trouble.

But a   computer couldn't be sure of the meanings unless  it already knew that dogs don't carry pencils,   and that you can't draw with bones.

So just  to get a computer to understand basic grammar,   you'd have to manually encode it with vast complex  rules of syntax and a broad understanding of how   thousands of different objects interact with each  other.

It's no wonder that in the 1990s computer   scientists kind of gave up on so-called symbolic  language processing in favor of a new strategy:   statistical language processing.

Instead of  trying to teach a computer the rules of language,   they developed algorithms that could analyze large  bodies of text, look for patterns, and then make   guesses based on statistical probabilities.

For  instance, when you ask Siri "What's the weather   looking like today?"

she doesn't bother parsing  your grammar.

She simply homes in on keywords and   guesses what you're looking for based on how  common the request is.

She'd probably give you   the same answer if you said "whether today" or  even just "weather."

Another application of this   technology is predictive text, which is what  your phone does when it tries to guess which   word you're going to type next.

For example,  if you type the word "good" the algorithm knows   from looking at thousands of pages of text that  the most likely word to follow is "morning."

It   doesn't know why or what those words mean--just  that it's a statistical probability.

Believe it   or not, there was a time when some thought this  was how human speech worked; That our brains   picked words one by one in order, deciding what  word was most likely to follow the one before.

20th century linguists like Noam Chomsky have  shown that human grammar is far too complex   to be constructed this way.

Take this sentence,  for instance: The fact that you went for a walk   this late at night without an umbrella even  after hearing the weather report on the radio   this afternoon.

You can probably tell that there's  something missing.

That's because the opening few   words obligate the speaker to finish the sentence  with a verb phrase, like "is unbelievable."

Your   brain subconsciously remembers this commitment,  even though there are 23 words in between.

Language is full of such grammatical promises,  like either-ors or if-thens.

A computer program   that only considers one word at a time would  never be able to fulfill them.

However,   recent advancements in digital neural  networks are raising expectations of   what predictive text can achieve.

In 2020,  the artificial intelligence company OpenAI   released a beta version of one of the most  sophisticated natural language processors   ever created, called GPT-3.

To find out how  GPT-3's extraordinary language capabilities   are achieved, I spoke with OpenAI technical  director Dr. Ashley Pilipiszyn.

How is GPT-3   different from previous NLPs?

Yeah, so unlike most  AI systems which are designed for one use case,   GPT-3 and our API provides a general purpose text  in text out interface, and it allows users to try   it on virtually any English language task.

Say  I have a piece of a legal document maybe an NDA   and I would ask GPT-3 to summarize this legal  document like a second grader and GPT-3 would then   be able to provide a couple sentences actually  compressing and making that legal document   into a much more understandable piece of  text.

So our model actually doesn't have a   goal or objective other than predicting the next  word.

Like most predictive text programs, GPT-3   is trained by feeding it a large body of text for  analysis, known as a corpus.

And GPT-3's corpus   is enormous.

Somewhere around 2 billion pages of  text, taken from wikipedia, digital books, and   a vast swath of the web.

It analyzes this text,  using hundreds of billions of parameters looking   for probabilities.

And because it does much of  it unsupervised, even its programmers don't know   exactly what patterns it's finding in our human  speech.

When GPT-3 is asked to complete a prompt,   it uses what it's learned to guess what should  come next.

But where your phone guesses words,   GPT-3 guesses "tokens": four character  blocks of text including spaces and symbols.

And can you tell us a little bit more about these  particular tokens?

So you and I as humans, when we   see a sentence we see a specific set of words with  spaces in between, etc.

When GPT-3 quote-unquote   "sees," they actually are seeing tokens, which you  can think of actually like a jigsaw puzzle--that   allows GPT-3 to process more text.

It's really  trying to predict what the next token is   going to be in a sentence, based on all of the  previous text it's seen before in that prompt.

Because of its longer memory it's able to complete  grammatical commitments like "the fact that..." And thanks to its huge corpus,  it actually does seem to know   that dogs are more interested  in bones than pencils.

It can even apply grammatical rules to unfamiliar  words.

A famous experiment from the 1950s asked   children to complete the following prompt.

Even  though they had never seen the word "wug" before,   the vast majority of children were able  to correctly apply an "-s" to make it   plural.

Similarly, even though GPT-3 has  probably never seen the word "ferfumfer"   in all its billions of pages of corpus, it still  knows to add an "-s" if you have two of them.

So   can GPT-3 pass the Turing Test?

Uh, most people  familiar with the traditional Turing Test would   probably say it comes very close.

It can actually  start to feel like you really are interacting,   you know, with that person.

But the longer  you talk with it and really begin to push it,   you do come across some mistakes.

And so that does  kind of indicate that okay yeah this isn't human.

So ultimately, it's it's coming close but it's  not quite there.

Despite GPT-3's often impressive   performance, there is a fundamental difference  between human speech and what GPT-3 does.

Humans   don't learn language by memorizing likely orders  of words, but instead word categories.

Chomsky   demonstrated this with a famously nonsensical  sentence: "Colorless green ideas sleep furiously."

Even though all your life you've probably never  heard any of these words follow the one before it,   you still know that the sentence is grammatically  correct, because the order of the word categories   is correct.

But GPT-3 does not make a  distinction between form and content.

It doesn't care that "colorless green ideas"  is a grammatically correct noun phrase.

Only that the likelihood of those tokens  going together is very, very small.

That's   why a slightly more advanced question from  the wug test can get... interesting results.

Our brains seem to be hardwired for grammar,  which ironically is closer to how the old SHRDLU   program worked.

We have thoughts about the world  that exist prior to and independent of language,   and we use grammar as the container to deliver  those thoughts to others.

For all its complexity,   GPT-3 is still guessing one token at a  time, based on statistical probabilities.

It has no plan or goal, unless given one  by a human.

That's why the developers   at OpenAI prefer to think of GPT-3 as a  writer's tool rather than a writer itself.

It's pretty astounding at faking  what human speech sounds like,