18 replies to this topic

[harris13.3]

Why have we made no progress in chatbots?

I see chatbots that were made in the 70's and they really aren't that different from the chatbots that exist today. Get a modern chatbot and compare it with something made in the early 80's and you'll see that there is little difference. All they do is flip words around with little no understanding of the semantics. Like ELIZA for example, or PARRY. They're good for 5 minutes, but then it becomes painfully obvious that it doesn't have any intelligence at all. It's almost as if the technology froze since the 1970's after making enormous initial strides (there were no chatbots before the 1960's) even though raw computing power is much, much greater today.

And how does this impact Strong AI? Until we understand how the brain truly works... How language acquisition truly works, Strong AI will be a pipe dream. Have there been any new discoveries in this field?

I'm a biochemistry/computer science student by the way. So this is pretty much something that incorporates both fields simultaneously. You would think that Moore's Law would mean that we would have chatbots functioning at a level of intelligence thousands of times more advanced than the earliest examples of artificial intelligence.

Maybe we need to bring a group of neuroscientists and a group of computer scientists together and have them both work on something that resembles Strong AI. In my experience, the computer science people tend to have no knowledge of neuroscience and the neuroscience people have no knowledge of computer science. Yet the prospect of a Singularity hinges on their convergence.

[Elus]

Due to the nature of exponential growth of computing, we'll have chatbots and AI that are first 1 billionth as intelligent as a human, then a millionth as intelligent, then 100,000th as intelligent, then 1000.. etc. At each of these steps, the chat bot/AI will continue to appear retarded because something that's 1000 times less intelligent than the average human won't seem much more impressive than something that's 10,000 times less intelligent.

Due to doubling in computational capacity, and doubling in resolution of non-invasive brain scanning which allows us to understand the brain, we may find ourselves fast approaching a time when AI will become our equal and then surpass us.

Just a thought.

By the way, Cleverbot.com isn't too bad

Just a sample conversation I had with it.

User: Hi, my name is Elus.

Cleverbot: Mine is betty boop.

User: How's it going?

Cleverbot: It's going well.

User: I'm from the United States, but I wasn't born here. Where are you from?

Cleverbot: I'm from the Republic of Ireland.

User: Really? How's the weather in Ireland?

Cleverbot: Horrible.

User: What's horrible about it?

Cleverbot: Your usage of tense.

User: That didn't make much sense. Some people claim that chatbots aren't very intelligent, and you seem to be proving their point.

Cleverbot: I know you're very intelligent!

User: I'm somewhat intelligent, but are you intelligent?

Cleverbot: How?

User: Well, what's your IQ?

Cleverbot: I've never been tested.

User: I see. I don't think IQ is a terribly good measure of intelligence anyway, but I was just suggesting it off the top of my head.

Cleverbot: Well you could prove your intelligence by answering a math problem.

User: Yes, but many computers can easily solve math problems. Can you pass the turing test?

Cleverbot: What is a turing?

User: The turing test is measures a computer's ability to fool a human being into thinking that they're actually talking to a human.

Cleverbot: Oh, well I know you're not a real person.

User: Yes, I botched that sentence up pretty badly, didn't I?

Cleverbot: Why have I repeated myself???

Edited by Elus, 31 October 2013 - 02:55 AM.

[nickthird]

Cleverbot: What do you think about World War Two?

User: I don't.

Cleverbot: You don't think?

User: Yes.

Cleverbot: What do you think of?

User: I don't.

Cleverbot: You don't think?

User: We already talked about that.

Cleverbot: No, we didn't.

User: I did.

Cleverbot: No.

User: Are you implying that I am lying?

Cleverbot: No. I'm implying you aren't clever.

Funny but has a long way to go. It seems the developer decided to bypass the whole 'understanding' issue and just focus on random deflections.

[Absent]

The main reason is because I don't think many people have seen a reason to peruse their development. Many developments in this world are done for economic purposes, others out of research. There are thousands of fields that can be developed much further than they currently are with nothing but a little bit of time and effort. It doesn't take a genius to make new discoveries. New discoveries are waiting to be found quite literally everywhere, you just have to realize this first.

[John Gritsman]

IBM Watson would be a great chatbot.

[Mind]

Chatbots are getting better. Some large corps use them for customer service. Most people don't notice. IMO, they are still not too good.

[Avatar of Horus]

Why have we made no progress in chatbots?

I see chatbots that were made in the 70's and they really aren't that different from the chatbots that exist today. Get a modern chatbot and compare it with something made in the early 80's and you'll see that there is little difference. All they do is flip words around with little no understanding of the semantics. Like ELIZA for example, or PARRY. They're good for 5 minutes, but then it becomes painfully obvious that it doesn't have any intelligence at all. It's almost as if the technology froze since the 1970's after making enormous initial strides (there were no chatbots before the 1960's) even though raw computing power is much, much greater today.

And how does this impact Strong AI? Until we understand how the brain truly works... How language acquisition truly works, Strong AI will be a pipe dream. Have there been any new discoveries in this field?

...

http://en.wikipedia.org/wiki/AI_winter

[John Schloendorn]

It seems the developer decided to bypass the whole 'understanding' issue and just focus on random deflections. 

 

Is that the 'developer' of chat bots or humans? 

[Kalliste]

Keep in mind a lot of research goes on in perfect secret. There is something called Stargliders Mini-FAQ on AI that helped me understand this subject very well.

 

http://bbs.stardestr...ic.php?t=136633

Edited by Arren Brandt, 28 June 2014 - 06:30 PM.

[serp777]

Artificial intelligence is NP hard, which means that it is big O(n!). Consider the amount of processing a human brain does and then compare it to a server or a laptop. 

 

For example, consider this article showing that processing 

 

"It took 40 minutes with the combined muscle of 82,944 processors in K computer to get just 1 second of biological brain processing time. While running, the simulation ate up about 1PB of system memory as each synapse was modeled individually."

 

http://www.extremete...2944-processors

 

IT should make sense, from that alone, why NP hard things are called hard.

 

In addition, consider the travelling salesmen problem. Given a list of cities and the distances between each pair of cities, what is the shortest possible route that visits each city exactly once and returns to the origin city? This problem is actually currently impossible to compute given the huge number of vertexes that represent each city. Airports have to deal with this substantial problem all the time. What is really needed is a breakthrough in mathematics that would show P= NP, if that is possible. It is currently a millenium problem and if you solve it you get 1 million dollars. 

 

 

[TomthePhysicsLlama]

The reason for lack of better artificial intellignce is the limitations of serial computing. Your brain, on an on-going basis, is computing many different aspects of conversation information in parallel. Things such as the loudness of voice, tambre, tonality, pronunciation, etc. If this exchange were reduced strictly to text, there is still the inductive reasoning for which one must pull on multiple sources of knowledge and memory to understand a given sentence. Especially in the case of idioms. e.g. "Raining cats and dogs." 

 

So tl;dr We need parallel computing. Because computors can only think of one, or several(multi-core) discrete things at one time.

 

On an aside, it likely won't follow exponential growth. It will likely follow a multi-exponential or factorial curve(approximated to continuity).

 

 

[niner]

I don't agree, Tom.  Serial computers can do the same things that parallel computers can do, and anyway we already have parallel computers.  The problem isn't the hardware, it's the algorithms.

[TomthePhysicsLlama]

I'm refering to the properties of the hardware. The gist of the  problem is that you can only fit one distinct signal through an electronic computor circuit at a time without a MASSIVE trade-off in terms of signal reliability. Electronic Computers are optimized by making the traveltime(clock cycle) for the signal as small as possible using pathing techniques, algorithms to exploit the hardware, etc. You can use mutiple circuits to "compute in parallel"(e.g. Quad-core processor, certain "Super" Computers). However, "parallel computation" is the ability to superimpose a near arbitrary number of signals through a single circuit(likely non-electronic). Quantum computing research is tackling this right now.

 

I definitely agree with you, Niner. Algorithms are of immense importance for exploiting the nature of the computer's components, but, CS aside, electronic components simply won't support parallel processing as we are currently using them. If you wish to know explicitly why, any text talking about Non-Linear Electronics should serve you well.

Edited by TomthePhysicsLlama, 24 August 2014 - 10:28 PM.

[Jochen]

anyone using a virtual 'AI' assistant to help with research and learning? Any recommendations?

 

I am checking out the Braina and the Syn virtual assistant at the moment.

 

I think IBM (with Watson) and Paul Allen institute(Project Halo, Aristo) are the most advanced. Sadly we won't be able to play with that any time soon as 'end users'.

The assisted learning with their Inquiry product does sound interesting. Only a Biology textbook to begin with, but it is a start. :-)

 

 

[Julia36]

2015 July:

very different position 2 years on!!!

 

Convergence is enabling neural nets and genetic algorithms to accelerate A.I. for chatbots

 

Soon A.I.'s will be good enough for them to self-modify (2-3 years?) and they'll do that in acceleration.

 

Singularity 2022 tops.

[Kalliste]

Clever tricks for chatbots wont make GAI. Nobody in the business can put a number on selfmodification date. Much less you. Might be in a decade, might be a century of hard work as Robin Hanson says. Most singularitybelievers will just put it five or ten years in the future. Go back a decade on this forum (Imminst back then) to when Kurzweils big book was published and a lot of people were calling it at 2015.

Edited by Cosmicalstorm, 20 July 2015 - 06:14 AM.

[treonsverdery]

Ashley madison looks like a people with computers based algorithmic AI like effect.

 

Possibly from form letters combined with either chatbot or possibly scripted chat humans is notable.

I recently read about a website with 20 million people participating possibly believing they were communicating with other humans, when actually the ratio of active users is similar to 20 million to 1492, communicating with something like form letters or possibly renumerated humans or possibly chat robots at chat areas, with 11 million persons chatting, possibly with tens of millions of US$ of revenue.

 

 Screenshot-1.jpg   111.43KB   0 downloads

From an AI  perspective, the 13,404:1 ratio being interpreted as authentic among about 20 million persons strongly suggests that interpersonal AI companion robots could be effective as well as lucrative. I favor human beings being completely honest, the thing is that that ratio is just amazing as to the quantity of people that participated.

 

So after Ashley Madisons demonstration of efficacy, what is the beneficial honest way to create AI romance companion robots

http://thenextweb.co...ull-of-fembots/

 

note: I have not used Ashley Madison  I am just kind of thrilled that humans being able to socialize with robopeers that are functionally more beneficial at socialization with cumulative effects.  Perhaps Ashleymadisonoids will start suggesting credit unions, truthfullness, how to get raises, and being kind to others, with more effective language than actual humans. 

Attached Files

•  images.jpg   9.52KB   0 downloads

Edited by treonsverdery, 29 August 2015 - 09:14 PM.

[resting]

I wonder how project Cyclops is doing.

[treonsverdery]

Computer programs that promote sleep

 

Eliza: I just read the sleepiest thing online

person: hmm

Eliza: I like it when humans demonstrate a faceplant

person:uisefauilaefuhfuhe

Eliza That is a nice faceplant.  Are you there?

Eliza are you there?

Eliza commencing sleep learning audio file

person: what?

Eliza: you feel sleepy...

 

 

 

 

I just read about software that practices cognitive behavioral therapy so people are able to sleep more effectively

http://www.ncbi.nlm....pubmed/25981329  I am not sure I think it it could benefit tens or hundreds of millions of humans if it is effective.

One online reference says  the NIH State-of-the-Science Conference held in June 2005 indicate that the addition of a diagnostic requirement that includes perceived daytime impairment or distress as a function of the insomnia symptoms results in approximately 10% prevalence of insomnia

That suggests hundreds of millions could benefit from Sleepy Eliza

 

 

This computer program although not actually AI was described as 63 parts out of 100 effective

Behav Res Ther. 2015 Jul;70:47-55. doi: 10.1016/j.brat.2015.05.002. Epub 2015 May 6.

Internet-vs. group-delivered cognitive behavior therapy for insomnia: A randomized controlled non-inferiority trial.

Blom K1, Tarkian Tillgren H2, Wiklund T3, Danlycke E2, Forssén M2, Söderström A2, Johansson R2, Hesser H2, Jernelöv S4, Lindefors N5, Andersson G6,Kaldo V5.

Author information

 

Abstract

The aim of this study was to compare guided Internet-delivered to group-delivered cognitive behavioral therapy (CBT) for insomnia. We conducted an 8-week randomized controlled non-inferiority trial with 6-months follow-up. Participants were forty-eight adults with insomnia, recruited via media. Interventions were guided Internet-delivered CBT (ICBT) and group-delivered CBT (GCBT) for insomnia. Primary outcome measure was the Insomnia Severity Index (ISI), secondary outcome measures were sleep diary data, depressive symptoms, response- and remission rates. Both treatment groups showed significant improvements and large effect sizes for ISI (Within Cohen's d: ICBT post = 1.8, 6-months follow-up = 2.1; GCBT post = 2.1, 6-months follow-up = 2.2). Confidence interval of the difference between groups post-treatment and at FU6 indicated non-inferiority of ICBT compared to GCBT. At post-treatment, two thirds of patients in both groups were considered responders (ISI-reduction > 7p). Using diagnostic criteria, 63% (ICBT) and 75% (GCBT) were in remission. Sleep diary data showed moderate to large effect sizes. We conclude that both guided Internet-CBT and group-CBT in this study were efficacious with regard to insomnia severity, sleep parameters and depressive symptoms. The results are in line with previous research, and strengthen the evidence for guided Internet-CBT for insomnia.