Novamente aiming for human-level AI in six years

Just saw this blog post by Simon over at Betterhumans. I am sure some people around these parts already knew of the timeframe, but I thought it was quite neat, and deserved a thread.

Six years until human-level AI is the premise. Go Bruce! [thumb]

It's seems quite bold to me. The main issue is, that I think verifying and validating AI is still a neglected part of AI systems development. Maybe I'm a bit of an "old-school" kind of system developer, but testing applications is, even for current systems, a cumbersome process. Add to that the political sensitivity, that already currently blurs development of even the simplest systems, gives me the feeling that on the positive side, maybe technological capabilities may be present in this timeframe, but that commonly accepted AI applications could still be extremely difficult.

I tried to start discussions about the verification and validation issue, but my feeling is that (young) AI enthusiasts just are a bit focussed on the positive possibilities and being enthusiastic as they are, do not want to discus practical implications of AI.

Hi all,

About the "six years" number -- although that came out of my mouth (I'm the chief designer of the Novamente AGI system, and the CEO of Novamente LLC, of which Bruce is now the President), I don't think the exact timeframe is the right thing to focus on.

The key point is that we have a design for an AGI which, when fully implemented, tested and tuned and taught, will yield a human-level intelligence (capable of going beyond the human level as well if this is deemed safe at that point). I believe that if Novamente LLC achieves a reasonable amount of funding this year, then the six year timeframe is quite doable -- but I wouldn't bet the ranch on the time estimate, only on the general feasibility of the project.

Still, the general order of the time estimate is important. It's not 6 months and it's not 25 years either.

AGI is possible, and it's possible for us to make it happen SOON -- the main obstacle for us at this point is not ideas nor intelligent, trained programmer/scientists but rather funding to pay salaries and buy machines.

If any of y'all know of any investors interested in funding radically advanced technology projects, let me know! This is not something that will yield instant wealth, but it will change the future of intelligence on Earth --- and in the shorter term, after 2-3 years it should be possible to create powerful and profitable spin-off applications from the in-development technology, for instance in the area of natural language question-answering...

I am well aware that the history of AI is full of false optimism and false promises. But, so was the history of human flight before the Wright Brothers. We really can do this....

-- Ben Goertzel

Thanks, Ben. I am glad that smart people like you and others are working on it. It would be terrific if you guys could find some investors. You would think finding investors would be easier (at least in theory) than finding philanthropists

[thumb]

Peter Voss over at A2I2 is saying 2 and a half.

I wonder what the time frames are for the other groups working on this. I, for one, hope DARPA is way behind the others.

Sebastian Thrun predicted they would have self-parking cars that could obey all California traffic laws within 2 years or so.

Hi Hank,

Do you have the link handy to Voss' prediction?

I suggest you tell your prospective investors that one of the first applications you will focus your emerging AGI on will be the virtual corporate executive. If you could supplant many of the highly compensated CEOs of the world's companies, the financial benefits would be HUGE. Well, it's just an idea, but I would think about it.

If you could supplant many of the highly compensated CEOs of the world's companies, the financial benefits would be HUGE.

Not if a lot of those potential investors are the highly overrated and over compensated CEO's. )

For a more detailed discussion of the Novamente project, please see ImmInst topic, Viability of AGI for Life Extension & Singularity.

Hank, do you have a reference on the Thrun thing? At the SSS he was giving substantially more conservative estimates.

The probability of successful AGI within the next six years is quite low. I base this statement on the capabilities of current systems, and the time it will surely take to improve them to the point of anything approaching open-ended learning.

If you were to follow around an AGI project leader all day for a couple weeks, and watch exactly what they are doing, and see how much the software improves as a result of their programming, then you would agree that progress is slow and six years is slightly too optimistic. Ten years, or fifteen years, or twenty years is more in the ballpark of realisticness. Surprise nanocomputers in 2010 would vastly accelerate things, however.

Here's the link to A2i2's March 2006 "2 years" to human-level notice:

Work on our prototype is progressing well - we are on schedule to achieve
robust human-level learning and cognition within 2 years
http://lists.extropy...rch/000042.html

Novamente is one of the few projects I've tracked down so far that have provided some crunchy technical details that offer a path to what might actually be a workable system for representing general intelligence. It seems a lot of AI companies are holding their cards close to their chest. I wonder if AI companies share information amongst themselves? Is there a living community out there actually knee deep in real details of implementation and problem solving? Because it certainly seems like a lot of projects are blowing smoke. Novamente seems to have a real approach that might provide real success or at the least, insight into a more successful next-generation approach after Novamente is complete.

What generation would Novamente be considered in terms of AGI development (as opposed to the generation of the discipline of AI R&D, which is very old and focused on problems that aren't really related to general intelligence)? Second? Isn't Novamente based on a previous design by Dr. Goertzel?

If you were to follow around an AGI project leader all day for a couple weeks, and watch exactly what they are doing, and see how much the software improves as a result of their programming, then you would agree that progress is slow and six years is slightly too optimistic.

Notes from this kind of process are exactly the kind of thing I want to read. What are examples of real technical problems facing these teams: anything from debugging an unexpected behavior to establishing the primacy of subsystem implementation, etc. What lessons are being learned and what areas of design are being iterated on over an average week of development? Do the designers and implementors of these systems try to draw in lessons learned from commercial development, or is there a disdain held for commercial developers and their methods? One example is an emphasis on robust tools. I can't see how you would be able to work with a hypergraph based AGI system without extremely robust tools for the analysis of that state. Tools are critical to commercial development, because of the financial pressure toward efficiency. When you have a much longer development arc and no commercial pressure, it seems like it would be easy to dismiss the development of tools until later in the process -- for fear of having to maintain the tools, in the belief that there'll be plenty of time to make the tools better later, etc.

Part of the process of the evolution of the game industry as a whole is the dialog of developers sharing what may seem like mundane details, but from which are derived a lot of insight and progress.

Is this happening in the AGI developer community and if so, where?

How does the Novamente team structure milestones and internal deliverables? How are weekly objectives managed? What is the internal process of changelist review? I ask these questions because as a software developer I am interested in how a much more "high-minded" project approaches software development. In terms of the corporate values of Novamente, what is the mix of pragmatism with commitment to ideology?

I'm at the point now, in my interest level and understanding of singularity based issues, that I want more meat and less retread of the basics.

Also, I think if AI companies provide a more realistic picture of how they operate and how the science and philosophy of AGI is translated into solid code, the general software development community would start to take AGI more seriously.

I know that as I've read Dr. Goertzel's papers and began to intuit elements of the object model that might drive a system like Novamente, I began to see how concepts could be represented that I didn't necessarily take seriously before. This is a natural result of education, but I think that the density of information isn't yet high enough to make it easy. (As someone who is not able to study AGI 24/7.) Although now I am equipped with a massive list of ref's to material I need to find, largely thanks to papers and abstracts published by SIAI or Novamente.

The efforts of the SIAI and Novamente, over the past six months alone, have been significant in helping translate conceptual theorizing into semi-technical detail. The next step is real technical explanation: to elevate verbal understanding to technical understanding.

ala http://yudkowsky.net.../technical.html

I suppose as a Popperite, what I want to read is more info on the crunchy mechanics that are driving to prove or substantiate theory and less overview.

Part of the process of the evolution of the game industry as a whole is the dialog of developers sharing what may seem like mundane details, but from which are derived a lot of insight and progress.

Is this happening in the AGI developer community and if so, where?

http://www.agiri.org/email

"The [singularity] list is for more philosophical and general discussions, while the [agi] list is for more technical discussions about current AGI projects."

I'm on SL4, EXI, and AGI, but not on this new list. I'll sign up and listen in.

I don't expect AGI before 2015-2020 but if novamente feels their progress is substantial enough to shout it from the rooftops, I'd be REAL careful who you shout it to. [mellow]

this stuff isn't a game as you all know, and there are more than a few nations/people/organizations ect, who would give ANYTHING to get their hands on it.

Best idea is to keep it extremely quiet until your protection of the client is far superior to fort knox, because in the end, the value of AGI is basically priceless; unlike a building full of shiny gold bars.

oy... I think the 2020s are going to be completely insane. Everything seems to converge there...

although I wouldn't be surprised if we saw AGI in the teens.

Hank, do you have a reference on the Thrun thing?

That was from my memory of the Singularity Summit at Stanford. Did I get that wrong? [:o]

How does the Novamente team structure milestones and internal deliverables? How are weekly objectives managed? What is the internal process of changelist review?

Brandon, I've sent you a PM with more information on this.

In addition to the technical documents, we've recently added a "Brief overview" of the NAIE here: http://www.novamente.net/engine

Best idea is to keep it extremely quiet until your protection of the client is far superior to fort knox, because in the end, the value of AGI is basically priceless; unlike a building full of shiny gold bars.

AGI is priceless and Novamente is keeping private the details of our design. However, there are benefits to be had from collaboration, as through our AGI-SIM project for example.

Hey Bruce, what about the webforum for the link you provided (AGI-sim). What do you think about it aesthethically?

Don't you think that these forums should be moving in that direction? It looks more professional

What generation would Novamente be considered in terms of AGI development (as opposed to the generation of the discipline of AI R&D, which is very old and focused on problems that aren't really related to general intelligence)? Second? Isn't Novamente based on a previous design by Dr. Goertzel?

Dr. Goertzel was instrumental in starting a similar company (WebMind) in the late 90's... a review of this project (and its demise) can be read via Waking Up from the Economy of Dreams.