Innovative Applications of Early Stage AI
Posted by Jeriaska on November 5th, 2007
Neil Jacobstein is chairman and CEO of Teknowledge Corporation, a 25-year old software company. Since 1992, he has served as chairman of the Institute for Molecular Manufacturing, a not-for-profit research group focused on the long-term feasibility, embedded safeguards, and applications of molecular nantotechnology. He was the leading co-author of the Foresight Guidelines for Responsible Nanotechnology Development and is a senior research fellow in the Digital Visions Program at Stanford University. At the 2007 Singularity Summit, he spoke to how early stage artificial intelligence has already produced a wide range of valuable knowledge systems applications in industry and government.
The following transcript of Neil Jacobstein’s 2007 Singularity Summit presentation “Innovative Applications of Early Stage AI” has not been approved by the author. An audio version of the talk is available at the Singularity Institute website.
“Eventually, technologies like AI do deliver more than expected, and that’s because technology acceleration, particularly of the double-exponential kind Kurzweil talks about, is not intuitive, including to practitioners in the field. We often underestimate how quickly a technology can arrive after 25 or 30 frustrating years of slugging it out in the trenches. ”
Innovative Applications of Early Stage AI
I don’t think we have an either/ or choice between AGI and narrow, specialized AI applications. We’ve had fifty-plus years of artificial intelligence as a field and we’ve had fifty years of impressive hardware acceleration, and we still don’t have AGI, not even a weak one like the HAL 9000 from 2001. In spite of this, I’m quite sure that the question is “When” and “How” we’ll have AGI, not “If.”
NASA has already built many excellent early AI systems that monitor spacecraft system health and control spacecraft operation. Alas, these systems don’t lip read, they don’t ask charming new questions, and they don’t have psychiatric problems. Early AI applications were largely built to solve specific industry or government problems. Many performed complex tasks that only humans, fairly capable humans, once did. And most are not attempts at artificial general intelligence. There are a few exceptions: Soar, Cyc, Novamente, Cog, others. But most of these systems demonstrate limited and narrow learning ability, if any, even when they incorporate machine learning modules in their operation. They are minimally self-reflective, but they are often goal-directed. They use goal-directed inference. Many of these systems are no longer narrow expert systems, as you’ll see. Many have complex components and hybrid designs, and they have now migrated to the web: Google search, American Express credit check for risk analysis, Babel Fish’s language translation.
There are lots of beliefs floating around about AI. Here are three of them. One is, if it works, it isn’t AI. Many applications work quite well using simple AI techniques. In addition, specialized AI applications are useful, effective, and, yes, limited. That’s okay. Another belief is that AI is all hype. It’s boom and bust from the 1980’s. Although there was a lot of that, early AI is in routine use in many domains. It’s already produced billions (with a ‘b’) worth of value. Writing off AI is a little bit like writing off e-commerce after the dot com meltdown. Other people believe that if we are not talking about human-level-and-greater general intelligence, then we’re really not talking about AI, and to that I would say, specialized AI applications routinely outperform humans on narrow tasks. That’s interesting.
There are many types of AI - a whole continuum of types of AI with different capabilities - just like I think there will be a continuum of different types of AGI’s with different capabilities. I believe there is no one true path to artificial general intelligence. A lot of this is about expectation management. New technologies tend to arrive late. The problems are hard, often harder than we thought. As Paul would say, clarity of vision is not proximity to goal. Sometimes it’s easy to confuse the two. Many technologies start out overhyped, AI certainly did, and our planning horizons are woefully short-term. Eventually, technologies like AI do deliver more than expected, and that’s because technology acceleration, particularly of the double-exponential kind Kurzweil talks about, is not intuitive, including to practitioners in the field. We often underestimate how quickly a technology can arrive after 25 or 30 frustrating years of slugging it out in the trenches. We tend to see future possibilities through today’s biases and filters. New technologies lead to transforming applications, not just faster and more of what we currently have. It took nature 600 million years to go from multicellular organisms to human-level AI. From that perspective, all the arguments about will it take five years or ten years, 10, 25, 50… are in the roundoff error.
The Association for Artificial Intelligence’s Innovative Applications of AI conference showcases deployed and emerging applications of AI. They’ve been doing this since 1989. 384 applications have been showcased during that period, and it’s primarily about hybrid knowledge systems. Let me tell you what they have in common. They all utilize some form of inferential reasoning and knowledge. They get their problem solving power via domain and task-specific knowledge. Unlike AGI’s, the knowledge is engineered by humans, rather than learned. That’s also true when machine-learning components are incorporated. Some of that knowledge is gathered by machines, but often a lot of the initial conditions are set up by people. These systems are now integrated with enterprise components. They are embedded in products and services. They cover a wide range of domains. They are globally distributed. They span a large number of tasks, a diverse set of users. They are implemented on client/server PC’s, PDA’s, and web services. They are just useful applications with no particular AGI agenda.
Here is a sampling of application sponsors. I think there were 154 by 2006. They include a large number of Global 1000 companies, as well as government agencies. There are quite a few countries that have submitted papers to IAAI, and this actually under-represents all the talented researchers from all over the world who have published under the auspices of U.S. institutions.
I did a study of the application domains, or the areas of focus of these applications. The ones between 1989 and 2006, there were 362 of them. The most common area was computers and software engineering (not surprising) followed by manufacturing, military, finance applications, business operations, telecommunications, arts and media, healthcare, space, ground transportation, airlines, education, sales, biotech, energy, emergency management, insurance, security, intelligence analysis, construction, law, agriculture, chemical engineering, paleontology, treaty verification, configuration and cost estimation. The next slide is on the tasks that these systems perform and they include this long list. Lots of different tasks, and you can see that they are clustered primarily around planning and scheduling, data interpretation, information retrieval, and classification and performance optimization, etc.
When we consider the value added by AI in these applications, they are augmenting or replacing human skills, which allows you to improve accuracy and consistency. They accelerate process timing. They improve products and service quality. That increases productivity. It allows you to decrease costs. It expands the range of the possible. It doesn’t just do more, faster, better. It also helps institutions manage their knowledge. Let’s look at just two of the very early expert systems. Here’s one that was built for British Petroleum that does micro-fossil identification associated with rich oil fields to drill in. This system was basically constructed to eliminate delays in the process of figuring out where to drill. One day of delay costs $1 million. British Petroleum reported that one particular delay cost them $15 million. The VIES system, visual identification expert system, acts as a domain expert assistant or member of the team and has virtually eliminated identification delays. It is implemented on LISP and KEE on a Sun workstation.
Here’s a pitch expert. Pitch is about 4% of cellulose. It’s a sticky, resinous substance that tends to gum up paper mills. This system was built for the Canadian pulp and paper industry. It’s a process control adviser. It includes a mill model and a diagnostic problem solving model. It recommends actions to reduce costs, largely costs of repairs and delays. It was implemented in a knowledge engineering tool called ART and also using LISP on a Sun workstation. It was implemented in 36 paper mills at the time of publications, saving an estimated $28 million.
In terms of the very early parts of the conference, a lot of the systems were delivered stand-alone on LISP and high-end workstations. As Rodney pointed out, LISP is still used today and it is very productive. A lot of these systems were built from the ground up and used very complex tools that subject matter experts really didn’t like. Even if an application has AI in it, it often has 85% system engineering code and other conventional components, and only 15% AI. A lot of the early systems were delivered with AI as the central focus. A lot of these systems are not properly integrated with mainstream application, they tend to be narrow and brittle expert systems. They generated considerable value added, however. Many of the systems were hard to maintain. They were delivered by researchers from the outside - kerplunk - and they were delivered into weak knowledge management cultures.
Here is a system that was not in the IAAI conference, but a really interesting system called the dynamic analysis and replanning tool. DART is a military logistics system that was built for DARPA by a team of contractors that included BBN and ISX and others. It was used in the ’90s during the first Gulf War. It focused on the logistics of material transfer between Europe and Saudi Arabia. It had three key components in it: an AI planner, a database forms generator, and a linear programming module. It was the poster child of DARPA’s 1990 planning initiative, and a comment from a former DARPA director, Vic Reis, was that the DART’s scheduling application paid back all of DARPA’s thirty years of investment in AI. It’s now been 40-plus years, and DARPA continues to be a very big investor in AI and in applications that have particular military significance. Rod mentioned some of them. You can also look at the DARPA Grand Challenge as another direction that the agency is going in.
Here is an example of a system that has been around for awhile. This system was published on its tenth anniversary in 2004. It’s a GE color plastics formulation tool. It’s a case-based system that matches color to GE’s vast inventory of polymers and dyes. It saved 22,000 hours annually. It’s in active use all over the world. They ported it to the web under the ColorXpress Select name. It saved millions of dollars in productivity in materials. It’s an example of an incremental innovation in a company that pays huge attention to process issues. Here is one from the 2006 conference, CombineNet ASAP. This system is focused on the multi-trillion dollar procurement industry, and it used heuristic constraints and sophisticated research to optimize the procurement process. Specifically, what it did was increase the expressiveness, the efficiency, and the granularity of the specification of supply and demand for very complex auctions. They hosted 230 procurement events, $16 billion worth of transactions, with documented cost savings of $1.8 billion. I think they’re getting somewhere. Here is a system that did on-demand web-based ASP product, a number of different ASP products, and they hosted on a 64-bit shared server farm. In terms of where we are today, AI is often delivered with other components as well as multiple types of AI components, and in addition to LISP there is a heavy use of mainstream languages like C++, Java, Ruby on Rails, Python, etc. And the use of Web 2.0 standards, deep integration with mainstream systems (instead of grafting them on from the side), browsers as the standard GUI, delivered on a wide variety of delivery platforms, now user-centric applications, and more attention to the target context for delivery, the process issues.
So, let’s talk about what didn’t work. A focus on AI tool features versus the underlying power of the knowledge. There was difficult knowledge-based reuse across time and institutions. That’s still a problem today. There were high maintenance costs, particularly if the knowledge changes rapidly and AGI could help solve that problem. Open loop R&D was not particularly productive. It was better when it was driven by real-world feedback. Poor integration with mainstream software, grafting it on from the side caused all kinds of problems. Rapidly evolving heterogeneous infrastructure and platforms causes everyone problems, including AI applications. Weak business processes in the target institutions, including rotation of champions, weak knowledge acquisition tools were a problem, and limited learning capabilities forcing a lot of manual input. Again, something that AGI could address.
So, we have talked about knowledge engineering and system engineering tools. Let’s talk about some of the cultural issues. I like to use the analogy of Frank Lloyd Wright’s falling water home. Here’s a brilliant designer working with highly skilled technicians, using exotic materials, doing one-off, very complex integrations. They delivered the system late, it was over-budget, and the system was buggy. Of course, the designer attributed any weaknesses in the system to the weaknesses in the user. That may sound familiar to you. Let’s contrast that with the continuous improvement methodology that Toyota uses to develop the Prius and the Lexus. They have documented incorporating over 20 million suggestions over just a 40-year period. It’s now a 50-year period, I don’t know what their new number is. But here is a culture that uses both AI and 3″ by 5″ cards to manage their knowledge, and they take it very seriously. It’s a very big deal, and they end up with products that reflect the kind of continuous improvement that they do.
In terms of what worked with the IAAI applications, real-world testing and experimentation, solving customers pain-point problems, capturing stable problem-solving knowledge because of the limitations in the technology, systems that perform one-or-more specific task. It could be more than one task, but then you have to have deep domain knowledge to go after the additional tasks. Combining Ai with other technologies, like statistics, optimization, simulation, using standard application infrastructure. Getting rapid, high-volume, signal-to-symbol interpretations, so you get very rapid transformations and users don’t have to wait. Attending to user, cultural and process issues. Finding champions and top management commitments, and improving systems continuously.
So, let’s look at the evolution of semantic web services. We already see that web data are becoming tagged, active and reusable. We have dynamic semantic service layers that are beginning to do interpretation and recognize patterns. They are going to be building context models of users, tasks, and domains. We will see web service assistants and associates proliferating. They will do association, complex modeling, and prediction. We will see the extensive use of HTM’s: hierarchical temporal memories, like of the kind that Numenta’s NuPIC program provides. We’ve heard about other cortical software here. And we will have immersive experiences in 3D semantic space. Not just zooming around in VR, but being able to interact with and query the objects in the space, talk to the Sphinx and learn about Egyptian history.
One of the things that we see now, Ben will be talking about this a little bit later, is that the ability to embed these systems is very important. Embed knowledge application systems in these VRs. So, ten rules for building AI applications. Balance long term vision and pragmatics. It’s still important, even if you are thinking about AGI’s. Distinguish technologies from market applications. Also important, because the pressure that researchers get to bring products to market very quickly, like in a three to five-year time frame, test ideas and code against real world problems. Leverage web economics standards and open systems. Utilize complementary technologies and hybrid applications. Leverage domain and task-specific knowledge. Assemble systems from robust components, instead of rolling them all individually. Develop applications that learn and improve. Do extensive user and consequence testing and develop AI roles where the AI gets increasing responsibility based on its behavior, going from an apprentice, to an associate, to a partner, and then, perhaps, more.
So, one possible scenario for the future, we might see a rapid increase in hybrid AI web agents. Intense economic and social selection pressure on the web. Agent cooperation and competition simultaneously. Applications that learn and improve continuously. A vast increase in machine-to-machine agents without humans in the loop. Maturation of neuroscience-inspired agents. Selection for useful new non-human tasks that are completely out of human-range. Embedded values, ethical principles, and layers of control. Agents as apprentices, associates and partners, as I’ve mentioned. And the emergence of different AGI’s with vastly different capabilities: radically new applications, risks, and opportunities. Thank you.
