EFTA01103465.pdf

Jeffrey Epstein IV Foundation AGI Initiative R&D Funding Proposal Ben Goertzel December 2, 2013 1 EFTA01103465  2 Executive Summary This proposal outlines a research initiative aimed at creating human-level, generally-intelligent thinking ma- chines within an 8 year period. During the last few decades, the Al field has wandered from its initial focus on human-like general intelligence, and has devoted nearly all its attention to the creation of highly specialized, task-specific AI. While this focus on "narrow Al" has borne impressive fruit, the time has come to redirect attention toward the field's original goals. Given recent advances in computing hardware and algorithms, cognitive science and neuroscience, the goal of building powerful Artificial General Intelligence (AGI) is far more achievable now than in the 1950s when the AI field was founded. In the proposed research, we will create a general-purpose cognitive engine, and demonstrate and test it on tasks from a number of domains: • Intelligently controlling animated characters in a 3D video game style world • Controlling humanoid and wheeled robots in an indoor physical environment • Engaging in natural language dialogue utilizing information on the Web • Analyzing genomics datasets relating to the longevity of various organisms • Automated program learning and theorem proving From a narrow AI perspective, these endeavors might seem to have little in common — but from an AGI perspective they are fairly similar, as they all all require general intellectual abilities such as those found in the human brain. Of course, like all revolutionary scientific advances, once advanced AGI has been achieved it will have a wealth of practical applications. The five areas mentioned above are only examples, which we have chosen largely due to our own prior experience applying Al in those domains. Creating something as complex as a human-level mind requires a comprehensive, coherent design based on sound science and engineering principles. This is something that did not exist in the early decades of the Al field, when scientists greatly underestimated the problem. But we have learned that lesson, and we have completed a full design, which we call the CogPrime AGI Design. CogPrime is a high level design for a human-level AGI system. It leaves many medium and lower level algo- rithmic and implementation problems open; but it gives a clear, coherent conceptual and software framework for AGI engineering, detailed design and experimentation. For a detailed description and an explanation of why we believe the CogPrime AGI Design has the capability to create true general intelligence, we refer the reader to the overview available at http : / / wik i . opencog . ors./ w /CogP ime_Overview. For a more complete and fully technical description, we refer to the books Engineering General Intelligence Vol. 1and 2, by Ben Goertzel, Cassio Pennachin and Nil Geisweiller, to be released in December 2013. We project that this project can be completed in an eight year period, by a team of 80 scientists and engineers, with a budget of USS80 million. This cost could be reduced by a factor of 3.4 via offshoring some of the R&D. In the first five years of the project, we will develop a thinking machine with the general intelligence of a 3 to 5 year old child, and highly powerful, practically useful capabilities in the areas of natural language dialogue, game Al, toy robotics, computer algorithm design, theorem-proving and genomic data analysis. In the following three years, we will give the thinking machine the ability to modify its own Al algorithms and to act independently as a research scientist. It will also be able to control robots moving commonsensically through everyday human environments, including attending and passing college classes. EFTA01103466  3 Contents 1 Introduction 4 1.1 Why Now? 4 1.2 The CogPrime Design 5 1.3 Application Foci 6 1.4 Modular Design & Development 7 1.5 Open Source Development 8 1.6 Potential for Commercial Spin-offs 8 2 Incremental Development Milestones 9 2.1 Phase 1 9 2.2 Phase 2 14 3 Staffing and Costs 18 3.1 Advisors and Technical Leads 18 3.1.1 Technical Leads 18 3.1.2 Advisors 20 3.2 Estimated Costs 20 3.2.1 Lower-Cost Alternatives 21 4 Scientific and Technical Development 21 4.1 Integrated Cognition 21 4.1.1 Unified Rule Engine 22 4.1.2 Probabilistic Reasoning 22 4.1.3 Motivation and Emotion 22 4.1.4 Procedure Learning 22 4.1.5 Procedure Execution 23 4.1.6 Pattern Mining 23 4.1.7 Planning 23 4.1.8 Language Processing 23 4.1.9 Attention Allocation 23 4.1.10 Concept Formation 24 4.1.11 Perception 24 4.1.12 Action 24 4.2 Distributed and Multicore Processing Infrastructure 24 4.3 Sensation (Vision, Audition & Haptics) 25 4.4 Robot Movement Control 25 4.5 Game World Development 26 4.6 Genomics Data Analysis 26 4.7 Text, Image and Video Mining 27 4.8 Automated Programming 27 4.9 Automated Theorem Proving 28 4.10 Teaching and Intelligence Testing 29 4.11 Software Integration and Testing 29 EFTA01103467  4 1 Introduction A best-selling book a few years ago claimed that "Al11 Really Need to Know I Learned in Kindergarten.- This maxim has some truth in an AI context. Arguably, what current AI programs are lacking is the kind of commonsense that every normal human toddler has. If one could create an AI with toddler-like common sense, alongside the specialized calculating and problem-solving ability that today's intelligent software already displays, one would be well on the way to creating software systems with general intelligence at the adult human level and beyond. This is precisely the thrust of the present proposal. Leveraging a new Artificial General Intelligence (AGI) design called CogPrime and an allied open source software system, OpenCog, we plan to begin by creating an AI software system that I. displays the commonsense knowledge and everyday creativity of a young human child, in the context of controlling an animated character in a game world, and controlling a humanoid robot in an indoor environment 2. displays impressive practical prowess at a variety of intellectual tasks: algorithm design, geometric theorem- proving, and genomic data analysis This combination, in itself, will not constitute the "end game" of our AGI work. However, we believe this "Phase 1" achievement will encapsulate solutions to the hardest problems in AGI design and engineering, and leave us poised to take the next step — toward a "Phase 2" AGI system that possesses the general intelligence of a human adult, enhanced by nonhuman calculational and problem-solving capabilities that digital computers bring. The Phase 2 system will be a genuine "AGI Scientist", which can self-modify and improve its own intelligence, alongside other science and engineering capabilities We propose that a "Phase 1" early-stage AGI system, displaying the dual capabilities described above as aspects of its unified cognitive functionality, can be produced within 5 years at a rough cost of US$9 million per year. Further, we propose that once Phase I has been achieved, Phase 2 can be accomplished within the same design and software system, and will actually be a smaller leap: 3 further years of effort at the same rate. 1.1 Why Now? Human-level general intelligence was, of course, the original focus of the founders of the AI field in the 1950s and 60s. But the hardware, software and conceptual frameworks of that time were not adequate to the task, and in subsequent decades the Al field shifted focus to narrower problems. Today, however, our hardware, software and understanding have advanced considerably, so that human-level Artificial General Intelligence (AGI) is an initiative whose time has finally come. Today, task-specific narrow Al currently pervades nearly every area of industry, within various forms of back- end software. It is also achieving an increasing public profile with achievements like self-driving cars, IBM Watson, online recommendation systems and chatbots like Siri and Google Now (to name just a few). Complementarily, neuroscience and cognitive science are providing us an ever-deeper understanding of the human brain and mind each year. While pursuit of human-level AGI was marginalized within AI academia during the 1980s and 90s, now it is increasingly becoming accepted as a valuable R&D direction once again. We now have annual conferences on AGI, Cognitive Systems and Biologically-Inspired Cognitive Architectures (BICA), and also an increasing number of sessions related to human-level AGI and AI within AAAI, IJCAI, IEEE and other generic AI-oriented conferences. Al visionary Ray Kurzweil has recently taken a position as a Director of Engineering at Google; and industry luminaries like Intel CEO Justin Rattner now forecast the arrival of human-level AGI within decades. The time is ripe for a serious frontal assault on the AGI problem. We have the tools and the knowledge; all that's needed now is the courage and persistence to confront the problem head-on. AGI is not a trivial problem by any means, combining as it does multidisciplinary R&D with large-scale software engineering. But given the technology and science of 2013, human-level AGI is an eminently reasonable near-term development goal. EFTA01103468  5 1.2 The CogPrime Design To create something as complex as a human-level mind requires a comprehensive, coherent design based on sound science and engineering principles. The foundation of the proposed research is the CogPrime AGI design, described in the books Engineering General Intelligence Vol. I and 2, by Ben Goertzel, Cassio Pennachin and Nil Geisweiller, published in December 2013 [10, 11]. CogPrime provides a high level design for a human-level AGI system. It leaves many medium and lower level algorithmic and implementation problems open; but it gives a clear, coherent conceptual and software framework for AGI engineering, detailed design and experimentation. Exposition of CogPrime and why we believe it has the capability to yield human-level general intelligence would extend this proposal excessively, so we refer the reader to the online CogPrime overview available at http : //wiki opencog org/w/CogPrime_Overview. There have also been several academic conference papers published on CogPrime and its potential, most recently [8] pre- sented at the 2013 IEEE Symposium on Hunan-Level AI, but the online article gives the clearest concise exposition of the design at this stage. The conceptual foundation of CogPrime is the "pattemist- theory of mind developed in Ben Goertzel's work during the 1990s and elaborated at length in [5], which views an intelligent system as concerned with recognizing patterns in itself and the world, focused substantially toward patterns regarding what actions will achieve its goals in observed contexts. On a practical level, the high-level architecture of CogPrime involves the use of multiple cognitive processes associated with multiple types of memory to enable an intelligent agent to execute the procedures that it believes have the best probability of working toward its goals in its current context. In a robot preschool context, for example, the top-level goals would be simple things such as pleasing the teacher, learning new information and skills, and protecting the robot's body. Each of CogPrime's cognitive processes is biased to recognize particular sorts of patterns, and the particular assemblage of cognitive processes is chosen based on a careful analysis of human cognition, with input from neuroscience, linguistics, philosophy of mind, computer science and other disciplines as well. CogPrime's memory types are the declarative, procedural, sensory, and episodic memory types that are widely discussed in cognitive neuroscience [18), plus attentional memory for allocating system resources generically, and intentional memory for allocating system resources in a goal-directed way. Table 1 overviews these memory types, giving key references and indicating the corresponding cognitive processes, and also crudely indicating which fun- damental cognitive dynamics each cognitive process corresponds to (pattern creation, association, etc.). The essence of the CogPrime design lies in the way the structures and processes associated with each type of memory are designed to work together in a closely coupled way, yielding cooperative intelligence going beyond what could be achieved by an architecture merely containing the same structures and processes in separate "black boxes:' All OpenCog memory types are implemented using a common weighted, labeled hypergraph knowledge store called the Atomspace; and all OpenCog cognitive processes are implemented as software objects called MindAgents, which interact with the Atomspace. The inter-cognitive-process interactions in OpenCog are designed so that • conversion between different types of memory is possible, though sometimes computationally costly (e.g. an item of declarative knowledge may with some effort be interpreted procedurally or episodically, etc.) • when a learning process concerned centrally with one type of memory encounters a situation where it learns very slowly, it can often resolve the issue by convening some of the relevant knowledge into a different type of memory: i.e. cognitive synergy Obviously this sort of high-level sketch merely serves to evoke the rough nature of the CogPrime system, and the curious reader should peruse the above references to get a fuller picture. The open source OpenCog software project (see opencog . org) provides a foundation designed explicitly for the implementation of the CogPrime design, and currently contains partial implementations of many of the algo- rithms constituting CogPrime. OpenCog has been used for commercial applications in the area of natural language processing and data mining. It has also been used for research involving controlling virtual agents in virtual worlds, and humanoid robots. EFTA01103469  6 General Cognitive Memory Type Specific Cognitive Processes Functions Probabilistic Logic Networks Declarative (PLN) [4]; conceptual blending pattern creation [3] MOSES (a novel probabilistic Procedural evolutionary program learning pattern creation algorithm) [16) association, pattern Episodic internal simulation engine [9] creation Economic Attention Networks association, credit Attentional (ECAN) [14] assignment probabilistic goal hierarchy refined by PLN and ECAN, credit assignment, Intentional structured according to pattern creation MicroPsi [2) association, attention Supplied by the DeSTIN Sensory allocation, pattern component creation, credit assignment [1) Table 1: Memory Types and Cognitive Processes in CogPrime. The third column indicates the general cognitive function that each specific cognitive process carries out, according to the patternist theory of cognition. The present proposal is aimed at completing the detailed design and implementation of the CogPrime AGI architecture within the OpenCog framework. 1.3 Application Foci It is important that an AGI system must do something even as its development proceeds; intelligence is centrally about engagement in a variety of complex tasks in complex environments. But choice of any one specific application runs the risk of overfitting the work to that application and ending up with a more specialized system than intended. Hence we propose six different application foci, to be pursued concurrently using the same integrated intelligent system: I. Control of intelligent animated characters in a 3D "video game" style world 2. Control of (humanoid and wheeled) mobile robots in an indoor environment 3. Natural language dialogue in the context of information available on the Web 4. Analysis of genomics datasets related to longevity in various organisms 5. Automated program learning 6. Automated theorem proving Each of these applications stresses different aspects of CogPrime and relates preferentially to different aspects of human intelligence. Pursuing them concurrently with the same developing AGI system guarantees generality of focus on the application level, alongside the generality of capability existing on the software level due to the nature of the CogPrime design. EFTA01103470  7 These applications have also been chosen because, with the exception of automated theorem proving, they are all areas that have already been explored using the OpenCog system, either in commercial or prototype applications; to wit: I. Animated character control: In the period 2008-2013, a variety of research prototype systems have been built using OpenCog to control virtual characters in virtual worlds [9]. Currently this initiative has research funding from the Hong Kong government, aimed at creating a toolkit enabling OpenCog-controlled non-player characters to be used in commercial games. 2. Mobile robot control: In 2009, at Xiamen University, a research project was conducted involving the use of OpenCog to control a humanoid Nao robot [15] [13]. This work was documented in the award-winning film Singularity or Bust, see http: //singularityorbust.com. Currently the Hong Kong government is providing research funding aimed at extending this work via using OpenCog to control David Hanson's Robokind humanoid robots, during 2014 and 2015. 3. Natural language dialogue: OpenCog's language processing tools have been used on the back end of several practical applications, such as an online language-teaching site, and a US government information system aimed at intelligence analysts. While these have not focused on dialogue, they have stressed most of the same NLP components that will be used in dialogue. These language processing tools have also been used in research together with the NIH on information extraction from PubMed abstracts [12]. A Hong Kong government grant has been obtained to support extension of this work toward the creation of an OpenCog dialogue system, in the specific context of a smartphone-based dialogue agent focused on media consumption. 4. Genomics data analysis: OpenCog tools, primarily MOSES but also PLN and clustering tools, have been extensively used to analyze genomics data for commercial, government and academic customers. This has led to various successes such as discovering the first genetic basis for Chronic Fatigue Syndrome (7], learning highly accurate diagnostics for Alzheimers and Parkinson Disease [17], and understanding the means via which calorie restriction impacts longevity [6]. 5. Automated program learning: OpenCog's MOSES component, which performs automated program learn- ing, has been used for numerous custom commercial data mining jobs (via the consulting firm Novamente LLC), to learn small programs constituting patterns in data ranging from biology to finance to market research and power transformer performance 1.4 Modular Design & Development The CogPrime architecture is structurally modular but dynamically unified. This means that, from an engineering perspective, it subdivides human-level AGI into a set of discrete modules. However, the intended intelligent oper- ation of the whole system is dependent on synergetic interactions between the modules. Each module is intended to display meaningful intelligent behaviors on its own but it's expected that these behaviors will be less scalable and more narrowly-scoped, than the behaviors the same modules will display in the context of the overall integrated CogPrime system. The modular architecture of CogPrime naturally supports development by a distributed team of teams, with teams focused on particular modules, teams focused on infrastructure tools useful across multiple modules, and then a central integrative team focused on putting all the pieces together to achieve overall generally intelligent behavior. While the module-specific teams may be purely research oriented, and the infrastructure teams may be purely engineering oriented, the central integration team must combine research and engineering capabilities. We propose that development be broken down according to the following teams, some of which correspond specifically to application areas as outlined above, and some of which correspond to important support tasks: I. Integrated Cognition 2. Distributed and Multicore Processing Infrastructure EFTA01103471  8 3. Sensation (Vision, Audition & Haptics) 4. Robot Movement Control 5. Game World Development 6. Genomics Data Analysis 7. Text, Image and Video Mining 8. Automated Programming 9. Automated Theorem Proving 10. Teaching and Intelligence Testing II. Software Integration and Testing In Section 4 below we summarize the basic approaches and tasks we propose each team to undertake. While it would be viable to colocate all the teams, it would also be viable to spread the teams among different locations based on the existence of relevant expertise. Some of the larger teams could potentially be split among more than one location, in themselves. 1.5 Open Source Development The OpenCog system has been developed as an open source software platform since 2008, and we propose to continue development of the CogPrime design in this vein. The advantage the OSS methodology presents is the capability of leveraging dramatic additional intellectual and software development, debugging and testing resources, via leveraging of the academic and OSS software development communities. On the software level, the advantages of OSS for ensuring software robustness, scalability and stability are well known. On the intellectual level, there are obviously very considerable benefits to be achieved via open involvement of the academic research world in the ongoing improvement of various aspects of a human-level AGI system as it matures. While a traditional business perspective would suggest that open sourcing the proposed software development is a negative from the perspective of ultimate monetization of the results, this is not necessarily the case. There are many viable, and potentially highly lucrative business models that the funders and developers of the proposed AGI software could pursue, that would not be negatively impacted by the open source nature of the underlying software. 1.6 Potential for Commercial Spin-offs The R&D project described here is proposed primarily for the transformative effect it would have upon science, technology, society and the evolution of intelligence. However, as a side-effect, numerous possibilities will arise along the way for leveraging the technology developed in various business domains. As AGI has the potential to transform every single area of commerce, giving a comprehensive list here would not be viable. However, a few of the possibilities closest to the specific AI applications to be pursued in the proposed work are: • AGI non-player characters for video games, or (depending on the game design) AGI for controlling whole game worlds. This could be provided as game-AGI middleware, or on a custom per-game basis in partnership with game companies • AGI toy robots, home service robots or elder care robots • Combining game characters, robots and other possibilities, it would be viable to develop a cloud-based facility for serving OpenCog based intelligence to various online software applications — e.g. games, robots, consumer electronic devices, specialized information systems. The developers of the first generally intelligent OpenCog system would obviously have a substantial "first-mover advantage" in setting up such a facility. And once EFTA01103472  9 such a facility were operational, the cloud-resident AGI would gain dramatic knowledge from its customers in the course of its operation, setting up an "increasing returns" dynamic that would make it very difficult for competitors to catch up (similar to but quite likely more substantial than the increasing-returns based advantages enjoyed by current firms such as Google and Facebook). • A host of genomics-related biomedical opportunities, including — Discovery of targets for pharmaceutical, nutraceutical and gene therapy interventions for age-associated and other diseases. — Integration of AGI software with rational drug design software to enable creation of novel molecules targeting combinations of genes highlighted by AGI genomic/proteomic analysis. — Predictive toxicology, to identify the human body's reactions to substances prior to their synthesis — Integration of AGI with systems biology simulation software to enable simulation of organismic response to therapies • Conversational personal assistants — like Siri, but with genuine understanding of what they're talking about • Conversational agents for helpdesk and customer support By its very nature a general intelligence can be applied in multiple domains, for great benefit and also potentially substantial financial profit. But first we must meet the challenge of creating a core of generally intelligent software, which is the focus of the proposed R&D work. 2 Incremental Development Milestones 2.1 Phase 1 Here we list high-level development milestones in each of the identified application areas, year by year for Phase 1 of the proposed project. Much more detailed milestones will be established year by year as the work proceeds. Year 1 Capability Milestone Animated Agent Conception and execution of plans to achieve complex movement and building tasks in blocks-focused game world Mobile Robot • Effective recognition of a closed class of objects and events • Navigation in dynamic indoor environments Dialogue System Simple dialogue about objects , events and goals in the game world Genomic Analysis Construction of an integrated Atomspace-based knowledge base containing gene expression, SNP and protein-protein interaction data, pathway and Gene Ontology data, and information extracted from PubMed abstracts Automated Pro- MOSES-based learning of sorting and searching algorithms gramming Automated Theo- Effective importation of Mizar formalized math database into rem Proving OpenCog knowledge representation EFTA01103473  10 Year 2 Capability Milestone Animated Agent • Linguistic communication about needs and desires in game world • Event recognition • Recognition of never-before-seen objects • Social reasoning in game world; theory of mind; basics of empathy, and social manipulation and deception Mobile Robot • Recognition of unfamiliar objects and events • Simple, goal-directed reaching and grasping Dialogue System Dialogue about objects , events and goals in the game world, in- volving complex sentences with multiple clauses Genomic Analysis • Supervised (MOSES) and unsupervised (MOSES, cluster- ing, pattern mining) learning based analysis of multiple ge- nomic datasets utilizing integrated information • Extrapolation of consequences for drug/nutraceutical target discovery and diagnostics for age-associated diseases Automated Pro- MOSES-based learning of simple Al heuristics for solving puz- gramming zles and narrow-AI problems Automated Theo- Simple set theory and geometry theorem proving within rem Proving OpenCog EFTA01103474  11 Year 3 Capability Milestone Integrated Cogni- Ability to pass "3 year old child" variant of AIQ test in the game tion world Animated Agent • Construction of complex objects • Group creativity among several AGI agents in game world • Following and giving of multi-step instructions Mobile Robot • Perception-guided object manipulation with robot hands • World-understanding based on integrating acoustic and vi- sual data Dialogue System • Game-world conversations involving roughly human child like understanding of context and intention • Dialogue about objects, events and goals in the physical world, in the robotics context Genomic Analysis • integration of information extracted from research article bodies • conception of novel hypotheses regarding relationships be- tween biological entities and processes, via PLN inference Automated Pro- Learning of modular programs, combining other learned pro- gramming grams in judicious ways Automated Theo- Use of PLN probabilistic reasoning to guide set theory and geom- rem Proving etry theorem proving EFTA01103475  12 Year 4 Capability Milestone Integrated Cogni• Ability to pass "4 year old child" variant of AIQ test in the game lion world Animated Agent Carrying out of simple "scientific" experimentation in game world Mobile Robot • Building structures from blocks and other simple objects • Supplementation of third-party speech-to-text with deep learning based speech-to-text Dialogue System • Ability to understand and produce sequences of sentences embodying a coherent, contextually relevant thought (in the game world and physical world • Dialogue about information extracted from texts (not only about directly experienced events) Genomic Analysis "Artificial bioinformatic scientist" functionality involving iterated automated generation of hypotheses and testing of hypotheses against datasets Automated Pro- Learning of simple programs that involve interaction with the gramming Atomspace Automated Theo- Use of intuitions gained from the game world to guide geometric rem Proving theorem proving EFTA01103476  13 Year 5 Capability Milestone Integrated Cogni- tion • Ability to pass "5 year old child" variant of AIQ test in the game world • Ability to pass "3 year old child" variant of AIQ test in the robotic embodiment Animated Agent Solution of complex game-world puzzles based on a combination of formal and intuitive reasoning Mobile Robot Creative, child-like play with physical objects Dialogue System Ability to understand and produce sequences of sentences em- bodying a coherent, contextually relevant thought, regarding in- formation extracted from text Genomic Analysis • Generation/testing of more complex hypotheses • Integration of more complex information from research ar- ticle bodies Automated Pro- Learning of programs involving Al-based heuristics that interact gramming with the Atomspace Automated Theo- rem Proving • Use of intuitions gained from robotics to guide geometric theorem proving • Extrapolation from experientially grounded (game-world- related) to ungrounded theorem-proving via PLN analogi- cal inference; e.g. from geometry to indirectly geometry- related set theory algebra EFTA01103477  14 2.2 Phase 2 Phase I development will be focused on implementation and testing of new CogPrime functionalities, and explo- ration of the implications of these functionalities in the chosen application domains. Phase 2, on the other hand, is intended to encompass 3 years of teaching the AGI system, and watching it learn and explore, and making modifi- cations to the system as merited by observing its progress. It is difficult, at this stage, to project the learning progress of a CogPrime system of this level of sophistication. However, based on the nature of the CogPrime design, we can conjecture with reasonable solidity as to what kind of functionality the system may acquire after roughly 3 years of learning from its environment and its human teachers: EFTA01103478  15 Year 6 Capability Milestone Integrated Cogni- tion • Ability to pass "8 year old child" variant of AIQ test in the game world • Ability to pass "5 year old child" variant of AIQ test in the robotic embodiment Animated Agent Ability to robustly take knowledge gained in the game world and port it to the physical world, and vice versa Mobile Robot Leaving the lab and learning to navigate and socially interact in the city streets (with human assistance at first Dialogue System • Ability to hold intelligent conversations at the rough level of a 7 year old human child • Ability to robustly learn new words and lin- guistic expression patterns from experience • Ability to read and understand general writ- ten information aimed at children aged 7 or younger • Ability to effectively correlate words with im- ages and videos, as required for understanding e.g. children's books or educational videos Genomic Analysis • Formalized understanding of experimental de- signs and how they relate to datasets • Robust inferential connection between bio- logical domain knowledge and general knowl- edge of the everyday world Automated Pro- Ability to automatically create simple MindAgents gramming to perform aspects of OpenCog reasoning, based on formal descriptions of MindAgent requirements Automated Theo- Ability to prove theorems in more abstract areas of rem Proving (still elementary) geometry or set theory, beyond what the game world provides grounding for EFTA01103479  16 Year 7 Capability Milestone Integrated Cogni- tion • Ability to pass adult-level variant of AIQ test in the game world • Ability to take ordinary human IQ test and correctly under- stand and answer a majority of questions Animated Agent Ability to enter a variety of game worlds, understand the prop- erties of these worlds, and figure out how to achieve the relevant goals there Mobile Robot More robust interaction in a variety of social and physical situa- tions outside the lab Dialogue System • Ability to hold intelligent conversations at the rough level of a 10 year old child — about science, and also about human relations and the system's own mind-state • Ability to read and understand most Web pages, except those with highly specialized or informal content • Ability to learn aspects of new languages based on experi- ence and teaching Genomic Analysis Creation of hypotheses, analysis of data and design of experi- ments at the level of a human biology undergraduate student Automated Pro- gramming • Ability to write simple scripts to carry out functions in the Linux operating system • Ability to create MindAgents performing aspects of OpenCog reasoning, based on informal description of MindAgent requirements Automated Theo- Ability to prove theorems in more advanced undergraduate ge- rem Proving ometry, set theory, topology and calculus EFTA01103480  17 Year 8 Capability Milestone Integrated Cogni• Ability to pass an ordinary human adult IQ test with a strong score Lion Animated Agent Ability to enter an essentially arbitrary new game world, under- stand the properties of the world, and figure out how to achieve the relevant goals in the world Mobile Robot • Ability to navigate and manipulate objects in unfamiliar sorts of environments (e.g. outdoors, in a basement, etc.) • Ability to automatically adapt to new robotic body parts Dialogue System • Ability to hold intelligent, though not necessarily precisely human-like, adult-level conversations — about science, and also about human relations and the system's own mind- state • Ability to adapt all aspects of language comprehension and generation based on linguistic experience • Ability to read and understand most Web pages • Ability to learn new languages based on experience and teaching, in the manner of human language learners Genomic Analysis Creation of hypotheses, analysis of data and design of experi- ments at the level of a human biology graduate student Automated Pro- gramming • Ability to write simple scripts to carry out functions in the Linux operating system • Ability to modify its own MindAgents for superior func- tionality Automated Theo- Ability to prove theorems in all areas of undergraduate mathemat- rem Proving ics, guided when needed by analogy to its grounded experience with geometry, set theory and arithmetic EFTA01103481  18 3 Staffing and Costs 3.1 Advisors and Technical Leads 3.1.1 Technical Leads Key to the success of the proposed work will be the involvement of individuals already expert in OpenCog software and its application, e.g. EFTA01103482  19 Name Role Dr. Ben Goertzel founder of the OpenCog project Dr. Lines Vepstas principal engineer of the OpenCog software system for the last several years Dr. Nil Geisweiller • current main developer of OpenCog's MOSES subsystem • prior developer of OpenCog's PLN subsystem and OpenCog's connection to the Nao robot • coauthor of Engineering General Intelligence Dr. Eddie Monroe OpenCog machine learning engineer at Novamente LLC Dr. Matthew We' co-developer of the mathematics underlying OpenCog's attention allocation and probabilistic logic modules Dr. Joel Pitt former OpenCog Hong Kong team lead, developer of OpenCog's attention allocation module Ruiting Lian lead OpenCog natural language developer David Hart OpenCog IT/infrastructure guru since the project's start Shujing Ke lead OpenCog planning, pattern mining & game-AI developer Lake Watkins current developer of the game world used for testing OpenCog Scott Jones OpenCog core system developer, current team lead of OpenCog Hong Kong project Cosmo Harrigan OpenCog Al developer Alex van der Peet OpenCog game world and AI developer Jade O'Neill OpenCog AI developer (principal developer of OpenCog's current Probabilistic Logic Networks implementation) Ted Sanders DeSTIN expert Michel Drenthe DeSTIN expert Misgana Bayetta OpenCog (MOSES) developer Teddy Habtegabriel DeSTIN developer Rodas Solomon and OpenCog language processing specialists Amen Belayneh Keyvan Sadeghi author of OpenCog temporallspatial reasoning Mike Duncan Biomind LLC bioinformaticist, currently applying MOSES to genomics datasets Angus Griffiths lead developer of Mathics, key tool for automated theorem proving EFTA01103483  20 The availability of a relatively large team of experienced AI software developers who are "ready to go" on OpenCog applications is a valuable asset. 3.1.2 Advisors The proposed work will also benefit substantially from the part-time participation of a set of AI-expert advisors, including many who have collaborated on OpenCog work in one way or another in the past, and some who have developed the