EFTA01115115.pdf

Project PrimeAGi -sealing a human level o. greater Artificial General Intelligence lung Goal: by 2025 to 2030 ,.)Ing it as an Open Source project for the good of all humanity Doing it Safely Raising funds for Phase 1: 3 years. SUS 6 million EFTA01115115  Artificial General Intel'. ence (AGI A system capable of choosing actions to achieve a set of complex goals in a variety of complex The ability to transfer learning from one domain environments using a realistic amount of to other domains. computation Human Level AGI The Turing Test: A text dialogue Al that is indistinguishable from a human in online conversation. Better Examples of Human Level AGI: • Wozniak's Coffee Test: when a robot can go into an average American house and figure out how to make coffee. including identifying the coffee machine. figuring out what the buttons do. finding the coffee in the cabinet. etc.' Robot University Student Test 'when a robot can enroll in a human university and take classes in the same way as humans. and get its degree. then we've created artificial general intelligence What is an Artificial General Intelligence (AGI) ? "a system capable of choosing actions to achieve a set of complex goals in a variety of complex environments using a realistic amount of computation" "general intelligence is the ability to transfer learning from one domain to other domains." Human Level AGI The Turing Test: A text dialog Al that is indistinguishable from a human in online conversation. Better Examples of Human Level AGI: Wozniak's Coffee Test: "when a robot can go into an average American house and figure out how to make coffee, including identifying the coffee machine, figuring out what the buttons do, finding the coffee in the cabinet, etc." Robot University Student Test: "when a robot can enroll in a human university and take classes in the same way as humans, and get its degree, then we've created 2 EFTA01115116  artificial general intelligence" TOO many references to Robots?? 2 EFTA01115117  Can on drive a Is TRAINED or • Deep Blue • AlphaGo • Facebock Face • Google Self Driving • IBM Watson for IBM Watson for Medical Compare General with Narrow Artificial Intelligence Narrow: Deep Blue AlphaGo Facebook Face Recognizer Google Self Driving Car IBM Watson for Jeopardy IBM Watson for Medical A Narrow Al can only do one thing. Eg: Deep Blue can't drive a car, nor Google Car play chess. A General Al can in principle learn to do anything a human can learn and more. A Narrow Al is TRAINED or just PROGRAMMED, a General Al is EDUCATED 3 EFTA01115118  3 EFTA01115119  Why Now? AGI has been the implicit long term goal of Al Research since at least 1950 Early Al researchers sought human-level Al. but underestimated the problem • From the mid-1970s through the mid-aughts. Al research focused on narrow problems. • Over the last 10 years. narrow Al heavy application in industry, and serious interest in AGI I= has resurged Today, for the first time. there exists the necessary convergence of research progress and ever increasing computational power and looks to make a serious assault on the AGI problem: • Computi,g hardware and networking have advanced massively • Software tools and libraries are far more reMed • Cognitive science has given us an overview of human mind architecture • Narrow Al and CS now provide a formidable arsenal of algorithms and representations • Others think so too: • Kurzweil. Hinton. etc to Google. Andrew Ng to 8aidu. Yann LeCun to Facebook. DeeplAnd acquired by Google for $500M. Vicarious: $40M funding from Zuckerberg. GoadAl: $10M from Marek Rosa diectly for human4evel AGI. OpenAl — $18 committed for future work on open-source AL.. Why Now? AGI has been the implicit long term goal of Al Research since at least 1950 Early Al researchers sought human-level AI, but underestimated the problem From the mid-1970s through the mid-aughts, Al research focused on narrow problems, Over the last 10 years, narrow Al heavy application in industry, and serious interest in AGI R&D has resurged Today, for the first time, there exists the necessary convergence of research progress and ever increasing computational power and tools to make a serious assault on the AGI problem: Computing hardware and networking have advanced massively Software tools and libraries are far more refined Cognitive science has given us an overview of human mind architecture Narrow Al and CS now provide a formidable arsenal of algorithms and representations Others think so too: 4 EFTA01115120  Kurzweil, Hinton, etc to Google, Andrew Ng to Baidu, Yann LeCun to Facebook, DeepMind acquired by Google for $500M, Vicarious: $40M funding from Zuckerberg, GoodAl: $10M from Marek Rosa directly for human-level AGI, OpenAl — $1B committed for future work on open-source Al... 4 EFTA01115121  Why Us? Ben Goertzel • leading thilker. evangelist and organizer in AI community Probabilistic • 15 books. 100+ papers Logic • Founder and leader of Networks • annual AGI Conference • Open Source AGI Foundation • OpenCog project • 2 decades practical AI application experience in robotics. Artificial Mance. biology. NLP. data minng. etc. General Intelligence Core team with 20 year learni-ig curve working together • Webmi-id -> Novamente -> Opencog -> PrimeAGI • Opencog is a solid foundation on which to build: significant open source community with 50 current participants plus 'alumni association" Forming General • A specific point of view: an approach and a plan — not just a general Intelligence, Pan I AGI research effort — Reasonably well documented in Engineering General Intelligence. Vols. 1 and 2 • Firm commitment to Open Source Why Us? Ben Goertzel leading thinker, evangelist and organizer in Al community 15 books, 100+ papers Founder and leader of annual AGI Conference Open Source AGI Foundation OpenCog project 2 decades practical Al application experience in robotics, finance, biology, NLP, data mining, etc. Core team with 20 year learning curve working together: Webmind -> Novamente -> Opencog -> PrimeAGl Opencog is a solid foundation on which to build: significant open source community with 50 current participants plus "alumni association" A specific point of view: an approach and a plan — not just a general AGI research effort — Reasonably well documented in Engineering General Intelligence, Vols. 1 and 2 5 EFTA01115122 Firm commitment to Open Source 5 EFTA01115123 Open Source AGI is too important for humanity to be the property of one commercial or governmental entity. Creating safe. beneficial AGI is a difficult problem and the best solution is more Ikely to emerge if • a wide variety of bright minds are working together on it • It is created within an effort with beneficial AGI as the central goal. rather than e.g. profitability of a particular corporation or watery success of a particular country High leverage on I-westment through growing network of volunteers. academics. and commercial Open Source community participants: • Directly funding 10 developers may ildireclly bring 1000 additional developers funded by other sources VS open source initiative 6 EFTA01115124 Project PrimeAGi Use and extend the existing Opencog AGI Framework to drive towards at least Human level AG in three phases: 'V2 Rocket of AGI" "Sputnik of AGI" -.Apollo of AGI" First system to show exciting First system to show Human-level AGI functionality that demonstrates incontrovertible evidence of AGI the core principles needed to make AGI work 7 EFTA01115125 Robot head that holds a simple Robot head that learns via Robot head that holds a simple conversation with a human. interaction with humans and the conversation with a human. displaying genuine world. similady to a human child displaying genuine understanding of its embodiment (including learni-ig new understanding of its embodiment and context languages and social skills via and context experiential interaction) Once an AGI can be taught university-level mathematics and CS, we begin moving toward profoundly self-improving, self-modifying AGI. 8 EFTA01115126 "V2 Rocket" Goal of this fundraising round A realistic-looking. emotionally and socially interactive robot head which holds a simple conversation with you. The conversation isn't just a °chat bar but an embodied intelligent dialogue system Robot dialogue displays genuine understanding of: • vAio and what it is. including its own history • who you are • wt at its physical surroundings are System able to learn and reason about its environment and interactions • Learning/reasoning ability broadly similar to a young human chid • In some cases more advanced — 'idiot savant' Ike 9 EFTA01115127  OpenCog: our unfair platform advantage Given the current state of software and hardware Other proto-AGI projects lack a platform of comparable tech. creating a human-level AGI is both appropriateness and quality • A substantial research project. comprisrig • Academic projects are mainly implemented by grad multiple coupled sub-research-projects students. with minrnal attention to scalattlity. code quality. software architecture. etc. • A large-scale software engineering project • Corporate projects tend to be implemented on powerful Executrig on these two aspects in parallel presents distributed frameworks origrially designed for non-AGI unique challenges purposes, hence not fully apropos for the goal We address this issue via inked but separate development of • PrimeAGI, a comprehensive. Wei fleshed out AGI design with deep theoretical grounding • OpenCog. a software framework and platform designed for scalable implementation of hybrid Al systems — very suitable for PrimeAGI. but not only for PrimeAGI 10 EFTA01115128  11 EFTA01115129  Procedural Library Cognitive Synergy I was a aly. I krow Iron lima memory Mel. Swat avallatiro la lha Miran. Vial be *Won was brernd Me door. and Meld maid walk on be Pvn Ito Itc 4:100t I walk to the door Multiple cognitive processes, operating on different types of UnIfIcol Rub memory stored in Atomspace. Ebro providing each other with real-time cognitive guidance Atomspace Integrative Technologies 12 EFTA01115130 Resources and Timeline Aim: the 'VT PrimeAGI system: embodied dialogue with deep reasoning and understanding Approximately 20 full time technical staff, composed of four teams: • Artificial General Intelligence R&D • Product • System and Tools • Community and Process Ongoing collaborations with external researchers • Three year project with three phases. broadly described as setup. research. delivery. 13 EFTA01115131 Team overview AGI R&D: Largest team. split into subteams (Inference. language, learning. attention. cognitive control, perception, action, memory) • Product: OSS releases, integration. tuning Im0 • System & tools: infrastructure and architecture projects, visualization. etc. • Community: demos. outreach. docs System and Tools 14 EFTA01115132  Project Phases Team Sizes • taon.,• . <44.0,404Y •SY401, h," • knamma tan >a 8,40•4•.coM W al. as I: Setup.,mos 2: Research. 21 mot 3: Daily . 9 r.>: KM aro! Vain Fundamental nod appled research • Fina in eechAGI Mem • Woe renetth SOMA • ChgeMaIrm tests art melautc, • OtaXSC oiler/Mx eM evelX6en • OVMCDO Saw • Firmil prOdott retneSeS !Reed • commas amprcramants comet mown) obscfm. Roma and swat 15 EFTA01115133 Key Goals by Phase • Opanagroady topo. as an OSS °prisai*, • Pire lui Øm ?rd bing 'hen UP b nottd • Reno minou:htitra 1T tordarranisl retesel. and break te/Rama:0 research Mina, oie tx.ralions isOa)ing aØ undoeslaading of Mo rotors 'nitr-utero. contas,. tacots and ol toll 'tractera usa! Io Omo tcd‘wee and eue* roatang so toa robot n tata abbto0123t4 uses and ache,* goals lion ace ,nrcent modulos crenno goal acreomnoni al b ablo domonsbalo corral«bahayet and «plain lis aises and molecanons ard moWaydl et, Green getfon, end suileol OSS protecl‘ earacting n •x/en communly Qmo reenerches end tiers O • Cleet.SliOngclererolnette «AGI achnownerls • Sectes &tenu k« mot MW, inhalera • COMM. fteclernente, ard nçied ReO • CommeOpaeo9Patien•wdOSS oxnntinity grisin 16 EFTA01115134 Core Team Members Dr. Ben Goertzel Jim Run. former Chairman of the Santa Fe Institute. CEO of Network Solutions. Inc.. CTO Thomson-Reuters. investor and startup entrepreneur Cassio Pennachin. co-founder and chief architect of OpenCog. co-author of PrimeAGI design. Dr. Goertzel's AI collaborator since 1998 Dr. Nil Geisweiller. PhD in probabilistic logic. developer on OpenCog and predecessors since 2005. co-author of PrimeAGl design Dr. Linas Vepstas. PhD in physics. developer on OpenCog and predecessors since 2007. lead OpenCog developer since 2006. lead developer for GnuCash. experience developrig for Linux kernel and numerous major corporate and OSS projects Andre' Luiz de Senna. co-founder of OpenCog and lead developer of original OpenCog code, Goertzel and Pennachiris AI collaborator since 1999 Dr. Eddie Monroe. PhD in psychology, developer on OpenCog and predecessors since 2006 17 EFTA01115135 Additional Material To be added to the presentation as appendices. covering the following topics: • Atomspace architecture and application • PrimeAGi design and concepts • Project plan specifics • Comparison to compelmg projects 18 EFTA01115136