Here are my messy notes from Sunday… and here are the powerpoints.

Here at the Crown Plaza in Palo Alto for the Artificial General Intelligence Research Institute’s AGI Workshop… people are just starting to show up, and it’s almost 2:00PM. The audience appears to be a balanced mix of AI researchers and general transhumanists/futurists, with very substantial overlap, of course. The first presenter is Ari Heljakka, a programmer with Novamente LLC and the CEO of Finnish AGI company GenMind Ltd. His talk is entitled “AGI & the Singularity”. It’s a general overview of Singularity ideas from Vernor Vinge, Eliezer Yudkowsky, Ben Goertzel, and Ray Kurwzweil… Eliezer walks in while a slide of his is being shown, he’s wearing the same shirt as in the slide, everybody laughs.

AGI vs. Narrow AGI is the second segment, it a short talk by Ben. Look at the powerpoint if you’re interested.

Third segment is a panel with myself, Christine Peterson, Ben Goertzel, and Eliezer. Christine seems to be unclear on the power of a superintelligent AI, thinking in terms of police rather than utility fog in every cubic centimeter of the earth’s crust. Stay tuned for the release of this video, it’s interesting.

Fourth segment:

Ben Goertzel, PhD
Novamente LLC
Biomind LLC
Artificial General Intelligence Research institute
Virginia Tech, Applied Research Lab for National and Homeland Security

Biggest commenters: Jef Allbright, Brad Templeton, Eliezer Yudkowsky.

Ben Goertzel again. Topic: Novamente, a Practical Architecture for Artificial General Intelligence. This focuses in on the AGI part, and temporarily puts aside the Friendly AI part. Goes over AI estimates - mentions that the estimates have come down a lot. Used to be hundreds or thousands of years, now it’s 50-100, even with academic AI researchers. Ben mentions that he’s more optimistic than Kurzweil, who says 2029, and says that if it takes that long, it’ll be due to sociopolitical reasons, not technical reasons. Mentions that predictions are tied to outcomes - self-fulfilling prophecies. Early 80s - time travel, unification of physics, AI, origin/synthesis of life. Thought that his impact in AI would be greatest - “all it does it figure out the right code to write”.

Brad Templeton: decide to work on time travel, if your future self doesn’t come back, it’s definitely intractable.

Long-term goal: AGI. Approach based on computer science algorithms, though human psychology is utilized. Working on a couple of narrow AI projects, reusing code from AGI. Scalable C++ run on linux machines. Biomind OnDemand for bioinformatic data anlysis, ImmPort:NIH Web biomind/novamente backend.

Patternist philosophy of mind: Recognizing patterns in world and itself. Probability theory used to quantify and relate patterns. Logic (term, predicate, combinatory), used as a base-level language for expressing patterns. Reflectivity of recognizing self-patterns and improving those patterns. Philosophical approach matters for how you approach it technically. Carrying out procedure P in context C will achieve goal G.

Ben books: Probabilistic Term Logic, Engineering General Intelligence, Artificial Cognitive Development. Open source: AGISim, embodiment for AI. Goal-driven inference, 45%, sensor processing 35%, background inference, 20%. Ben is not a purist about embodiment, cares about the ingestion of databases (WordNet, FrameNet, Cyc, quantitative scientific data, etc.); approach called ‘post-embodied AI’. Non-body AI is okay, but it would have a hard time getting together with humans. Eliezer: AI is made of math, so even if it seems easy, it’s actually that hard to do Friendly AI all the time.

Formal stage: Reasoning about situations different than anything you’ve ever experienced? Reflexive stage: deep understanding and control of self structures and dynamics. Full self-modification: something humans can’t do. Ben lists various types of databases which could be integrated at various stages. Cyc, Mizar database, etc.

As long as it rewrites the schedulers and knowledge representation, it’s still reflexive, when it rewrites the C++ itself, then you’re in a new domain. Eliezer: Capability is more important than whether or not it actually is called a different thing. Atoms = nodes or links. Atoms have truth values (probability + weight of evidence) and attention values (short and long term importance). Atomspace weighted, labeled hypergraph. Steve Omohundro: do you keep track of the source? Ben; in many cases you can’t, but you can still represent in terms of nodes or links. Jef Allbright: what about context? Ben: most links are context links. In any finite system: ‘alienation of knowledge’ - it can’t always be put directly in context. Atomspace is not a neural net, nor a semantic net. Markov networks, Bayes networks - this is not one of those either. What processes are taking place? This is cyclic rather than acyclic. Accurate Bayes - too computationally intensive, heuristics always required for approximation. Meant to approximate Bayesian judgements.

Low short-term importance and long term = useless. Low short, high-long = remembered but not used right away (mother’s phone #), high short, low long = used then forgotten, like precepts. High long, high short = used and remembered. Truth values. Weight of eveidence low, strength low = weakly suspected to be false. Strength high, weight low = weakly suspected to be true.

Mind agents implement cognitive processes, atom space is passive. Gigabytes of knowledge in the machine - can be likened to a giant blackboard. Simplified workflow - precepts, active memory, feelings and goals, execution management, active schema pool, the world, and back again. Attention allocation - “simulated economy” approach. Pattern mining of the AtomTable, embody as predicates. SystemActivityTable, which MindAgents did what to which Atoms at which times. Nodes which fire together, wire together, language generation, probabilistic chaining, importance propagation. ABC, deduction, induction, abduction.