Among the many AGI designs and proposals mentioned in this thread, it was refreshing to see some actual results from Peter Voss's Aigo. (Also entertaining as my Alexa was listening and answering back while I played the demo videos). Experimental results are a lot more work to obtain than ideas, which is why most publishers and reviewers require them. I realize this is difficult for AGI, which I guess is why 85% of the papers accepted to the AGI conference still lacked a results section the last time I looked.
My last 20 years of research can be summarized as finding experimental evidence (not proof) supporting the following hypotheses: 1. The best language models are based on neural networks. 2. Intelligence grows logarithmically with CPU time and memory. 3. Automating all human labor with AGI will probably cost $1 quadrillion. We recently learned that the best vision models are neural networks. My work suggests this is true of language too. It is based on testing thousands of versions of 200 compression programs since 2006 on a 1 GB text benchmark, found at http://mattmahoney.net/dc/text.html Text compression measures text prediction or modeling by adding a coder, which is a solved problem. The top models use dictionary preprocessing to convert words into tokens followed by PAQ style compression predicting one bit at a time using ad-hoc context features and shallow neural networks. They implement essentially toddler level language models with hard-coded lexical features, proximity based semantics and flat (n-gram) simple grammars and dictionaries sorted by grammatical role (i.e. grouping "monday" with "tuesday" or "brother" with "sister"). The models so far lack advanced grammars necessary to understand math, software, or complex sentences. Prior to my work on PAQ based compression, the best models were PPM (prediction by partial match) until about 2003. PPM predicts bytes rather than bits using the longest matching contexts. I started work on neural based compression in 1998, 5 years before achieving this result. The second hypothesis has several caveats. By intelligence, I mean text prediction accuracy. I show that human level prediction (which we have not yet achieved) implies passing the Turing test. Not everyone accepts the Turing test as general intelligence since it lacks non-text based processing like vision, music, and robotics, all requirements for AGI or automating labor. Also, my tests (with the same benchmark) only show a logarithmic trend over the range of a few bytes up to 32 GB and 1 to 10^6 operations per byte. If we assume that 10% of the human brain is used to process language, then the goal figure is 10^13 bits of memory and 10^14 operations per character. For my third hypothesis, please note I am estimating the cost of several billion human level intelligences, not just one human level AGI. The two pieces of evidence I produced in support of my claim are: 3A. My 1998 masters thesis where I showed the scalability and robustness of distributed indexing using computer simulations. Distributed indexing is an essential feature of an AGI design consisting of lots of independently developed and competing narrow AI such as my 2008 proposal. (The thesis is here: https://cs.fit.edu/~mmahoney/thesis.html ). 3B. I showed that recursive self improvement in a closed environment (boxed AI, sometimes proposed as a shortcut to AGI or a singularity) is impossible. http://mattmahoney.net/rsi.pdf Of course none of this disproves the possibility of other, less expensive routes to AGI. But logic based AI is probably not one of them (per my first result) and early progress does not predict success (per my second result). -- -- Matt Mahoney, mattmahone...@gmail.com ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T731509cdd81e3f5f-Mda6e59327c21a47a77423b17 Delivery options: https://agi.topicbox.com/groups