On Thu, Jul 13, 2023 at 6:21 AM <[email protected]> wrote: > > On Thursday, July 13, 2023, at 4:28 AM, Matt Mahoney wrote: > > Organizing disorganized thoughts begins with a goal. Why build AGI? We can > group the goals into roughly 4 categories. > > 1. Scientific curiosity, understanding the brain and consciousness. > 2. Automating labor. > 3. Uploading, immortality. > 4. World domination, launching a singularity, creating utopia. > > 5. Creating an artificial descendant
Nice one. Yudkowsky was right. AI will kill us all and we won't even put up a fight. But probably not in this century. Let's put a timeline on this. Assume Moore's Law continues doubling global computing power every 2 years. This is uncertain because clock speeds stalled at 2-3 GHz in 2010 and transistor sizes are likely to stall this decade because we are close to the ~5 nm spacing limit between dopant atoms in silicon. A RAM capacitor stores a bit using 8 electrons. Further advances will require nanotechnology, moving atoms instead of electrons, to solve the power problem. In about 60-70 years we will stall at the Landaurer limit 4 zJ per bit operations at room temperature, still a 10^9 improvement over transistors. 1. The AGI algorithm is mostly understood. LLMs pass the Turing test. We understand how neural networks succeeded where symbolic processing failed. Language evolved to be efficiently learnable one layer at a time in the order of phonemes, word segmentation, semantics, and grammar. Symbolic models, like those used for compilers, failed because they put grammar before semantics (e.g. how to parse "I ate pizza with Bob/olives/a fork"). Fully connected neural networks like transformers can learn arbitrarily deep hierarchical concepts like mathematics and world models of physics and social interaction. Human knowledge is half inherited and half learned (about 10^9 bits each), but LLMs can learn the inherited part, like human emotions, from an appropriately large enough unlabeled corpus. It knows how to model feelings without having feelings. It knows that it is an LLM. It is self aware without being conscious, in the sense that it understands how humans have an irrefutable sense of being conscious (as part of our evolved fear of death), without having this sense itself. 2. Automating labor requires more than language. Vision and robotics are advancing but not at human level yet. It will take about 30 years to reduce the cost of producing a movie from $1M to $10. The value of labor is world GDP divided by interest rates, about $1 quadrillion. We should expect investment on this scale. Modeling 10^10 human brain sized neural networks will require 10^26 OPS, 10^25 parameters, and 10^17 bits of human knowledge collected no faster than 5-10 bits per second per person at a cost of > $100 trillion. This is slow enough for humans to adapt to the changing job market without massive unemployment. Instead, AI will make us more productive, improve our lives both at work and home, and increase our income. But the big change is we will have little need or desire to interact with other humans because AI will be far more helpful. You can have everything you want, but this is not where happiness comes from. A state of maximum utility is static, without feeling. 3. We already have the technology and enough personal data to construct an LLM that claims to be you, happily living in a virtual utopia. All that remains is to construct a world where nobody else knows or cares that you exist in a human body. 4 and 5. To transform the world, technology has to catch up to biology. The biosphere has 10^37 bits of DNA storage and executes 10^29 DNA copy and 10^31 amino acid transcription operations per second. Human evolution was the result of 10^48 operations over the last 10^17 seconds (3 billion years). Photosynthesis generates food at a rate of 500 TW, out of 90,000 TW available solar power. We already have solar panels that are 30% efficient. Global computing power is now about 10^19 OPS and 10^26 bits. At the current rate of Moore's Law and with intelligent design, our self replicating, non DNA based descendents will be ready to displace DNA based life around 2100. -- -- Matt Mahoney, [email protected] ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T59f369fee7febd6d-Mcb1ab97f0526ccce0d01fb1b Delivery options: https://agi.topicbox.com/groups/agi/subscription
