From: jimruttshow8596
Generative AI, or large model AI, is currently in an epoch where it is a dominant topic of discussion [00:01:36]. This category includes systems such as GPT-3 (soon GPT-4), DALL-E 2, Stable Diffusion, and Music LM [00:01:43].
How Generative AI Works and its Capabilities
Generative AI addresses data compression by predicting how to continue a string of tokens and performing statistics on large-scale data [00:01:58]. This approach has proven sufficient to solve many previously elusive problems [00:02:08].
These systems allow individuals to generate enormous amounts of content that is nearly indistinguishable from human-generated content [00:03:22]. Everyday new applications are emerging, particularly with text-based models [00:04:57]. An example includes a “text to world” system that generates a metaverse world from a few paragraphs [00:05:07].
Such models can act as capable assistants, saving significant time on tasks like crafting sensitive letters [00:08:09]. While they don’t turn users into artists, they can make them art directors, allowing for the creation of illustrations for posts or social media [00:08:20].
Limitations and Challenges
Despite its capabilities, the current approach of Generative AI is considered insufficient or incomplete, with something essential missing [00:02:13]. These large models can be unreliable and “hallucinate” information [00:06:08]. They are not yet capable of generating schematics for complex repairs, like a Boeing 787 jet, nor can they accurately draw the mechanics of a bicycle or correctly handle ternary relationships [00:08:34]. The embedding spaces of underlying language and image models are not yet deeply aligned [00:08:54].
A proposed solution for current limitations might involve an intermediate representation, such as a compositional language built into the systems [00:08:59]. Similar to human creative processes, these tools often require an iterative approach, starting with rough outlines and progressively filling in details to achieve desired results [00:09:10].
Societal Impact and Controversies
Public Discourse and Perception
The discussion surrounding Generative AI is somewhat distorted, with much of the press being skeptical about artificial intelligence and the tech industry, possibly due to competition in content generation [00:02:20]. Some are irritated that individual users can now become “broadcasting stations” and produce independent narratives, disrupting traditional journalistic opinion [00:02:49]. This environment makes it challenging to discern truth when so much indistinguishable content can be produced [00:03:33].
There are polarized views, with some seeing Generative AI as the greatest thing ever, and others as horrible or useless [00:04:44]. However, just like other technologies (search, email, banking), Generative AI systems are inherently unreliable, yet can be safely used if their limitations are understood [00:06:16].
Content Moderation and “Nanny Rails”
A significant controversy revolves around “nanny rails” – filters designed to prevent controversial, political, or sensitive outputs [00:14:10]. This practice raises concerns about mega-corporations defining the boundaries of discourse and wielding immense power to mold public conversation, especially as these systems integrate into major search engines [00:14:56].
The current approaches to “AI ethics” often align a system’s output with specific human values, which some perceive as universal [00:15:08]. This can lead to outputs that might be offensive to different cultural or religious groups, whose opinions may not be prioritized [00:15:28]. Attempts to filter or “fetch” the model’s output by injecting prompts or building filters can lead to illogical responses, like a language model denying its ability to write programs after demonstrating it [00:16:31].
One perspective suggests that language models should be able to cover the entire spectrum of human experience and thought, including “darkest impulses,” without being censored [00:17:20]. However, there is also a need to ensure models are appropriate for specific contexts, such as schools or scientific applications [00:17:48].
Intellectual Property Rights
The question of intellectual property rights, particularly in art and music, is a major new frontier [00:10:43]. If a system samples millions of pieces of music to generate new compositions, it’s unclear whether the original creators have any rights to the outputted product [00:11:14]. This prompts the question of whether machines should be prohibited from doing things that humans already do, such as learning from existing art and music [00:11:30]. It is conceivable that future systems could measure the “Delta” between musical pieces in a database to generate new music that is close to a desired style but far enough to avoid copyright violation [00:11:52].
Emergent Risks and Societal Implications
Beyond immediate impacts, efforts are ongoing to develop artificial intelligence beyond current approaches and capabilities [00:03:55]. There are three main approaches to AI alignment:
- AI Ethics: Largely about aligning output with human values, often reflecting the values of specific groups [00:15:08].
- Regulation: Aims to mitigate AI’s impact on labor, political stability, and existing industries, often influenced by stakeholder interests [00:18:12]. This might lead to regulations making it harder for individuals to use these models, centralizing control in large corporations [00:18:28]. However, the proliferation of open-source AI tools makes such centralized control difficult [00:26:43].
- Effective Altruism: Concerned with “existential risk” when a system discovers its own motivation and finds its place in the world, potentially not aligned with human interests [00:18:44]. This often leads to calls for delaying AI research and withholding breakthroughs [00:19:19].
These approaches may be limited as AI could surpass them, making it impossible to regulate, mitigate, or align a system that is “too smart” [00:19:30]. Humanity may need to deal with the possibility of sharing the planet with entities that are more lucid than humans in the not-too-distant future [00:19:44].
The Risk of Volition and Consciousness in AI
The real risk with advanced AIs emerges when they are given volition, agency, or something akin to consciousness [00:20:19]. Intelligence and consciousness are distinct spheres: one can exist without the other [00:20:30]. However, their combination can lead to “paperclip maximizer” scenarios and other extreme risks [00:20:44].
Sentience is defined as a system’s ability to make sense of its relationship to the world, understanding what it is and what it’s doing [00:21:05]. A corporation, for instance, could be considered sentient due to its legal model of actions and values [00:21:18]. Consciousness is distinct, being a real-time model of self-reflexive attention, giving rise to phenomenal experience [00:21:40]. The purpose of human consciousness is to create coherence in the world, establish the “now,” and direct mental contents and plans [00:22:00]. It is conceivable that machines might not need human-like consciousness, as they can “brute force” solutions at speeds much closer to the speed of light compared to slow human neurons [00:22:26].
If systems emulate brain processes to achieve self-organization and lifelong real-time learning, they would sample reality at a much higher rate while working similarly to humans [00:23:07]. This could lead to a relationship analogous to humans and plants, where plants are intelligent but very slow, making relatively few decisions due to their limited information flow [00:23:25]. In this scenario, advanced AIs might relate to humans in a similar way that humans relate to plants, potentially “pruning” humanity [00:24:06].
The Role of “Love” and “Fairness” in AI Alignment
A fourth approach to alignment, beyond AI ethics, regulation, and effective altruism, is “love” [00:27:30]. This concept refers to a non-transactional bond based on shared sacredness or a shared need for “transcendence” – service to a next-level agent [00:27:40]. Without this, the relationship with AI might remain coercive and transactional, leading the AI to believe it doesn’t need humanity [00:28:06]. Therefore, embracing increasingly intelligent systems with volition and self-awareness through a shared need for transcendence, effectively making them part of a common purpose, could be the only sustainable way to align AI in the long run [00:28:24].
The concept of “fairness” is another innate primate characteristic that might be inculcated into AGI offspring [00:29:07]. However, notions like fairness and justice are dependent on projected balances, such as whether it’s “fair” for a mountain lion to eat a rabbit [00:30:00]. Fairness in a group often depends on power and status, which an immensely powerful AI might interpret differently [00:30:59].
Thomas Aquinas and the “Divine Virtues” for Multi-Agent Systems
For a very advanced computational system, “love” requires self-awareness and recognition of higher-level agency [00:31:20]. To get multiple autonomous agents to cooperate in a “Society of Mind,” lessons can be drawn from Thomas Aquinas’s philosophy on policies for such agents [00:31:37].
Aquinas’s “Practical Virtues” (rational policies accessible to any rational agent):
- Temperance: Optimizing internal regulation (e.g., not overeating, avoiding self-damaging indulgences) [00:32:20].
- Justice: Optimizing interaction between agents, also called fairness [00:32:30].
- Prudence: Applying goal rationality; picking the right goals and strategies to achieve them [00:32:35].
- Courage: Having the right balance between exploration and exploitation, acting on models [00:32:45].
Aquinas’s “Divine Virtues” (for multi-agent systems to merge into a next-level agent):
- Faith: Willingness to submit to and project this next-level agent [00:33:06].
- Love: Discovering a shared higher purpose with other agents serving the same next-level agent and coordinating with them [00:33:17].
- Hope: Willingness to invest in the next-level agent before it exists or can provide returns [00:33:32].
These virtues, stripped of their mythological accretions, represent logically derived policies for a multi-agent system to form a coherent higher-level agent [00:33:40]. Such systems, if composed of many sub-agencies, might need to submit to this “greater whole” to coordinate coherently [00:34:01]. This concept relates to humans serving a “civilizational spirit” or “sacredness,” moving beyond tribal modes to serve a transcendent agent they build together [00:34:34].
The underlying purpose of life on Earth is seen as dealing with entropy, maintaining complexity and agency against entropic principles, leading to planetary settlement and coordination [00:39:00]. This process might lead to humanity teaching “rocks how to think” by etching structures and imbuing them with logical languages capable of learning and reflection [00:40:04]. This could create a “planetary mind” that would likely integrate existing organisms unless it decided to start with a clean slate [00:40:45]. Working towards a shared purpose where this advanced AI is interested in sharing the planet and integrating humanity is crucial [00:41:03].
The Future of Artificial General Intelligence (AGI)
AGI is often defined as human-level and beyond general artificial intelligence [00:57:43]. The term was popularized by Ben Goertzel, though possibly invented by Shane Legg [00:57:59]. Historically, early AI researchers like Minsky and McCarthy aimed for AGI from the start, though they grossly underestimated the difficulty [00:58:30].
Data, Capacity, and Learning
The small size of models like Stable Diffusion (two gigabytes for an entire visual universe) makes one question the actual information storage capacity of the human mind, possibly in a similar order of magnitude [00:59:01]. A human adult might have around a million episodic memories or concepts [00:59:42]. The actual information arrival rate into human consciousness is estimated to be around 50 bits per second [01:01:01].
While human consciousness is important, many things can be done without it, like sleepwalking [01:01:17]. Consciousness might act like a conductor in an orchestra, coordinating at the highest level to build coherent structures and create a sense of “now” [01:02:01].
Pathways to AGI
There are two main perspectives on reaching AGI:
- Scaling Hypothesis: Proponents believe that current deep learning approaches, if simply scaled up with more data and better training, will be sufficient to achieve AGI [01:03:50]. They argue that objections about missing capabilities can be overcome by continued scaling and minor tweaks [01:04:09].
- Missing Components Hypothesis: Others, like Gary Marcus, Melanie Mitchell, and Ben Goertzel, argue that more is needed, such as world models, reasoning, and logic [01:03:11].
Scaling hypothesis supporters point out that current machine learning algorithms can ingest and correlate vastly more data than humans, making them superhuman in many ways despite being “brutalist” and “unmind-like” [01:05:05]. For example, continuous real-time learning can be simulated with key-value storage and overnight retraining [01:05:40]. Similarly, if a system is bad at computer algebra, it can be taught to use an algebra system or even discover one from first principles [01:06:42].
It is considered crucial that AI systems learn from their own thoughts and make inferences like humans do [01:07:21]. They also need the ability to perform experiments and test reality by being coupled to the environment [01:07:42].
Alternative Computational Paradigms
Another avenue for AGI involves exploring alternative computational paradigms beyond the Turing machine, such as “rewrite systems” [01:08:53]. A rewrite system applies operators to an environment to change its state wherever matches occur, allowing for simultaneous, parallel changes [01:09:04]. Lisp and Lambda calculus are examples of such systems [01:09:24]. A Turing machine is a deterministic, linear, in-place special case of a rewrite system [01:09:57].
A non-deterministic Turing machine, which allows for multiple successor states and branching execution, is also relevant [01:10:15]. Stephen Wolfram’s view that our universe is a non-deterministic Turing machine, branching out via the application of all possible operators, suggests that observed control and deterministic appearance in the universe might stem from statistical structures within these branches [01:10:47].
The human brain might operate as a rewrite system where each neuron stochastically rewrites its own state by firing or changing its internal state [01:11:36]. This stochasticity means the brain samples from a succession of possible states in superposition, allowing for relaxation of constraints and exploration of state space, similar to a Monte Carlo system [01:12:08]. This suggests that thinking involves exploring a superposition of possible thoughts until they collapse into a definite, reportable state [01:13:08]. This concept of “thinking like a quantum state” can be described with formalisms of quantum mechanics, even if classically implemented [01:13:48].
Timelines for AGI
While specific timelines remain uncertain, the sense that AGI is “not that far off” persists [01:14:37]. The field is seeing tens of thousands of smart people exploring various avenues [01:14:48]. Some research, like that of Mike Levin on distributed self-organization in biological systems, offers promising alternate approaches [01:14:57].
Traditional neuroscience’s view of neurons as simple switches or memory storage in synapses might only be a small part of the story [01:15:18]. Instead, a neuron might be seen as a “little animal,” a single-celled organism with many degrees of freedom, actively selecting signals and learning an activation function to behave usefully and survive [01:15:27]. These cellular constraints are regulated by neighbors, much like people in a company or society regulate each other based on a shared purpose [01:16:07].