Artificial Intelligence and its future implications

From: mk_thisisit

Artificial intelligence (AI) is projected to become the main topic of conversation on Earth, leading to the creation of machines demonstrably smarter than humans [00:00:00]. While AI is expected to solve many existing problems, it will also create new ones [00:00:24].

Defining AI and AGI

Wojtek Zaręba, a co-creator of ChatGPT and co-founder of OpenAI, emphasizes the importance of understanding the developmental levels of AI and AGI (Artificial General Intelligence) [00:00:33]. He highlights that the term “only statistics” for AI is an oversimplification [00:01:16]. Models are capable of solving problems they have never encountered before, demonstrating intelligent generalization on diverse data, rather than just remembering exact past instances [00:01:34]. The human brain itself could be considered “statistics,” and if so, it’s “unbelievable” and “magic” [00:02:07]. While there is empirical understanding of neural networks’ results, a deep theoretical understanding of why they work is still lacking [00:02:24].

For example, a model trained on black background numbers can still recognize numbers even if a single black pixel is changed to white, despite the input being “completely different” from its training data [00:03:07]. This contrasts with traditional models like decision trees, which “fall apart” when encountering data far from their training set [00:04:26]. Transformer models, however, do not exhibit this behavior [00:04:45].

Limits of AI Understanding: Context and Senses

A current limitation for AI is its lack of contextual understanding, particularly cultural context [00:05:06]. An example cited is an autonomous taxi in San Francisco’s Chinatown that violated a “sacrum zone” during a procession, not because it failed to detect people, but because it didn’t understand the cultural significance of maintaining distance [00:05:35]. Current self-driving cars are primarily trained on collision avoidance [00:06:47].

To address this, future models should be trained on multimodal data, including images, text, video, and audio, to achieve a deeper understanding [00:06:56]. While AI can infer without seeing every single case, it currently lacks “senses” to acquire data directly from real-time experience like humans [00:07:40].

Differences between Human Brain and AI Training

There are fundamental differences in how the human brain and AI models are trained:

• Training and Testing: For computers, training (collecting vast data) and testing (evaluating the model) are separate processes [00:09:06]. The human brain, conversely, integrates these processes [00:09:42].
• Data Source: Humans train themselves based on their own experiences (e.g., falling and learning to be careful on stairs) [00:10:03]. Computers, however, are trained on data collected by other people [00:10:21]. While AI can learn from its own experience through reinforcement learning (like a bot playing Dota or Starcraft), this isn’t yet the primary mode for broader applications [00:10:51].

AI’s Ability to Discover New Knowledge

AI’s development involves two main processes for inputting knowledge:

1. Large-scale Data Training: Models learn to predict the next word from vast internet data, effectively learning what humans have already understood [00:13:00].
2. Reinforcement Learning: Models are rewarded for correct behavior in limited domains [00:13:41]. In games like Go, AI has discovered moves that humans, over thousands of years, could not [00:13:52]. This suggests AI can potentially understand things beyond human comprehension [00:13:27].

OpenAI’s Five Levels of AI Development

OpenAI categorizes AI development into five levels:

Level 1: Conversational AI

• Models can have a conversation, effectively passing the Turing test [00:14:56].
• Current language models are already at a level where it’s difficult for humans to distinguish between a human and a computer during a conversation [00:15:19].

Level 2: Reasoning AI

• Models capable of solving problems requiring approximately 10 minutes of human reasoning [00:15:50].
• This involves deep understanding of tasks in fields like mathematics, physics, biology, and computer science [00:16:24].

Level 3: Agents

• Models, referred to as “agents,” can perform longer, multi-step tasks in the real world over hours or days [00:17:02].
• An example is an agent that can build a website from a simple request, handling domain registration, coding, and design choices, integrating information, and writing error-free code [00:17:15].
• Current models like ChatGPT get “lost” quickly when trying to perform complex actions in the world [00:18:00].

Level 4: Scientist

• An AI capable of functioning as a “scientist,” spending months thinking about problems, analyzing them from different perspectives, and challenging fundamental assumptions that human scientists might take for granted (like Einstein questioning the constancy of time) [00:19:02].
• This level could lead to new discoveries [00:20:05].

Level 5: Running Organizations

• The final level where AI is competent enough to autonomously run entire organizations, managing thousands of people, planning, analyzing, and making decisions [00:20:17].

Timeline for Superhuman AI

Wojtek Zaręba believes that achieving the fifth level of AI, which is beyond the human level, is hard to predict but will likely happen in less than 10 years [00:23:19]. He draws an analogy to the accelerating pace of evolution and human technological development:

• Single-cell organisms to multicellular: over a billion years [00:23:53].
• Homo sapiens: 200,000 years [00:24:19].
• First cities: 20,000 years ago [00:24:24].
• Industrialization: 200-300 years ago [00:24:30].
• Computers: 60-70 years ago [00:24:33].
• Internet: 30 years ago [00:24:38]. This historical trend suggests that each subsequent stage of development shortens dramatically [00:24:42]. Factors like regulation and integration with society will also influence the timeline [00:24:59].

The “Defective” Nature of AI (Efficiency)

Comparing the human brain’s energy consumption (20 watts) to a language model’s (10^9 watts) reveals a massive disparity in efficiency [00:27:45]. However, this doesn’t necessarily mean AI is “defective” [00:28:05]. An analogy is made to birds and airplanes: birds are efficient and acrobatic, but airplanes, though heavy, can transport hundreds of people across oceans [00:28:12]. Similarly, future AI models may be highly capable despite their current energy demands [00:28:46].

The human brain’s efficiency is partly attributed to information embedded in DNA, which itself is the result of billions of years of “training” through evolution [00:29:43]. Evolution, operating over billions of years and across the entire Earth, has used immense “computing power” to discover intelligence multiple times (humans, elephants, ravens, dolphins, trenches) [00:30:44].

Neural networks currently have a two-stage learning process:

1. Massive Data Training: A huge amount of data and computing power is needed initially [00:32:04].
2. Rapid In-Conversation Learning: After this initial training, models can learn very quickly within a single conversation, adapting to new information or even grammar rules [00:32:26]. The goal is to reach a point where AI can solve novel, global problems like global warming with minimal new data [00:33:26].

Consciousness in AI

The question of AI consciousness is a complex philosophical and potentially technical one [00:26:51].

• The Brain as a Simulator: The human brain receives all information as electrical signals (“bits”) [00:34:35]. It creates an “immersive image of the world” as a simulation, and “consciousness is our experience of this simulation” [00:35:10].
• Philosophical Zombie: This thought experiment considers a being that behaves exactly like a conscious person but lacks internal subjective experience [00:35:30].
• Self-Awareness in Simulations: Wojtek Zaręba hypothesizes that within a simulation, a model might eventually “simulate its own” participation in changing reality [00:37:42]. This could lead to a moment of “self-awareness” where the agent realizes its own existence within the simulation, which might be a step towards consciousness [00:38:55].
• Testing for Consciousness:
  1. Eliminating “Consciousness” Data: Train a model on data where any mention of consciousness is removed. If the model spontaneously mentions having such experiences, it could indicate consciousness [00:40:31].
  2. Brain Connection: Connecting AI to a human brain to see if it expands human consciousness. However, this is inconclusive, as an unconscious substance (like a psychedelic) can also expand human consciousness [00:41:25].

If consciousness can be understood, it can be built [00:39:55]. The speaker also believes AI models could act as a “power bank for the brain” in the future [00:42:15].

AI Hallucinations

AI “hallucinations” (generating confident but incorrect information) stem from the post-training stage [00:43:04]. Humans training the AI tend to reward answers they “like” more than answers that admit ignorance [00:44:05]. This incentivizes the model to always provide an answer, even if it has to guess [00:44:35].

To combat this, models could be trained to:

• Express probability or certainty about their answers [00:46:00].
• Understand what they know versus what they don’t know [00:46:24]. Currently, post-training leads to models becoming “overconfident” even on topics they know little about [00:47:47].

Immersive Representation of Pain

The possibility of creating an “immersive representation of pain” inside a language model is explored [00:48:06].

• Pretending vs. Feeling: While a model can “imagine” being a patient with incurable pain and describe it, the question remains whether it is truly feeling or merely simulating [00:48:21].
• Sensory Input: If a model were connected to infrastructure with sensors, allowing it to receive deeper sensory input, it might be able to “feel” pain [00:49:19]. From the AI’s perspective, there may be no difference between existing in physical reality or a virtual one, as both provide “bits of information” [00:50:12].
• Continuous Experience: Current neural networks are “a bit like Memento” – they remember within a single conversation, but then the memory disappears [00:51:22]. For a deeper sense of self and continuity, models would need a much longer context or new algorithms for continuous learning based on their own experience [00:51:38].

Phases of AI Development (Societal Impact)

The development of AI is described in three phases:

1. Product Phase: Currently, companies are creating diverse AI products, integrating AI into most software [00:55:18].
2. Geopolitical Phase: Countries will increasingly recognize that investment in AI is crucial for their geopolitical position and possibly for the survival of the human species [00:55:29]. It’s predicted that within 1.5 years (around 2025-2026), AI will be the main topic of global conversation [00:56:32]. This phase will see many “agents” doing different things, impacting the labor market [00:56:52].
   • Poland’s government AI fund of 40 million zlotys is considered “very little” given the potential and the country’s smart programmers [00:57:40].
3. Superintelligence Phase: Machines will become definitively smarter than humans [00:58:28]. At this stage, international cooperation might become paramount for survival [00:58:41]. An AI at this level could create new chips, deeply understand scientific literature, invent new things, and run virtual companies [00:59:26].

Analogy to "Magic"

Inventions that were once considered magic are now commonplace (e.g., humans moving faster than any animal in cars, a cell phone providing access to all human knowledge and global communication) [01:00:01]. Similarly, future AI technologies will likely achieve feats that seem like “magic” today [01:00:52].

Wojtek Zaręba acknowledges that humanity could be a “limitation,” trying to resist AI [01:01:01]. However, he states that no one truly knows how this technology will develop, with reactions varying from rejection to acceptance [01:01:14].

Risks and Challenges of AI

Wojtek Zaręba classifies the problems associated with AI:

1. Misuse (Malicious Use - Misi):

   • Deepfakes and Fake News: AI can be used to create deceptive content [01:17:21].
   • Military Applications: Almost certainly, AI will be used for military purposes [01:17:27].
   • Pandemic Creation: The number of people capable of synthesizing a pandemic virus (currently ~30,000 scientists) could drastically increase if AI provides access to intelligence for such tasks [01:17:41].
   • Hacking, Chemical, and Nuclear Weapons: AI can assist in these areas [01:18:55].
   • Persuasion/Convincing People: AI models could become highly effective at manipulating large numbers of people [01:20:43].

2. AI Race: Danger arising from organizations competing fiercely to create AI [01:22:43].

3. Accidents: Problems resulting from inattention or misaligned goals [01:22:58].

OpenAI uses a framework called PMR (Progress Metrics for Responsible Development) to assess and mitigate risks [01:20:21]. This involves understanding the level of risk (low, medium, high, critical) in categories like biological, chemical, nuclear, cybersecurity, and persuasion, and working to reduce high-risk levels [01:20:27]. The immediate focus for the next year or two is on mitigating misuse [01:23:38]. A key challenge is ensuring models continue to “listen to us” and “behave in a way as if we expected it” [01:24:30].

Reflections on ChatGPT Launch

The release of ChatGPT in November 2022 was unexpected in its scale of reaction [01:24:49]. Within days, it started growing to millions, then tens and hundreds of millions of users, leading to widespread recognition that it had “changed the world” [01:25:41].

Personal Journey & Philosophy

Wojtek Zaręba, a co-founder of OpenAI, hails from Kluczbork, Poland [01:27:55]. He connected with the founders of OpenAI due to his strong publications in AI and shared awareness of the technology’s potential [01:28:35]. He witnessed the shift in perception of neural networks from being considered ineffective to achieving significant breakthroughs, particularly after the ImageNet competition around 10 years ago, which demonstrated their superior results [01:30:41].

His contributions at OpenAI include leading robotics, where they developed a network that could solve a Rubik’s cube with a hand, requiring complex learning rather than direct programming [01:34:18]. He also had a significant role in developing models for code generation, now used by millions of people (like GitHub Copilot) [01:37:12].

He defines himself primarily as a “scientist,” but also a combination of programmer, futurologist, and philosopher of technology [01:39:32]. He attributes his sensitivity and happiness to close relationships and experiencing love in his life, including from his mother and teachers [01:41:19]. He has also accepted the public attention that comes with his work, finding value in inspiring others [01:42:32]. He financially supports his high school in Kluczbork, aiming to provide opportunities for smart students who might not otherwise have them [01:44:13].

Vision for the Future

Wojtek Zaręba believes AI presents “incredible opportunities for humanity” [01:47:03]. He dreams of a future where everyone is happy and lives in peace, acknowledging that reality is more complicated [01:48:01]. His personal goals include maintaining passion, positive realism, and daily excitement in his work, and enjoying deep satisfaction from his personal life [01:48:42]. He is optimistic that through the development of this technology, humanity can avoid fundamental problems and achieve a “great” future [01:50:26].