From: mk_thisisit
Currently, AI systems are considered “very stupid in many ways” [00:00:02]. People are often fooled into considering them intelligent due to their ability to manipulate language effectively [00:00:09]. However, they lack fundamental human-like capabilities that limit their understanding and interaction with the world [00:05:35].
Core Deficiencies
Current AI systems are deficient in several key areas:
- Understanding the Physical World They do not comprehend the physical world [00:00:13], a crucial aspect that even young animals like cats and dogs grasp intuitively within months [00:14:27]. This contrasts sharply with their proficiency in language manipulation [00:05:42].
- Permanent Memory Unlike humans, they do not possess permanent memory [00:00:15].
- Reasoning and Planning They cannot genuinely reason or plan, which are key features of intelligent behavior [00:05:53]. Human-like reasoning involves internal mental simulation, anticipating consequences of actions, and hierarchical planning towards goals [00:28:11]. Current large language models employ primitive and computationally expensive methods like generating numerous token sequences and selecting the best one, which is not how humans think [00:27:14].
Unsolvable Problems for Current AI
The physical world is significantly more complex to understand than language [00:12:35]. While language is discrete (a sequence of symbols), making predictions uncertain but manageable through probability distribution [00:13:03], the continuous and high-dimensional nature of the physical world makes exact predictions impossible [00:47:52]. For example, predicting every detail in a video frame is an “unsolvable mathematical problem” in high-dimensional space [00:48:02].
Moravec’s Paradox
This challenge is exemplified by Moravec’s Paradox, observed by robotics expert Hans Moravec [00:15:01]. It notes that while computers excel at tasks difficult for humans, like playing chess or solving mathematical puzzles, they struggle with seemingly simple physical tasks like manipulating objects or jumping, which animals perform easily [00:15:15]. The reason is that discrete symbol manipulation is easy for computers, but the real world is too complicated [00:15:33]. The amount of information received through senses (sight, touch) is “absolutely huge” compared to language [00:16:01]. This explains why LLM chatbots can pass law exams and solve math problems, but we lack realistic robots or fully autonomous cars that learn to drive in just 20 hours like a human [00:16:16].
Limitations in Learning Paradigms
Machine learning has evolved through several paradigms:
- Supervised Learning [00:08:51]: The system is given correct answers during training (e.g., “This is a table” for an image) [00:09:00]. It’s effective for specific tasks like image recognition but relies on labeled data [00:09:29].
- Reinforcement Learning [00:09:51]: The system receives feedback (good or bad) without explicit answers, mimicking how humans learn from trial and error [00:09:55]. However, it’s “extremely ineffective” for real-world scenarios due to the need for millions of trials [00:10:17]. For instance, training a self-driving car this way would require thousands of crashes [00:10:42].
- Self-Supervised Learning [00:10:57]: This method, behind recent advances in natural language processing and chatbots, trains systems to capture the structure of input data (e.g., predicting missing words in text) [00:11:11]. While successful for language, it’s insufficient for understanding the physical world [00:12:30].
Challenges in AI Understanding of the Physical World
A major hurdle is the difficulty in training AI systems to understand complex sensory data, particularly sight [00:16:46]. A child’s visual data intake in the first four years of life (around 10^14 bytes) is comparable to the training data of the largest linguistic models, highlighting the vast amount of sensory information involved in learning about reality [00:18:34]. To achieve human-level AI, systems must be taught to understand the real world, not just text [00:18:50].
Limitations in Human-like Traits and Consciousness
While AI systems might develop emotions like excitement or joy related to goal prediction [00:06:30], negative emotions like anger or jealousy will not be permanently built into them [00:07:16]. The concept of consciousness itself is not clearly defined or measurable, making it difficult to determine if something is conscious [00:07:37]. Obsession with the nature of consciousness may stem from asking the wrong questions, similar to how historical misunderstandings about inverted retinal images were resolved by re-evaluating the problem [00:23:13].
Current State of Robotics
Despite significant progress in AI, autonomous cars and advanced robots still face major limitations [00:31:59]. Existing industrial robots are only effective in highly controlled environments where objects are always in place [00:31:43]. The reason is not a lack of physical abilities, but that they are “just not smart enough to deal with the real world” [00:34:06]. Claims of imminent level five autonomy for self-driving cars, like those made by Elon Musk, have been repeatedly proven false [00:32:38]. The coming decade is anticipated to be the “decade of robotics” [00:42:44] due to expected AI advances that will enable more versatile robots [00:34:39].
Data Resources and Future Prospects
While some believe global resources for training AI models have run out, much textual knowledge remains undigitized or is not publicly available (e.g., medical, cultural, historical data) [00:21:56]. Therefore, there is still “a lot of data” to leverage [00:21:56]. The challenge is not necessarily data quantity, but how AI systems learn from it to understand complex sensory input [00:16:46].