The Limits of Data: Why AI Needs Knowledge, Not Just More Data
Modern AI chatbots demonstrate remarkable abilities, from generating creative content to tackling complex tasks. However, these large language models (LLMs) are prone to fundamental errors – failing at basic math, misunderstanding simple rules, and exhibiting what’s been termed “jagged intelligence.” This instability isn’t merely a quirk; it’s a critical limitation that fuels skepticism about the current AI boom.
The Problem with “Jagged Intelligence”
The core issue lies in how current AI models learn. They don’t possess inherent understanding of how the world works. Instead, they infer patterns from massive datasets. This reliance on probabilistic inference means they essentially “guess,” and those guesses can range from harmlessly amusing to dangerously incorrect. Enterprises are understandably hesitant to entrust critical operations – like supply chain management, human resources, or financial processes – to AI systems that can exhibit such unpredictable behavior.
From Pattern Recognition to Codified Knowledge
Human learning provides a useful analogy. Infants initially learn through pattern recognition, but this is quickly supplemented by explicit, codified knowledge – the rules and principles taught to them. From basic arithmetic to operating machinery, humans rely on established knowledge to learn efficiently and avoid mistakes. Current AI lacks this foundation.
Recent advancements demonstrate the power of integrating explicit knowledge. For example, researchers have improved LLMs’ mathematical abilities by incorporating actual mathematical rules, enabling them to solve complex problems like those found in math Olympiads. [1]
Why More Data Isn’t the Answer
Simply feeding AI models more data, even diverse data like video, won’t resolve the problem of jagged intelligence. Because these models are fundamentally probabilistic, they will continue to make errors, regardless of the dataset size. The solution isn’t more data; it’s knowledge – clearly defined concepts, constraints, rules, and relationships that ground AI behavior in reality.
Building a Public Knowledge Base
To equip AI with a human-like stock of knowledge, a publicly accessible database of formal knowledge spanning various disciplines is needed. This resource could be accessed by developers and AI agents to provide verifiable insights, from everyday tasks to complex regulations. Such a system would reduce errors and potentially require less data and energy for training, though further research is needed to confirm these claims.
Transparency and Control
Unlike the “black box” nature of current AI models, where knowledge is embedded within billions of parameters, a formalized knowledge base would be transparent, examinable, and controllable. Regulators could verify a model’s knowledge, and users could be assured of its reliability.
The Role of AI in Knowledge Modeling
The creation of such a knowledge resource isn’t a new ambition in AI, but renewed efforts are warranted. Just as algorithms accelerate protein modeling in biology, generative AI can aid in knowledge modeling, accelerating the process of translating human expertise into a machine-readable format. The emergence of companies like Scale AI, which specializes in high-quality data for AI training, signals the growth of a profession dedicated to this task. [4]
Beyond the AI Bubble
While AI models continue to improve with more data and advanced algorithms, overcoming jagged intelligence requires a fundamental shift in how AI learns and relates to the world. Data-driven algorithms enabled us to communicate with machines, but knowledge – not data – is the key to unlocking the full potential of AI and ensuring its long-term viability. [3]