A Google DeepMind paper titled “LLMs can’t jump,” published by Tom Zahavy, discusses an obvious truth that AI software lacks a physical body, which, as a limitation, introduces a cognitive barrier. The paper also states that Large Language Models cannot replace human beings and their creative genius, with software-based intelligence only mastering one distinct type of inference.
The research states that the three types of inference are induction, deduction, and abduction. Thanks to statistical pattern recognition and massive data compression, AI has successfully mastered induction, with systems like AlphaProof quickly mastering deduction using derived logical proofs from established rules. Unfortunately, the impenetrable ceiling arrives with abduction, which is the ability to generate explanatory hypotheses when faced with scarce data.
German computer scientist Jürgen Schmidhuber, most notably known for his work in artificial intelligence, claims that scientific discovery is just an advanced form of data compression. The DeepMind paper refutes this assessment, countering it with Albert Einstein’s theory of General Relativity. When Einstein developed his breakthroughs, visual data was extremely scarce, giving him little to no dataset to analyze or compress using statistical pattern matching.
Not being fazed by the lack of data, Einstein instead relied on “embodied thought experiments,” such as visualizing what a person would feel when falling inside a sealed elevator. As a result, new mathematical principles came into discovery just by employing physical intuition to connect sensory experiences. An LLM, in comparison, could seamlessly handle the complex mathematical deductions after being presented with Einstein’s initial findings, but the AI model could never invent those original premises on its own simply by analyzing existing text.
Current AI discovery frameworks, which include automated coding tools, remain limited to certain rulebooks and thresholds, rather than creating entirely new frameworks. Since LLMs lack physical grounding in the real world, it’s impossible to create intuitive models based on cause and effect. Also, expanding the AI infrastructure with data centers and compute won’t magically enable these LLMs to develop genius-level intellect capable of revolutionary scientific breakthroughs.
The DeepMind paper says that bridging logical calculation and true scientific invention will require future AI architectures to evolve beyond text and image processing and move towards physical multimodal world models. Giving AI systems the ability to run embodied simulations in virtual environments will allow them to build real-world physical intuition before generating equations or code. You can read the entire paper by clicking on the source link below, and let us know if you found anything compelling in the comments.
News Source: Tom Zahavy, Google DeepMind (LLMs can’t jump)
Follow Wccftech on Google to get more of our news coverage in your feeds.