Fran Litterio (@fpl9000.bsky.social)

Large reasoning models almost certainly can think The article argues against the notion that large reasoning models (LRMs) cannot think, countering Apple's claim that they merely perform pattern-matching. It refutes the idea that LRMs fail because they struggle with complex algorithmic tasks, drawing a parallel to human limitations. The author posits that LRMs almost certainly can think, defining thinking in terms of problem-solving and linking it to brain functions like problem representation, mental simulation, pattern matching, monitoring, and insight. Chain-of-thought reasoning in LRMs is compared to inner speech in humans, with LRMs demonstrating backtracking capabilities. The author states that next-token prediction, the basis of LRMs, is a powerful form of knowledge representation, aligning with the expressive power of natural language. A next-token predictor requires world knowledge and the ability to represent intermediate steps, which proves that it is thinking. Even though a perfect auto-complete would seem omniscient, LRMs meet the criteria to represent the thinking process. Open-source LRM benchmarks show successful problem-solving, demonstrating reasoning abilities, even outperforming untrained humans in certain instances. The author concludes that LRMs possess thinking capabilities by meeting the requirements to do so across the board. https://venturebeat.com/ai/large-reasoning-models-almost-certainly-can-think