LLMs Are Short-Circuiting. What Comes Next? (www.forbes.com)

A growing chorus of researchers and critics are declaring that the current era of large language models (LLMs) has hit its practical and conceptual limits. The piece argues LLMs’ visible failures—hallucinations, brittle reasoning advances, legal battles over training data and soaring infrastructure costs—signal diminishing returns from the “more data + more compute = AGI” thesis. High‑profile voices (Gary Marcus, Mounir Shita, Marc Fawzi) contend investors and engineers should stop treating scale alone as the pathway to general intelligence; demand for enterprise models may be softening as models prove costly, fragile, and insufficiently grounded. Technically, the critique centers on three points: LLMs are offline statistical priors that model language about the world, not goal‑conditioned causal models that act in it; scaling laws imply exponentially rising compute/data needs (and massive data‑center CAPEX); and fluency does not equal structured understanding. Alternatives suggested include cognitive‑science and physics‑first approaches, systems that combine four coherent layers—statistics, structure, inference, objectives—plus grounding via sensors, simulators, tools and verifiers, and models that update from interaction and causal feedback. In short, the next phase likely shifts from ever‑larger sequence predictors toward integrated, causal, goal‑directed architectures that can act, test hypotheses, and adapt in real time.