As of 2021, neural networks still struggle with notions that are obvious to babies.

Judge as of 2021-04-13. True: published research or benchmarks from around 2021 show that state-of-the-art neural networks performed poorly on basic concepts that infants understand, such as object permanence, intuitive physics, causality or numerosity (e.g. results on IntPhys, infant-inspired AI benchmarks). Mostly true: the research shows widespread failures, but some systems handled certain such notions reasonably well. Mixed: the evidence was evenly split, with neural networks doing well on about half of the commonly cited infant-level notions. Mostly false: neural networks broadly handled such notions by 2021, with only narrow failures. False: the evidence showed neural networks reliably mastered infant-level concepts.

Set before the research began.

1. E1 Baby Intuitions Benchmark (BIB): Discerning the goals, preferences, and actions of others arXiv

   “Nevertheless, recently proposed, deep-learning-based agency reasoning models fail to show infant-like reasoning, leaving BIB an open challenge.”

E1 reports that recently proposed deep-learning models for reasoning about agents fail to show infant-like reasoning, but it is undated, so it cannot be placed relative to 2021-04-13. It also covers only one area (reasoning about agents) and does not show how neural networks did on the other infant-level notions named in the criteria, so it does not settle the claim.