AI currently lacks a way of building reusable skills.

Judged as of 2019-11-14. True: the research literature of that time shows that AI systems generally could not acquire skills that they could reliably reuse and combine across tasks, with transfer learning, meta-learning and hierarchical reinforcement learning still widely seen as unsolved for this purpose. Mostly true: this was the broad consensus, though there was some limited success (e.g. pretrained models fine-tuned for other tasks) that the claim glosses over. Mixed: substantial reuse already existed in some areas (e.g. ImageNet or BERT-style transfer) while it was clearly lacking in others, such as robotics and reasoning. Mostly false: methods for reusing skills were well established and widely used by 2019, and the claim turns real limits into a total absence. False: the evidence shows that reusable skill-building was a solved, standard capability by then.

Set before the research began.

1. E1 Do Better ImageNet Models Transfer Better? CVPR 2019 (Open Access), Simon Kornblith, Jonathon Shlens and Quoc V. Le, Google Brain

   “Together, our results show that ImageNet architectures generalize well across datasets, but ImageNet features are less general than previously suggested.”

2. E2 Do Better ImageNet Models Transfer Better? CVPR 2019 (Open Access), Simon Kornblith, Jonathon Shlens and Quoc V. Le, Google Brain

   “ImageNet pretraining accelerates convergence and improves performance on many datasets, but its value diminishes with greater training time, more training data, and greater divergence from ImageNet labels.”

3. E3 A Review of Robot Learning for Manipulation: Challenges, Representations, and Algorithms arXiv (Kroemer, Niekum, Konidaris)

   “However, despite access to these excellent generic machine learning methods, the challenges of learning robust and versatile manipulation skills are still far from being resolved.”

4. E4 AI Hasn’t Found Its Isaac Newton: Gary Marcus on Deep Learning Defects & ‘Frenemy’ Yann LeCun | Synced Synced · 15 February 2019 Own record

   “Almost everything that anybody has ever written in AI is brittle and narrow — it works for the circumstances for which it was built, and it can't adapt.”

A 2019 CVPR paper by independent authors found that ImageNet pretraining speeds up convergence and improves performance on many datasets, and that ImageNet architectures generalize well across datasets, though the features are less general than previously suggested (E1, E2). So substantial reuse of learned representations existed in computer vision. A robot-learning review from the same period says that learning robust and versatile manipulation skills was still far from resolved (E3). Under the criteria, real transfer in some areas alongside clear gaps in others such as robotics fits "mixed".