Links for 2023-05-01
Prompting Unlocks Theory-of-Mind in LLMs: “GPT-4 performed best in zero-shot settings, reaching nearly 80% ToM accuracy, but still fell short of the 87% human accuracy on the test set. However, when supplied with prompts for in-context learning, all RLHF-trained LLMs exceeded 80% ToM accuracy, with GPT-4 reaching 100%.” https://arxiv.org/abs/2304.11490
ChatGPT outperformed doctors at answering patients' online questions https://gizmodo.com/chatgpt-ai-doctor-patients-reddit-questions-answer-1850384628
GPT Code interpreter: ChatGPT for data science https://twitter.com/emollick/status/1652170706312896512
“Neural Systematic Binder...learns to automatically decompose common factors shared across objects and thereby support systematic generalization to unseen composition.” https://sites.google.com/view/neural-systematic-binder
Meta-Learned Models of Cognition https://arxiv.org/abs/2304.06729
Multi-Party Chat: Conversational Agents in Group Settings with Humans and Models https://arxiv.org/abs/2304.13835
What's going on when neural networks "grok" or rapidly gain qualitatively new abilities during training? https://twitter.com/s_r_constantin/status/1651998993407258625
A diffusion model that generates proteins with desired secondary structures. https://news.mit.edu/2023/ai-system-can-generate-novel-proteins-structural-design-0420
Concerns about AI that go beyond political correctness have entered mainstream discourse: "We are advancing the capabilities of these [AI] systems faster than... our ability to control them" UK Government tech adviser Matt Clifford says it is vital to invest in technology needed to "make AI safe" https://twitter.com/BBCPolitics/status/1652606279364358146
Discovery of carbon-capturing organism in hot springs could lead to efficient way of absorbing climate-heating gas https://www.theguardian.com/environment/2023/apr/19/volcanic-microbe-eats-co2-astonishingly-quickly-say-scientists
A new crop of biotech startups want to revolutionize human reproduction. [The New Yorker] https://archive.is/jovC2
A neuromorphic visual sensor can recognise moving objects and predict their path https://www.aalto.fi/en/news/a-neuromorphic-visual-sensor-can-recognise-moving-objects-and-predict-their-path
Broadcom releases chip for wiring together AI supercomputers https://www.reuters.com/technology/broadcom-releases-chip-wiring-together-ai-supercomputers-2023-04-18/
Researchers have mimicked the human hippocampus to improve autonomous navigation https://www.humanbrainproject.eu/en/follow-hbp/news/2023/04/17/brain-model-learns-drive/