May

1 events

Editors' summary

Google separates what it can build from what it will ship

On 23 May, Google announced Imagen. It claimed results ahead of DALL·E 2 and said in the same breath that it would not release the model, citing bias in the training data and the risk of misuse. Only a curated set of demonstration images was shown.

Drawing a line between what it could build and what it would ship marked where Google stood at this point.

On the 24th, researchers at the University of Tokyo and Google published a paper showing that appending a single line — "Let's think step by step" — was enough to make a model work through a problem in stages. Google's January paper on chain of thought had relied on worked examples; this one needed none.

This block is written by the editors. It is kept separate from the sourced record below.

Record1 events

23
ResearchGoogle
Google unveils Imagen — but does not release it
Google published Imagen in the paper "Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding," combining text understanding from large Transformer language models with high-fidelity image generation from diffusion models. It set a then-best FID score of 7.27 on COCO without ever training on it, and human raters judged its samples on par with real COCO data for image-text alignment. It came six weeks after OpenAI put DALL-E 2 into limited release, but Google declined to make Imagen publicly available, citing concerns about harmful content and the reproduction of social bias. Holding back a product while leading on the research was characteristic of Google in this period.

↑ 2022

Google unveils Imagen — but does not release it