Dall-E and ChatGPT: Google prepares its answers

A little overwhelmed by the buzz around generative AIs such as Dall-E and ChatGPT, Google is preparing a counter-attack based on the research carried out by its researchers and those of DeepMind with Muse and Sparrow.

The dazzling progress made by OpenAI’s generative models, including the Dall-E 2 image-generating AI and the ChatGPT conversational AI, has amazed the world and, apparently, taken Google by surprise. The search leader immediately perceived ChatGPT as a dangerous competing technology and triggered a “Code Red” to quickly bring its own technological advances in these areas.

On the one hand, Google AI researchers have just published a new research paper around Muse, a new generative image-to-text AI. Unlike Dall-E or Stable Diffusion, two AIs based on so-called diffusion techniques, Muse relies on a new MGT (Masked Generative Transformers) model capable of producing photorealistic images and whose results could prove to be more original. .

Above all, Muse promises to be significantly faster than current image generation AIs such as Dall-E, Imagen or Google’s other generative AI, Left, using regressive models. With a single TPUv4 accelerator, it takes half a second for Muse to generate an image of 256×256 pixels compared to 9 seconds for Imagen for example. According to the researchers, Muse would also be three times faster than Stable Diffusion.

At the same time, DeepMind (one of Alphabet’s startups, acquired 9 years ago and largely financed by Google) has confirmed that it is working on an artificial intelligence competitor to ChatGPT and known as Sparrow. Demis Hassabis, the CEO of DeepMind, has thus confirmed to Time prepared a “Beta” version in the course of the year 2023.

While OpenAI has caught the attention of the general public, DeepMind has attracted the attention of specialists, in particular with its code generation AI (Alpha code) and more with its AI AlphaFold capable of predicting the 3D structures of all known proteins.
His Sparrow looks very promising. But DeepMind does not want to go too fast and wants to avoid certain pitfalls highlighted by ChatGPT. Typically, Sparrow should be able to cite the sources that inspire his “answers”. Just like ChatGPT, Sparrow does not try to impersonate a human and multiplies the rules of control to avoid being rude, rude, discriminatory, or simply giving inappropriate answers. However, much like ChatGPT, Sparrow seems to have a tangential approach to truth and can also fabricate or give off answers.

Also read:

SantaCoder, an AI coder that avoids violating copyrights

Microsoft wants to strengthen its OpenAI partnership and integrate ChatGPT into Microsoft 365

Microsoft wants to integrate ChatGPT with Bing to compete with Google

AI at the heart of the future Windows 12?