[ad_1]
Soon after Google Gemini spit out ‘historically inaccurate’ images, the company swung into action, suspending the image generation capabilities of the AI chatbot. While CEO Sundar Pichai called it completely unacceptable, Google co-founder Sergey Brin said that the company “definitely messed up” with the image generation capabilities.
At that time, Prabhakar Raghavan, Google’s senior vice president, said that Gemini’s image generation feature was built on top of an AI model called Imagen 2. He said that since the users come from all over the world, the model was trained to include a diverse range of people, and not just one type of ethnicity (or any other characteristic).
The model apparently took this training a bit too over the top when it returned ‘people of colour’ when asked to create images of ‘Nazi people in Germany’ and Founding Fathers of the US. However, recent reports suggest that a lack of (enough) testing and rush led to this slip.
‘Code Red’ may be responsible
Citing sources at Google, a report by The Verge says that the “bad Gemini responses slipped through testing because everyone felt rushed to ship.” This makes sense because the company issued a Code Red soon after OpenAI launched ChatGPT.
Meanwhile, Brin’s recent comment – “We definitely messed up on the image generation, and I think it was mostly due to, just like, not thorough testing” – also imply that the launch was hurried. Furthermore, Google reportedly wanted to get the text-to-photo model feature out of the door faster as competitors had already made inroads on this front.
Lack of alignment within teams
Another reason discussed is the lack of alignment between Google DeepMind chief Demis Hassabis’s research team building the underlying models and Raghavan’s search organisation that is putting them into user-facing products. Pichai had already hinted at “structural changes” to address the issue.
At that time, Prabhakar Raghavan, Google’s senior vice president, said that Gemini’s image generation feature was built on top of an AI model called Imagen 2. He said that since the users come from all over the world, the model was trained to include a diverse range of people, and not just one type of ethnicity (or any other characteristic).
The model apparently took this training a bit too over the top when it returned ‘people of colour’ when asked to create images of ‘Nazi people in Germany’ and Founding Fathers of the US. However, recent reports suggest that a lack of (enough) testing and rush led to this slip.
‘Code Red’ may be responsible
Citing sources at Google, a report by The Verge says that the “bad Gemini responses slipped through testing because everyone felt rushed to ship.” This makes sense because the company issued a Code Red soon after OpenAI launched ChatGPT.
Meanwhile, Brin’s recent comment – “We definitely messed up on the image generation, and I think it was mostly due to, just like, not thorough testing” – also imply that the launch was hurried. Furthermore, Google reportedly wanted to get the text-to-photo model feature out of the door faster as competitors had already made inroads on this front.
Lack of alignment within teams
Another reason discussed is the lack of alignment between Google DeepMind chief Demis Hassabis’s research team building the underlying models and Raghavan’s search organisation that is putting them into user-facing products. Pichai had already hinted at “structural changes” to address the issue.
[ad_2]
Source link