A photo of a used potato press gave DeepMind’s Flamingo model a chance to show how an AI system might combine image analysis with conversation. The demonstration also raised a larger question: if chatbot interfaces change how people search, what would that mean for Google?
From an image to a conversation
DeepMind introduced Flamingo in May as a multimodal AI model, combining image processing with language. During training, it processed images alongside text, developing a basic ability to identify what images contain.
That combination lets Flamingo respond to questions about an image through a dialog interface. A person can ask for more detail or background information, making the system’s image analysis interactive rather than a one-time description.
In a demonstration shared on Twitter by Oriol Vinyals, a DeepMind research director and deep learning lead, the starting image showed his father’s used potato press. Flamingo first identified it as an ice crusher. After two rounds of user feedback, it reached the correct answer and then explained how a potato press works.
The exchange was a small but useful illustration of how feedback can guide a model toward a better interpretation. It also made the limits visible: the first answer was wrong, and the user had to help steer the system.
Understanding a picture means more than naming objects
Another Flamingo example focused on a photograph whose humor depends on understanding what people are doing together. President Obama secretly puts his foot on a scale so that the person being weighed appears heavier, while people in the scene laugh.
Recognizing a scale is one thing; identifying the hidden action and understanding why the scene is funny requires connecting details and context. The article recalls that Tesla’s former AI chief, Andrei Karpathy, had written about a decade earlier that AI was still “very, very far from” understanding the image.
Roman Ring demonstrated Flamingo’s response to the photograph in May 2022. Karpathy called that demonstration “not exactly convincing, but cute,” pointing to incorrect or imprecise answers and questions that supplied strong guidance. He said it was not clear the model understood the joke, though it was “clearly on track to.”
That distinction matters when judging a conversational AI demo. A fluent exchange can look like understanding, but the model may still rely on clues supplied by the person asking questions. The examples show a system making progress while also showing why a few impressive answers do not settle how reliably it interprets images.
ChatGPT sharpened the search question
At the time of the article, some commentators were describing ChatGPT as a threat to Google’s core search business. The New York Times had reported that Google issued “Code Red” over ChatGPT. Yet the article also noted that ChatGPT could produce fictitious or generic text and help with simple code, while struggling to provide reliable answers.
That leaves several challenges for any chatbot positioned as a search alternative: answers need to be dependable, current, and transparent about their sources. The article also raises copyright and scaling as unresolved issues. An engaging interface, by itself, does not answer those questions.
DeepMind had Sparrow in development, while Google was likely working on Assistant 2.0 with LaMDA, according to the article. Alongside Flamingo, those projects suggested that Google had technical options for responding to chat-based AI. Yann LeCun argued that Google was better positioned to bring language technology to search than a language-model company was to build a search engine, and that Google had been doing this work for years.
The business model is part of the challenge
Technical capability is only one part of the competitive picture. Google earns money from advertising, and the article said about $39 billion out of about $70 billion in the third quarter of 2022 came from ads in Google Search.
If people shift toward chatbot search, the replacement would need to support revenue at a level comparable to Google Search. Google might retain a strong position in online search and still face financial pressure if it had to change how search works without replacing the advertising income.
Microsoft’s support for OpenAI could also shape the stakes: the article argued Microsoft had more to gain in internet search and could therefore take the risk. Flamingo’s demos did not show how Google would answer that business challenge. They did, however, give a concrete preview of multimodal conversation—and of the gap between an impressive demonstration and a dependable search product.