Comment by observationist

There's a huge difference between demonstrating something with fuzzy accuracy and playing something off as if it's giving good, correct answers. An honest way to handle that would be to highlight where the bot got it wrong instead of running with the answer as if it was right.

Deception isn't always outright lying. This video was deceitful in form and content and presentation. Their product can't do what they're implying it can, and it was put together specifically to mislead people into thinking it was comparable in capabilities to gpt-4v and other competitor's tech.

Working for Google AI has to be infuriating. They're doing some of the most cutting edge research with some of the best and brightest minds in the field, but their shitty middle management and marketing people are doing things that undermine their credibility and make them look like untrustworthy fools. They're a year or more behind OpenAI and Anthropic, barely competitive with Meta, and they've spent billions of dollars more than any other two companies, with a trashcan fire for a tech demo.

It remains to be seen whether they can even outperform Mistral 7b or some of the smaller open source models, or if their benchmark numbers are all marketing hype.