Hey, Google: Preliminary AI Presentations are the Coward’s Exit

Almost every major technological event today includes updates of artificial intelligence, often with a sheet of live demonstrations – and sometimes these demonstrations fail. But some companies avoid these pitfalls by pre -recording their basic presentations. And I call these movements cowardice.

Last year Made by Google The event, twins failed twice during a live demonstration. Although moments like this are undoubtedly disturbing to companies, they add a layer of authenticity that you do not receive with a pre -recorded main event. But unfortunately Google selected the pre -recorded route for Tuesday Android Show: I/O EditionS The format felt too placed and polished to my taste, and it took a sense of reality that came with alive, warts and all demonstrations.

During the Android show: I/O Edition, we saw a demonstration of Gemini’s makeup tips, helping someone find the time to take lunch in their busy schedule and give a summary of Jane Austin’s pride and prejudice. As these were pre -recorded interactions, twins dealt with the requests with aplomb – without hiccups or problems in the eyes. But the tests show that AI models routinely break things.

According to the AI ​​test site LivebenchGoogle’s Gemini 2.5 Pro Preview In general, about 79% of the time is correct. It’s not bad, but it’s not great. And despite this result, this twin model is still one of the best AI models that the site has been tested, losing only two other models: O3 High and O4 Model of Openai.

Of course, nothing is perfect, and the devices and software have errors. But if you give me a calculator and promise that it works continuously, but in fact it is wrong 20% ​​of the time, it feels like a big discrepancy.

As the twins were superior to most other AI models tested Livebench, there is a great chance of still using twins, even if the live demonstration stopped. But since Google has chosen as a super -thes demonstration, it’s hard for me to know what to believe.

Look, I understand why a company would like its product to work properly at its own event. But displaying AI tools that make mistakes feels more honest than behaving as the tool is perfect. These options are insufficient and this is good, but be honest with people about these shortcomings and show your new functions in action. Don’t sell me smoke and mirrors.

For more Google information, here’s what to know Android 16 and Material 3 expressive designS