Recently, Google CEO Sundar Pichai released the artificial intelligence (AI) large language model "Gemini", but it has been suspected of "fake" only in the past day, but Google has denied this. Let's take a look!
It is understood that this model is a native multimodal large language model, including three orders of magnitude: Gemini Ultra, which is the most capable, Gemini Pro, which is suitable for multitasking, and Gemini Nano, which is suitable for specific tasks and devices. Google also released a 6-minute demo** on the same day that it released Gemini, and it was launched with ChatGPT 40 as a comparison, I want to use this to reflect "more powerful than ChatGPT". Musk also commented below, "(Gemini) is impressive".
In this 6-minute demo**, Gemini showcases its powerful multimodal capabilities, such as spoken conversation cues combined with image recognition. In a demo where a man does slow motion to lean back and dodge, Gemini immediately replies:This is the scene from the performance of "Bullet Time" in The Matrix.
In another demo, Gemini was able to "quickly" infer that the drawing was a crab based on a few random points, demonstrating the inference capabilities of large language models that Google's Deepmind AI Lab had trained for years. All in all, if you just look at the demo**, you will definitely be amazed by how "smart" and how fast the Gemini is.
However, after the launch, a Bloomberg op-ed by Parmy Olson noted that Google had "misretorted" Gemini's true performance in a demo. Olson believes that in this paragraph released by Google,Gemini seems a little too powerful。And, Olson pointed out that some of the reasoning abilities of Gemini shown in this paragraph of Google are actually also available in ChatGPT Plus.
Subsequently,Google respondedThis demo of Gemini's performance is really not liveInstead, still image frames from the original footage were used, and text prompts were written to get Gemini to respond. It is understood that Google's ** for Gemini performance demonstration ** has an important "disclaimer" saying that in order to achieve the demonstration purpose, ** has been artificially reduced, and at the same time, Gemini's calculation and feedback time has also been shortened for "brevity".
This means that Google does not directly recognize GeminiPerformance "fake"., which simply means that the extra part of ** was deliberately cut out in order to show the desired effect, and this part is the calculation and feedback time of gemini.
Although Google has denied the "fake" claim, there are still doubts about the authenticity of its demo**. Machine Xi instructor Santiago Valdarrama posted on the X platform suggesting that Google's above-mentioned ** "disclaimer" seems to be trying to convey that "in order to show carefully selected good results, clipped** is used, not in real time." He also said bluntly"This is misleading the user. ”
In addition, Google's MMLU multi-task language understanding dataset test shows that Gemini Ultra surpasses GPT-4 and even human experts. However, industry experts have found that in the MMLU test, the Gemini Ultra results are marked with small gray letters cot@32, indicating that the best result was selected after 32 attempts using the Chain-of-Thought Prompting technique. GPT-4, on the other hand, has no prompt word skills and has only been tried 5 times. This shows that Gemini Ultra is no smarter than ChatGPT-4, and thus refutes Google's claim that it is "the strongest".
According to a report in the United States, in order to show the "best side" of their products, technology companies "often" fake when making presentations**. For example, when Steve Jobs released the first Macintosh in 1984, in order to show that it was more intelligent and different, he also faked it at the press conference - adding a human voice of "hello" to the boot screen, but in fact, Macintosh did not have this feature at that time. In the "Steve Jobs" movie released in 2015, this dramatic scene was also restored.
Overall, although Google has denied the "fake" claim, the outside world still has doubts about the authenticity of its demo** and Gemini's performance. However, for the "blooming" of artificial intelligence large language models, it is a good thing for users, after all, there is one more option.
Welcome to like, comment, **Favorite, thank you for your support!
Google's most powerful language model has been accused of fraud