A ghost of GPT (Generative Pre-trained Transformer) is wandering over Gemini.

A specter of GPT (Generative Pre-trained Transformer) haunts Gemini.

Author: Luo Yihang

  1. Every time Google makes a big move in the field of generative artificial intelligence, it can make people feel a hidden and huge emotional power: forbearance, unwillingness, and competition.

  2. At Google I/O in May, Google released the LianGuaiLM series models, which are seen as a strong challenge to the then-popular GPT-4. At the end of the year, they unexpectedly unveiled the Gemini series large language models (including the mobile version Nano, the basic version Pro, and the advanced version Ultra), which directly target GPT in terms of key performance and benchmark evaluations.

  3. Google officially claims that in 32 academic benchmarks widely used in large language model research and development, Gemini Ultra outperforms the current state-of-the-art level represented by GPT-4 in 30 of them. Gemini boasts “native multimodality,” which means it is pretrained on a combination of modalities such as text, images, audio, video, and code from the start, potentially delivering stronger performance in complex comprehension and reasoning, especially in solving math and physics problems.

  4. Google emphasizes the aforementioned advantages tirelessly—OpenAI follows a “progressive multimodality” approach, starting with textual corpora and gradually integrating code, images, and videos. Gemini, on the other hand, was trained on multimodal data from the beginning, followed by fine-tuning with multimodal data, making its training method “more advanced” than OpenAI’s GPT.

  5. Theoretically, advanced multimodal training can lead to stronger performance. The detailed results of Gemini Ultra’s academic benchmark evaluations, which surpass GPT-4 to a large extent, seem to support this. However, academic benchmark tests themselves are part of the theory and do not truly reflect the application’s effectiveness. Many people criticize Chinese large language models for being overly focused on benchmark testing for scoring, but we should treat them all equally, and Google’s approach of striving to surpass GPT for scoring is essentially no different from large models in China.

  6. Currently, users on social platforms have already tested Gemini Pro (Bard chatbot only supports Pro version) and have contributed quite a few critiques. For example, it confuses the winners of the 2023 and 2022 Oscars and does not know how to write a simple function in Python to find the intersection of two polygons. We also discovered that it cannot recognize the number of leaves or solve simple acute angle geometry problems correctly. Even though Gemini Pro is benchmarked against GPT-3.5, it still falls a bit short.

  7. Another prominent issue pointed out by people is the “deception” in Gemini’s promotional video: Gemini Ultra quickly responds to a set of gestures and says it’s a rock-paper-scissors game, but the functionality documentation that the video doesn’t display provides at least two prompts: “What am I doing?” and “Hint, this is a game.” Some tests even require more prompt questions to assist in generating results, but this process is omitted in Gemini’s official video. As a result, most people who don’t pay much attention overestimate Gemini’s understanding and response speed, which can be misleading.

  8. I still remember being at Google I/O in 2017, where a Google employee demonstrated how to book a restaurant directly through Google Assistant voice assistant. Thunderous applause erupted, and I joined in, thinking it was amazing. But a month later, news broke that it was a prearranged scenario. Google did not deceive, but it often eagerly presents its AI optimism, so much so that it frequently glosses over the process and inadvertently exaggerates the effect.

  9. Simply put, the exaggerated presentation in Gemini’s video only shows how much Google cares about Gemini outperforming GPT. It is in a

    We will continue to update Blocking; if you have any questions or suggestions, please contact us!

    Share:

    Was this article helpful?

    93 out of 132 found this helpful

Discover more

Market

Bitcoin, Ethereum, Solana, Binance Coin, XRP, Cardano, Avalanche, Dogecoin, Chainlink, Polkadot price analysis for 2/16

Bitcoin's strong upward momentum may encounter resistance around $52,000, but any potential decrease is expected to b...

Market

Bitcoin Battles Between Bulls and Bears: Understanding the Current Market

Although analysts have shown a negative outlook in the short term, the launch of new Bitcoin ETFs has brought about s...

Market

Cardano Price Prediction: ADA Set for Spectacular Gains Post-Bitcoin Halving – Future Millionaire Maker?

Cardano gains 4% in 24 hours, reaching $0.383747 amidst a 2.5% marketwide increase.

Blockchain

Cardano (ADA) Breaks Out: Is a New All-Time High on the Horizon? 🚀

Cardano (ADA) has made a notable price breakthrough, surging to $0.710 for the first time since May 2022 on Thursday....

DeFi

🤩 The US Cryptocurrency Conundrum: Regulating in Chaos 😱

Charles Hoskinson, creator of Cardano, expressed to Cointelegraph that regulators in the US, like the SEC, have effec...

DeFi

Cardano's DeFi Ecosystem Flourishes Total Value Surges to Almost $450M as ADA Rockets 17% on Layer 1 Push!

This week, the total value locked (TVL) of Cardano-based tokens skyrocketed to more than $440 million, surpassing the...