Gemini Ultra cost way more to train than GPT-4, but is still worse

  • Google reportedly spent $191 million to train its best AI, Gemini Ultra.
  • This is almost two times more than OpenAI invested to train its best AI yet, GPT-4.
  • Despite Google’s benchmarks, independent research and user reviews consistently claim GPT-4 is still better than Gemini Ultra.

At its opening presentation for Google I/O 2024 yesterday, the Alphabet-owned search giant revealed upgrades to its Gemini generative AI, including the introduction of a version that can produce responses faster called Gemini 1.5 Flash.

The most advanced version of AI that Google currently has is called Gemini Ultra, which is available to users via the Gemini Advanced subscription model. Ultra can be compared to GPT-4 in that it is the leading model for Google, as GPT-4 is to OpenAI – except now with the reveal of GPT-4o.

According to data from Stocklytics, Gemini Ultra has been the most expensive generative AI model to train since 2017, with Google investing $191 million just on the training.

“Training” in this sense is to give the model massive swathes of data to consume, analyse and then tweak a number of complicated algorithms in how the AI processes and then uses the data to generate responses.

In comparison, OpenAI only spent $78 million to train GPT-4, often considered the most advanced generative AI out there. While both models cost huge amounts of money to train, Google’s spending is emblematic of the company’s current plans and trajectory around AI: Google is playing a game of catch up.

“The astronomical training costs pegged on Google’s new foundation model underscores the underlying cost pressure of developing new advanced technologies. However, the cost burden is a pain worth investing in as more tech giants augment their efforts to expand their AI network and develop new language models, including Microsoft and Meta,” says Stocklytics financial analyst Edith Reads.

It has had less time than OpenAI to invest and train generative AI, even if Google was the pioneer in the space with its DeepMind projects. But what has that $191 million gotten the Big G for its trouble?

As per Google’s own benchmarks, the company gives GPT-4 the general capability of 86.4 percent, while it gives its own Gemini Ultra AI the capability of 90.0 percent. It says that Gemini Ultra surpasses GPT-4 in reasoning, generating code, and in solving both simple and complex math problems.

The only parameter where Google believes GPT-4 is better is in common sense reasoning for everyday tasks.

Unfortunately for Google, not everyone agrees with its benchmarks. Independent comparisons online of Gemini Ultra and GPT-4 almost always point out that OpenAI’s model is superior. One research paper from Ukraine even goes as far as calling Gemini Ultra’s capabilities “purely business PR campaign that mainly misleads users.”

Another real-world comparison claims that GPT-4 still has the edge over Gemini Ultra in “tasks such as writing, document creation, summarization, general-purpose image generation, and learning through conversations.”

Many comparisons still point out that both models have their strengths and weaknesses, but with OpenAI already on the upper hand in terms of human-like responses with their latest version of its AI, the question becomes just how much more will Google spend on future training of models.

It is likely that Google is already working on an even better version of Gemini, one that surpasses Ultra, but can it launch it to the public before GPT-5 hits the net? Maybe for another couple hundred million.


About Author


Related News