Listen to this story
Google wanted to win the generative AI race against OpenAI and Microsoft so badly, that it temporarily put on hold its grudges with DeepMind, and started working on their next-generation project called Gemini. The project is being led by Jeff Dean, the head of Google Brain, and Demis Hassabis, the founder of DeepMind. The team is confident that Gemini will eclipse OpenAI’s GPT-4, but is that really possible?
Firstly, the moment a company starts comparing its projects and its capabilities with GPT-3 or GPT-4, it becomes pretty clear who is the actual leader in the race. For example, even Meta’s plan to release open source Llama 3 was in the limelight because the person compared it with GPT-4, and was confidently against the alignment policy.
Speaking of alignment, Google has also been seriously concerned about making AI more “bold and responsible”. It made sure that its approach of building AI this way remains at the centre of every announcement it makes. This definitely comes with a lot of downs, and only a few ups, as we have seen how models start hallucinating the more we try to align them.
Subscribe to our Newsletter
Join our editors every weekday evening as they steer you through the most significant news of the day, introduce you to fresh perspectives, and provide unexpected moments of joy
Too late to the grind
Apart from the claims that Gemini is going to be multimodal, there is very little information about how exactly Google plans to surpass GPT-4 in terms of capabilities. For the model to have different modalities, Google definitely has the golden data of YouTube, which OpenAI still does not have, but that is all.
Interestingly, DeepMind recently also published a paper about Reinforced Self-Training (ReST) for Language Modeling. This paper aims to remove humans from RLHF by letting LLMs build their own policy with just one initial command. If this is integrated in Gemini, it might possibly be able to improve the current capabilities of LLMs.
Another advantage DeepMind gives Google is the dedication to building AI agents based on reinforcement learning through game-based models like AlphaGo, which might actually be a game changer here. On the other hand, possibly learning from Google DeepMind, OpenAI has also acquired Global Illumination, a company building creative simulations and training AI agents within it.
It is too hard to break OpenAI’s monopoly at the moment. Meta has been trying very hard to do it with Llama 2. But as soon as people start to move away from GPT-based models, OpenAI pulls out another trick, like the most recent release of ChatGPT Enterprise. This time, OpenAI is ready to even take on Microsoft Azure OpenAI Service and Google VertexAI in the generative AI battle.
Amid all the GPU crunch, a recent trending blog by SemiAnalysis said that Google is the most compute-rich firm in the world. It further adds that Gemini is already in training and given the massive infrastructure that Google has, it might actually be unbeatable. Interestingly, the GPU Rich vs GPU Poor conversation was not taken seriously by anyone.
Can’t take Google seriously enough
Given Google’s failed attempts of the past with LaMDA, PaLM, and the chatbot Bard, it is not entirely clear why someone would expect the company to be able to actually take the lead with Gemini. Even though they have an arsenal of algorithms this time, and possibly a lot of computational power, it does not matter if they come up with something on par with GPT-4 by the end of the year, when OpenAI is already planning to build GPT-5.
In the meantime, Google has recently also integrated Meta’s Llama 2 in its cloud platform, Vertex AI. This is possibly after realising that Microsoft’s partnership with OpenAI hasn’t been working well, and even it is partnering with Meta for offering smaller and open models on Azure Cloud. This might be a good chance for Google DeepMind to decide to kill its project Gemini — boldly and responsibly.
Google is predominantly known for making tall claims but ends up releasing experimental products, and services, and eventually taking them down.
At this point, no one cares if Gemini is going to perform better than GPT-4 or not. All they want is an accessible API, instead of a locked-out model like Bard or ChatGPT. For this, OpenAI is already on its way to make its models enterprise friendly by making it more data secure. Seems like OpenAI is also becoming “bold and responsible”, and leaving Google behind.