Will DeepMind’s Gemini Supersede ChatGPT?

Google DeepMind is a Google subsidiary tasked with AI development. Born out of a merger between Hassabis's startup, DeepMind, purchased by Google in 2014, and Google's own venture, Google Brain, it has been making strides since its inception in April. 

For years, DeepMind dedicated efforts to AlphaGo, a model that eventually triumphed over the world champion in the ancient game of Go. Hassabis's repertoire also includes AlphaFold — an AI with expertise in protein folding processes. This may not ring bells for the layperson, yet there isn't a researcher in the field of biology or medicine who doesn't utilize this AI model today. Every pharmaceutical behemoth leverages AlphaFold to expedite the development of new drugs. 

Meanwhile, Google Brain has centered its efforts around more conventional AI tools, such as chatbots and photo editing, among others. 

However, the tides are shifting as Google DeepMind specialists now turn their expertise towards crafting a new AI, dubbed Gemini. This substantial language model will be tailored to handle text, mirroring GPT-4, yet surpassing it considerably. The team intends to embed Gemini with additional capabilities that are already functional in AlphaGo and AlphaFold. The goal? To empower the system to plan and tackle specific challenges across a broad spectrum of scientific and economic disciplines.

At a high level you can think of Gemini as combining some of the strengths of AlphaGo-type systems with the amazing language capabilities of the large models. We also have some new innovations that are going to be pretty interesting,

hints Demis Hassabis.

Google DeepMind CEO brings to mind the surprise rippling through the research circles due to the furor ChatGPT sparked. As he points out, the level of commotion exceeded even OpenAI's expectations. At that juncture, the potential of most market models was approximately on par. The unexpected turn of events underscores a realization that has been dawning on industry insiders over the past two or three years: AI is poised to transition out of research labs and contribute to the creation of products that serve a practical purpose for people. 

For Google, this practical application manifests in projects like Bard and Search Generative Experience (SGE), as well as in the integration of AI in Google Photos. As Hassabis puts it, certain operations are happening “under the hood.” For instance, the company has long relied on AI systems to orchestrate the cooling mechanisms in their massive data processing centers, resulting in a 30% reduction in energy consumption. Concurrently, the market-oriented projects under the tech giant's umbrella are swiftly gaining momentum. 

So I think, in a year or two’s time, we are going to be talking about entirely new types of products and experiences and services with never-seen-before capabilities,

says Hassabis.

Current chatbots still lack features like memory, planning, and the ability to reason. Moreover, they're prone to “hallucinate”. Hassabis candidly discusses the fact that today's AI models could fabricate an answer to questions such as, “Which resources should I read up on for a particular topic?” They can create a seemingly credible list of references, using the names of acknowledged industry experts, and compiling titles of non-existent publications from newspaper headlines. The most significant issue arises when there's a risk of the AI learning from its own hallucinations. At Google DeepMind, they are devising a solution in the form of digital watermarks on AI-generated data, intending to integrate them into the generative process to prohibit subsequent removal. 

Additionally, researchers fear that simply amplifying the scale of existing systems might eventually lead to a reduction in performance indicators. Therefore, the creation of AGI (Artificial General Intelligence) necessitates substantial investments into fundamentally new approaches for developing large language models. To clarify, AGI is the next step in AI evolution. It's not just capable of solving, but also framing a diverse range of tasks as effectively as a human. Regardless, Hassabis remains quite optimistic.

I would not be surprised if we approached something like AGI or AGI-like in the next decade,

forecasts the CEO of Google DeepMind.

Well, Demis, let's circle back to this conversation in 10 years!