AI Showdown: The Rise of Gemini series and the Challenge to GPT-4’s Throne

KZ Lim
5 min readFeb 7, 2024

--

Not too long ago, Google dropped some pretty big news with their AI model, Gemini, and before we knew it, the online version, Gemini Pro, was making waves. Now, this isn’t just a ripple in the pond; we’re talking about major splashes. Gemini Pro has climbed the ranks, and it’s giving OpenAI’s GPT-4 a run for its money, almost neck and neck with the GPT-4 Turbo — and that’s saying something!

So, what’s the big deal? Well, Gemini Pro isn’t even the final thing; there’s talk of an even more powerful version called Ultra waiting in the wings. This is like Google signaling a high-stakes challenge to OpenAI. After a bit of a slow start, it looks like Google’s really stepping up its game in the generative AI arena.

Remember when OpenAI came out with ChatGPT? That was a game-changer. Suddenly, Microsoft had its hands on the AI crown, and Google was left playing catch-up. Fast forward a bit, and OpenAI, with just a touch over a billion dollars in revenue, is rocking a valuation that’s through the roof — we’re talking $90 billion! The reason? They’re at the forefront of something truly groundbreaking: foundation models. These are the kind of AI models that can tackle a whole bunch of tasks, even stuff they weren’t specifically trained on, thanks to the magic of large language models (LLMs) and their in-context learning prowess.

What’s in-context learning, you ask? It’s like this awesome ability of AI to take brand new info on the fly and use it right then and there to come up with accurate responses. It’s the secret sauce that makes ChatGPT, and now Gemini, so darn impressive.

But all this success and sky-high valuation have put OpenAI on the hot seat. They’ve got to keep proving they’re worth it, especially since their daily active user numbers aren’t exactly shooting off fireworks. And with rumors of Meta’s LLaMa 3 lurking around the corner, the pressure’s on.

Daily usage of ChatGPT far from impressive. Source: Sequoia Capital

So, where does that leave OpenAI? Well, it’s no secret that they’re probably cooking up a new model as we speak. But what’s this next-gen model going to look like? That’s where things get really interesting.

Let’s talk about thinking — and not just any old thinking, but the way these AI models process stuff. There’s this concept, made popular by Daniel Kahneman, about two types of thinking: System 1 and System 2. System 1 is all about quick, gut-reaction thinking, while System 2 is your more thoughtful, logical type of thinking. Right now, unless you’re really specific with your prompts, LLMs like ChatGPT tend to default to System 1, giving you fast answers without much deep thought.

But here’s the kicker: what if we could get LLMs to use System 2 thinking more often? Some folks believe that the more brainpower — or computational power — the model puts into each answer, the better the results. So, how do we get our AI to think deeper?

There are a couple of ways people are trying to make this happen. One is through something called Process-Supervised Reward Models (PRMs). It’s a fancy way of saying we’re teaching AI to consider each step of its thinking process, kind of like showing your work in math class. This could lead to AI that thinks things through more thoroughly.

Interface showing step by step reasoning in a PRM. Source: OpenAI

Then there’s the Tree-of-thought approach, where the AI explores different paths before settling on an answer, kind of like how you’d work through a tough problem. It’s like giving the AI a scratch pad to brainstorm on before it gives you the final answer.

Source: Google Deepmind

Now, all this sounds great, but it’s not cheap. More thinking means more computing, which means higher cost. However, the results could be well worth the investment. It’s like AI is finally getting to the point where it’s not just about the cool tricks it can do but how well it can think and solve real problems.

So, we’re at the starting line of a whole new race in AI. The industry’s been buzzing with excitement over what these models can do, and now it’s time to deliver on those promises. The big players need to push the limits, and getting AI to think more like us seems like the next big hurdle.

While all this is going on, AI labs need to keep an eye on safety and alignment. We’re reaching a point where AI can outperform humans in various tasks, and that’s both amazing and a bit concerning. It’s crucial that as we develop these super-smart models, we also invest in making sure they play nice and stay aligned with our values and ethics.

As we look ahead, 2024 is shaping up to be a watershed year for AI. This is the year that could see AI stepping out of the pages of science fiction and firmly into reality. It’s not just a bunch of hype anymore; we’re on the cusp of having AI that can truly think, learn, and maybe even understand in ways we’ve only dreamed of.

But let’s not forget that with great power comes great responsibility. The leaps we’re making in AI aren’t just technical — they’re going to touch every part of our lives. From how we work, learn, and play to the bigger questions about what it means to be human in a world where machines can think. This is a journey we’re all on together, and it’s critical that we navigate it with care.

So, as we wrap up this little chat, remember that the story of AI is still being written. It’s a story of challenge and opportunity, of risks and rewards. And whether you’re an AI researcher, a tech enthusiast, or just someone who’s curious about the future, you’re part of that story. Let’s make it a good one.

--

--

KZ Lim

🚀 AI Enthusiast & Practitioner | Transforming Ideas into Intelligent Solutions