Comparing popular large language models: AI race heats up

Comparing popular large language models: AI race heats up

Anyone who isn’t currently living under a rock knows that the Large Language Models (LLMs) landscape is evolving rapidly.

We’re witnessing a fascinating shift in both technical capabilities and business strategies that’s reshaping the future of artificial intelligence. And here’s what you need to know about it.

Web 4.0?! The AI Revolution

The journey from GPT-2 to GPT-4 and beyond represents one of the most rapid technological advances in recent history.

However, we’re beginning to see some limitations surrounding the technology. Compute power and processing speed have been persistent bottlenecks, which is what makes GPT-4’s performance improvements so significant. The model shows enhanced capabilities in several critical areas:

  • Multimodal processing (text, audio, and visual inputs).

  • More nuanced understanding of conversational context.

  • Improved processing speed and performance.

  • Better grasp of tone and sentiment.

But these improvements also highlight a crucial reality: we may be approaching the limits of what’s possible with current training methodologies. And it may be time to pump the brakes a bit.

But to really understand this groundbreaking technology, let’s start by looking at what it can actually do, starring me and Anthropic’s cash cow.

Here’s an example of an interaction with Claude, demonstrating how modern LLMs can engage in nuanced conversation while maintaining context and providing detailed, relevant responses:

different large language models

This exchange showcases how LLMs can process complex queries and provide structured, coherent responses while maintaining a natural conversational flow.

Make sure to notice the model’s ability to understand context and provide detailed, relevant information without losing focus on my core question.

AI BattleBots: Google vs OpenAI

The competitive landscape between Google and OpenAI has evolved into a fascinating strategic battle.

OpenAI initially capitalized on the first-mover advantage with ChatGPT’s consumer rollout, dominating headlines and garnering public use. However, Google has not gone down without a fight, and is now making waves in the space with Gemini (and more).

Google’s strategy centers on integration, weaving LLM capabilities throughout its existing product ecosystem. This approach leverages their massive user base and established infrastructure. Its decision to offer these AI features for free through their ad-supported model has forced OpenAI to reconsider its subscription-based strategy. This cat-and-mouse game has been fun to watch.

OpenAI’s counter-move into search, while seemingly logical, plays directly into Google’s strengths. As the disruptor becomes the challenger, OpenAI must now compete on Google’s home turf. Their partnerships with (Microsoft’s) Bing and potentially Apple represent crucial steps in building the distribution network and user relationships necessary to compete effectively.

For a more holistic view: This Labellerr infographic shows a number of different large language models in the AI space.

Expensive Training Wheels: The Data Dilemma

The most significant challenge facing LLMs today isn’t computational power or business models. It’s the limitations of training data.

The internet-scraping approach that powered early advances is showing its constraints. Modern LLMs face several persistent challenges. Let’s take a look at the biggest ones:

Output Originality

The models often struggle to produce truly novel content, instead generating variations of their training data. This limitation becomes more apparent as users demand more creative and unique outputs.

Quality Control Issues: Hallucinations Not Caused By Psychedelics

The information provided may sound correct, due to how the chatbots present it, but that doesn’t mean it actually is. And this is a major concern.

Models also struggle with expressing appropriate levels of uncertainty in their responses. So it’s hard to discern what’s accurate and what isn’t. Be aware of this drawback.

Contextual Understanding

Despite improvements, LLMs still face challenges with deep reasoning, sentiment analysis and maintaining consistent context in longer conversations.

The Path Forward: Synthetic Data

The solution to these limitations may lie in synthetic data generation. This approach offers several advantages over traditional internet-scraped data. Let’s take a look at a few of them:

  • Controlled Quality: Synthetic data can be generated with specific parameters and quality controls.

  • Targeted Coverage: It can fill specific gaps in training data.

  • Scalability: Larger models can generate training data for smaller, specialized models.

  • Customization: Data can be tailored to specific use cases or industries.

The Competitive Landscape

The current state of competition in the LLM space is driving rapid innovation, but also raising important questions about sustainable business models. OpenAI’s subscription approach and Google’s ad-supported model represent different visions for AI monetization.

And it’s unclear who will win in the long run. Maybe a hybrid approach, integrating the strengths of both models?

At any rate, there are a number of key factors that will determine success. Let’s start with integration capability: specifically how well companies can embed AI into existing workflows.

Distribution networks are also important, as they need to reach and retain users. Furthermore, anyone who understands modern technology knows adaptation speed is vital. Companies must have the freedom to pivot on a dime if need be, in response to market changes.

And let’s not overlook resource management. Computational costs are ridiculously expensive, and growing. Data centers don’t grow on trees!

Reading The Tea Leaves: A Bright Future

The future of LLMs likely lies in specialized models trained on synthetic data, rather than increasingly large general-purpose models trained on internet data.

This shift could democratize AI development, allowing smaller players to create highly effective, domain-specific models.

Companies will need to focus on developing better synthetic data generation techniques, first and foremost. This will allow them to create more efficient training strategies and tools.

And reducing hallucinations, as well as model reliability, is key to keeping users on board.

As for sustainable business models: No one is sure what this will look like. Google, OpenAI, Microsoft and Anthropic likely have some ideas, but good luck hearing about it. We can only speculate. And it’s likely going to be a bit of the legacy search model, mixed with the current chatbot experience we see today.

The winner in this race won’t necessarily be the company with the largest model or the most data, but rather the one that best addresses these fundamental challenges while building a sustainable business model and user base.

Don’t forget about the picks and the shovels, either. NVIDIA, AMD, ARM, Arista Networks, Microsoft and Meta will play a big role here, and will be major players as well.

Traditionally, we’ve seen a dominant tech oligopoly, with a few behemoth companies that rely on lobbying and donating to political campaigns. And they don’t lose. So I expect a few big winners, while everyone else fights for whatever scraps are left, unfortunately.

But it will be fun to watch, that’s for sure. Buckle up!