To receive industry-leading AI updates and exclusive content, sign up for our daily and weekly newsletters. Learn more
Where once we speculated about when we would have software that could consistently pass the Turing Test, we now take it for granted that not only does this amazing technology exist, but that it is rapidly improving and becoming more powerful.
It's easy to forget how much has happened since ChatGPT was released on November 30, 2022. Since then, innovation and power have continued to emerge from publicly available large-scale language models (LLMs). It seemed like every few weeks we saw something new that pushed the boundaries.
Now, for the first time, there are signs that the pace may be slowing significantly.
To understand this trend, consider the releases of OpenAI. The jump from GPT-3 to GPT-3.5 was big and put OpenAI in the public's attention. The jump to GPT-4 was also impressive, a huge step forward in terms of power and capacity. Then came GPT-4 Turbo, which improved speed, and GPT-4 Vision, which unlocked GPT-4's existing image recognition capabilities. And then, just a few weeks ago, GPT-4o was released, which offered enhanced multi-modality but relatively little in terms of added power.
Other LLMs such as Anthropic's Claude 3 and Google's Gemini Ultra have followed a similar trend and now appear to be converging to similar speed and power benchmarks as GPT-4. We haven't yet reached a plateau, but we do appear to be entering a period of slowdown. The pattern that is emerging is one of less advances in power and range with each generation.
This will shape the future of solution innovation
This is very important. Imagine you have a disposable crystal ball. Your crystal ball can tell you anything, but it can only ask one question. If you were trying to predict the future of AI, that question would probably be: How fast will the power and capabilities of LLMs continue to improve?
As LLM evolves, so does the world of AI. Every significant improvement in LLM capabilities makes a huge difference in what your team can build and, more importantly, what they can work reliably with.
Consider the effectiveness of chatbots. With the original GPT-3, responses to user prompts could be hit or miss. Then came GPT-3.5, which made it much easier to build persuasive chatbots and improved the quality of responses, but they were still spot on. It wasn't until GPT-4 that we got consistently spot-on output from an LLM that actually followed instructions and showed some degree of reasoning.
GPT-5 is expected to arrive soon, but OpenAI seems to be carefully managing expectations. Will its release surprise us with a major leap and spark a new surge in AI innovation? If not, and progress continues to dwindle on other public LLM models, we expect there will be significant implications for the broader AI field.
Here's how it unfolds:
Further specialization: If existing LLMs are not powerful enough to handle nuanced queries across topics and functional areas, the most obvious response for developers is specialization. We may see more AI agents developed to address relatively narrow use cases and serve very specific user communities. Indeed, OpenAI's launch of GPT can be interpreted as a recognition that it is not realistic to have one system that can read and respond to everything. The rise of new UIs: The dominant user interface (UI) in AI so far has undoubtedly been chatbots. Will that continue to be the case? Although chatbots have some clear advantages, their apparent openness (users can type any prompt) may actually disappoint the user experience. We may see more forms in which AI works but with more guardrails and restrictions to guide the user. For example, consider an AI system that scans documents and offers the user some possible suggestions. Open source LLMs fill the gap: LLMs are believed to be prohibitively expensive to develop, so open source providers like Mistral and Llama, who do not have a clear commercial business model, are likely at a significant disadvantage. However, that may not be as much of a problem if OpenAI and Google stop making significant progress. They may be able to hold their own if the competition shifts to features, ease of use, and multimodal capabilities. The race for data heats up: One of the reasons LLMs are starting to fall into the same feature range is the lack of training data. As the end of publicly available text-based data approaches, LLM companies will have to look for other sources. This may be why OpenAI is focusing on Sora. Leveraging images and videos for training means that not only could the way models handle non-text inputs be significantly improved, but they could also be more nuanced and subtle in understanding queries. New LLM architectures emerge: So far, all major systems use the Transformer architecture, but there are other architectures that show promise. However, with the rapid advances being made in Transformer LLMs, these models have not been adequately explored or invested in. As these advances begin to slow, we may see increased interest in Mamba and other non-Transformer models.
Final thoughts: the future of LLMs
Of course, this is speculation; no one knows where LLM capabilities or AI innovation will go next. But what is clear is that the two are closely related, which means that all developers, designers, and architects working with AI need to be thinking about the future of these models.
One pattern we may see emerge in LLMs is increased competition on levels of functionality and ease of use. Over time, we may see some degree of commoditization, similar to what we've seen elsewhere in the technology industry. For example, consider databases and cloud service providers. There are significant differences between the various options on the market, and while some developers have clear preferences, most developers consider them roughly interchangeable. There will be no clear, absolute “winner” in terms of which is the most powerful and capable.
Cai GoGwilt is co-founder and chief architect at Ironclad.
Data Decision Maker
Welcome to the VentureBeat community!
DataDecisionMakers is a place where experts, including technologists working with data, can share data-related insights and innovations.
If you want to hear about cutting edge ideas, updates, best practices, and the future of data and data technology, join DataDecisionMakers.
You might also consider contributing your own article.
Learn more about DataDecisionMakers
