Written by Here we’re going to dive into the current state of generative AI and (not just large) language models and where they might be headed in 2024, a topic that’s gaining a lot of buzz as we approach the year’s end.
The realm of generative AI and language models, a field as complex and intricate as the human brain itself, has been a hotbed of innovation over the last year. Just like our brain, which is composed of various specialized regions — from the cerebrum to the occipital lobe to the hypothalamus — each handling distinct functions like vision, language, and sensory data, generative AI is also evolving towards a similar complexity. It’s a shift from viewing the brain as a singular entity to appreciating it as a network of interconnected, specialized mini-brains, each contributing to our overall cognitive abilities. This intricate network enables us to perform everyday tasks with a level of complexity that we often overlook.
This analogy is particularly apt when we think about the trajectory of generative AI. The initial years, starting around 2017 with Google’s introduction of transformers, have been characterized by a race towards larger and more intricate models. Each new iteration of language models has boasted more parameters, weights, and tokens — striving for bigger and more robust architectures. It’s like watching the brain’s evolution, but in the realm of artificial intelligence.
Andrej Karpathy of OpenAI highlighted an intriguing point in one of his talks — that the transformers architecture is bound only by compute power, suggesting that there may be potentially limitless gains to be had. This suggests a future in which the size and complexity of AI models will continue to expand, while mirroring the multi-faceted nature of the human brain. The takeaway is that bigger models equate to better performance, a parallel to how our own brains enable more sophisticated processing.
As we look towards 2024, the journey of generative AI seems to be on a path where integration and specialization of various models could be key. Just like the brain’s specialized regions work in harmony, future AI might see an amalgamation of different models, each fine-tuned for specific tasks, working together to create a more powerful and efficient system. This approach could be the key to unlocking the full potential of AI, moving beyond the current paradigm of simply building larger models.
This vision for generative AI and language models is not just a technological leap but a conceptual shift. It’s about recognizing and harnessing the power of specialized, integrated systems. The future of AI, therefore, might lie in creating a symphony of specialized models, each contributing its unique strength to the collective intelligence.
As we delve deeper into the intricacies of generative AI, especially considering recent advancements in language models, it’s essential to address the elephant in the room: the increasing demand for compute power. This aspect, often overshadowed by the more glamorous front-end applications like ChatGPT for text generation or DALL-E for image creation, plays a critical role in the functioning of these systems.
To provide a glimpse behind the curtain, let’s consider the mechanics of operating these AI models. First off, there’s the model itself, which usually comes in the form of single, massive file, often housed on external storage due to its size. Then, there’s the model loader, which loads the model and provides an interface for you to interact with it. For open-source models interfaces like LM Studio are good for general operations and KoboldCPP is great for creative writing. Setting these models up properly involves some pretty intricate configurations, requiring you to determine aspects such as memory usage, compute power, the size of the model’s environment, and its working memory. This setup process, while invisible to end users interacting with consumer interfaces like ChatGPT, is going to be crucial for anyone deploying AI within their organization.
Once operational, however, these systems showcase their complexity in real-time — parsing prompts, generating outputs and managing working memory usage. It’s a symphony of computational processes, but fundamentally, these models are just sophisticated word prediction machines. Seeing them in operation makes it immediately obvious that they lack self-awareness, sentience, or genuine reasoning skills. Their proficiency (for now at least) lies purely in predicting and generating language-based outputs, and that is their sole function.
Despite their sophistication, even the most advanced models are confined to their primary function of word prediction. This limitation becomes apparent when tasks extend beyond language-based requests. Which is why many models still fall short here, unable to effectively handle non-linguistic tasks.
Looking ahead, I believe the the focus will shift towards developing what are known as agent networks. These networks could potentially address the limitations of current language models by integrating various specialized models, each adept at different types of tasks. This approach could pave the way for more versatile and efficient AI systems, capable of tackling a broader range of challenges beyond just language processing. Such advancements would mark a significant step in the evolution of AI, moving us towards a more integrated and multifaceted approach more akin to the structure of the human brain.
The concept of an agent network in the realm of AI is an intriguing and potentially revolutionary approach, representing a significant shift from the traditional model-centric view. Think of it as a well-orchestrated kitchen in a restaurant, where each component plays a specific role, contributing to the overall functionality and efficiency of the operation.
In this metaphor, the language model is akin to the waiter — the agent that interfaces with the customer. However, it’s essential to remember that the waiter isn’t the one cooking the meals. Similarly, a language model shouldn’t be expected to perform tasks beyond its linguistic capabilities. When we interact with tools like ChatGPT, we often mistakenly assume that they are the entire restaurant, capable of handling every aspect of the process, and leading to disappointment when they fall short, when in reality they’re just one piece of a much larger system.
And this is where the beauty of an agent network lies. It’s a comprehensive ecosystem encompassing both AI and non-AI components, such as databases, web browsers, custom code, and APIs, all working in tandem to achieve a single goal. In an agent network, different models or agents each perform tasks that they are specialized in, much like the division of labor in our restaurant’s kitchen.
Recent advancements and research, such as the work being done by Mistral, has shown that smaller, more efficient models can be highly effective. These smaller models, fine-tuned for specific tasks, can rival larger models in performance while being more cost-effective and consuming less power. The analogy that generating a single AI image uses the same amount of power as charging a phone, while context-dependent, underscores the importance of tackling efficiency in AI model design.
In an agent network, the synergy of multiple models, each adept at a particular task, creates a more efficient and effective system. For instance, in the writing process, we, as humans, have distinct roles — writers, editors, and publishers. Each person may have the skills to perform the other roles, but they usually specialize in just one. Similarly, in an agent network, one model could generate output, another could critique and refine it, and a third could oversee the entire process, ensuring adherence to objectives and quality standards.
The future of generative AI, therefore, may increasingly rely on the development and integration of these agent networks. By leveraging the strengths of specialized models within a collaborative framework, we can achieve more sophisticated, efficient, and effective AI systems, capable of handling complex tasks with greater precision and less resource consumption. This approach holds the promise of transforming the entire landscape of AI, moving us towards a more modular, integrated, and scalable paradigm.
Imagine you’re at the helm of a burgeoning business. In the early days, it’s just you, a jack-of-all-trades, tackling everything from product design to sales, and even the more daunting legal and accounting tasks. It’s a bit like being a one-person band, isn’t it? Even if you manage to play all the instruments, there’s a limit to the complexity and harmony that you can achieve on your own.
Now, think of this scenario in the context of AI, specifically the transformers architecture that underpins language models. Initially, these AI models are like solo practitioners, handling a vast array of tasks but within certain boundaries. They’re programmed to avoid harmful language, for instance, but that’s just skimming the surface. There’s an underlying complexity in human interaction and behavior that goes beyond mere words. It’s about nuance, context, and the intricate dance of social norms and ethics.
But here’s where it gets interesting. Just like a growing business, AI is evolving. No longer is it about one monolithic model trying to do it all. The future lies in creating a network of specialized AI agents, each adept in its own field, much like your business evolving to include a sales expert, a legal advisor, and a bookkeeper. Each one brings a unique skill set to the table, and when they work in concert, the result is far more powerful and efficient than any one agent working alone.
Again, this concept mirrors the complexity of the human brain. Our brain isn’t just one big processor. It’s a symphony of specialized regions, each playing its part in the grand concert that is human consciousness and capability. When these regions communicate and collaborate, they create something far greater than the sum of their parts — the human experience.
Similarly, we can envision a future where AI systems are a tapestry of interconnected models, each an expert in its domain, working together to form a more cohesive, intelligent, and resource-efficient whole. It’s an exciting direction for AI, one that promises not only more advanced capabilities but also a framework that’s more aligned with our natural cognitive processes.
This particular evolution reflects a fundamental truth about growth and specialization. As your business expands, you bring in experts, each a master of their craft, and your business transforms into a network of agents, all communicating and working towards a common goal. This network is more robust, more capable, and more adaptable than any one person, or in the case of AI, any one model, could ever be.
So, in essence, the journey of AI is not just about making smarter models, but about creating a smarter system of models, much like building a successful, multi-faceted business. The future of AI, like the future of business, lies in collaboration, specialization, and the magic that happens when diverse skills and perspectives come together.
Just as an artist would never rely solely on one brush for an entire masterpiece, the savvy use of AI involves a diverse array of tools, each tailored for specific aspects of the creative process. The big players in the industry — OpenAI, Meta, Google — they’re like the manufacturers of these brushes, constantly innovating and producing ever-more sophisticated tools. But the true artistry? That lies in how these tools are employed.
Even in your own personal or professional AI endeavors, adopting a specialized approach can transform your entire experience. Imagine your AI toolkit as a set of finely-tuned instruments, each with a distinct role. Your library of prompts, for instance, should become a curated collection, some designed for drafting, others for editing, and yet others for sensitivity checks. Each prompt is like a specialist in your organization, ready to tackle a task as the point that it’s best suited for.
For the more advanced AI user, the concept of Custom GPTs opens up a whole new world of possibilities. Instead of a one-size-fits-all “Content Creation GPT,” consider a suite of specialized models that include a Write GPT for creative drafting, an Editor GPT for refining content, and a Critic GPT for critical analysis. The approach is akin to an assembly line in a workshop, where each step is handled by an expert in that particular phase, ensuring a more polished and precise outcome.
As we embrace the new year, it’s time to shift our perspective on AI. It’s not just about finding the best tool for a job; it’s about assembling the ideal toolkit for a job, creating a collection of specialized instruments that, when used in concert, can achieve far more than any single, all-encompassing tool ever could.
In the ever-evolving AI landscape, success and efficiency will favor those who adopt a more nuanced, specialized approach. It’s about recognizing that there is no ‘One Model to Rule Them All’, but rather a symphony of specialized models, each playing its part to create a more harmonious and effective whole. Embrace this approach, and you’ll find yourself at the forefront of AI utilization, leveraging the full spectrum of its capabilities to achieve your goals with much greater speed and precision.
Happy 2024!