The Era of Specialized AI - Part II
Part II: The Reality - These Are Early Days of Model Proliferation
We are in the very early stages of model proliferation, with hundreds of AI breakthroughs expected in the coming months and years. This is similar to when businesses transitioned from paper-based accounting to spreadsheets in Microsoft Excel. Two decades later, there are still businesses that run on Excel sheets loosely cobbled together where a software application could generate significant operational cost savings and efficiency. But that ship has sailed, and many businesses struggled to manage and went out of business because they put off their digital modernization for years. I personally witnessed this during my career in retail with a long list of retail businesses that evaporated over a period of a couple of years because they couldn't provide a consistent experience to their customers online and in their stores. Your customers will expect intelligent services, AI-powered experiences and interactions, combined with human touchpoints—and no AGI is going to magically solve this for your business domain.
Even in early 2025, the model breakthroughs are incredible. The pace of innovation is faster than anything I've witnessed in my 25 years in the field of statistical, machine learning, and deep learning models. Here is brief summary of recent model releases:
Claude 3.7: Anthropic's updated AI model featuring a hybrid approach to give both fast responses and engage in more in-depth reasoning when needed. It also boasts significant improvements in coding abilities.
Grok 3: xAI's model focused on understanding the true nature of the universe.
DeepSeek-R1: DeepSeek's model showcasing the potential of GRPO for enhanced reasoning and problem-solving.
Cosmos: Nvidia’s AI platform designed to accelerate AI physical models for robotics and autonomous vehicles
GPT-4o: A large language model by OpenAI, known for multimodality, advanced reasoning, and real-time interactions.
Gemini 2.0: Google's updated multimodal model with enhanced capabilities and real-time processing.
Bard Advanced: Google's updated language model with enhanced reasoning and creative writing capabilities.
Llama 3.3: Meta's open-source language model, offering a more accessible and customizable alternative to proprietary models.
Orca 2: Microsoft's smaller, more efficient model demonstrating comparable performance to larger models.
We have models that are showing very early signs of logical thinking, including multistep thinking, reasoning, problem-solving, demonstrating the ability to solve mathematical problems, generate code, and engage in strategic games. We also have early versions of models trained on real-world data enabling a better understanding of the physical world that will open up possibilities for robotics, autonomous vehicles, etc. And we have gone multimodal, meaning we have gone beyond text incorporating images, videos, and audio. This enables image captioning, AI-assisted video editing, and creating a new medium for creativity with generative art.
However, it's important to recognize that these advancements don't magically equate to AGI. Powerful models alone cannot translate to autonomous workflows, systems and operations. It is an important building block, but a number of other areas are highly underdeveloped to truly achieve intelligent autonomous systems that are capable of replacing jobs, amplifying jobs and creating new jobs.
The other issue I see with generalized models and the social review that follows each one of these model releases is that everyone is saying more or less the same stuff, making me wonder how much of the review itself is generated by AI. This level of convergence on ideas and innovation this early is not a good thing. We should want and expect more divergence in these early stages of discovering how to do real-time reasoning or how to solve problems or write code.
On the other hand, most businesses currently lack the infrastructure, talent, or ecosystem to truly extract value from these powerful models making it impossible to show any real impact despite billions of dollars spent in developing these models. Therefore, it's crucial to start your AI adoption journey now by empowering your organization to create the environment that allows for continuous testing and iteration of use cases with AI. If software adoption was iterative over the past two decades, AI is poised to do the same in the next five years, and it is at least 10 or 100x more iterative compared to software development. I will expand further on this in part IV.