How the GenAI Revolution Reminds Us of 1990s Computing
Generative AI (GenAI) represents a significant step forward in the evolution of computing. We have seen how computing evolved over the last few decades, from mainframes (1980s) to client/server (1990s), web (2000s), and cloud and mobile (2010s). It has the potential to transform the computing landscape.
The 2020s will witness the infusion of GenAI into almost every aspect of computing, both personal and enterprise.
The emergence of GenAI marks a paradigm shift comparable to the technological revolutions of the 1990s. Let’s explore the parallels between the two eras, highlighting how today’s trends are shaping the future of computing.
|Current GenAI Era
|CPUs based on DEC Alpha, SUN SPARC, Intel x86, Power PC
|AI Accelerators from NVIDIA, AWS Trainium, Azure Maia, Google TPU
|OEMs such as Compaq, Dell, HP, IBM and Sun
|Public cloud platforms such as AWS, Microsoft Azure, and Google Cloud
|Commercial OS like Windows and Mac; Open source OS like Linux
|Open Foundation Models such as Llama 2 and proprietary models like GPT-4
|Built-in OS utilities and commands
|Intrinsic commands and tools bundled with the OS
|AI utilities based on vector databases, retriever models, and semantic search
|CLI of BASH and GUI of Windows and Mac
|APIs exposed by AI cloud platforms and model providers
|Built-in applications such as a calculator and a word processor
|AI assistants like Google Duet AI, Microsoft Copilot, and Amazon Q
|Integrated development environments
|Borland Delphi, Visual Studio, and Eclipse
|Tools to build custom AI assistants, such as Microsoft Copilot Studio, Google Generative AI Studio, and Amazon Bedrock
From Diverse CPU Architectures to AI Accelerators
Then: The 1990s saw a diversity in CPU architectures, with DEC Alpha, SUN SPARC, and Intel x86 leading the market.
Now: The spotlight is on GPU and AI accelerator chips, exemplified by Amazon, Google, Microsoft, and NVIDIA.
Similar to the diverse CPU architectures of the 1990s, AI accelerators are evolving as the workhorse of GenAI workloads. Amazon’s investment in Trainium and Inferentia chips, Google’s emphasis on tensor processing units (TPUs), and most recently, Microsoft’s official entry with Azure Maia, act as evidence of this trend. It is also expected that OpenAI will have home-grown chips to train foundation models.
Along with custom AI hardware, the software layer responsible for interacting with it is evolving. Though NVIDIA’s CUDA dominates this space, the AWS Neuron SDK, Azure Maia SDK, and ONNX are gaining traction. To get the most out of the custom hardware, deep learning frameworks like TensorFlow, PyTorch, and JAX are optimized for these layers.
The custom AI accelerators remind us of the diverse CPU architectures available in the 1990s.
The Shift in OEM Dynamics
Then: Traditional original equipment manufacturers (OEMs) like Compaq, Dell, HP, IBM and Sun were the hardware powerhouses.
Now: Public cloud platforms such as AWS, Azure, and Google Cloud have emerged as the new OEMs. They play a pivotal role in hosting and deploying AI technologies.
OEMs such as Compaq, HP, IBM and Sun shipped servers based on a specific CPU architecture in the 1990s. In the current context, public cloud providers can be compared with OEMs.
The proprietary hardware, bare metal or virtual servers, and custom software layer closely resemble how some vendors, such as Sun, shipped end-to-end stacks based on SPARC processors, Solaris OS and other components required to run workloads.
Operating Systems to Foundation Models
Then: Linux and Windows were the core operating systems, fundamental to computing.
Now: Foundation models have become the kernel of the AI operating system, with some models being open source and others proprietary.
The debate about open source versus commercial software dates back to the 1990s, which saw the rise of GNU and FOSS. Fast forward to 2024, and we are still discussing the pros and cons of open and closed foundation models.
With Meta leading the pack with LLaMA, other players, such as Mistral, are gaining importance. On the other hand, we have GPT-4 from OpenAI, Gemini from Google, Titan LLM from AWS, and a slew of other models such as Claude 2, Jurassic 2, and Command from Anthropic, AI21, and Cohere, respectively.
Large language models (LLMs) will become so important that they will be integrated into the OS kernel to provide generative AI capabilities and even self-healing of the operating system.
From Built-In Utilities to AI Tools
Then: Operating systems were equipped with essential utilities and commands.
Now: Vector databases, search, and orchestration tools form the backbone of AI utilities, enhancing the capability and efficiency of AI platforms.
Almost all operating systems come with in-built utilities and commands to manage the system. From basic file management to advanced optimization tools, operating systems bundle tools with them.
Similar to these utilities, vector databases with retrievers and ranking models will become an essential part of the AI stack. The new AI stack will have this as a layer sitting on top of the LLMs to influence their responses and provide contextual inputs through the prompts. High-level applications, such as agents, will use them to automate various tasks that rely on storage, search and retrieval.
Shell and Graphical Interfaces to AI Platforms
Then: Command-line interfaces powered by BASH and Zsh, and sophisticated graphical interfaces built into Windows and MacOS democratized access to computing.
The APIs provided by AI platform providers provide access to the foundation models in the same way that Shell provides access to the OS. They offer a simple interface for pre-training, fine-tuning, versioning, deploying models, and performing inference in the new OEM environment — the public or private cloud.
The low-code and no-code tools resemble the GUIs available in Windows and macOS. They democratize AI by enabling non-developers and power users to consume foundation models and build modern applications.
Software Applications to AI Assistants
Then: Software applications were either developed by OEMs or third parties.
Now: AI assistants, like Google’s Duet AI, Amazon Q, and Microsoft Copilot, are the new applications that are increasingly integral to both consumer and enterprise environments.
If AI platforms are the new Shell, AI assistants are the new applications. Similar to how OS providers ship built-in applications while enabling developers to create custom applications, the new platforms ship with an embedded assistant while offering the development environment and tools to build custom AI assistants.
Duet AI is integrated with Google Workspace, while Microsoft is embedding Copilot in almost every business application. Amazon Q is tightly integrated with the AWS Management Console to enable users to perform common tasks.
Development Environments: Then and Now
Then: IDEs like Borland Delphi, Visual Studio and Eclipse were the standard for software development.
Now: Microsoft Copilot Studio, Google Generative AI Studio, Amazon Step Functions and similar environments represent the new generation of development tools tailored for AI and machine learning.
Developers relied on tools such as Visual Studio, Eclipse, and XCode to build custom applications. In the GenAI era, cloud-based tools such as Microsoft Copilot Studio, Google Generative AI Studio, and Amazon Bedrock + AWS Step Functions have become the preferred IDEs to develop AI assistants. They enable developers to integrate diverse data sources, LLMs, prompt engineering and guardrails to build enterprise-grade AI assistants.
The GenAI era is redefining the computing landscape, mirroring the transformative changes of the 1990s but with a focus on AI and cloud technologies. As leaders, embracing these changes and understanding their implications is vital for driving your organizations forward in this new era of computing.