The Wild Juxtaposition of AI's Evolution: What Has Technology Crossed From Start to Now?

The Industry Transformation Behind One Comparison

A tweet that recently went viral on Twitter used the classic "How it started / How it's going" format to vividly illustrate the astonishing leap AI technology has made in an incredibly short time. The original author described it as a "Crazy juxtaposition" — a phrase that perfectly captures what so many people in the tech industry are feeling right now: AI is advancing faster than almost anyone anticipated.

The reason this kind of "then vs. now" comparison resonates so deeply is that it compresses abstract technological progress into an instantly perceivable visual shock. Whether it's model capability, generation quality, or the sheer breadth of applications, the changes over just a year or two have been nothing short of generational leaps.

Why the AI Evolution Comparison Is So Striking

The Non-Linear Acceleration of Technological Iteration

Traditional technology tends to progress linearly and incrementally — refining a single feature can take years. But AI, and especially generative AI, exhibits a distinctly non-linear acceleration.

Generative AI refers to artificial intelligence systems capable of producing new content (text, images, audio, video, etc.), with deep learning-based generative models at its core. From early GANs (Generative Adversarial Networks, introduced by Ian Goodfellow in 2014) and VAEs (Variational Autoencoders), to today's dominant Diffusion Models and Large Language Models (LLMs), generative AI has undergone multiple generations of architectural revolution. The Transformer architecture proposed by Google in 2017 was the pivotal turning point for modern generative AI — its Self-Attention mechanism enabled parallel processing of sequential data, dramatically improving training efficiency and generation quality. Around 2022, products like ChatGPT and Stable Diffusion brought this technology into the public eye, igniting the current generative AI wave.

Take image generation as an example: early models produced distorted, blurry outputs that couldn't even reliably render the correct number of fingers on a human hand. Today's models produce work approaching photographic realism, with detail, lighting, and composition that meet professional standards. The same pattern holds for text generation, code assistance, video synthesis, and more. Each domain represents a massive leap on its own; together, they create a collectively dizzying impression of overall progress.

From "Toy" to Genuine Productivity Tool

The deeper shift lies in AI's fundamental change of role. In the early days, many AI tools were more like gadgets for tech enthusiasts or demonstration pieces — interesting, but hard to integrate into real workflows. Today, AI is deeply embedded in real production processes like software development, content creation, and data analysis, becoming an indispensable part of many people's daily work.

This leap from "Demo" to "Product" is what truly makes the contrast feel "crazy." It means AI is no longer confined to labs or social media showcases — it's actively reshaping how we work and how businesses operate.

The Core Drivers Behind AI's Rapid Evolution

The Triple Resonance of Compute, Data, and Algorithms

AI's rapid evolution is inseparable from the combined breakthroughs in compute, data, and algorithms.

On the compute side, modern AI training demands computational resources far beyond what most people imagine. Training GPT-3 (175 billion parameters) is estimated to have consumed roughly 355 GPU-years of computation. NVIDIA's GPUs evolved from graphics rendering chips into the core infrastructure for AI training — the H100's FP8 training throughput is nearly 6x that of the previous-generation A100. Cloud computing has further democratized access to compute: developers can now provision thousands of GPUs on demand without building their own data centers, a critical infrastructure prerequisite for compressing AI innovation cycles.

On the algorithm side, the Transformer architecture's Multi-Head Self-Attention mechanism enables direct modeling of relationships between any positions in an input sequence and is natively suited for parallel GPU training — fundamentally changing how AI models are designed. Building on this foundation, researchers introduced alignment techniques like RLHF (Reinforcement Learning from Human Feedback) to make model outputs better match human intent, as well as parameter-efficient fine-tuning methods like LoRA and QLoRA that dramatically lowered the barrier to customized training.

These three forces don't simply add up — they create a positive feedback loop: greater compute enables larger models, larger models unlock more applications, more applications drive investment and research, which in turn expands compute and data resources further.

The Open-Source Community as an Accelerant

The open-source ecosystem has played a pivotal role in this acceleration. Meta's release of the LLaMA series in 2023 was a landmark event — its openly released model weights quickly spawned dozens of derivative models including Alpaca, Vicuna, and Mistral, forming an open-source ecosystem that competes with closed models like GPT-4. At the tooling layer, the Hugging Face platform has aggregated over 500,000 public models and datasets, effectively becoming the "GitHub" for AI developers; open-source frameworks like PyTorch and LangChain have significantly lowered the technical barrier for model integration and application development.

The wide availability of model weights, training frameworks, and toolchains allows developers worldwide to iterate rapidly on each other's work. This community-driven, collectively optimized innovation environment means that the output of a single research team or company can be quickly reproduced, improved, and applied by millions of developers globally — generating an innovation velocity that traditional closed R&D models simply cannot match. This is also why many people feel that AI progress is happening on a nearly week-by-week basis.

A Measured Look at This "Crazy Juxtaposition"

As exciting as this contrast is, we should stay grounded. The visual intensity of a stark before-and-after comparison can amplify the perception of progress, while true technological maturity often requires sustained effort on less glamorous dimensions like reliability, safety, and cost control.

Leaps in generation quality are undeniably significant, but AI still faces substantial challenges in real-world deployment: hallucination, bias, and interpretability, among others. AI hallucination refers to the phenomenon where large language models generate content that sounds plausible but is factually incorrect or fabricated. Mechanistically, this stems from the nature of language models themselves: they predict the next token based on statistical probability rather than truly "understanding" or "retrieving" facts. When a model lacks relevant knowledge from its training data, it tends to "confidently make things up." Current mitigation approaches being explored by the industry include Retrieval-Augmented Generation (RAG), Chain-of-Thought prompting, and other technical paths — but fundamentally solving hallucination remains a key obstacle on the road to truly reliable AI systems. Beneath the impressive surface of "How it's going," a great deal of engineering and governance work remains unfinished.

Conclusion: We Are Living Through a Historical Acceleration

The reason this brief tweet struck such a widespread chord is that it articulates a shared industry conviction: we are at a historical inflection point of technological acceleration. The vast gap between "How it started" and "How it's going" is both a reflection on past effort and a glimpse of future possibility.

If just one or two years can bring such transformative change, what unimaginable picture will the next "How it's going" screenshot reveal? That may be the most thought-provoking question this "crazy juxtaposition" leaves us with.

Key Takeaways