From GANs to AI Agents: The Evolution of Generative AI
A living timeline of the research breakthroughs and product launches that transformed generative AI from experimental image synthesis into systems that write, reason, see, hear, create video and act through software tools.
-
Research
GANs turn generation into a competition
Ian Goodfellow and colleagues introduced Generative Adversarial Networks, training a generator against a discriminator.
The generator learned to create samples that could fool the discriminator, while the discriminator learned to detect synthetic output. GANs became influential in realistic image synthesis, although training instability and limited diversity remained recurring problems.Sources & references 1 source
-
Research
The Transformer rewrites the architecture of language AI
“Attention Is All You Need” introduced an architecture based on attention rather than recurrent processing.
Transformers process relationships between tokens in parallel, making training easier to scale than earlier recurrent approaches. The original paper focused on translation, but its architecture became the foundation of most modern large language models and many multimodal systems.Sources & references 1 source
-
Research
GPT shows the power of generative pre-training
OpenAI showed that a Transformer trained broadly on unlabelled text could be adapted to multiple language tasks.
The first Generative Pre-trained Transformer was modest by later standards, but its two-stage recipe became highly influential: learn general language patterns through prediction, then fine-tune for specific tasks. The approach helped shift the field toward reusable foundation models.Sources & references 1 source
-
Research
GPT-3 demonstrates learning from prompts at scale
A 175-billion-parameter language model performed many tasks from instructions and examples without conventional task-specific fine-tuning.
GPT-3 made few-shot and zero-shot prompting a central interface for large language models. Its fluent output attracted broad attention, but the research also documented unreliable answers, bias and substantial computing requirements. Scale improved capability without solving factual reliability.Sources & references 1 source
-
Research
DALL-E links natural-language prompts to image creation
OpenAI demonstrated a model that generated and edited images from textual descriptions across varied concepts.
Illustration: Justin Jay Wang / OpenAI DALL-E helped show that large-scale generative modelling could connect language and vision. Early results were constrained and curated, but the project anticipated the rapid growth of text-to-image systems that followed.Sources & references 1 source
-
Open Release
Stable Diffusion brings image generation to local hardware
The public release of Stable Diffusion widened access to text-to-image generation and enabled a large ecosystem of adaptations.
Unlike image generators available only through hosted services, Stable Diffusion’s released weights allowed developers and many users to run and modify the model. That openness accelerated experimentation while intensifying debates over training data, consent, copyright and misuse.Sources & references 2 sources
-
Product Launch
ChatGPT turns a language model into a mass-market conversation
OpenAI released a conversational research preview trained to follow instructions and respond through dialogue.
ChatGPT did not introduce the Transformer or large language models, but its accessible interface changed adoption. Millions of people could iteratively ask, correct and refine requests. The launch also made hallucinations, educational disruption, labour impact and AI governance public issues rather than specialist debates.Sources & references 1 source
-
Model Release
LLaMA accelerates the open-model ecosystem
Meta introduced a family of smaller foundation language models intended to make high-performance research more accessible.
LLaMA and its successors helped fuel an ecosystem of fine-tuned and locally deployable models. “Open” remained a contested description because licences, training data and release conditions differed from conventional open-source software, but the shift reduced dependence on hosted proprietary systems. -
Model Launch
GPT-4 brings image understanding into a leading language model
GPT-4 accepted text and image inputs while generating text, marking an important step toward mainstream multimodal AI.
The model could analyse documents, photographs, diagrams and screenshots alongside text. OpenAI released limited technical detail about its size and training data, illustrating a broader reduction in transparency as frontier systems became more commercially valuable.Sources & references 2 sources
-
Model Launch
Gemini is designed as a natively multimodal model
Google introduced Gemini as a model family trained to work across text, images, audio and video.
Rather than adding separate perception components only after language training, Gemini was presented as multimodal from the start. The launch intensified competition around models that can understand and generate across several forms of information, although access and capabilities varied by model version.Sources & references 2 sources
-
Research Preview
Sora demonstrates longer, coherent text-to-video generation
OpenAI previewed a video model capable of generating complex scenes and extending existing footage from prompts.
Sora highlighted rapid progress in generative video and the use of diffusion transformers. Early demonstrations were selected by the developer, and the model still struggled with physical consistency and cause-and-effect. Its significance was capability direction, not proof of reliable world simulation.Sources & references 1 source
-
Model Launch
GPT-4o unifies real-time text, vision and voice interaction
GPT-4o processed and generated combinations of text, audio and images with lower latency than earlier stitched-together systems.
OpenAI The “o” stood for “omni,” reflecting a shift from text chat toward more natural multimodal interaction. Voice and vision made AI assistants more immediate, while also raising stronger questions about impersonation, privacy, emotional dependence and the safeguards needed for live media generation.Sources & references 1 source
-
Research
Reasoning models spend more computation before answering
OpenAI introduced o1-preview, using additional inference-time computation to improve performance on multi-step problems.
The release shifted part of the scaling race from larger pre-training runs toward “test-time compute,” where a model performs more internal processing before responding. Better benchmark performance did not guarantee truthfulness or remove the need to verify outputs, but the approach influenced the next generation of frontier models.Sources & references 2 sources
-
Model Launch
GPT-5 integrates automatic reasoning into the default experience
OpenAI made GPT-5 the default ChatGPT system, routing between direct responses and deeper reasoning according to the task.
The launch reflected a broader product shift away from asking users to understand a growing model menu. A unified system could decide when additional reasoning was useful. It also showed how generative AI products were evolving from standalone models into orchestrated systems combining routing, tools, safety layers and multiple modes of computation.Sources & references 1 source
What comes next?
Generative AI has moved from producing plausible samples to participating in complex workflows across text, images, audio, video and software. Yet fluent output is not the same as reliable knowledge or independent judgement. Hallucinations, bias, copyright disputes, security risks, energy use and concentration of computing power remain central limitations. The next phase will be shaped less by whether a model can generate content and more by whether it can reason, use tools, preserve context and act reliably under human control.
