From transformers to diffusion models — explore how AI architectures work, how they compare, and which one fits your needs. Fully interactive.
+ +Explore the most influential AI models across language, vision, audio, and generation tasks.
+ +A frontier multimodal model excelling at reasoning, coding, and creative tasks. Optimized for both text and vision inputs with reduced latency.
+ +Known for nuanced reasoning, long context handling (200K tokens), and strong safety alignment. Excels at analysis and coding.
+ +Google's most capable model. Natively multimodal — processes text, image, audio, and video in a single architecture with 1M+ context.
+ +The largest open-weight model. Competitive with proprietary models on major benchmarks. Fully customizable and self-hostable.
+ +European frontier model with strong multilingual capabilities. Efficient architecture with 128K context and tool use support.
+ +Open-source image generation using a Multimodal Diffusion Transformer (MDiT). excels at photorealism and prompt adherence.
+ +Segment Anything Model 2 — real-time video and image segmentation. Works across domains with zero-shot generalization.
+ +State-of-the-art speech recognition. Trained on 680K hours of multilingual data. Near-human accuracy on English transcription.
+ +Generates high-quality music from text descriptions. Uses a single-stage auto-regressive Transformer with codebook pattern modeling.
+ +A combination that understands text and images natively, then generates images with DALL-E 3. Full round-trip multimodal reasoning.
+ +Best-in-class artistic image generation. Known for stunning aesthetics, coherent text rendering, and photorealistic outputs.
+ +Open-reasoning model using reinforcement learning for chain-of-thought. Excels at math, code, and logical reasoning tasks.
+ +Click on any architecture to learn how it works, its strengths, and where it's used.
+ +Self-attention mechanism that processes all tokens in parallel. The foundation of modern LLMs.
+Convolutional Neural Networks extract spatial features through sliding filters. King of computer vision.
+Recurrent networks process sequences step-by-step with memory. LSTMs solved the vanishing gradient problem.
+Generates data by learning to reverse a gradual noising process. State-of-the-art for image/video generation.
+Two networks — generator and discriminator — compete in a zero-sum game to create realistic outputs.
+Sparse activation — only a subset of parameters processes each token. Enables massive models at lower compute cost.
+Toggle between two leading models to see how they stack up across key dimensions.
+ +Key milestones in AI model development from early neural networks to today's frontier models.
+ +Google introduces the transformer architecture, replacing RNNs with self-attention. This becomes the foundation for every major AI model today.
+OpenAI's GPT-1 and Google's BERT show that pre-training on massive text data produces general-purpose language understanding. The era of large language models begins.
+175 billion parameters. GPT-3 demonstrates that scaling model size dramatically improves capabilities, including few-shot learning without fine-tuning.
+OpenAI connects vision and language. CLIP learns visual concepts from text, and DALL-E generates images from descriptions — pioneering multimodal AI.
+Stable Diffusion democratizes image generation with open source. ChatGPT takes the world by storm, reaching 100M users in 2 months.
+Frontier models push boundaries. GPT-4 is multimodal. Claude introduces 100K context. Gemini processes text, image, audio, and video natively.
+Llama 3, Mistral, and DeepSeek prove open models can compete with proprietary ones. o1 and DeepSeek R1 introduce chain-of-thought reasoning.
+Models that can browse, code, and take actions autonomously. Sora and similar models simulate physical worlds. AI models become AI agents.
+Answer a few questions and we'll recommend the best AI model for your use case.
+ +Common questions about AI models, answered.
+ +AI models differ in their architecture (transformer vs CNN vs diffusion), training data, parameter count, training methods (supervised, RLHF, etc.), context window size, modalities supported, and optimization objectives. Even models with similar architectures can produce very different results based on how they were trained and fine-tuned.
+Open-source models (like Llama 3, Mistral) release their weights publicly, allowing anyone to download, run, fine-tune, and study them. Proprietary models (like GPT-4, Claude) are only accessible through APIs — you can use them but can't see how they work internally. Open models offer more control and privacy; proprietary models often have better performance and ease of use.
+Parameters are the numerical weights within a neural network that are learned during training. More parameters generally means the model can learn more complex patterns, but also requires more memory and compute. A model with 7B parameters has 7 billion learnable weights. However, architecture and training quality matter as much as raw parameter count.
+The context window is the maximum amount of text a model can process in a single conversation. GPT-4 handles 128K tokens (~96K words), Claude 3.5 handles 200K, and Gemini can handle over 1M. A larger context window means the model can reference more of the conversation, documents, or code at once, which is critical for long documents, codebases, and complex analysis.
+Mixture of Experts is an architecture where a model has many "expert" sub-networks, but only activates a few for each input. A router decides which experts to use. This means the model can have trillions of total parameters while keeping inference fast because only a fraction are active. Mixtral and possibly GPT-4 use this approach to balance scale with efficiency.
+Consider: (1) Task type — language, vision, audio, or multimodal. (2) Budget — proprietary APIs charge per token; open models require hardware but have no per-use cost. (3) Latency requirements — smaller models are faster. (4) Data privacy — self-hosted models keep data local. (5) Fine-tuning needs — open models allow customization. Start with benchmarks but always test with your actual use case.
+Reinforcement Learning from Human Feedback (RLHF) is a training technique where human evaluators rank model outputs, and a reward model is trained on these preferences. This is then used to fine-tune the language model via reinforcement learning. RLHF is what makes models like ChatGPT and Claude helpful, harmless, and honest — bridging the gap between raw language modeling and useful assistant behavior.
+