2.6 AI models historical
1 This is actually a useful list
because it roughly follows the historical evolution of AI models. For ZiptieAI, I’d treat most of these as: • interesting side roads while:
- Core NN
- CNN
- Transformer
- JEPA / World Models
are probably your main road.
RNNs / LSTMs (1990s-2010s)
Purpose: Process sequences one step at a time.
Examples:
- speech
- text
- time series
Mechanism:
- token 1 ↓ token 2 ↓ token 3.
- each step passes memory to the next.
- Problem: slow , hard to train, long-range context problems.
- Mostly replaced by transformers.
Autoencoders
Purpose: Compress data and reconstruct it.
Mechanism: input ↓ compressed representation ↓ reconstruction
Examples:
- anomaly detection
- dimensionality reduction
- feature extraction
Interesting historically because they introduced the idea of learning latent representations.
GANs (Generative Adversarial Networks)
Purpose: Generate realistic synthetic data.
Mechanism: Generator vs Discriminator
Examples:
- fake faces
- image generation
- deepfakes
Very famous around 2015-2020. Mostly superseded by diffusion models.
Graph Neural Networks (GNNs)
Purpose: Learn from graph structures.
Examples:
- social networks
- fraud networks
- chemical molecules
- knowledge graphs
Input:
- nodes
- edges
- relationships
These are still important. Drug discovery often uses GNN-like approaches.
Diffusion Models
Purpose: Generate images, audio, video.
Mechanism:
- add noise
- learn to remove noise
Examples:
- Stable Diffusion
- Midjourney-like systems
- image generation
Current state-of-the-art for many image-generation tasks.
JEPA / World Models
Purpose: Learn how the world works.
Mechanism:
- observe
- predict hidden structure
- learn representations
Examples:
- robotics
- autonomous systems
- planning
This is the area associated with Yann LeCun and is one of the reasons you became interested in NCU and robotic AI.
My recommendation for ZiptieAI
Main path: Core NN ↓ CNN ↓ Transformer ↓ JEPA / World Models ↓ Agentic AI
Side roads (learn later if needed):
- RNN/LSTM
- Autoencoder
- GAN
- GNN
- Diffusion
The reason is simple:
- RNNs -> mostly historical
- GANs -> mostly historical
- Autoencoders -> useful but niche
- GNNs -> important in some industries
- Diffusion -> important for media generation
- JEPA -> directly related to your robotic AI interests
So if your goal is:
- AI models
- → robotic AI
- → agentic AI
then I think your current roadmap is already pointing toward the most relevant technologies.
2 i just want to study the old list
because when i see these talked about in the future i will understand how they are irrelevant to what i am doing. or how they led to the tech that i am using.
RNN/LSTM
- Early sequence models.
- Used before transformers.
- Important historically.
- Mostly replaced by transformers.
GAN
- One of the first successful image generators.
- Interesting historical step.
- Largely replaced by diffusion models.
Autoencoder
- Learns compressed representations.
- Helped establish the idea of latent spaces.
- Many later architectures borrowed these ideas.
GNN
- Still relevant.
- Used when the data is naturally a graph.
- Not a competitor to transformers.
- Different problem class.
Diffusion
- Current dominant image generation approach.
- Not particularly relevant to PAL, agents, or workflow automation.
JEPA / World Models
- Potential future direction for robotics.
- Directly relevant to my robotic AI interests.
historical landmarks
↓
MLP / Core NN
↓
CNN
↓
RNN/LSTM
↓
Transformer
↓
Agentic AI
CNN
↓
GAN
↓
Diffusion
Transformer
↓
JEPA / World Models
↓
Robotic AI
26.0603 (v1 26.0603)