Today, we explore the AI of the non-LLM form.
Examples include:
Slides
The class plan:
- Ethics and LLMs: 2 remaining areas for discussion.
- What is the range of non-LLM AI?
To understand “AI without Large Language Models (LLMs),” it is helpful to view LLMs as just one specific tool in a vast toolbox. While LLMs have recently dominated the headlines due to their ability to generate text and code, they represent only a fraction of the artificial intelligence ecosystem.
AI without LLMs is primarily focused on predictive analytics, computer vision, robotics, and decision-making. It is the “invisible” engine that powers much of the modern world.
Here is a breakdown of what AI looks like without the hype of LLMs:
1. Predictive Analytics and Machine Learning
This is the workhorse of the industry. Instead of generating new text, these models analyze historical data to predict future outcomes or classify existing data. * Recommendation Engines: The algorithms driving Netflix, YouTube, Spotify, and Amazon are not LLMs. They use collaborative filtering and matrix factorization to predict what you will watch or buy next. * Fraud Detection: Banks use AI to analyze millions of transactions in real-time, flagging anomalous patterns (like a credit card being used in two different countries an hour apart) to prevent fraud. * Search Engines: Google’s core ranking algorithms rely on complex machine learning models to determine relevance and quality, long before they integrated conversational LLMs.
2. Computer Vision (The “Eyes” of AI)
Computer Vision is arguably the most mature and widely deployed form of AI. It allows machines to interpret visual information from the world. * Medical Imaging: AI algorithms analyze X-rays, MRIs, and CT scans to detect tumors or fractures with a speed and accuracy that often rivals human radiologists. * Autonomous Driving: While a car’s “brain” might use an LLM to navigate, its “eyes” use Convolutional Neural Networks (CNNs) to identify pedestrians, traffic lights, lane markings, and other cars in real-time. * Facial Recognition: Used for unlocking phones, airport security, and social media tagging.
3. Robotics and Control Systems
AI in the physical world is distinct from the digital world of text. This involves control theory and motion planning. * Industrial Automation: Robotic arms in car factories use AI to optimize their movements for speed and precision, welding and assembling parts. * Drones and Drones: Autonomous drones use AI to maintain stability, avoid obstacles, and map terrain without human input. * Warehouse Logistics: Amazon and other logistics companies use AI to optimize the paths of thousands of robots moving goods through a warehouse.
4. Reinforcement Learning
This is a type of AI where a machine learns by trial and error, receiving rewards for good actions and penalties for bad ones. * Gaming: Before LLMs took over the news cycle, AI like DeepMind’s AlphaGo and AlphaStar had already defeated world champions in complex games like Go and StarCraft II. * Resource Optimization: AI is used to optimize data center cooling systems (like Google’s), significantly reducing energy consumption by learning the thermal dynamics of the building.
5. Non-Text Generative AI
While LLMs generate text, other models generate different modalities of data: * Image Generation: Tools like DALL-E, Midjourney, and Stable Diffusion use “Diffusion models” to create images from noise. * Audio and Music: AI models (like Suno or specialized speech synthesis engines) generate music, sound effects, and realistic synthetic voices.
6. Symbolic AI and Expert Systems
Before the “Deep Learning” revolution (which spawned LLMs), the dominant form of AI was Symbolic AI. * Rule-Based Systems: These systems use a strict set of logical rules (If X happens, then do Y). They are still used today in complex financial trading systems, legal compliance software, and specialized medical diagnosis tools where accuracy and explainability are more important than creativity.
Summary
If LLMs are the new “front-end” interface for AI—providing a conversational, creative layer—the world of AI without LLMs is the “back-end” infrastructure. It is the mathematical engine used to see, predict, calculate, and physically manipulate the world.
Readings: