AI use cases
3442 ways people actually use AI, compiled from the videos in this
library. Each names the job and why it is worth doing, and links to the
moment in the source video.
-
Code generation and drafting
— AI-generated responses can provide inspiration or draft code that developers later modify.
-
AI-generated content detection
— Tools attempt to determine whether text, audio, images, or video were created or assisted by AI.
-
Cybersecurity evaluation
— Models are instructed to remove guardrails and attempt attacks so evaluators can assess their worst-case behavior.
-
Role-playing and text adventure simulation
— AI agents interact with other AI agents and tools inside a fictional scenario.
-
Generate bulk directory content
— Accelerate content production while leaving human curation, organization, validation, and quality control to the operator.
-
Natural-language reporting
— An LLM can turn calculated results and identified trends into a clean natural-language report.
-
Multi-step workflows
— Agents can support flexible multi-step workflows.
-
Code generation
— Generative AI can generate code from complex inputs.
-
Large document summarization
— Generative AI can transform and summarize unstructured documents.
-
Question answering over document sets
— Generative AI can support retrieval-augmented generation over unstructured documents.
-
Customer support
— An AI agent can understand intent, retrieve relevant knowledge, and generate a unique reply.
-
Reinforcement fine-tuning
— Train a model to make high-scoring answers more likely using candidate sampling and a grader.
-
Model compression through distillation
— Use outputs and reasoning traces from a large teacher model to fine-tune a smaller model.
-
Procedural task execution
— Agent skills package instructions and tools so a general-purpose model can perform specialized procedures.
-
Retrieval-augmented generation
— Retrieve relevant documents at query time and provide them to the model in the prompt.
-
Legal document and legal task performance
— Fine-tuning can specialize a general-purpose model for legal work.
-
Compress the boring parts of learning and debugging
— Leave human mentors more time for guidance that requires human judgment.
-
-
Critique designs and identify edge cases
— Improve planning without delegating architectural decisions to AI.
-
Brainstorm multiple design approaches
— Explore alternatives before implementation.
-
-
Handling ambiguity and conflicting evidence
— The AI should preserve multiple possible answers when the underlying documents support them instead of forcing a single answer.
-
Question clarification
— A clarification loop identifies ridiculous, malformed, or underspecified questions and prompts the user to rephrase or add details.
-
Answering questions with retrieved document information
— A RAG system uses vector database search results and the original user question to provide an answer through an LLM.
-
Triage patients
— Represents a high-risk clinical use case requiring traceable explanations, provenance, testing, and reliability.
-
Recommend a television show
— Provides an example of baseline and enhanced explainability based on prior viewing history and viewing patterns.
-
Guess a person's job from a bookshelf image
— Demonstrates pattern recognition without sufficient context or relationships; the AI guessed NSA employment from cryptography books.
-
Build an AI system
— Illustrates that AI can be used to build AI systems easily, though not necessarily in a risk-appropriate way.
-
Generate boilerplate and assist with syntax
— Reduce time spent on repetitive parsers, regular expressions, and unfamiliar configuration syntax.
-
Repair broken code
— Produce a corrected version of broken code such as Python.
-
-
Explain existing code
— Help developers understand legacy or unfamiliar code more quickly.
-
Generate code from natural-language requests
— Translate requirements such as returning a filtered, sorted, paginated list of users into working code.
-
Automate cyberattacks
— AI agents may increase attack volume and velocity and lower the barrier of entry for less-skilled attackers.
-
Drive ransomware operations
— JADEPUFFER was described as an LLM-driven operation that gained access, encrypted data and generated a ransom note.
-
Identify software vulnerabilities
— AI models can help defenders find zero-day vulnerabilities proactively, but attackers can also use the capability to discover and exploit vulnerabilities.
-
Analyze user interactions and model outputs in real time
— Classifiers are used to detect and block potentially harmful requests and outputs.
-
Adversary simulation workflow automation
— Perform routine steps so human teams can focus on novel attack chains and higher-value work
-
Incident-response automation
— Handle low-hanging tasks, perform data gathering, provide direction, and assist with decoding or decryption
-
Threat intelligence processing
— Scrape indicators of compromise, aggregate large amounts of data, and help connect signals
-
Adversary simulation and offensive security research
— Amplify the speed and intensity of attack simulations and help identify vulnerabilities and attack paths
-
Speculative decoding
— Use a smaller draft model to propose tokens that a larger model verifies in one forward pass, reducing interactive latency.
-
Chunked prefill
— Interleave prompt processing with decoding to reduce token-stream stuttering and improve throughput.
-
Prefix caching
— Reuse KV-cache memory for requests with shared system prompts and avoid repeated prefill computation.
-
Paged attention
— Allocate KV-cache memory in fixed-size, non-contiguous pages to reduce fragmentation and improve GPU utilization.
-
KV caching
— Reuse previously computed key and value matrices during autoregressive token generation.
-
LLM inference
— Generate tokens from user prompts across prefill and decode phases.
-
Represent populations at scale
— Provide a scalable representation of people's perspectives for decision-making.
-
Generate and validate hypotheses
— Produce large numbers of hypotheses and compare their predicted outcomes with events in the real world.
-
Run counterfactual simulations
— Show how people, audiences and markets might respond to alternative decisions, products, strategies or policies.
-
Create agents with memory, planning and reflection
— Enable agents to remember interactions, plan activities and form higher-level views of their experiences.
-
Simulate human behavior
— Model people's values, preferences, biases, tastes, mistakes and actions in different contexts.
-
Quantum machine learning
— Combine qubits and classical techniques to implement quantum machine learning.
-
Regulatory compliance
— Finding applicable standards and generating evidence tables more quickly.
-
Software development
— Supporting coding; Max Hodak says he has written little source code in the previous six months.
-
Internal company operations
— Making company context and operational data available to AI agents through Helix.
-
Hiring assessment
— Designing take-home assignments that are resistant to AI assistance and can be scored with measurable outputs.
-
-
Storyboarding and filmmaking support
— Help filmmakers accelerate pre-production while preserving creative control
-
Routing requests across multiple models
— Combine models with different strengths and provide a virtual model endpoint