AI summary by TwinMind
Luis Serrano + Josh Starmer Q&A Livestream!!!
StatQuest with Josh Starmer
57m 30s
Introduction and Global Audience
- Live stream collaboration between Josh Starmer (StatQuest) and Luis Serrano (Serrano Academy)
- 56+ viewers from worldwide locations: Missouri, Colombia, India, Switzerland, Poland, Greece, Saudi Arabia, Brazil, Italy, Argentina, Taiwan, Pakistan, Bangladesh, Tunisia, South Africa, Syria, Palestine, Morocco, Vietnam, Uzbekistan
- Josh in Chapel Hill, North Carolina; Luis in Toronto, Canada
Getting Jobs in AI
- Background diversity: AI welcomes people from all fields - psychologists, philosophers, various disciplines
- Learning resources: Coursera, Udacity, YouTube channels, hands-on practice through Kaggle
- Portfolio development: Essential to create solid GitHub repositories demonstrating:
- Coding skills and commenting ability
- Collaboration on group projects
- Self-initiative and teamwork
- Professional presentation as resume substitute
- Interview preparation: HackerRank, Kaggle, LeetCode, Euler Project recommended
- Career paths: Josh's experience at Google, Apple, Cohere; Luis's zigzag path through different companies
StatQuest Character Origins
- Normal Saurus: Created from t-shirt design by friend's young daughter - dinosaur with normal curve as backdrop
- Stat Squatch: Represents Josh's learning persona - asking basic questions when learning new material
- Suggested new character: "Less Monster" with logarithm neck, living in Scotland
Bessel Correction (n-1 vs n) Explanation
- Traditional problem: Why divide sample variance by n-1 instead of n when estimating population variance
- Luis's alternative approach: Define variance as average squared distance between pairs of points
- N points can form n × (n-1) unique pairs (excluding self-pairs)
- Dividing by n × (n-1) naturally produces the correction term
- Makes Bessel correction intuitive rather than mathematical derivation
- Upcoming video on Luis's channel explaining this concept
Eigendecomposition in Machine Learning
- Primary application: Principal Component Analysis (PCA)
- Purpose: Dimensionality reduction and denoising
- Process: Takes high-dimensional data (multiple measurements per subject) and reduces to handful of useful eigenvectors
- Modern implementation: Now typically done with Singular Value Decomposition rather than eigendecomposition
- Related techniques: UMAP, t-SNE use PCA-like denoising steps for nonlinear visualization
Multimodal LLMs and Agents
- Current limitation: LLMs only do text-in, text-out communication
- Agent evolution: Moving beyond conversation to action-taking
- RAG as first agent: Retrieval Augmented Generation - searches databases/internet before responding
- Tool integration: Agents can use calculators, compilers, email, calendar apps
- Multimodal workflows: Combining images, sounds, text in chain-of-thought processes
- Future direction: Agents collaborating with each other in complex workflows
Transformers in Omics Technology
- Omics definition: Large-scale biological data (genomics, transcriptomics, proteomics)
- Josh's background: Former genetics professor at UNC Chapel Hill
- Application example: Luca Pinello's Harvard research using LLMs to generate gene control sequences
- Process: Training BERT-like models on genomic sequences (A, T, G, C patterns)
- Capability: Creating novel sequences to turn genes on/off in specific conditions
- Medical relevance: Potential cancer treatment by controlling cell identity
LoRA (Low-Rank Adaptation)
- Problem: Fine-tuning large models (375B+ parameters) is expensive and time-consuming
- Solution: Create small adjunct models with few parameters for specific tasks
- Approach: Matrix factorization - represent large 10×10 matrix as 10×2 and 2×10 matrices
- Benefit: One foundation model + multiple small specialized adapters
- Applications: Customer service, legal assistant, translation - each with dedicated LoRA adapter
XGBoost Implementation Details
- Variable selection: Both approaches possible
- Original paper: test every variable for each split
- Practical implementation: random subset like Random Forest
- Configuration: Parameters control how many variables used per tree
- Hybrid approach: Combines XGBoost boosting with Random Forest-style randomization
Data Quality and Hallucinations
- Xerox problem: Concern about AI-generated content contaminating training data
- Hallucination nature: Not a bug but a feature - models always hallucinate, sometimes accidentally correct
- Solution approaches: RAG, heavy fine-tuning, tool use
- Federated learning: Generate synthetic data to protect privacy while training
Model Size Considerations
- Traditional advice: "Too big with regularization" better than "too small"
- Analogy: Oversized pants can use belt, undersized pants unusable
- LLM context: Decision primarily driven by budget and computing resources
- Recommendation: Go as big as budget allows, then optimize
Domain-Specific LLM Applications
- RAG limitations: Good for retrieval but struggles with comparative analysis
- Data enhancement: Creating summary pages and fact sheets improves RAG performance
- Medical example: Skin cancer classification using image databases
- Key insight: Enhancing data often more effective than enhancing models
Career Advice for Non-AI PhDs
- Opportunity: Huge head start in applying AI to non-tech fields
- Advantage: Domain expertise + AI skills = powerful combination
- Recommendation: Dive in immediately, learn AI fundamentals
- Market position: Be the forerunner in your field's AI adoption
- PhD value: Demonstrates deep knowledge and persistence/grit
Hosting and Deployment
- Recommendation: AWS for permanent deployment due to widespread adoption
- Integration: Works with multiple LLM providers (Anthropic, Cohere, OpenAI)
- Cost consideration: Bulk negotiations possible for long-term hosting
- Development: Lightning AI offers free GPU credits for experimentation
Internship Strategies
- Dual approach: Apply to both large companies (Google, Facebook, Apple) and startups
- DIY internships: Look for local opportunities requiring computer help
- Example project: Neonatal ICU nutrition calculator - high-impact, resume-worthy
- University connections: Ask professors for summer research opportunities
- Impact focus: Projects helping others provide infinite energy source
Future Plans
- Next session: Early June 2024
- Question submission: Post on social media (YouTube, Twitter, LinkedIn) for advance consideration
- Format: Continue collaborative Q&A format
- Preparation: Advance questions allow better preparation and RAG processing
