Advice for beginners in AI: How to learn and what to build | Lex Fridman Podcast

Implement small models on single GPU

Start by coding a simple LLM from scratch that runs on one GPU to understand pre-training, attention mechanisms, and supervised fine-tuning, rather than using complex production libraries.

Avoid Hugging Face Transformers initially

While Hugging Face is the industry standard for loading models, its codebase (400+ models, hundreds of thousands of lines) is too complex and intertwined for learning fundamentals; use it only for verification.

Self-verify through reverse engineering

Load pre-trained weights from Hugging Face into your scratch-built model and match outputs to unit test your implementation, providing verifiable rewards for correct architecture.

Embrace the struggle

Developing 'taste' requires struggling through mathematical derivations and debugging (like DPO algorithms) rather than using LLMs to skip steps, though AI can provide hints without spoilers.

Go narrow after fundamentals

Instead of trying to keep up with everything, specialize deeply in specific areas like character training (making models funny/sarcastic), RLHF, or reasoning models, as many topics have only 2-3 key papers.

Character training and model specs

OpenAI's published Model Spec reveals intended behaviors versus training failures, but methods for curating data to achieve specific personality traits remain underexplored research areas.

RLHF and unsolvable preferences

RLHF assumes preferences can be quantified and aggregated into single values, but this compresses complex philosophical and economic trade-offs (related to social choice theory) that make the problem fundamentally unsolvable.

Evaluation as low-compute entry

Researchers without GPU clusters can build careers by creating evaluations that expose failures in frontier models (Claude, GPT-4); if labs address your findings in release notes, it provides career momentum without massive compute.

Academia vs. Industry compensation

Frontier labs like OpenAI offer average compensation exceeding $1 million annually in stock, while PhD students earn essentially nothing, creating extreme opportunity costs for academic research despite greater publication credit.

The 996 culture warning

Leading AI labs increasingly adopt '996' culture (9am-9pm, 6 days/week) with intense leapfrogging competition, causing burnout, whereas professors report higher average happiness despite grant-writing stress.

Publication vs. proprietary work

Academia offers clear portfolio building and public recognition, while industry researchers become 'cogs in the machine' with limited publication rights but massive real-world impact.

Compute constraints define options

Academics with limited resources should focus on inference-only research, fine-tuning 7B parameter models with LoRA, or long-term bets on what matters in 10 years, rather than training foundation models from scratch.