Companion videos

Picking the right model for the job — companion videos

The article explains how to compare candidate models on one task whose result you can check. The videos provide model-selection examples and comparison methods to examine against that task; they do not establish a current provider ranking or a permanent choice for your work.

Primary pick

2:11:12
How I use LLMs

Andrej Karpathy

Karpathy spends explicit chapters on "Be aware of the model you're using, pricing tiers" and "Thinking models and when to use them," then keeps switching between ChatGPT, Claude, Gemini, Grok, and Perplexity throughout the rest of the walkthrough. Use those model-selection examples as context for the article's controlled comparison on your own checkable task. The named products and tiers remain dated examples.

What you should get from this: Learn to route work across fast, cheap, deep-reasoning and source-grounded tools instead of using one model for everything.

Watch or know first: Comfort using at least one mainstream chatbot and comparing outputs on the same task.

AI Expert note: The model-picker habits are valuable; exact model names, prices, context limits and rankings are not stable. Re-run the comparison on your own tasks before standardizing a team recommendation.

Open video page

Also worth watching

16:50
The New, Smartest AI: Claude 3 – Tested vs Gemini 1.5 + GPT-4

AI Explained

Older than the article (March 2024), but the methodology is what's useful: a single careful reviewer running the same hard tasks — OCR, theory of mind, instruction following, math — through three frontier models side by side and showing exactly where each one cracks. The model names are dated, the framework for comparing models is not.

What you should get from this: See a structured comparison method you can reuse when deciding which model is good enough for a task.

Watch or know first: Know that Claude 3, Gemini 1.5 and GPT-4 are no longer the frontier baseline.

AI Expert note: Treat this as a testing-method video, not a current ranking. The concrete winners are historical; the useful part is running the same examples, checking failure modes and resisting vague "best model" claims.

Open video page