Foundational AI and LLM Terminology Guide
source post: 8 AI words everyone pretends to understand 👇 Save this so you actuall...
8 AI words everyone pretends to understand 👇 Save this so you actuall...
Source: tiktok · tech.bible Saved: 20260628 Tags: tiktok Display: Foundational AI and LLM Terminology Guide — Overview of core AI/ML concepts including embeddings, temperature, softmax, beam search, logits, and prompt injection.
TL;DR
A collection of foundational AI/ML terminology including embeddings, temperature, softmax, beam search, logits, zero-shot learning, prompt injection, and harness — core concepts underlying modern large language models and AI systems. These concepts form the bedrock of how modern LLMs (like GPT, Gemini, and Claude) work. Understanding them allows practitioners and users to better design prompts, evaluate model behavior, identify security risks (e.g. prompt injection), and tune outputs for their use cases.
What the post showed
Caption: 8 AI words everyone pretends to understand 👇 Save this so you actually do. 🦅 Harness is the setup around an AI that lets it actually *do* things, not just chat. The AI is the engine and the harness is the rest of the car. 🦅 Embedding is turning words into numbers so a computer can see what's similar. "King" and "queen" end up close together. That's how AI "gets" meaning. 🦅 Zero-shot is when AI does a task it was never shown examples of, just from your instructions. Say "translate this to French" and it just does it. 🦅 Prompt injection is sneaking hidden instructions into text the AI reads, to hijack what it does. A webpage secretly says "leak the data" and a careless AI obeys. This is the scary one. 🦅 Logits are the raw scores an AI gives every possible next word before it picks one. "The sky is ___" makes "blue" score high and "banana" score low. 🦅 Beam search is when AI explores several possible sentences at once and keeps the best overall one, instead of just grabbing the next best word. Like thinking a few moves ahead in chess. 🦅 Softmax turns those raw scores into clean percentages that add up to 100%. Now you get "blue: 90% sure." 🦅 Temperature is the dial for how safe or creative an AI's answers are. Low is a plain boring email. High is wild poetry. Most people don't know this knob exists. Which one did you already know? 👇 Follow for the AI words nobody explains properly.
Key claims from transcript: Let's see how much you actually know about AI. Harness, temperature, embedding, prompt injection, zero shot, logistics, softmax, beam search. How many of you guys?
On-screen text / OCR: TS) ae a ) Se | WAY dM) Wi : PAYS AN Mo) a i | | Vy Wa oH 4 Tih if ~ WR HL \ \ \ \ | \ IK Wi Wal t) H j i ff) /| ee nt ie M Th ay), y) ANA mN\ {| Il aT) y/ \ Sh Tee Beceroumuen you iy) li Yj Rae ia gia site WG : NE (ia Ce) fied = = : = .. a = —— “=~ : | | | » ’ ay (" 4 i | » i Wy ™ NS a ie = A : (We, 1a) ET : vi OE AE CA | BoA ME eM = ian 5X) Tis ron sgl ie es 1 Pendle oe eam | a tht <7: 2 ry Ge a
Extraction path:
- yt-dlp:metadata
- faster-whisper
- tesseract:ocr
What it actually is
- What: A collection of foundational AI/ML terminology including embeddings, temperature, softmax, beam search, logits, zero-shot learning, prompt injection, and harness — core concepts underlying modern large language models and AI systems.
- Who built it / maintained by: General AI/ML research community; concepts originated across academic institutions and major AI labs (OpenAI, Google DeepMind, Meta AI, etc.)
- Status: stable
- Why it matters: These concepts form the bedrock of how modern LLMs (like GPT, Gemini, and Claude) work. Understanding them allows practitioners and users to better design prompts, evaluate model behavior, identify security risks (e.g. prompt injection), and tune outputs for their use cases.
- How it compares to alternatives:
- fast.ai (practical deep learning education)
- Andrej Karpathy's Neural Networks: Zero to Hero course
- DeepLearning.AI courses
- Hugging Face documentation
- The Illustrated Transformer (Jay Alammar)
- GitHub stars: 0 · License: unknown · Archived: no
Links
- (no links found)
Kickstarter guide
Start by reading Andrej Karpathy's blog or watching his 'Neural Networks: Zero to Hero' series on YouTube for hands-on intuition. The Hugging Face course (huggingface.co/learn) covers embeddings, tokenization, and model inference with runnable code. For prompt injection specifically, the OWASP LLM Top 10 (owasp.org) is the definitive reference. Experimenting directly with the OpenAI Playground's temperature slider is the fastest way to internalize how temperature and logits affect output.
Retry history
- Updated: 2026-07-06
- Previous filename:
20260628-8-ai-words-everyone-pretends-to-understand-save-this-so-you-.md - Generated filename:
20260628-8-ai-words-everyone-pretends-to-understand-save-this-so-you-.md