46 AI Models Powering
Every App You Use
The complete map of every AI model and classical tool behind photo, voice, and video processing — with the real production apps using each one and the exact API payloads.
46
categories mapped
35
require AI models
11
are classical tools
3
media domains
Photo Processing
16 categories
◆ 11 AI Models
■ 5 Classical Tools
Voice & Audio
14 categories
◆ 14 AI Models
■ 0 Classical Tools
Video Processing
16 categories
◆ 10 AI Models
■ 6 Classical Tools
The rule nobody talks about
Machine learning isn't always the answer. 13 of 46 categories are best served by classical deterministic tools — same input, same output, no inference cost, no confidence score. Knowing when not to use AI is half the architecture. Voice/audio is the exception: all 14 categories require trained models.
Photo Processing
16 categories from compression to visual reasoning — tap any row to see the model and a real app using it.
Voice & Audio
14 categories — and unlike photo or video, every single one requires a trained AI model. No classical tools.
Video Processing
16 categories combining the complexity of photo and audio — from deepfake detection to semantic search.
What this map actually tells you
01
Voice is 100% AI — no exceptions
Every single voice/audio task requires a trained model. There's no deterministic shortcut for transcription, diarization, emotion, or deepfake detection. Budget accordingly.
02
Safety is almost always a TOOL, not a model
CSAM detection (photo and video) uses hash-matching tools — PhotoDNA, Thorn — not inference models. This is deliberate: deterministic matching = no false negatives, no confidence thresholds.
03
Embeddings are the secret layer
Semantic duplicate detection, visual search, and video search all share the same architecture: embed → store → retrieve → rerank. One pipeline, three product features.
04
Deepfakes need a model for each medium
Photo, audio, and video deepfakes each require a domain-specific model. Reality Defender, Pindrop, Hive Detect — different problem, different solution, even when the underlying threat is identical.
05
FFmpeg is still everywhere
Compression, transcoding, thumbnail candidate extraction — FFmpeg powers all three. The most deployed video tool in existence isn't AI. It's a 30-year-old C binary.
06
Provenance is the next mandatory layer
SynthID and C2PA are showing up in photo AND video. AI-content watermarking is transitioning from 'nice to have' to regulatory baseline. Build it in now.
Start a Critical Discussion
These questions don't have consensus answers. Share one to LinkedIn or X and see what your network actually thinks.
"Is AI infrastructure more like railroads in 1870 or the internet in 1999 — and does the distinction matter for your career bets?"
"Jensen says Layer 5 (applications) hasn't exploded yet. What's actually blocking it — talent, trust, or tooling?"
"The trades (electricians, cooling engineers) are the #1 AI job category by volume. Is this fact criminally underreported?"
Share this analysis
If this changed how you think about something, share it. The AI workforce conversation needs more data and less hype.