Learn AI
◆ TIER 1 · ORIENT

Multimodal & Voice

Same next-token loop, richer input and output. Images and audio get encoded into tokens the transformer already consumes — so the new skill isn't the model, it's knowing when a multimodal model beats a classic OCR / ASR pipeline. That call is all this page installs.

IF YOU KNOW APIS
Multimodal is the same model that now also accepts file uploads, not just text — like an API endpoint that takes images and audio, encoded to tokens the same transformer already consumes.
The pain: you built an OCR-then-LLM pipeline; it drops tables, mangles handwriting, and breaks every time the document layout changes.
01 · EVERYTHING BECOMES TOKENS

Tap a stage. The whole trick is that pixels and sound become tokens too.

One architecture, richer input — vision and voice are not new models, just new token streams into the loop you already know.

02 · VISION: MODEL vs OCR PIPELINE

Tap a card to flip it. Two honest columns.

◆ USE THE MODEL
▣ USE A PIPELINE

The classic line: reach for the model when the input is messy or needs reasoning, the pipeline when it's clean and high-volume.

03 · VOICE: THE ASR → LLM → TTS LOOP

Tap a stage. Three boxes — but the words were never the hard part.

The latency budget is everything: every hop adds delay, and a beat too slow feels broken. The hard part is turn-taking and interruptions — knowing when to stop and listen — not transcribing the words. Newer speech-to-speech models collapse the three boxes into one, cutting the lag.

04 · SHOULD I GO MULTIMODAL?

Pick the situation you actually have. Get the boring, correct answer.

MY SITUATION
THE CALL
A vision model
Unpredictable, human-shot images need reasoning, not extraction. Send the pixels straight to a multimodal model and let it read like a person.
vision and voice are the same loop with richer tokens — the skill is the model-vs-pipeline call · last reviewed 2026-06
Fine-tuningFoundations