You've bolted sight onto a language model, watched video blow up the token budget, folded every modality into one backbone, learned what makes an embedding's cosine similarity mean something, and walked speech in — one chunk, one long file, one room full of speakers. Prove it stuck before the sound comes back out.
Eight questions. Score 80% or better to unlock streaming TTS, speech-to-speech, and the image- and video-generation lessons. Every option explains itself, and you can take another pass any time.
01 · Checkpoint
Perception & speech-in — VLMs, embeddings, ASR.
Vision-language and video models, omni-modal backbones, embedding architecture and throughput-bound inference, and the ASR trio: single-chunk latency, long-file stitching, and diarization.
Check your understanding0 / 8
Question 1
A vision-language model (VLM) is, at its core, built from:
Question 2
What is the central problem when a VLM ingests video, and the standard answer?
Question 3
What distinguishes an omni-modal model from a cascade of separate models?
Question 4
An embedding model is an encoder plus a pooling head. What makes its cosine similarity actually meaningful?
Question 5
Compared with LLM decoding, embedding inference at scale is best characterized as:
Question 6
In Whisper-style ASR on a 30-second chunk, which stage dominates latency, and why?
Question 7
Serving a 90-minute file means chunk → VAD → transcribe → stitch. Where does the hard engineering live?
Question 8
Speaker diarization answers "who spoke when?" Which describes the classical pipeline and its production integration?