CLM-8B: Contrastive Language Models
Read the Portuguese translation →
Author: Prompt Engineering · Watch on YouTube
Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.
Original language: English
Original video: CLM-8B: Contrastive Language Models · Channel: Prompt Engineering · 00:30 · Watch on YouTube
Topics: AI agents, Data, databases and embeddings, AI models
Summary
Section titled “Summary”The video introduces CLM, a contrastive language model from Stanford and Nvidia research, and claims it can be up to nine times faster than comparable models. Unlike a generative model that processes the situation and every option together, CLM encodes the situation and candidate actions separately, then selects the closest match. Because actions often remain the same across steps, their embeddings can be cached, reducing each new step to one embedding and a dot product. The video also notes that CLM does not compare options side by side.
Key points
Section titled “Key points”- CLM is described as a contrastive language model from Stanford and Nvidia research.
- The video claims CLM is up to nine times faster than comparable models.
- CLM is described as having around eight billion parameters.
- A generative model processes the situation and all options together on each call.
- CLM uses one encoder for the situation and another for each possible action.
- CLM selects the action whose embedding is closest to the situation embedding.
- Action embeddings can be cached when the actions do not change between steps.
- At a thousand options, the video reports 44 milliseconds for CLM versus 579 milliseconds for Jeff.
- Because options are encoded separately, CLM does not compare them side by side.
Chapters
Section titled “Chapters”| Start | Chapter | Summary |
|---|---|---|
| 00:00 | Introducing CLM | The video introduces CLM and contrasts its claimed speed and size with Jeff-style models. |
| 00:30 | How CLM selects actions | CLM encodes the situation and each action separately, choosing the closest action embedding. |
| 00:30 | Caching and the trade-off | Cached action embeddings speed up repeated decisions, but CLM does not compare options side by side. |
Quotes
Section titled “Quotes”It’s called CLM, a contrastive language model from Stanford and Nvidia research. — 00:00
CLM never generates an answer at all. — 00:00
Every new step is just one embedding plus a dot product. — 00:30
CLM never compares options side by side. — 00:30
Concepts and entities
Section titled “Concepts and entities”CLM, CLM 8B, CLIP, Stanford, NVIDIA, embedding, produto escalar
Full transcript (original language)
Section titled “Full transcript (original language)”00:00 After Jeff, everybody’s talking about system one models, but there’s a new model that was completely different. It’s called CLM, a contrastive language model from Stanford and Nvidia research. And they’re claiming it’s up to nine times faster than Jeff. So, how is it that possible? It’s the same size as most Jeff clones, at around of eight billion parameters. Now, here’s the difference. With a Jeff style model, your situation and all your options go through the model together on every single call. More options means more to read. CLM never generates an answer at all. The easiest mental model is clip
00:30 but for decision. One encoder reads the situation, another reads each possible action, and the answer is whichever action lands closest. Now, here’s the trick. Your actions usually don’t change from step to step. So, CLM embeds them once and caches them. Every new step is just one embedding plus a dot product. At a thousand options, that’s 44 milliseconds against 579 for Jeff. But, there is a catch. Because each option is encoded on its own, CLM never compares options side by side. To learn more, watch this video.