CS 471/671 (How to grow language models) - #18 - Fall 2026

Item: CS 471/671 (How to grow language models) - #18 - Fall 2026
Rating: 4
Author: PostSingularity Institute

Title history

CS 471/671 (How to grow language models) - #18 - Fall 2026 — observed on 20261104-0808 UTC (YouTube)

by Center for Language & Speech Processing (CLSP), JHU · View the channel on PSI

Formal & Physical Sciences AI — Models & Research UYQArtificial intelligence (AI)UYQLNatural language and machine translation

🎙 Center for Language & Speech Processing (CLSP), JHU 👥 4K 📅 November 4, 2026 ⏱ 47 min 👁 20 📄 tutorial 🧭 2026-11-04

Available in: English (current) Français

Keywords

LoRAinstruction tuningpost-trainingfine-tuningreinforcement learning

Summary

This is the 18th lecture of a graduate-level course on self-supervised learning at Johns Hopkins University. The professor reviews previous adaptation methods (supervised fine-tuning, head tuning, and LoRA) and analyzes their computational trade-offs. He clarifies that LoRA reduces memory usage but not FLOPs, and that the efficiency gains come from allowing larger batch sizes. He then contrasts LoRA and head tuning in terms of trainable parameters, noting that LoRA’s parameter count depends on rank and placement, and encourages empirical verification. Next, he introduces the concept of post-training, which aims to make a pretrained model broadly useful and aligned with human values. The lecture covers instruction tuning (essentially SFT on instruction-response pairs), discusses its limitations (annotation variability, mimicry, off-policy data causing hallucinations), and then pivots to reinforcement learning, explaining its origins in cognitive science and its role in post-training. The professor emphasizes that post-training practices are messy and evolving, with a shift toward verifiable feedback (e.g., code execution, math proof checking) and AI-generated feedback. He also mentions multi-tenant serving as a practical extension of LoRA for personalized models.

177 words

Critical Evaluation

Value of the Information & Strength of the Argument

The lecture provides valuable clarifications on the efficiency of adaptation methods. The professor argues that LoRA’s benefits are primarily in memory, not computational speed, and supports this with a back-of-the-envelope calculation comparing trainable parameters for head tuning vs. LoRA. He openly expresses uncertainty when appropriate, which strengthens credibility. The argumentation is logical, step-by-step, and grounded in concrete examples, such as a model with specific dimensions. The discussion of post-training is well-motivated, contrasting pre-training and post-training, and highlighting the non-standardized nature of current practices. The value lies in demystifying common misconceptions and providing a framework for thinking about adaptation and post-training.

Scientific Rigor, Source Quality, Title Accuracy

The scientific rigor is high for an educational setting: the professor references a specific paper he co-authored (FLAN) without giving the full citation, and points to HuggingFace’s adapter list for further exploration. The title accurately matches the content, as the lecture is part of a course on growing language models. The sources cited in the description are limited to the course website, which contains additional materials. The lecture itself relies on established concepts in the field, with no apparent fabrications. However, since it’s a live lecture, some claims are not fully verified, and the professor’s admission of uncertainty (e.g., on FLOPs comparison) reflects intellectual honesty. Overall, the title is appropriate and the content is reliable for pedagogical purposes.

233 words

Title / Content Match

The title accurately reflects the content: the lecture is part of a course on language models and covers aspects of growing/training them, focusing on adaptation and post-training.

Quality & Reliability

8/10

Lecture by an expert in the field, based on established research and academic knowledge. The professor clearly explains concepts, acknowledges uncertainties (e.g., 'I'm only 95% sure on this claim'), and provides concrete numerical examples. The content is educational and rigorous, but not peer-reviewed and presented informally.

Key Moments

Markers derived by PSI from the transcript: the creator did not define chapters.

Cited Sources

Course website: Self-Supervised Learning (Fall 2026) — Course materials and syllabus

Contribution & Novelties

This lecture adds pedagogical value by systematically comparing adaptation methods and clarifying common misconceptions. It emphasizes that LoRA’s efficiency is memory-based, not FLOP-based, and that its parameter count depends on rank and placement, requiring empirical testing. It also frames post-training as a broader concept than task-specific adaptation, highlighting the shift toward verifiable feedback and the messy reality of industrial practices. The discussion of multi-tenant serving as a practical extension of LoRA is a forward-looking insight.

Pour aller plus loin :

LoRA: Low-Rank Adaptation of Large Language Models — The original LoRA paper, foundational for understanding the method discussed.
Finetuned Language Models Are Zero-Shot Learners (FLAN) — The paper on instruction tuning scaling, likely referenced in the lecture.
Training Language Models to Follow Instructions with Human Feedback (InstructGPT) — Key paper on RLHF for post-training and alignment.

136 words

Radar Profile

The profile is fairly balanced with high scores across all dimensions, reflecting a technically dense and reliable lecture. The low number of views (20) suggests a niche audience, but the content quality is high, with a strong quantitative information score and solid rigor.

Reliability 8/10