Congrats! Big props unifying vision language datasets and making it easy to access them. We also recently released a massive VQA dataset of 30M samples(22M filtered), spanning 42 countries and 39 languages. The dataset contains factual visual questions and answers about world
@jeande_d
-
New Stanford Course on Transformers and Large Language Models
By
–
[New course] Transformers & Large Language Models, CME 295 Stanford Yet another excellent course on transformers and large language models with a consolidated curriculum starting right away from transformers. Topics/lectures:
– Intro to transformers
– Transformer-based models & -
CMU Advanced Courses: NLP, Inference, and LLM Systems
By
–
Exactly the same 2 courses I thought seeing the post. There are more across other schools, look at:
– advanced NLP at CMU(
https://
youtube.com/playlist?list=
PLqC25OT8ZpD2-RuhyacIsODl5iJVgMjI3
…)
– new inference class of its kind in academia CMU (
https://
youtube.com/playlist?list=
PL8PYTP1V4I8DY15Ob83nmwBAVyNtUQadX
…)
– LLM systems CMU(
https://
llmsystem.github.io/llmsystem2025s
pring/
…)
– large -
Karpathy’s Nano Repo Brings Joy to X Timeline
By
–
When Karpathy drops a nano repo, X timeline is great again! Everyone is basically happy.
-
Learning from Post-Training Guides and Continuous Development
By
–
Happy to share. That's a tough question and I don't have an answer… But you can learn so much about postraining using the guide and keep the learning trajectory going.
-
Comprehensive Guide to LLM Post-Training: SFT, RLHF, and RL Algorithms
By
–
An excellent technical guide on LLM post-raining covering SFT(supervised finetuning), RL rewards such as RLHF/human preferences, RLAIF/constitutional-AI, RLVR/verifiable outcomes, process-supervised and rubric rewards. Also covers common RL training algorithms from PPO, GRPO, and
-
Andrew Ng’s Deep Learning Course 2025 Lectures Now Available
By
–
Deep learning course by legendary Andrew Ng is public now. New 2025 lectures are out!!
-
LoRA Matches Full Fine-Tuning Performance With Proper Implementation
By
–
LoRA Without Regret – Recent Blog from Thinking Machines TL/DR: LoRA actually matches full supervised fine-tuning(SFT) when you get the details right. Nearly same sample efficiency, loss(or better), same final performance. Some plain points:
– Apply LoRA to ALL layers, -
Multimodal Models and Vision-First Foundation Model Scaling
By
–
Multimodal, world models, recent works on scaling up vision-first foundational models…You have unique views/experience on these things.
-
ImageNet’s Human Baseline and Deep Learning’s Foundation
By
–
ImageNet is also the only large-scale benchmark that had real accurate human baseline accuracy, Karpathy was the human baseline hh. I read/watched somewhere that he went through every image and label in ImageNet and that's very very cute. Without ImageNet efforts, deep learning