This was my first time experiencing Machine Learning on a large scale. long-horizon agent runs, Scheduling GPU jobs, Finetuning transformers on my local machine for 4 hours... these 3 days were insane. Learned a lot of stuff. Haven't had this much fun in ages.
@neural_avb Great video. Asking out of curiosity, would you still call it RL if you use the ground truths to generate the reward or would it fall more into supervised learning (but different from SFT as SFT trains on token-level loss)
@DevanshuXi I think you feel this way because you already might be familiar with some fundamentals, but for a student new to tech, I think focusing a little bit more on the fundamentals is better. What do you think?
@DevanshuXi IMO, technology keeps evolving. The more "fundamentally deep" a concept is, the more it lives. If you are someone who works within these high level concepts, you might need to switch technologies once they get obsolete, and knowing the fundamentals helps in switching easily.
@neural_avb now that I think of it, evaluating pretrained bases definitely seems challenging. The true capabilities of the model come out during SFT/RL
@fchollet I've reached the NLP section and I'm loving this book so far. The way you explain things by giving examples that are intuitive and easy to understand is amazing. Thank you for writing such a masterpiece!
@neural_avb Wow I did not know that. It's hard to find the self hosting part in their documentation from their main website. So you have a forever free self hosted web searching MCP for your agents am I right? Are all the features included like crawling, scraping etc.?
@neural_avb Just wrapped up Computer Vision basics in my deep learning journey. I learnt about CNNs, image segmentation and object detection. Thinking of moving to time series models next, but I'm also kinda inclined to making a CV project, so I'm searching for some ideas at the same time.
@DataChaz@openclaw Openclaw does not need hundreds of gb of RAM. The LLM that runs inside requires the memory that comes in a mac mini. You could run openclaw with a lightweight LLM too.