General course chat

Part 1 2022
83

Hello. If one were to spend around 4 hours a day on this course, how long do you think it would take to complete?

Kind regards,
Ben

84

Three weeks. So you would be almost finished if you started the day you asked :slight_smile:

3 Likes
85

Thanks!

86

Greetings all! Not sure if this is the best place to ask this question, but figure it’s “general” enough. I’ve been lurking around fast.ai for sometime even working on some lessons and such. In recent days I’ve become really enamored with the capabilities of LLMs and the likes of GPT-4 for “copilot” usage, etc. Part 2 of the course looks awesome working with Stable Diffusion, but does it give any insight into how one would develop/train an LLM for use as a personalized copilot? I’d be grateful for any feedback from those that might have completed Part 2 or anyone that has knowledge of this. Thanks in advance for any help!

Brian

1 Like
87

In case anyone else is wondering about this, Jeremy’s appearance on this episode of the Gradient Dissent podcast, seems to have answered the question. This is discussed around the 32:00 minute mark. Paraphrasing, it seems like working with LLMs might come in a future part of the course as it’s different than image generation models. Although it was back in January, I assume this to still be relevant today. However, it does appear that Stabiliy AI has recently released its own LLM which perhaps may be a bit of foreshadowing regarding what students might get to tinker with soon. Cheers.

Brian

88

In meantime, the Cluster of Stars study group is currently going through NLP & LLM. You are welcome to join.

1 Like
89

Thank you Allen for the invite! I’ll take a look!

Brian

1 Like
90

Hi guys,

I made a notebook with a nice widget-thing to extract transcriptions from Jeremy’s videos and make them available for gpt prompt engineering - it’s called ‘Lazy Student’. There’s also an (ai-)auto-chapter feature and everything is yaml, oh yeah, it’s bilingual too.

Have a look. I have an animated gif on the github page.

enjoy

91

Hi @jeremy,

FYI

Chris Lattner: Future of Programming and AI | Lex Fridman Podcast [#381] / 2 jun 2023

see also timestamps at [2:25:30]- Jeremy Howard
Discussing Mojo

2 Likes
92

I didn’t know this thread existed! :face_with_monocle:

93

Hi Everyone, I have a question about lesson 1.

In the following example of birds vs forest prediction, why does it always return ‘bird’, even if the probability is low? thank you in advance for your help.

is_bird,_,probs = learn.predict(PILImage.create(‘bird.jpg’))

regards, jlam

94

This is the most common confusion from the first lesson. To make it easier to digest on a a first reading, is hardcoded to always output the probability of a bird (i.e. probs[0] is always first label) even when the prediction is not a bird. You need to make use of that ignored result…

Change… is_bird, _, probs =
to… pred, pred_ndx, probs =

95

Hello everyone,

Posting this here in case it’s useful for someone as beginner as I am (and also because I’m quite happy my Docker image got leaner).

I’ve been struggling deploying a Docker image that is not too large for a web app’ that provides a predict endpoint for inference using a fastai exported model; I’ve solved this by :

  • remembering that GPU is only for training, so I’ve installed the CPU version of the lib (there is a widget that can help you do that on the official website of the lib)
  • downgraded the Python version/base image to python:3.8.17-slim (I’ve been using a 3.12) before that otherwise the install would fail
  • upgrading pip setuptools wheel before installing torch

I’ve went from a 3.5GB image in the Google Artifact Registry to a 709MB one ! (nice storage savings here). Link to the PR => Merge pull request #45 from yactouat/cicd/leaner-image · yactouat/SWS_waste-sorter@0cf0b6a · GitHub

1 Like
96

Hi all. I know that LLMs aren’t really part of the fast.ai curriculum but as I wrestle with hype vs facts, I can’t help but think about the ethics surrounding the use of a few, mostly proprietary models, with the approach taught here of building our own models, fine-tuned to each domain and unique data set. After reading Jeremy’s recent blog post on AI Safety and the Age of Dislightenment, I struggle with how much of my time to invest in leveraging the ecosystem and evolving stack around LLMs or whether to invest more time in the fast.ai courses. I know they’re different approaches and useful for different scenarios, but I worry about the black boxes that are LLMs. Further, it’s concerning to me that data security and privacy concerns aren’t more prevalent in the discussions of adopting usage of these models. I guess in other words, a lot of what I’ve learned from this community, and Jeremy and Rachel in particular, seem to fly in the face of the hype machine around LLMs. I’m just wondering if others are wrestling with the same conflicts and how they are navigating these waters. TIA!

Brian