gdiamos

16 ms·

1. 331.▲
   by gdiamos 3y ago

   This enables you to chain and switch between thousands of finetuned LLMs. A free notebook that lets you train a model using this optimization: https://colab.research.google.com/drive/1QMeGzR9FnhNJJFmcHtm... A walkthrou
2. 332.▲
   by gdiamos 3y ago

   Alex's papers have been a great source of inspiration in the past. Even if the methods require scaling up and significant engineering to put into production. The parent is right to set expectations. Ideas like this often take 5-15 yea
3. 333.▲
   by gdiamos 3y ago

   How a tiny diffusion LLM generates a sentence. https://imgur.com/gallery/kZa6VuZ Visualization from the paper. Figure 20.
4. 334.▲
   by gdiamos 3y ago

   Memory bound token generation is a limitation of transformer decoders. In the past, hardware has motivation algorithm innovations. I’m curious how long it will take until we see more hardware friendly models.
5. 335.▲
   by gdiamos 3y ago

   Thanks that made me smile. Take my upvote
6. 336.▲
   by gdiamos 3y ago

   We can run llama 2 on an AWS vm if you have enough GPUs: https://lamini.ai/ Install in 10 minutes. Make sure you have enough GPU memory to fit your llama model if you want good perf
7. 337.▲
   by gdiamos 3y ago

   Is it the same api as the public OpenAI
8. 338.▲
   by gdiamos 3y ago

   I find it interesting to see how competitive this space got so quickly. How do these stacks differentiate?
9. 339.▲
   by gdiamos 3y ago

   we wrote a blog post about why companies do this here: https://www.lamini.ai/blog/specialize-llms-to-private-data-d... Here are a few: Data privacy Ownership of IP Control over ops The table in the blog lists the top 1
10. 340.▲
    by gdiamos 3y ago

    Can you fine tune it?
11. 341.▲
    by gdiamos 3y ago

    Excited to see the MLSys growing. Deep learning methods are so computationally intensive, many advances have come through new algorithms and optimization methods.
12. 342.▲
    by gdiamos 3y ago

    It's just so we don't mix up models and users. Anyone can create an account for free by pressing one button.
13. 343.▲
    by gdiamos 3y ago

    This initial public alpha release of Lamini LLM fine tuning includes several advanced optimizations: 1. Chinchilla recipe smaller models pretrained on data increases inference speed 2. Instruction fine tuning training on a small high qualit
14. 344.▲
    by gdiamos 3y ago

    I’m all for efficiency. How much energy is used on AC? How much would a 10% improvement in efficiency save?
15. 345.▲
    by gdiamos 3y ago

    The future is exciting and terrifying. Scaling laws have resulted in GPT-4 and they keep going. The design space for LLMs is bigger than just transformer architectures. It includes long prompts, retrieval, chains, data quality, compression
16. 346.▲
    by gdiamos 3y ago

    I'm generally a believer in trading compute for insight. So this makes sense. I'l also curious how adding 100x of compute into a longer prompt compares with using the compute for something else. I'm sure there is a design sp
17. 347.▲
    by gdiamos 3y ago

    Thanks, that's interesting.
18. 348.▲
    by gdiamos 3y ago

    One thing that seems to be overlooked with very long prompts is that the compute still scales at best linearly with the input size. So a context size of 100k requires 100x more compute than a prompt size of 1k. For which applications is tha
19. 349.▲
    by gdiamos 3y ago

    I’m just off Reddit at this point.
20. 350.▲
    Specializing LLMs on Databricks(databricks.com)

    3 points by gdiamos 3y ago|0 comments
21. 351.▲
    by gdiamos 3y ago

    Here are some examples: https://github.com/songanz/3D-LiDAR-annotator/blob/master/an... https://dataloop.ai/platform/lidar/ A good supervised learning process requires teaching
22. 352.▲
    by gdiamos 3y ago

    Current generation deep learning systems rely on supervised learning, which requires human labeling. This is one argument for vision alone. It’s easier for humans to teach the deep neural network what to do if they both see and label the s
23. 353.▲
    by gdiamos 3y ago

    I give Tesla a lot of credit for shipping features that I can buy. I use lane following, summoning (which saves me time), and the windshield wipers every day. Does anyone remember will it wipe? https://youtu.be/0SSYFMtdJ5k
24. 354.▲
    by gdiamos 3y ago

    Self driving hype and investment got ahead of the tech.
25. 355.▲
    by gdiamos 3y ago

    The demand is driven by scaling laws. Now that we see the light at the end of the tunnel to AIs that are more powerful than ChatGPT, the race to develop computers powerful enough to run them is on.
26. 356.▲
    by gdiamos 3y ago

    Nice. Will we be able to get to 1M tokens?
27. 357.▲
    by gdiamos 3y ago

    Fine tuning where you freeze the weights of a neural network has been used for a long time in computer vision. There are many variations of these methods. More recently there are some good libraries that make them easier to use. For examp
28. 358.▲
    by gdiamos 3y ago

    You are right, but Adam Cheyer (the Siri founder) worked on it for a long time before it was sold to Apple. https://en.m.wikipedia.org/wiki/CALO
29. 359.▲
    by gdiamos 3y ago

    It’s amazing what throwing a ton of computers at problem gives you. In the 80s-90s, Shannon’s predict the next token idea gave us spell check. In 2000-2010 it gave us Siri & Google Home. In 2020 it gave us ChatGPT.
30. 360.▲
    by gdiamos 3y ago

    These results are predicted by LLM Scaling Laws and the GPT authors knew it before they started.