leroman

8 ms·

1. 31.▲
   by leroman 2y ago

   Thanks for sharing, will look into adding this as a flag in the options!
2. 32.▲
   by leroman 2y ago

   Markdown being a very minimal Markup language has no need for much of the structural and presentational stuff (CSS, structural HTML), HTML has many many artifacts which are a huge bloat and give no semantic value IMO.. It's the goal he
3. 33.▲
   by leroman 2y ago

   Will add some side-by-side comparisons soon! the goal is not just to translate 1:1 HTML to markdown but to preserve any semantic information, this is generally not the goal for these tools. Some specific features and examples are in the REA
4. 34.▲
   by leroman 2y ago

   Author here- it's a good point to have some benchmarks (which I don't have..) but I think it's well understood that minimizing noise by reducing tokens will improve the quality of the answer. And I think by now LLMs are well
5. 35.▲
   Show HN: Convert HTML DOM to semantic markdown for use in LLMs(github.com)

   146 points by leroman 2y ago|56 comments
6. 36.▲
   by leroman 2y ago

   Mostly look for an operator for that
7. 37.▲
   by leroman 2y ago

   By “eco system” I mean all the charts that get shared, when ever I try to look under the hood I instantly regret it..
8. 38.▲
   by leroman 2y ago

   Kustomize is basically a higher level file for K8s deployments, I have all the resources as declarative code that gets deployed when I apply the relevant directory. I have istio + ssl certs + services and any other resource, multiple proje
9. 39.▲
   by leroman 2y ago

   Somehow I am able to get bye with Kustimise. Cant stand the mess that is helm and its eco system.
10. 40.▲
    by leroman 2y ago

    No idea if its a valid approach but possibly train with a hidden layer containing a “role”?
11. 41.▲
    by leroman 2y ago

    Thanks to investing into k8s I was able to migrate a non trivial production with minimal downtime between cloud providers 3 times now, who offered us free credits, without too much friction.
12. 42.▲
    by leroman 3y ago

    GPT-4 is already hugely useful. If they are able to lower the cost of GPT-4 further, say 5x or 10x that in itself would be useful and huge.
13. 43.▲
    Easy and low cost model training on HF "DGX cloud"(huggingface.co)

    2 points by leroman 3y ago|1 comments
14. 44.▲
    The Sphere in Las Vegas features 4PB of flash memory at 400 GB/s to deliver 16K(tomshardware.com)

    2 points by leroman 3y ago|0 comments
15. 45.▲
    Data-center power usage is straining the power grid(nytimes.com)

    3 points by leroman 3y ago|1 comments
16. 46.▲
    by leroman 3y ago

    Well we finally have motivation to move upwards in the Kardashev Scale
17. 47.▲
    by leroman 3y ago

    From my testing the two top models both can do stuff only GPT-4 was able to do (also Gemini pro 1.0 couldn't).. The pricing for the smallest model is most enticing, but it's not available to me on my account for testing..
18. 48.▲
    by leroman 3y ago

    Some tasks we are using LLMs for are performing very close to GPT-4 levels using 7B models, so really depends on what value you are looking to get.
19. 49.▲
    by leroman 3y ago

    I understand this part but it seemed that the 16->8->4 etc is similar to compression of the "net" and seemed to lower quality below 8.
20. 50.▲
    by leroman 3y ago

    - we have llama.cpp (could be enough or at least as mentioned in the paper a co-processor to accelerate the calc can be added, less need for large RAM / high end hardware) - as most work is inference, might not need for as many GPUs -
21. 51.▲
    by leroman 3y ago

    Can someone versed in the ways of math explain how this is different from previous quantization methods? And specifically, seeing how going from 16fp to 8bit mostly gives same perplexity while anything further seems to lose quality / d
22. 52.▲
    by leroman 3y ago

    DigitalOcean has been great, have managed k8s and generally good bang for your buck
23. 53.▲
    by leroman 3y ago

    Oh yes, I know of one sure way, works 100% of the time: A man buys an apple, walks around, and sells it. Then, he buys two apples, walks around again, and sells them. Next, he doubles his effort by buying four apples, walking around, and se
24. 54.▲
    by leroman 3y ago

    This is really cool! one use case I can think of is car to car communication to allow them to share some dangers up ahead for example
25. 55.▲
    by leroman 3y ago

    If we constrain our selves to a pool of words of say Wikipedia entries, minutes names and maybe some other stuff, and use a "super node" like "addition" to kind of act as a math operation.. maybe this makes more sense?
26. 56.▲
    by leroman 3y ago

    I thought they were using some kind of vector space searches like embedding.. no idea if that's the case
27. 57.▲
    by leroman 3y ago

    This is where I see K8s fit in, future proofing and un-vendoe-locking (mostly). My own experience running on DO and moving to AWS (free credit ) while using DO S3 clone, managed pg and K8s, the move was rather straight forward..
28. 58.▲
    by leroman 3y ago

    Yes, maybe the better wording would be "public cloud", still probably everyone reads this as "cloud" vs "private hosting"
29. 59.▲
    37Signals Left the Cloud(world.hey.com)

    53 points by leroman 3y ago|14 comments
30. 60.▲
    by leroman 3y ago

    Nice, years ago I wrote a TypeScript (originally CoffeeScript re-written to TypeScript) library to generate schedule times for a cron between two time points, should be easy to write a service like this with it. https://github.co