K0IN

No, I tuned a model on my (unique) table data, which does not take long, since the model is small.

My model seemed in my tests at least to hold up good enough, since its only used as a preselect to find "good enough" candidates to use Levenshtein later on.

But yes, a universal model (maybe a fine-tuned transformer / embedding model) might be better, but i did not have the time (and knowledge) to build one yet.

This is very interesting, i was building something similar, but i used https://github.com/K0IN/string-embed (embeddings based on a distance function - Levenshtein in my case) to generate embeddings, for deterministic matching.

I will follow your project, im interested in your ann search speeds :)

i need this as a chrome plugin asap :)

It looks really cool and clean!

so they are now close to where every float in the original gpt3 model, is worth 1$.

sadly this suggests invalid emails for example: for kein it suggests k0.in (which does not seem to be a valid Indian domain)

I am not that deep into react, but I always wondered, what exactly a vdom is and how the diffing works, the code is so easy to read and has a lot of comments which is really nice, thanks! I think I now understand react (or at least I hope so) a bit better.

Link redirects me to a 404

for me, the advantages are:

* In WSL2 I don't need to worry about, ports, mounts and data persistence (I think this is possible in docker but it needs more setup work)

* I don't need docker desktop running

in the end I think that both can solve the problem, I myself prefer wsl for my day to day tasks.

Hei interesting, thanks for sharing.

Also, the idea to do it the other way around WSL -> docker image is really nice.

I love WSL, but I have multiple systems and juggling apt packages is a pain.

So I had this in my head for a long time, that I can just build a rootfs using docker and load it in WSL.

I tried it yesterday, and now I have a CLI tool to build my WSL image using Dockerfiles.

I'm not sure if it helps anyone, but I wanted to share it, just in case.

Cheers.

I just did something similar with openstreetmaps last week. In my case I tried the other way around (finding similar addresses)

I benchmarked fts(sqllite) Vs typesense.

Which I can only recommend for such tasks.

repo: https://github.com/K0IN/open_streetmaps_typesense

btw I also tried the python import but it was unusable (for Germany) slow so I switched to osmtools and exported it to Jason first which is way faster.

this lives in my head rent free for the rest of my life

can you elaborate a bit on why? (noob here)

so for fast responses we usea a rtx4090 with vllm, but yeah it depends on your use case

The link seems to be broken

on android there is a Killswitch.

if it can make test from the decompiled code, we could reimplement it with our code style. might be cool to have some bunch of llms working together with feedback loops.

I wonder why the snapdragon does only slightly increase in compute power even tho they nearly triple the power consumption, or am I reading the charts in the source (linked on page) [0] wrong?

https://wccftech.com/qualcomm-unveils-more-snapdragon-x-elit...

I'm not sure how to do it "the right way" but you may have a look at Paramiko, it's a python library that helps you implement your own SSH server (easily).

maybe you can build your own wrapper that calls a subprocess.

you can "force" integer arithmetic with bit or (example your_float|0), this helps the JavaScript runtime to optimise calls and use (faster) int arithmetic. (https://stackoverflow.com/questions/23436159/how-bitwise-ope...)

always remember the fastest code is the code that is not executed at runtime.

we switched to librechat [0] it is a great app if you don't use gpt 4 often enough to hit the 20$ worth of tokens, also it supports plugins. If you use more than 20$ then just stay with gpt plus.

https://docs.librechat.ai/

when can we expect the first ~100+ million parameter models to run on raspberry pi Pico?

I build notify a small serverless app with a pwa, that allows you to send web push notifications to your devices.

It was a cloudflare worker in the beginning (you can find it in the branches) but i rewrote it for deno deploy, since I found hosting cf worker locally (which was one of my goals) a pain.

https://github.com/K0IN/Notify

I need this for GPUs.

One Problem I always see in such apps is that the ai can't see in to the database or into all entries, so queries without stating data EXACTLY as in the database run into issues.

example: give me the revenue for all logistics firms

but in the database these might not be called "logistics" and may be called "transport" (or anything)

maybe there are some counters to this like finding unique values per column or even better use a grammar based approach, wich will select only valid entries.

but the simple text to SQL is at this point not the "hard thing to solve"

what happens around 512 tokens? how can rope jump this far while fft is at its lowest?

Maybe I miss the point here, but it I see these sections as "why customers should choose us" so this point does not make any sense to me, since it's worse(or at least not adding value) from a customer view.

why cant they just use the old approach for anyone and the good stuff for loggedin users tho.

> Note that while Boomerang is optimized for low-latency performance, models like GTR-XXL, which weighs in at 4.8 billion parameters, are very challenging to productionize.

So what is the size of your model than, or did i miss something?

seems like it cant defeat the holy grail: tldr;