today, im excited to release a repository that implements llama3 from scratch -- every matrix multiplication from attention across multiple heads, positional encoding and every other layer in between has been carefully unwrapped & explained. have fun :)
today, im launching the largest ever minecraft q&a dataset with 700k question answer pairs -- they cover every topic, item, block, mechanic and craft in minecraft. the dataset was generated by extracting over 18,000 minecraft wiki pages and over 11 hours of llama3 on 8xh100s.
Today, I'm launching a search engine powered by Mixtral-8x7B. The website is 100% free and requires no login information. Give it a spin!
Also, the website relies on donations like Wikipedia. It costs $1.6 per million tokens, so every dollar goes a long way!
I wrote a comprehensive blog on latent consistency models (LCMs).
>
The blog explains everything you need to know about LCMs including their math, architecture, finetuning and code.
last night my llama3 implementation worked, from self attention, attention across multiple heads, multiple layers of the transformer with rotary position embeddings and all of the layers in between -- everything works one matrix multiplication at a time. dropping this weekend :)
Introducing WhatOnEarth, a search engine that lets you chat with the entire internet. As a free & open alternative to ChatGPT Plus, the website will only get better and cheaper in the months to come.
What on earth. Give it a spin, it's free.
im excited to announce that ive joined the founding team at
@glaiveai
, as an ml engineer. we are so fucking back. we are so fucking back. we are so fucking back. we are so fucking back. we are so fucking back. we are so fucking back. we are so fucking back.
I'm working on a detailed blog on llama.cpp - the goal is to cover the math, the code and pretty much every detail you possibly need to know about the codebase and architecture of LLMs.
I wrote a comprehensive blog on latent consistency models (LCMs).
>
The blog explains everything you need to know about LCMs including their math, architecture, finetuning and code.
🔍 Introducing a search engine powered by ChatGPT!
⭐ 30 minutes ago the ChatGPT API was released to the public. Do you want to know how well it performs?
🌎
Give it a spin + share bugs/features that want to see 👇
cc:
@_buildspace
@nexgenfounders
Today, I'm announcing FashionAI, a website where you can edit clothing in under 15 seconds.
The first feature we're releasing is editing using prompts, the model detects clothing and replaces it with a new clothing item described by you.
But FashionAI doesn't stop here.
this week im working on a mistral-7b inference implementation file from scratch where i will implement - attention, multi-head attention, transformers, rope embeddings, sliding window attention, rolling buffer and every other detail. step by step. one ipynb cell at a time.
Do you need Github Copilot on your Google Colab notebook?
Introducing an AI assistant that writes python code for you on Google Colab 🧠
shift + alt + enter = ✨ ai magic ✨
Check out the demo & set it up 👇
over the last three months (ongoing), i had completely stepped away from my side projects. my plan was to focus on myself, get absolutely cracked and build my taste and learn what i like. in my opinion, this is the best path to take. only one way to find out :)
I realized that function calling + zapier could automate my whole life, so I made a function calling playground --
You can set up your zapier webhook and send emails, setup calendar events, manage your notion etc. under a single interface.
In January, I obsessively maxed out my "reading code, papers & math" skill tree.
I'm excited to show off some of the stuff I've learnt, starting off with an (extremely detailed, hopefully useful) LLM blog dropping in late Feb :)
I'm working on a detailed blog on llama.cpp - the goal is to cover the math, the code and pretty much every detail you possibly need to know about the codebase and architecture of LLMs.
making a repository that contains llm code for - tokenizing, training, fine-tuning etc. honestly the fastest way to learn small details is to try everything out :]
I wrote a script that retrieves the source code of any public model made available on replicate. I'm able to setup, modify and run new models on my local system in a few minutes.
It's nothing too crazy, but it will save me many many hours in the future :]
Today, I'm launching an AI music playlist, containing 50+ songs and various themes.
>
With over 2 hours of music, there is a song for every single one of you.
🌃 Who doesn't love stories?
Stories can be a fun way to learn new things. And guess what? Reading them can be super addictive too :)
You guys loved the MVP, so I am officially launching it.
🌎
What did you learn today? Share your screenshots with us!
It's been 20 hours since this tweet, and I genuinely think we are onto something big!
updates from the last few hours
- whatonearth = mistral + internet
- 4000 new users
- 2 million tokens served
- $1,600 contributed
- i'm so grateful :')
- announcements soon
Today, I'm launching a search engine powered by Mixtral-8x7B. The website is 100% free and requires no login information. Give it a spin!
Also, the website relies on donations like Wikipedia. It costs $1.6 per million tokens, so every dollar goes a long way!
In ~10 hours, I might win $25,000.
N&W is a gameshow that lets people vote between 32 projects after 6 weeks (the top 4 share a prize pool of $100,000).
I've made it to the final 32 teams, it would mean the world to me, if you’d vote for me!
@_buildspace
@_nightsweekends
Today, I'm launching an AI music playlist, containing 50+ songs and various themes.
>
With over 2 hours of music, there is a song for every single one of you.
I'm excited to launch an AI music generation tool on StockMusic. Generate songs across 50+ genres for free.
>
Sign up for 120 music credits and generate your first song.
"don’t believe what your eyes are telling you. all they show is limitation. look with your understanding, find out what you already know, and you’ll see the way to fly."
managed to hit ~900 tokens per second on llama3, im currently generating a synthetic dataset with 750k rows. ive tried to parallelize everything as much as possible.
avg tokens generated per row: 50
rows generated: 18 rows/second
hardware: 8xh100s
estimated time: 11 hours
Today, I'm launching StockMusic. I want to make it easier and cheaper for anyone to create music.
🔗
With this project, I'm going to focus on generating a collection of non-copyright songs for creators/brands.
> exciting update
> context. i'm building a research lab.
> new rebrand. "aaaaaaaaaa".
> why? it's more my vibe.
> bye
> more updates soon, or not.
> i'm studying 8-9 hours a day
> AAAAAAAAAA
> i created the first chatgpt wrapper
> launched 30 min after API release
> 80k page views
> 100k+ searches
> iterated many times
> no login was required
> cost me many $100s
> archived project
> a smol part of
@OpenAI
lore
> just felt like showing off
> lol, bye
🔍 Introducing a search engine powered by ChatGPT!
⭐ 30 minutes ago the ChatGPT API was released to the public. Do you want to know how well it performs?
🌎
Give it a spin + share bugs/features that want to see 👇
cc:
@_buildspace
@nexgenfounders
Day 3/10 in Dubai, I'm convinced that I'm living in a simulation.
> my parents met farza!?
> graduation day was like hogwarts
> this cat likes me
> sight seeing time sheesh
> if you are around in dubai hmu
Dear builders, this is your daily reminder to post your app on Reddit. The comments can be brutal, but they mean love. Let them cook you, in return for valuable feedback. We hit 47k impressions and 1.5k new whatoneathers from Reddit yesterday.
Introducing WhatOnEarth, a search engine that lets you chat with the entire internet. As a free & open alternative to ChatGPT Plus, the website will only get better and cheaper in the months to come.
What on earth. Give it a spin, it's free.
I'm announcing another massive update to whatonearth - our Mixtral-8x7B is now connected to the internet.
You can query links, search code, get news and even ask the search engine to roast your profile.
it's been just over 2 hours here are some stats:
- one user every second
- 1,000 searches already
- $30 already contributed
- now using both neets and fireworks api
- shipping to music
- my mom is yet to try the app
brb gotta grab some dinner
Today, I'm launching a search engine powered by Mixtral-8x7B. The website is 100% free and requires no login information. Give it a spin!
Also, the website relies on donations like Wikipedia. It costs $1.6 per million tokens, so every dollar goes a long way!
one day at a time, one day at a time, one day at a time, one day at a time, one day at a time, one day at a time, one day at a time, one day at a time, one day at a time, one day at a time, one day at a time, one day at a time, one day at a time, one day at a time
It's been a week since launched, it's currently at 2,700 users, 30 feature requests and 8 people willing to pay (almost zero marketing, only 2 tweets).
Adding social sharing buttons worked for organic growth (even though it's ugly).
@_buildspace
@elonmusk
I think the font on x(dot)com is perfect. It displays confidence that the company can own the most simple logo and yet be recognizable. Kinda badass imo.
Also, looks classy and accelerationist.
stuff i'm doing this week - llms studying & fine-tuning, experimenting with ideas around with tracking models, trying comfyui, working on my thesis, going on walks & reading books :)
My weekend hack
@algoz_xyz
is live on the Polygon Mumbai Testnet! The docs are also available for you to play around with 👀
It started off as a weekend hack for the
@buildspaceIND
Bangalore event and is now shaping out to be a really useful project for teams! LFG. ❤️
I'm creating an affordable music library for creators and brands.
This is exciting because I'll be able to explore my creative juices with music generation while growing my first revenue focused project.
cc:
@_buildspace
@_nightsweekends
Bots are a huge problem on
@solana
NFTs and on-chain games can and will be scripted by bot developers. We built an MVP at the Solana Summercamp hackathon.
Every upvote counts towards our win, please support our submission 🐥
I'm going to try my best to study and share research in public. The plan - share blogs, make looms, share notes and code.
LCM blog dropping first and soon, no more delays.
Today, I am launching FashionAI 2.0, a tool that modifies your clothing based on custom prompts.
[🎀]
This version includes an improved UI, fixed prompts bugs, share options and a focus towards fashion brands.
Everyone should spend 3-4 hours/week working on a silly project. At the start, it might feel lame, but you learn a lot about how internet consumerism works.
my project
> its called stockmusic
> 220 daily active users
> only made $70
> need to fix paid user conversion rate
this is what a weekend of a professional matrix multiplicator looks like. my cactus and i are currently finishing up a mistral-7b implementation (from scratch) demo and then ending the day with dingboard diagrams to explain the math and architecture of mistral-7b.
Day 9&10/10 in Dubai, jumped off a plane.
> gotta catch up w/ research
> literature survey for my thesis
> current notpink project -> twitch chat controls a paint brush
> gradient updates (quote tweet)
> we are so back
> bye :3
just me in my room trying to figure out how and why my app did $75 in revenue this weekend while i was eating away at a potluck. anyway 400 users visited the website yesterday and 3 people paid not bad tbh :')
Twitter is the only social media that makes me more productive. 10 min of scrolling makes me second guess wasting time and get back to building stuff.
Side note if I follow you I am fking bullish that you are gmi. Keep building.
This week I'm going to share detailed notes on LLMs. It's probably going to cover: math and arch, detailed explanation of speed and memory optimizations and code.
I wrote a comprehensive blog on latent consistency models (LCMs).
>
The blog explains everything you need to know about LCMs including their math, architecture, finetuning and code.
2 weeks ago
@_buildspace
launched spaces. Which was an invite-only exclusive forum for builders.
They started off with a $100,000+ treasury to fund projects and builders.
We created a $5000 treasury request to fund
@algoz_xyz
. Could you help us get to 69 votes? (1/3)
🌠 Launching WhatOnEarth!
👯 Last week I created a search engine that uses ChatGPT.
It's completely functional + free + no login is needed.
🌎
Reply with 🌎 if you tried the search engine + share your thoughts below 👇
> hiii
> SURPRISE launch
> improved quality of AI music
> any length (upto 10 min)
> affordable pricing
> 120 free credits on signup
>
> byee, help with likes/comments/rts
> 🤍
example; ai meditation music (5 min)
🌎 24 hours ago I shipped an app using the new ChatGPT API, here is my daily update tweet.
🔗
💁♀️ We now have 13,000 MAU and >60,000 searches.
🔎 Added a Search on Google button as a fallback (this is what I usually do when ChatGPT does not work)
Officially done with 1 year of Masters, I feel so free and relaxed. I'm one quarter ahead which gives me roughly 5 months of holidays to build and create.
For the first time in my life I'm happy with my goals, missions and hobbies.
Time to build, but this time, all day.
Hey, as some of you may know I'm creating a search engine called whatonearth. You should now notice improved and detailed results. Also, the sources are clearly cited now.
updates coming next
- images along with search
- follow up questions/chat
Today, I'm launching a search engine powered by Mixtral-8x7B. The website is 100% free and requires no login information. Give it a spin!
Also, the website relies on donations like Wikipedia. It costs $1.6 per million tokens, so every dollar goes a long way!