Staff Software Dev. Learning 🧠 AI/Large Language Models, electrical engineering to build 🤖s (
@BehindEng
), 3D w/
#r3f
. Built Atom Beautify to 9mil installs
Working on a new Text Splitter for
@LangChainAI
and am delighted with the results so far! 😍
✅ Recursively splits chunks until small enough
✅ Prefers complete sentences
✅ Reduce # of chunks
✅ GPT-4 wrote almost all the new code!
Demo of the new splitter on the right:
Introducing Alvea!
An adaptive AI assistant creating unique user interfaces personalized to your tasks just in time. Your evolving tool, designed to enhance productivity and user experience 🚀
Won 🥇
@agihouse_org
w/
@sockcymbal
@narphorium
@vhurgenes
Developing an improved automated task planner for autonomous (
@Auto_GPT
/
@babyAGI_
style) agents with
@arindambarman
& ChatGPT4.
Using OpenAI's GPT4 API,
@LangChainAI
JS/TS, Next.js, Mermaid.js, and a lot of prompt engineering!
More progress updates coming soon 🚀
@SergioRocks
Why wouldn’t the manager encourage the rest of their team to start using ChatGPT then? 🤦♂️ Disappointing when ego and status quo get in the way of improving productivity and learning from others. IMO The OP should be looking for a new job at a better place with the right values.
Hi dev friends! We are exploring a couple of new ways to get your Spline scenes into your code projects.
Now you can export Spline scenes as:
- Vanilla JS
-
@threejs
- react-three-fiber
#r3f
- React
🙌 Here are some demos!
👉
👉
Boost your Semantic Search with Customized Embeddings 📈
1️⃣ Gather similar/dissimilar pairs
2️⃣ Train a matrix to highlight relevant aspects of your embedding
3️⃣ (Original Embedding)×(Custom Matrix)=(Custom Embedding)
Coming soon: Live Predictions Preview in
@weights_biases
during Axolotl LLM training!
✅ Easily see how your language model is improving
✅ Quickly compare predictions between steps
Super excited for my first substantial Axolotl contribution. To many more! 🍾
I had a blast at the
@agihouse_org
today! Met a lot of inspirational and incredibly talented people who presented ambitious projects for only a day’s work. Congrats to the prize winners! Can’t wait for the next hackathon together! 🚀
All righty we're at the end of the hackathon, votes are cast, someone is going to win a branded pair of
@weights_biases
AirPods, and a bunch of other prizes from
@togethercompute
(a bunch of credits!)
Also
@agihouse_org
is quite the place 👏
Also selfie w/
@nutlope
🤗
Definitely Solved - 🆕 Access Waitlist
What could you solve with a photographic memory of all open-source packages? Let's find out!
🙋You: Describe your problem
🤖 AI: Plan ➡️ Search updated packages ➡️ Solve with a live & editable preview
Sign up:
Customizing embeddings is incredibly useful!
I improved my semantic search accuracy from ~60% to ~96% 🤯🎉
Took me a bit of time to create a reasonable sized dataset, well worth it.
Huge thanks to
@OpenAI
for Ada-002 and writing these fantastic examples!
@subby_tech
@_buildspace
@fdotinc
@FarzaTV
It would be great if multiple songs could be selected to produce a more precise/mean embedding. I find similar search on embeddings don't always recognize what part to focus on, in this case what specifically in the song I like/dislike.
Learn how to train custom language models step-by-step!
🔗 New Repo:
Currently includes my scripts to:
1️⃣ ✅ Bulk download YouTube videos
2️⃣ ✅ Convert audio to text with
@DeepgramAI
3️⃣ 🚧 Convert unstructured text into chat format
More soon 🚀
@altryne
@AravSrinivas
Agreed. Both cringy but lovely. Loved the sentiment of being held accountable to Mother Nature. For those who finished watching, clearly see Apple holding to their commitments 🎯
Startup Mentor AI v0.2 is a massive leap forward in quality & helpfulness! 🎉
✅ 100% automated pipeline from raw text/YouTube to "textbook" high-quality synthetic dataset /w
@LangChainAI
📈 Output (right side in screenshot) appears much better to me!
What do you think?
Update for Edie, the AI Writing Editor 📝
🏎️ Near Real-time Feedback - Thanks to the fast, fine-tuned
@OpenAI
language model
🎯 Inline - Feedback above the relevant paragraph & hover to read in full
📹 Demo - Critiquing a story ChatGPT wrote by copying & pasting paragraphs
@zachtratar
I primarily use JSON output with GPT3 models. Did take a bit of tweaking to work with ChatGPT. Writing TypeScript also helps to enhance the prompt and enforce the output structure.
Try something like this screenshot. Hope it helps!
🎉 *Live Demo of V-HACD for Web*
I’m recently back to work on approximate convex decomposition for the web! 🎉
AKA how to quickly add physics collision detection to your 3D models using only your web browser.
Demo & details below. 👇
You can reproduce the Cybertruck example too!
@LangChainAI
The splitter config used is:
chunk_size = 50
chunk_overlap = 10
with length function set to gpt2 tokenizer from tiktoken.
The original text is a short story written by ChatGPT-3.5
For more info on TextSplitters see:
Existing AI chat apps boost productivity yet fall short in delivering optimal experiences due to their reliance on text chat UIs
Alvea's solution is integrating AI to develop dynamically rendering user interfaces that are interactive and tailored to the user's current task
I had a lot of fun building Your Devin with my teammates this weekend! 😃
One of the biggest & most impressive hackathons I’ve participated in and hopefully many more ahead 🚀 Thank you for hosting
@SHACK15sf
, and
@MistralAI
and any many others for sponsoring!
@shwnhll
“Doubt anyone would hire a job hopper”
As an interviewer, if an interviewee shares a *reasonable explanation* for why their current employer isn’t a good fit, I wouldn’t hold it against them. Buyers remorse can go both ways.
Want to help fine-tune an open source LLM that naturally formats as a
@LangChainAI
agent chain?
No more wasting tokens on formatting instructions in the prompt, and reliable output parsing baked into the model weights!
We’re building 🛠
@rasbt
Would you be interested in writing an article about reinforcement learning as an alternative to supervised training of language models? It is my next area of study and I’d love to see how
@LightningAI
could be used to help. Thanks for all you’re doing!
WIP: Automated pipeline from raw YouTube/etc to an AI/language model
🎯 Goal 1: AI Startup Mentor for All
1️⃣✅ Ordered/DAG Learning Objectives Extraction
2️⃣✅ "Textbook-quality" Lessons Generation
3️⃣🚧 Exercises Generation (Q&A, multiple-choice, etc)
4️⃣ Train AI model
@adamwathan
We’re building an online platform to help anyone learn to engineer and go from an idea to a physical & functional product 🚀🎉 For more details and demos follow
@BehindEng
and check out:
We're building an online platform to help anyone learn to engineer and go from an idea to a physical & functional product!
👉 Learn ➡️ Design quickly ➡️ Experiment in simulations ➡️ Build for real ➡️ 🚀
@shwnhll
Exactly. Their concern of being perceived as flight risk is valid. Fortunately salvageable. In fact, it could even be framed as a positive signal since the candidate (them) knows what they want, can articulate it, and interviewer can assess if there is a strong mutual fit. 🎯
It's ready! 🎉
🔥Startup Mentor AI v0.2 dropping soon🔥
Been working hard over the weekend, refactoring my prototype code to use
@LangChainAI
with
@OpenAI
GPT-3.5 so moving much faster now!
@AlexReibman
Absolutely agree! My favourite to learn from the cookbooks so far has been customizing embeddings — I posted my lessons learned in the link below, if you’re interested. Which OpenAI recipe has been the most informative or useful for you?
Boost your Semantic Search with Customized Embeddings 📈
1️⃣ Gather similar/dissimilar pairs
2️⃣ Train a matrix to highlight relevant aspects of your embedding
3️⃣ (Original Embedding)×(Custom Matrix)=(Custom Embedding)
@jerryjliu0
@gpt_index
Fantastic addition! Great minds think alike 😜 I’ve been developing a similar solution for problems decomposition, which also extracts parameters for tools:
🆕 Data2AITextbook 📈
Automatically convert unstructured data into a high-quality 'textbook' format, optimized for fine-tuning Large Language Models (LLMs) 🚀
Inspired by Textbooks Are All You Need paper.
Just getting started. Contributions welcome!
Adding
#r3f
#Rapier
#physics
preview to V-HACD for Web
✅ Automatically converting from arbitrary
#threejs
geometry to approximated & efficient colliders
✅ Spawning 100+ red spheres raining from above to demonstrate performance (60 FPS) even with detailed meshes
🚀 Meet Edie: Your AI Writing Editor📝
After my private 24-hour hackathon & 380+ training examples, I now have a fine-tuned Curie model which is:
📈 Higher-Quality
🤑 Inexpensive
🏎️ FAST!
I had a lot of fun learning to fine-tune language models! What should I learn next? 🤔
Any recommendations for cloud inference of custom Llama-2 fine-tuned models?
- GPT3.5 Turbo is an incredible price per 1K tokens
- Serverless GPUs ($/second) are still quite expensive even for 7B LLM with GPTQ+ExLlama
- 🤩 Hourly GPUs+7B+ExLlama can potentially match GPT3.5
Experimenting with gears in a simulated
#physics
environment (Cannon-ES) on the web.
I'd love to build something with gears and pulleys which runs in an educational web app.
#buildinpublic
I’ve been learning Python again solely to use packages like this. Finally I can fully utilize my TypeScript experience for building machine learning applications! 😍🚀🎉
Just released a new open source machine learning library for TypeScript 🔥
Node.js devs can now take full advantage of Python's powerful scikit-learn package – without needing to know any Python! 🤯
PCA, KMeans, TSNE, ... 257 classes in all!
→
@jacklangerman
@tanmay2099
@fchollet
@CVPR
@CVPRConf
💯. First to release code >> first to publish & post paper. Both ideas are great but without the code to the rest of the world they are just ideas. Biggest impact is other people leveraging these incredibly innovative solutions.
Requiring an OpenAI API Key for your app? Worried about user onboarding?
🚀 Share or embed my guide:
✅ High-Res Annotated Screenshots
✅ Comments enabled for questions: Open for logged-in Notion users
✅ Maintained: I'll be reusing this page, too!
Had a lot of fun building these incredible improvements with
@arindambarman
! I was a little surprised to see just how much better search could be by adding these features 🤯. Excited to see how others use
@clamy_ai
now & push it to the limit of what AI can do! More coming soon!
Amidst all the crazy stuff
@OpenAI
launched recently,
@glavin
and I worked on some fun features for
@clamy_ai
, which we thought we would share anyway.
Here's a demo where we ask the bot to write an email and it automatically searches the relevant documents and forms an answer. 🧵
🎉 I've gained access to the GPT-4 API! 🤖
Ready for a fun weekend pushing GPT-4's boundaries! 🚀
1st up: experimenting with reasoning & problem decomposition compared to 3.5 🧠
💡 What other mind-blowing capabilities or innovative projects should I explore next?
#GPT4
#AI
@jxnlco
Great solution! Ive been learning about coreferences lately to solve the same problem, instead of relying on GPT4. Maybe use GPT4 to create a dataset for training a new model?
🚀 Preview: Definitely Solved 🚀
Quickly find functions & packages to solve your problems.
✅ Updated: not 2 years old as ChatGPT
✅ Smaller & Faster: reduce code, don't reinvent the wheel
✅ Safer: popular packages already battle-tested
#buildinpublic
*Updates to V-HACD for web*
👉 Try it now:
✅ Button to export JSON for approximated convex hulls, so you can use with Cannon and
#Rapier
#physics
.
I'll provide docs to help you use this JSON data soon.
✅ Improved UX, control labels, etc.
#buildinpublic
@jxnlco
Excited to see your approach to answer questions like that!
I worked on some problem decomposition + RAG system a while back and learned a lot. Impressive the reasoning LLMs can do without domain or task fine tuning.
@davewanders13
@Jauwnio
> “People don’t have access to the same jobs you do.”
Isn’t it actually the opposite?
Ability to work remotely *enables* those people in small towns to have access to the same jobs 🎯
Introducing Alvea!
An adaptive AI assistant creating unique user interfaces personalized to your tasks just in time. Your evolving tool, designed to enhance productivity and user experience 🚀
Won 🥇
@agihouse_org
w/
@sockcymbal
@narphorium
@vhurgenes
@OpenAI
@mermaidjs_
@arindambarman
@clamy_ai
If you want to see more of this kind of research then follow, like, and retreat. There's a lot more I've learned and would be willing to share if there is interest.🙂
I’ve experienced the same, glad to know I’m not alone. Hope they fix soon. I now keep a Google Doc with GPT prompt versions. Btw you can paste multiple prompts into chat and explicitly tell it to merge into prompt, being careful to capture all details of both. Works well for me
PSA: careful with the GPT builder when editing. It tends to wipe all previous instructions and start over. We've lost quite a few good prompts out there...
@SaikiK66287209
@Teknium1
Same! Realized a couple weeks ago I spend more time talking to ChatGPT-4 than any other human. Told someone that and they pointed out I had personified GPT-4 by saying “any other human” 😮
@diego_gallovich
@marc_louvion
In this analogy for businesses: Painkiller products are essentials solving critical, immediate needs, while vitamins offer beneficial enhancements, more nice-to-have than essential.
@abacaj
@filipe_almeida
Could you elaborate on why starcoderbase? I’m not as familiar with <7B model options and their trade-offs. Recently heard of TinyLlama, have you tried it?
@rasbt
Thank you so much for sharing! Especially the hardware and configuration options used so we can reproduce. I’ve been learning fine-tuning OSS LLMs and struggle with tweaking those appropriately to maximize utilization of hardware and minimize training duration. Exciting work! 🚀
@steventey
@vercel
Any examples with Python on Vercel? Would love to combine Python + Next.js for the best of both worlds. Some AI libraries are only in Python, for now.
Teapot
#physics
rigid body collider
I like how the balls gradually roll off the teapot. V-HACD was able to reasonably approximate the curved shape. If I created the collider shapes manually it definitely would not look so good.
Start making $10k+ / month using ChatGPT4 to its best potential.
Spent this week curating a list of 80+ use cases of ChatGPT that you can build right now
Will DM it if you
1. Follow Me
2. Retweet this tweet
3. Reply "ChatGPT4"
@onirenaud
@CosyStudios
Is there an engineering blog we could read about how you’ve achieved this? Maybe there’s an opportunity for a new library which makes this easier for other
#webgl
projects to do the same.
@pmndrs
has so many great
#r3f
libraries already, this would be an excellent addition!
@amasad
Love this! What other industries do you think still need this advancement? With
@Replit
making software development much more accessible to newcomers and providing superpowers to existing coding experts. My friend and I are trying to help people become engineers with
@BehindEng
Custom LLM streaming at 61-72 tokens/sec on Replicate! 🔥
Building my go-to LLM deploy repo 🚀
1️⃣ repo to max code reusability and:
✅ deploy to
@runpod_io
&
@replicatehq
easily
✅ support AutoGPT, ExLlama, & future fast libs
✅ streaming
Try now: