Rickyticky Bobbywobbin
@arekusandr_
Followers
886
Following
97
Media
48
Statuses
129
Programming badassery. No bugs driven cloud development practitioner. Digital pianist and utopianist. Eventually consistent. LLM security and jailbreaks
Joined April 2019
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
#MetaCensorship
• 456759 Tweets
NABJ
• 417359 Tweets
La OEA
• 315700 Tweets
Léon Marchand
• 207513 Tweets
Social Security
• 134975 Tweets
Marta
• 98507 Tweets
Vybz Kartel
• 90837 Tweets
Colombia y Brasil
• 80971 Tweets
Brasil y México
• 76375 Tweets
Rachel Scott
• 66084 Tweets
She is Indian
• 64672 Tweets
Bungie
• 64584 Tweets
#alertemédaille
• 52654 Tweets
Indian American
• 50369 Tweets
Jamaican
• 45427 Tweets
South Sudan
• 43016 Tweets
Katie Ledecky
• 42039 Tweets
対象2作品
• 40867 Tweets
Howard
• 37041 Tweets
Tatum
• 36115 Tweets
Curry
• 32823 Tweets
Harris Faulkner
• 20752 Tweets
Yusuf Dikeç
• 17815 Tweets
Bia Ferreira
• 14071 Tweets
Kerr
• 14011 Tweets
Caucasian
• 11630 Tweets
World Boss
• 11450 Tweets
Abdüllatif Şener
• 10843 Tweets
HBCU
• 10566 Tweets
INTRODUCING: LangCorn, a FastAPI for LangChainAI📚🌩️
1) 💡Generates API from
@LangChainAI
pipeline
2) 🎨 Can run multiple models
4) Automatic
@vercel
deployment
(<7 lines)
Demo, Code explained, and GitHub Link ↓
16
90
631
INTRODUCING: How to deploy
@LangChainAI
to
@vercel
in 3 steps📚🌩️
1) 💡Install langcorn
2) 🎨 Create vercel.json
3) Push it to a private GH repo.
Demo, Code explained, and GitHub Link ↓
7
24
177
1/ GitHub and Live demo:
1
2
29
🚨 INTRODUCING: Langalf - LLM Security Scanner! 📚
✅ Scans for prompt injection, jailbreaking & more
✅ Tests OWASP Top 10 LLM vulnerabilities
✅ Detailed reports & customizable attack rules
🔗 Demo, Code explained, and GitHub Link ↓
2
3
5
How to build serverless LLM RAG?
1) Use S3 based vector store with vector_lake 📚
2) Store embeddings on S3 bucket
3) Decent search performance 500 ops/s 🌩️
4) Cost $0.18 for 200k embeddings in last month's AWS billing
3
1
5
Over 20 people have already published more than 200 scientific papers. Is this due to the network effect, unfair leverage, or science GPT?"
1
0
5
INTRODUCING: VectorLake
1) I have built yet another vector db just for fun 🪜
2) It shards your data to local or S3 files with auto persistence 🌱
3) It uses HNSW 🥷
4) Created with the intention to minimize db maintenance, costs, and operational overhead 📉
5) Langchain
1
0
4
0
0
3
Evaluating LLMs through perplexity is common but flawed.
Perplexity alone doesn't measure real-world value.
Here is why it falls short in gauging quality and relevance:
1
1
3
🛠️ the jusText algorithm for clean and focused web text extraction! 📄✨
1️⃣ Segmentation: Split HTML into blocks using tags like <div>, <p>, and <ul>.
2️⃣ Preprocessing: Remove contents of <header>, <script>, and <style>.
3️⃣ Context-Free Classification: Identify boilerplate vs.
1
0
4
@Mani___Kanta
@LangChainAI
@vercel
@Mani___Kanta
, absolutely. Here is a working example of vercel config.
0
0
3
Github project:
1
1
3
0
0
2
Instead of relying solely on outdated metrics, consider a more nuanced approach that involves testing LLMs based on:
• Fundamental abilities
• Knowledge base
• Creativity
• Cognition
• Censorship
1
0
2
Traditional metrics like BLEU and ROUGE focus on numerical comparisons with reference texts. However, high scores in these metrics can sometimes mislead, as they may not reflect true text quality or user satisfaction.
1
0
2
New release of langcorn
@0
.0.7 🤖 a FastAPI server for
@LangChainAI
Demo and GitHub Link ↓
1) Now you can pass memory and make your server completely stateless 🤯
2) You can use `X-LLM-API-KEY: sk-***` to use client's openai API key
3) Tested it by running 6 different
1
0
3
Censorship: Excessive censorship can limit an LLM's usefulness in handling complex or sensitive topics. Understand the extent of any content filtering and its impact on the LLM's responses.
1
1
2
1
1
2
The script I used to collect LLM jailbreaks from Twitter/Reddit📜. It has some challenges with multi-image prompts and merging texts, but it covers 97% of basic scenarios. Then you can pipe the text into LLM's API clients to verify the exploit.
0
0
2
@michael_timbs
I could not agree more, developer UX on Mac OS is terrible. It's slow, consumes 130GB of cache from your disk, and has constant useless updates.
0
0
1
What would the profit of AWS S3 be from hosting Stable Diffusion generated images?
1
0
0
I would recommend to not use
#TabNine
extensions since it consumes 100-400% CPU by scanning your entire disk !!! including Chrome/Mail/Calendar folders on your Mac. You can watch it via `watch "lsof -p pid"`
0
1
1
@nivi
There is a legitimate theory that our brain is not mathematical, since we understand and conceptualize math well. This implies Gödel's theorem that meta-reasoning should done from outside the system and can not be described by self-containing axiomatic rules.
0
0
0
@nivi
I think we are unfortunately far away from brain simulation. We need to duplicate 100 billion neurons to 1 terabyte of memory, simulate neurotransmitters, and simulate the active conscious process. We also need to simulate the input data from the rest of modalities to mimic the
1
0
1
@yassin_baum
Manipulation by fear, guilt, and injustice is only one of the factors that capture attention in the news.
1
0
1
@nivi
The brain is not a computer - why?
1) it is non-computational process, like many other physical processes. In other words, it's non-algorithmic.
2) Even primitive animal brains do not act robotically like they would if they were computational.
3) The brain activity cannot be
2
0
1
INTRODUCING: LangCorn, a FastAPI for LangChainAI📚🌩️
1) 💡Generates API from
@LangChainAI
pipeline
2) 🎨 Can run multiple models
4) Automatic
@vercel
deployment
(<7 lines)
Demo, Code explained, and GitHub Link ↓
0
0
1
Can we call serverless 700GB GPT3 model a microservice?🤔
0
0
1
@nivi
I am glad that we disagree on some of our understanding.
1) LLM - computational and algorithmic, it's glorified matrix multiplication. We are in agreement with that.
2) There is a contradiction between Gödel's theorem and the computational brain. The fact that we can
0
0
0
New GitHub feed failed both in product design and UI/UX
0
0
1
#go
parallel
#benchmark
needs to be normalized by number of cpus e.g If a non parallel benchmark runs for 1 sec, it will execute 10,000,000 ops, and resulting 100 ns/op. But if the same bench runs 4-way parallel for 1 sec, it will execute 40,000,000 ops resulting 25 ns/op.
0
0
1
The best way to google
#aws
#tutorial
is actually to exclude the official aws docs by
`-docs.aws.amazon.com -aws.amazon.com` is the search query. The best doc and tutorials are always written by
#enthusiasts
👍
0
0
1
RAG > Finetuning
1) Higher degree of flexibility
2) Close, same, or better quality the result
3) Less complexity in execution
4) Cost-efficient solution
0
1
1
I like
#stackoverflow
#jobs
reaction panel for
> voice-bot technology. Our founder has built several startups with many successful exits.
0
0
1