Clint J.
@SearchDataEng
Followers
912
Following
1,307
Media
611
Statuses
6,110
Explore trending content on Musk Viewer
Cheney
• 468785 Tweets
Texas
• 229611 Tweets
Michigan
• 189797 Tweets
England
• 148330 Tweets
Mustafa Kemal
• 133881 Tweets
Ireland
• 123315 Tweets
MSNBC
• 74358 Tweets
Lookman
• 47456 Tweets
Nelly
• 38557 Tweets
Grealish
• 32893 Tweets
Arkansas
• 29646 Tweets
Kentucky
• 29036 Tweets
#كاريزما63
• 24590 Tweets
Iowa
• 24026 Tweets
Musiala
• 21052 Tweets
Ohio State
• 16289 Tweets
Penn State
• 16106 Tweets
Pedro Almodóvar
• 15876 Tweets
#BaşkomutanErdoğan
• 14579 Tweets
Sabalenka
• 14004 Tweets
#الهلاليين_يرفضون_لودي
• 13010 Tweets
León de Oro
• 11538 Tweets
Auburn
• 11521 Tweets
Riya
• 10997 Tweets
Sellers
• 10271 Tweets
Potential LLM learning course for python developers :
1. Learn OpenAI ( assistants, function calling , multi-modal , Chat Complete, ChatGPT)
2. Learn Weaviate
3. Learn unstructured
4. Learn DSPy
5. Learn ollama and OpenSource AI
6. Learn Finetuning
7. Learn Weights & Biases
6
28
212
pydantic
Redis
FastAPI
Ansible
Weaviate
pandas
DSPy
These are a few of my favorite things,
All such beautiful pieces of art, and also code.
Such grace, such simplicity, such Power.
2
20
181
I spent a lot of time, months even, going back and forward on manually tuned prompts in a multi step AI chain, and when DSPy was first released, I understood the value right away.
This does not only optimize a single prompt,
it optimizes a chain of prompts!
"Self-Refining
Sunday morning read and code: DSPy Assertions: Computational Constraints for Self-Refining Language Model Pipelines by
@arnav_thebigman
et al. ⚠️
DSPy Assertions give you more control over the behavior of your language model system! The language model will output the desired
5
39
184
3
13
54
My favorite part of the new DSPy documentation...
CHEAT-SHEET ! Lots of code examples, and explanation for all components.
Thank ya very much DSPy team.
1
9
47
@WarrenInTheBuff
I would like to see the front end go away, and be replaced by a network of API's , which just serve JSON.
People can create their own personal front ends locally.
Bring back geocities, and add LLM chat bot.
6
1
40
🚀 Mind-blowing DSPy meetup! 🧠💥
📢 TL;DR of the talks:
1️⃣
@simigd
@cohere
: RAG updates 🆕
2️⃣
@cshorten30
@weaviate
: Generative Feedback Loops 🔄
3️⃣
@mikeldking
@arize
: Observability & Ops 🔍
4️⃣
@lateinteraction
@stanfordnlp
: DSPy deep-dive 🌊
It is all here, all that
1
11
41
Microsoft , Anthropic, and google trying to catch up to Groq speed.
1
1
38
I imagine that ChatGPT will integrate DSPy.
This will give the user the ability to correct the responses, and these corrections will act as new DSPy examples to compile on.
It will also be possible for the user to train ChatGPT with new skills.
"You should execute code like
3
7
36
@GaryMarcus
@WSJ
People are using LLM to apply.
Companies are using LLM to review.
Endless death loop
2
0
32
Claude likes XML , OpenAI likes JSON, and sometimes plain text which includes ## .. You can spend hours, days, months, and years trying to find the best prompt, and then the model changes in some small, or large way.
DSPy is fit for just this scenario, and Erika does an
One prompt does not fit all language models ☝️
Luckily for you, DSPy automates the task of prompt engineering!
Here is a thread with a few things to know about the collection of compilers in DSPy. It is also outlined in a new blog post from
@CShorten30
and I, “Your Language
8
65
251
1
6
33
Today I am watching (studying)
@ecardenas300
recent tutorial on Weaviate, and DSPy .. Their educational courses on DSPy began at a high level, months ago, now they are getting into the nitty gritty.
Thank you Weaviate Academy!
I will follow up, and report on what I learn!
1
5
30
Really useful and informative article on semantic chunking for RAG, and how Unstructured handles it.
One might not assume that it is such a challenge, but PDF and other data formats are not clean, and normalized as CSV or JSON.
0
7
29
High Quality - Vector Retrieval ready chunking is what makes
@UnstructuredIO
the perfect solution for PDF extraction!
4
4
28
@GaryMarcus
@seanmcarroll
I'm sorry as an AI created by OpenAI I am not able to solve all of physics at once, however if you would like to discuss recent theories published on Wikipedia, then I am here for you.
0
0
27
@GaryMarcus
After "borrowing" all intellectual property open to a web browser , in all of the world wide web, what is one more little voice here or there?
1
1
20
@bindureddy
Some people will use LLM so much that they will be incapable of doing "it", whatever it may be without the LLM.
Like developers that generate code that is beyond their ability to debug, or understand.
1
0
19
Using llama3 with ollama on my macbook pro, and a few observations :
It is fast. Accurate.
Most interesting... It asks follow up questions more than OpenAI or Claude, out of the box.
3
0
19
In yon
#Weaviate
's hallowed halls,
Where data's voice in silence calls,
A tale unfolds of search so grand,
That spans across th' electronic land.
The secret lies within the soil,
Where Ethernet and roots embed.
A mycelium of networks vast,
Ensures their growth will ever last.
4
3
18
As far as I can tell dspy is *like* Machine Learning for LLM prompting.
1
3
19
LLM pricing (per million output ) :
Llama 3 70B on Groq : .79
Claude Opus : $75.00
gpt-4-turbo-2024-04-09 : $30.00
1
2
19
On the topic of LLM pipelines.
@UnstructuredIO
seems to be the best parsing library for all varieties of input sources.
Here is a brief tutorial for PDF ingestion!
1
3
18
The python returns library is really cool, and I think a good design choice for LLM agents.
3
1
18
Multi Stage Vector Search - Weaviate
Sometimes one query is just not enough, especially when dealing with short to long text, with high degrees of similarity .
1
4
16
I still remember my first time calling a function with a LanguageModel.. I think at the time I was using Langchain, and I was calling some kind of SQL script, or maybe it was a pandas call .. no , matter .
The exhilaration was real, even a tad bit of adrenaline , only matched by
Language models paired with function calling (or tool use) is a powerful way to build AI systems 🤖
Developers define a set of functions/tools, what they do, and their input arguments. Then we leave it to the language model to select the right one.
For example, given the
4
29
105
2
3
18
@Scobleizer
@GoDaddy
They likely bought the domain because they know exactly who you are, and the value of the domain.. I believe that PageRank is still included with purchased domains.. so they get all of your SEO progress, and backlinks as their own.
They may be willing to sell it back to you for
1
0
17
from their example directory :
Notice the wonderful inclusion of lru cache
1
1
15
@GaryMarcus
That is what I heard about video LLM as well.
It is much, much more expensive to train, and to generate content with.
These LLM companies aren't even making money with text LLM, how in the world will they make money from video?
1
0
14
I will be spending much of my study time in the DSPy discord community now.
All of the stars of DSPy are there already!
Yessssssss
0
3
16
@AbHomineDeus
@ASM777__
Can they use the same hardware ( robot ) with new software ( AI ) ?
I don't know anything about robotics, but it seems like their robots are the most advanced in terms of agility.
2
0
14
@lateinteraction
explains how DSPy is an LLM model equalizer , and it shifts importance to the architecture, and the LLM pipeline!
1
1
15
Two of the functions that I am creating so far to clean headers, and remove English artifacts.. It is not there yet, but I am making progress.
This is part of a custom PDF to Audiobook tool.
1
1
15
Pre-Generative RAG goals :
1 . Remove filler text
2. Remove irrelevant text
3. Remove duplicate concepts
4. Categorize
5. Summarize
6. Embed
7. Normalize size
8. Create metadata for Vector filtering.
1
2
15
What might happen if OpenAI integrated DSPy into GPT-5 ?
From a user, and developer stand point.
If all prompts, and instructions trained the API KEY, and they were considered for future outputs.
Not just fine tuning though, recompiling ideal prompting, on hyper-paramaters for
5
2
14
Last clip of the night.
Billions, and trillions being invested into new & more powerful language models, and then listen to what he says.
Where is the investment in LLM system architecture primarily?
Makes ya think.
0
3
13
@Thewimo
@CShorten30
Before using Weaviate, I used MongoDB (not vector search), and it is powerful for its use cases (storing and filtering on nested structures), but in my opinion, it is also slow, bloated, and complicated.
I used Chroma for a while, and I did love and appreciate the simplicity of
1
3
12
@pwang
@IntuitMachine
@remisussan
@internetarchive
Yeah, and this is why we need Cooperative / Open Source data archival warehouses.. So that no one can modify these archives without our knowing.
1
0
11
@lateinteraction
LLM Stacking - Small Questions - Multi-Hop , these are all new concepts that I have learned since I started researching DSPy .
And I can say from personal experience.. prompt engineering in a pipeline, with more than maybe 5 steps is an endless nightmare.
Change, one comma..
1
1
12
Absurd to think about,
expectation :
A man, interacting with an OnlyFans model, sending her money, hoping to be noticed,
reality :
Unknowingly interacting with a custom trained LLM... never interacting with his favorite model even once.
6
2
8
@BenjaminDEKR
@DrJimFan
Probably by posting about a massively improved GPT-4.51-Turbo, without any specificity .
2
0
11
@thomasahle
pydantic integration with DSPy is coming, from someone, somewhere.
I would imagine an interface like instructor, but wrapping DSPy, and Colbert / Weaviate.
Perhaps it could even have instructor built into it, though I'm not sure how instructor plays with DSPy
1
2
11
The brittle nature of stringing together LLM prompts, and better methods with DSPy
0
2
11
This is a special one from Weaviate! 100th episode of the Podcast. Congratulations Connor, you are a model of passion, and persistence.
Here is a technical summary of the podcast.. They are breaking new ground with this discussion it seems.. Shaking paradigms, and introducing
I am SUPER EXCITED to publish the 100th episode of the Weaviate Podcast (💯🎙️🎉) with Lucas Negritto (
@lucasteez
) and Bob van Luijt (
@bobvanluijt
) on Generative UIs! 🖼️♻️
This is an amazing example of AI-native applications and the new generation of software! Rather than
12
13
61
0
3
11
When AGI (a point that cannot be defined, or described, only known when experienced )is reached ,
a few streams will combine into a river. Then all to the Sea.
Those streams are :
1. Larger models training smaller models.
2. Vector Search
3. Agentic chains, calling an open Hub
Agentic Reasoning and Acting is the missing piece in your RAG applications 🤖
RAG systems are usually hard-coded pipelines with just retrieval and generation, whereas Agents utilize LLMs for decision-making.
The ReAct framework by
@ShunyuYao12
et al. further changes the Agent
5
76
365
2
3
10
Yesssssss , this might be the front end solution for my absurd Search Engine.
1
0
10
1
0
10
I am continuing to learn the new
@LangChainAI
and
@OpenAI
Function - Chain integration, I have added a custom tool to the agent. The tool is from the uszipcode library, & it takes a ZIP code, or a City and State, and returns information about the city. Demo of multi step chain!
2
1
10
@GaryMarcus
Why should we expect AGI to return the best response in one step, and not in many steps?
Why aren't we looking to develop prompt hubs? Millions of open tools accessible to the LLM.
Access to the Internet in the form of a uted distribknowledge graph.
Constant learning (
1
0
8
@philipvollet
@weaviate_io
@femke_plantinga
@etiennedi
@aestheticedwar1
I don't see my favorite Weaviate teacher ={
2
1
10
I feel like the OpenSource LLM community has already surpassed GOOG, MSFT, ChatGPT ,combined. (in the past 2 months)
Sure, it may be possible to get 10% better general accuracy from GPT-4-turbo, better images, and that is about it.
OpenAI vector retrieval is not great, there is
0
1
8
@AISafetyMemes
If only Samael would have thought twice, before in his act of curiosity he created all of this. And he probably did think twice, but alas curiosity, maybe even greed got the best of him.
1
0
8
LLM Development is hard.
That warm and cozy sense of certainty that exists for standard software development, is absent in LLM dev.
You just never know what might come back.
1
0
9
I'm digging into Weaviate Generative Search, and exploring the benefits over standard Vector Retrieval.
I have to head straight to the Weaviate podcast, before proceeding!
1
2
9
@hwchase17
Ragas looks pretty interesting,
"Faithfulness: measures the factual consistency of the generated answer against the given context.
Relevancy: measures how relevant retrieved contexts and the generated answer are to the question"
0
2
8
Software Developers, using LLM to program.
You will do well to learn the notebook, and the pen.
Get yourself a nice pen, one you like.
Walk through all steps in the program, write them down as basic doc strings in your notebook.
Then double check the whole flow, bring it by
0
0
9
My ideal AI home assistant would have these qualities :
1. Cool and high quality box designed by TE.
2. Local small LLM running on ollama.
3. Connection to Groq API running llama3 , or Mixtral.
4. Vision capabilities.
5. Integration with Whisper, and Elevenlabs.
6. Personal
2
1
8
1
3
9
@ns123abc
@GaryMarcus
It could learn to count.
It could learn what grammar rules mean, and how to utilize them.
It could handle JSON input and output of 100's of rows.
It could plan ahead, and anticipate challenges in multi step pipelines.
It could replace only a single word , or phrase in a text
2
0
9
What it means to add depth (layers) to a DSPy program, and most important how to do it.
Check it out, and you will not regret it.
Adding Depth to DSPy Programs!! I am SUPER excited to share a new video exploring what it means to add depth to DSPy programs! 🚀
One of the headlines of DSPy is the comparison with PyTorch, whereas layers == sub-tasks as we decompose complex tasks like writing blog posts or
8
59
242
1
3
9
Now I embark into the latest and greatest tutorial from
@CShorten30
, as he has described it "Hello World" for DSPy.
When learning tools like this, I first like to identify all of the most important concepts, in this case "teleprompters" , "assertions" , "suggestions" ... and so
2
1
9
Weaviate ( the best community in all of LLM ) has a hero community . Very cool!
I found this when searching
"where can I buy Weaviate stickers"
0
2
9
@CShorten30
I'm making my way through your "DSPy Explained" video, and it is a good one, thanks for putting that together.
0
1
9
@BrianRoemmele
It even gets the ridiculous humor element of the scene right.. This is the best that I have seen yet!
2
0
9
@hellokillian
Show it a series of bluetooth devices with no further context, and have it switch connection to the devices.
1
0
7
I tried Claude function calling for the first time, and I am pleased to report that it works!
This is exciting.
To fully take advantage of the power of an LLM,
func calling, and JSON formatting ( or pydantic ) is absolutely necessary.
By integrating structured input, and
1
0
8
Some books that explore concepts of Intelligence, AI , and Humanity.
"Accelerando" by Charles Stross
"Neuromancer" by William Gibson
"Permutation City" by Greg Egan
"Diaspora" by Greg Egan
"Blindsight" by Peter Watts
"The Culture" series by Iain M. Banks
"The Quantum
2
0
7
@GaryMarcus
Just give it 6 months, until we have AGI, then it will be able to count and understand the concept of time.
2
0
8
@semrush
Google Search I guess, but I am trying to change my habits to use Perplexity AI instead.
2
0
8
Ollama is a tool that seems too simple to be able to do so much.
I spent many days on the documentation, trying to find the difficulty in the installation, or in running it.
Then, finally, I overcame my complexity neurosis, installed it, and ran it from the command line. After
1
1
8
Today is a
#Weaviate
kind of day for development.
I have just discovered Embedded Weaviate, and it is precisely what I needed!
Before finding the Embedded solution I was using ChromaDB, but I had to create a custom Hybrid model, and that just won't work.
0
4
8
The LLM model is just one smallish part of the whole, when it comes to a production LLM Pipeline.
LLM are great with some tasks, and terrible or mediocre at other tasks.
What about SQL, what about Cloud tools, how to put them all together? That is no small task.
Weaviate
Although your data is distributed throughout numerous databases, you can still use ALL of it for your RAG application 🍱
In this notebook, we build an end-to-end RAG pipeline that uses
#Google
's Big Query and
@weaviate_io
, using DSPy!
Context Fusion with Agents will first
7
40
143
1
3
8
I think that people want a cool open source "Siri" like hardware device, more than they want a hand held phone competitor.
I think that people would like for their home smart device to ask questions, and to locally remember conversations, and to always be learning.
2
1
8
If it is true what he says that AI is not intelligent, then it seems that a library like DSPy gives the AI intelligence.
This probably relates to
@lateinteraction
views on AGI, and the need for deterministic tools, agents, and chains, as opposed to "ever increasing amounts of
François Chollet says AI language models have a very low level of intelligence, as measured by the ARC challenge
31
38
217
1
1
8
In addition to LLM :
Reinforcement Learning (RL)
Graph Neural Networks (GNNs)
Self-Supervised Learning (SSL)
Symbolic AI and Knowledge Graphs
Evolutionary Algorithms and Neuroevolution
Federated Learning and Privacy-Preserving AI
Wetware and Biological Computing
1
1
8
In building an AI startup.. it is absolutely essential that you do something difficult. Likely including scraping, complicated extractions, custom API integration.
Elsewise ChatGPT or other big-LLM will simply put you out of business in an upcoming update.
2
0
8
The Community that surrounds a software library is almost as important as the library itself.
A good community will welcome, and help new people learn, and they will continue to develop, and refine on top of the original.
0
1
8