Karthik Kalyanaraman’s Post

Hello peeps, I just tested OpenAI's brand new o1 model with James Murdza. The model claims to spend more time "thinking" before responding, and it was interesting to test it quickly based on how it reacts to certain problems that previous GPT models have struggled with. We started with the notorious: "How many occurrences of 'r' can we find in the word 'strawberry'?" GPT-4-O counted only 2 'r's, while in reality, there are 3. Although we could work around this by asking it to count based on a Python script (which did return the correct result), OpenAI's o1 mini took a "chain of thought" approach, processing for 5 seconds before coming up with the correct number of 'r's. It was fascinating to see how it broke the problem down step-by-step, which helped it get the correct answer, letter by letter. Next, we tackled the infamous "Tribonacci sequence" problem, which GPT-4 and other GPT models typically struggle with. These models are often misled by the similarity to the Fibonacci sequence, resulting in incorrect outputs. However, o1 mini broke down the function, mapped it correctly, and validated the sequence, producing the right result. This demonstrated how crucial step-by-step reasoning is for solving such problems. In 10 tests, models like Code Llama, GPT-3.5, and GPT-4 had a 0% accuracy rate, but surprisingly, o1 mini solved it correctly in just 10 seconds. We also experimented with some JavaScript components using GitWit's sandbox (I love seeing preview, deploy, and terminal all on the same page) and got some interesting results. Feel free to check out the stream here:   [Stream Link](https://lnkd.in/d4Fbzs8R) And maybe test out some of these interesting questions to compare with previous models using this repository:   [Github repo with questions](https://lnkd.in/dYTiNQ4Y) Lastly, big thanks to James for having me over! #o1 #openai

𝐋𝐚𝐧𝐠𝐂𝐡𝐚𝐢𝐧 𝐋𝐋𝐌 𝐋𝐢𝐛𝐫𝐚𝐫𝐲 𝐎𝐯𝐞𝐫𝐯𝐢𝐞𝐰 𝐖𝐡𝐚𝐭 𝐢𝐬 𝐋𝐚𝐧𝐠𝐂𝐡𝐚𝐢𝐧? A library for developing LLM powered apps -  LangChain gives structure and tools for connecting LLMs to knowledge sources and guiding the LLM's reasoning process. You can think of LangChain as the bridge between raw LLM power and real-world applications. 𝐖𝐡𝐲 𝐔𝐬𝐞 𝐋𝐚𝐧𝐠𝐂𝐡𝐚𝐢𝐧? 𝐓𝐨 𝐦𝐚𝐤𝐞 𝐋𝐋𝐌𝐬 𝐦𝐨𝐫𝐞 𝐩𝐨𝐰𝐞𝐫𝐟𝐮𝐥 𝐚𝐧𝐝 𝐯𝐞𝐫𝐬𝐚𝐭𝐢𝐥𝐞 - LangChain allows you to connect LLMs to external knowledge sources (databases, documents, websites). This feature makes LLMs more powerful and unlocks applications beyond just basic LLM responses. 𝐓𝐨 𝐬𝐚𝐯𝐞 𝐝𝐞𝐯𝐞𝐥𝐨𝐩𝐦𝐞𝐧𝐭 𝐭𝐢𝐦𝐞 - LangChain offers pre-built components, integrations, and templates. These pre-built components and templates reduce the need for low-level coding, allowing you to focus more on your application's unique logic. 𝐊𝐞𝐲 𝐋𝐚𝐧𝐠𝐂𝐡𝐚𝐢𝐧 𝐂𝐨𝐦𝐩𝐨𝐧𝐞𝐧𝐭𝐬 𝐂𝐡𝐚𝐢𝐧𝐬 - These are sequences of LLM calls combined with other tools. For example: A chain could (1) search a knowledge base, (2) summarize the information found, and (3) have an LLM generate a response using the summary. 𝐀𝐠𝐞𝐧𝐭𝐬 - You can think of agents as AI-powered decision-makers. An agent uses an LLM to choose a series of actions, executes them, and adjusts behavior based on results. Example: an agent for writing a marketing email could generate text, select images, and proofread its own work. 𝐌𝐞𝐦𝐨𝐫𝐲 - LangChain allows state and information to be retained between different calls to your app. This lets apps have "ongoing conversations" and adapt based on past interactions. 𝐋𝐋𝐌 𝐖𝐫𝐚𝐩𝐩𝐞𝐫𝐬 - LangChain simplifies working with different LLMs (like OpenAI, Google, etc.) by offering a standardized interface. To summarize, LangChain is a superb LLM library that simplifies the task of building applications powered by large language models (LLMs). Refer to the first comment for the playlist details. #llms #generativeai #nlproc #ai #datascience #deeplearning #ai #langchain #library #python

🌟 Excited to share my completion of DataCamp's "Introduction to Embeddings with the OpenAI API" course! 🌟 I dived into the world of embeddings, creating semantic search engines and recommendation systems. Learned to use ChromaDB for efficient text embedding storage and querying. A big leap forward in my AI journey! Onward to new AI projects now! 👀 #AI #MachineLearning #Embeddings #OpenAI #DataScience #Python #ArtificialIntelligence #DeepLearning #TechSkills #dcdonates

💥 🚀 OpenAI just released the much much anticipated o1 model (rumored as Q*/ strawberry/ Next etc.) Full blog: https://lnkd.in/gNBxXiHz The Aha! 👍🏽 Exceptional at complex reasoning questions, crushed benchmarks 👍🏽 As expected, the LLM uses Chain of Thought to slowly think thru the problem and improve accuracy. O1 generates various paths it could follow and evaluates progress at each step. OpenAI used Reinforcement Learning to train o1 to abandon a path that's not working and try an alternative approach. 👍🏽 o1 is not yet using agentic features eg. web browsing, file support, python etc. This is just the base LLM without adding all the addtn capabilities. 👍🏽 scored 89th percentile on competitive programming questions (Codeforces) 👍🏽 Ranks among the top 500 students in the US for the USA Math Olympiad qualifier (AIME) 👍🏽 First model to exceed human PhD-level accuracy on physics, biology, chemistry problems (GPQA) diamond benchmark 👍🏽 Self reflection at each step gives it a chance to check it's safety alignment, and meaningfully prevents jail breaks 👍🏽 It was the first model I have used that has been able to code up Tetris game in one shot flawlessly The Uh Oh! 😏 A bit overhyped to call it "reasoning". But this is significant progress towards it. 😏 No mention of how they would price this. There were rumors that they could demand $2,000/month for a human-equivalent digital agent and justify the ROI 😏 They havent talked about the increase in Latency, or time to first token casued by chain of thought etc. I am seeing it take 4 sec to answer how many ‘r’ in strawberry. A coding task took 23 sec. 😏 I didn’t see ways in which enterprises may be able to influence the thought process within o1. 😏 Didn't mention if these could support finetuning 😏 Some tasks are still just as well addressed by GPT4o 😏 Significantly increases the cost of inference. Both in # of output tokens and compute cost. -------- 🔔 If you like such content, I encourage you to connect, comment and repost ♻️ Subscribe to my free ‘AI with Whiskey’ newsletter https://lnkd.in/gK4zrSYv --------

Simplifying API Calls with OpenAI's Structured Outputs 🚀 OpenAI’s latest structured outputs feature streamlines function calling by replacing complex JSON schemas with easy-to-use SDK tools. In our new blog, we walk you through the transition from traditional API integration to this more efficient method. With examples like fetching stock prices and currency exchange rates, you’ll see how these improvements save time and reduce errors. Follow our blog " Simplifying OpenAI Function Calling with Structured Output: A 2024 Guide: https://lnkd.in/eN8_3Q7W " We're here to help you with any questions you may have. #AI #NLP #MachineLearning #OpenAI #JSON #FunctionCalling #StructuredOutputs #APIs #Python #PragnakalpTechlabs

Google’s "Gemma is not Open-source" Google just released Gemma, their first open-weight LLM based on Gemini, which outperforms Mistral AI 7B! https://lnkd.in/dVfcXyYE Before diving into the nitty-gritty, Here's the TL;DR: - Trained on 2T tokens (2B) and 6T tokens (7B) - Default context window - 8192 tokens - Used RLHF for instruction model - 7B trained on 16 TPUv5e pods for 23 days - 2B can run on Phones - Beats Mistral 7B, DeciLM 7B, and Qwen1.5 7B - Available for use with JAX or Pytorch - Allowed for commercial use - 7B scores 64.56 on MMLU and avg. of 63.75 on the Leaderboard Overall, Amazing performance! You can see the base 7B model on the Open LLM Leaderboard here: https://lnkd.in/dhBYESDg And here's my piece: Google's latest models are open-weight, not open-source. has made Gemma an "open model," which means that the pre-trained parameters, or weights, are available for everyone to access. However, the source code and training data are not available to the public. According to an announcement by Google, the Gemma models are now being released as "open models". This means that individuals can use them for free and access the model weights. However, there are some rules about how the models can be used and shared, depending on the specific terms of use for each model. But, the good news is that the Gemma models are available to developers, researchers, and commercial users for free, and they can also create and publish their own versions of the models. 📚 The Gemma models’ terms of use: https://lnkd.in/dWgmkbh5 Back in the days, from the original Transformers paper to Tensorflow (both open-sourced by Google), open-source development has allowed researchers to iterate, improve, and adapt. This has encouraged more people to use it and has made it more useful in many different ways. But that is not the exact case with Gemma. However, there are still many benefits to making its model weights available to the public. These benefits include the ability to collaborate and transfer learning from one model to another. Even though it's not open-sourced, you should definitely give it a try. Here are the some resources: Kaggle: https://lnkd.in/dtrZfkVZ HF: https://lnkd.in/dmHtGgGv Colab: https://lnkd.in/dZ5W-uXK Github: https://lnkd.in/dVfcXyYE Happy Reading!

Just cracked the code on text chunking for Rag. After testing multiple strategies, here's the surprising result: RecursiveCharacterTextSplitter: Performs poorly. Avoid if possible. Semantic Chunking: Better results but slow. And still didn't work consistently. Perplexity based chunking: A novel strategy that looks at model perplexity instead of sentence similarity (more on this soon). Performed very well but my naive implementation was too slow! Markdown Splitter: Decent for markdown texts, but limited use case. The game-changer: Using GPT-4o-Mini for chunking. Here's why it's superior: 1. Quality: Significantly outperforms traditional methods in coherence and relevance. 2. Cost-efficiency: Can process millions of documents for only a few hundred dollars. Beats most commercial solutions. 3. Speed: Faster than semantic chunking. We're talking dozens of hours saved on large datasets. 4. Context-awareness: The AI provides contextual information for each chunk. This is crucial for maintaining meaning when chunks are processed individually or used for embeddings. 5. Flexibility: Works on any text type, not just markdown. 6. Scalability: Easily handles large documents and datasets. Implementation detail: We ask the model to identify the necessary context for each chunk. This ensures that when an LLM later reads a chunk in isolation, it has the required background info. Unexpected bonus: The chunks make more sense in embedding space! Tech stack: Python, OpenAI API, Streamlit for the demo interface. If you want to see it in action or contribute: Demo: https://lnkd.in/gUvR3yZZ Code: https://lnkd.in/g4ERwEFy No API key needed to try it out. Pull requests welcome. Next steps: I'm exploring more robust settings for LLM based chunking, where you can input your own prompts/data. Curious about your experiences with text chunking. Any strategies I missed? How does this compare to your current approach?

In this tutorial, Tahreem Rasul explains why LLMs struggle with math, and how we can resolve these limitations using LangChain Agents, OpenAI and Chainlit:

Been test-driving Andrew Ng's new library, and wow... 🚀 Think of it as the ultimate "universal remote control" for AI models - aisuite lets you seamlessly switch between different LLMs (OpenAI, Anthropic, Google, etc.) without changing a single line of code! 🎮 🔥 What makes it special: - One interface to rule them all (just like a Swiss Army knife for AI) - Support for major players: OpenAI, Claude, Google, AWS, Groq, and more - Super clean Python implementation - Perfect for A/B testing different models As an Andrew Ng fanboy, I'm especially excited about how this simplifies AI development. No more juggling between different APIs - it's like having a universal translator for AI models! 🌎 The best part? It's open source and actively maintained. Whether you're building a startup or just experimenting with different LLMs, this is a must-have tool in your AI toolkit. 🛠️ Check it out here: https://lnkd.in/gfD3wfdZ #AI #MachineLearning #OpenSource #AndrewNg #ArtificialIntelligence #Python #Development