Vahid Tavakkoli’s Post

Introducing MCT Self-Refine (MCTSr) algorithm! 🚀 This cutting-edge blend of Large Language Models and Monte Carlo Tree Search boosts mathematical reasoning and decision-making. Achieving success in complex Olympiad-level math problems, MCTSr sets a new standard for AI applications. 🌟 #AI #MachineLearning #Mathematics #Innovation https://zurl.co/7xIA

Just wrapped up an insightful lecture from Stanford's CS229, diving deep into the mechanics of building large language models (LLMs). The discussion covered the evolution of LLMs, from core architectures like transformers to fine-tuning strategies that optimize model performance. Key takeaway? Beyond technical prowess, designing impactful LLMs demands a balance of data curation, computational resources, and ethical considerations. Curious about the future of AI? Check out the full lecture https://lnkd.in/db9px8xu #MachineLearning #AI #LLM #Stanford

🚀 Exploring the Future of Language Models with Multi-Token Prediction 🚀 I am thrilled to share insights on a groundbreaking new model that leverages multi-token prediction. This innovative model trains by predicting multiple future tokens simultaneously, resulting in higher sample efficiency and improved performance on generative benchmarks such as coding tasks. Key Features of Multi-Token Prediction: Enhanced Training Efficiency: Predicts multiple future tokens at once, improving sample efficiency without increasing training time. Improved Performance: Outperforms traditional models by significant margins, especially in large model sizes and generative tasks. Faster Inference: Models trained with multi-token predictions are up to three times faster at inference, even with larger batch sizes. Relevant Data: The 13 billion parameter model achieved a 12% higher problem-solving rate on the HumanEval benchmark and a 177% improvement on the MBPP benchmark compared to traditional next-token prediction models. This advancement is poised to influence the future architecture of language models, offering substantial benefits in both natural language and coding applications. For more detailed information, visit: https://lnkd.in/ggiXd8Gj #AI #MachineLearning #LanguageModels #Innovation #Tech Feel free to share your thoughts and experiences with multi-token prediction models in the comments below!

Delving into the realm of AI models recently, one standout has been DeepSeek-R1-Llama-8B. Utilizing it primarily via tools like Cursor (deepseek-v3), its remarkable potential shines, especially when operated on my robust machine (RTX 4090 with 24GB VRAM!). Impressively, this open model adeptly handles intricate tasks, spanning from code generation to natural language processing. DeepSeek showcases that smaller, specialized architectures can indeed make a significant impact, challenging the dominance of larger counterparts in the field. Though not yet multimodal like Claude, the performance I've witnessed is undeniably remarkable. The capacity to produce top-tier code and execute tasks locally has revolutionized my workflow. Here's a toast to the future of AI! Excited to witness the ongoing evolution of models like DeepSeek. P.S. Seeking a potent, open model for efficient results on your hardware? Give DeepSeek-R1-Llama-8B a spin—it's already streamlining my processes! #DeepSeek #Agents #OpenModels

OpenAI "o1" is a key turning point - signaling high reasoning performance comes with search, and a whooping cost of inference. But the bigger point is that performance bar for such workload gets raised by reward signals - and they will NOT always come from text, but from retrieval sources, or user interactions or other feedback mechanisms. So, just the off-the-shelf "o1"s will not be enough. This path is not new - AlphaGo demonstrated it. We have shown in our work over the last year how a Monte Carlo Tree Search [1] or even a beam search guided by accurate signals can outperform GPT-4 for specific application domains like chemistry [2]. Every hardware player has been focused on pushing down the cost of single inferences. The entry of search on top of inference, will change the game. [1] https://lnkd.in/gWcgXpew [2] https://lnkd.in/g7TtcCxQ #reasoning #openai-o1 #ai

✨ Exciting insights from Yann Dubois's recent lecture at Stanford University on building Large Language Models (#LLMs) 🟢 Components of LLMs: Focus on architecture, training algorithms, data collection, evaluation methods, and system optimization for efficient model deployment. 🟢 Pre-training vs. Post-training: Understand the classical language modeling paradigm and the emerging trend of fine-tuning models like ChatGPT to function as AI assistants. 🟢 Importance of Tokenization: Effective tokenization strategies are crucial for handling diverse languages and ensuring model performance. 💎 This lecture is a must-watch for anyone interested in #AI advancements! 👉 https://lnkd.in/dcevp3cJ

𝟐𝟎𝟐𝟐 𝐰𝐚𝐬 𝐚𝐧𝐨𝐭𝐡𝐞𝐫 𝐢𝐦𝐩𝐨𝐫𝐭𝐚𝐧𝐭 𝐲𝐞𝐚𝐫 𝐟𝐨𝐫 𝐂𝐨𝐦𝐩𝐮𝐭𝐞𝐫 𝐕𝐢𝐬𝐢𝐨𝐧 𝐚𝐟𝐭𝐞𝐫 𝟐𝟎𝟏𝟐. As the transformer took over, ViT showed good performance on large datasets, CNN community wanted to understand the reality. After Swin Transformers and ConvNext papers, we understood that a combination of CNN and ViT is the best for computer vision tasks. ConvNext is based on CNN but it borrowed the design structure of the ViTs. I made this time series of computer vision models: CNN-based and ViT-based, and then a slow amalgamation of them. Naturally, I have to give LLM (transformer-based) [green], and vision language multimodal models (blue) to help you understand where the research is going towards. Now is the time for Multimodal models and the creation of Foundational Models, which proved extremely helpful for downstream tasks in NLP. I believe multimodal models can be used for better image classification due to the enhanced feature representation. With the Segment Anything Model, the idea of Visual Prompting sounds to be a scalable approach towards segmentation. #statistics #machinelearning #deeplearning #dataanalytics #datascience

Based on the insights from the paper "Attention is All you Need" by Google Researchers in 2017, the transformative 'Transformers' architecture has revolutionized the realm of General Artificial Intelligence (Gen AI) and Large Language Model (LLM). Here is a simplified overview of the different components of this architecture for everyone's comprehension: 1. **Tokenization:** The model processes prompts by breaking down sentences into words or subwords, assigning a unique token to each. Think of this as creating a dictionary for the model, where every word and subword has a distinct position for easy reference. The size of this dictionary can vary across different models. 2. **Word Embedding:** Following tokenization, each token is embedded into a multi-dimensional coordinate based on its features, reflecting the word's meaning from its various appearances in sentences. In essence, each word/subword is transformed into a unique vector, with its magnitude as the token and direction as the word embedding. 3. **Positional Encoding:** An array is generated based on the positional order of each word/subword in the prompt, aiding in maintaining the positional information. 4. **Multi-Head Self Attention:** This component enables the model to grasp the context of the current word by considering other words within the sentence. 5. **MLP (Multi-Layer Perceptron):** Drawing on pre-training data from sources like Wikipedia and Quora, the model incorporates knowledge gained during training, enhancing its understanding and predictive capabilities. 6. **Prediction:** Utilizing context, knowledge, and temperature settings, the model predicts the subsequent word in the prompt. The temperature setting, akin to a TV tuner, influences the randomness of the machine's response, with lower values yielding more deterministic predictions and higher values introducing greater variability. This foundational architecture forms the cornerstone of various Gen AI concepts, paving the way for advanced language processing capabilities. Special thanks to Sumit Mittal Sir for the enlightening course "AI - The Ultimate Masters Program for Data Engineers #Sumitmittal #AIforDataEngineers #AI #LLM #AIBASICS #GENAI #DataEngineering #AI #ContinuousLearning

Sharing Some Insights from a Stanford Online Lecture on Building LLMs: I came across a publicly accessible lecture from Stanford Online that some might find worthwhile, especially if you’re interested in large language models (LLMs). Here’s a concise overview: • Pretraining and Data Filtration Shows how large web-crawled datasets are carefully filtered, de-duplicated, and balanced to improve model quality. • Alignment and Fine-Tuning Explains methods like Reinforcement Learning from Human Feedback (RLHF), which transform a raw model into a more user-oriented assistant. • Scaling Laws Demonstrates why more parameters and more data often result in better model performance, and how to manage these resources efficiently. • Systems Optimizations Highlights techniques (e.g., low-precision arithmetic, operator fusion) to handle the immense computational demands of LLMs. This is a straightforward, high-level look at the core steps involved in building and optimizing modern language models. It may be relevant for those exploring AI research or industrial deployments. As always, credit goes to Stanford Online for making this lecture available. https://lnkd.in/dt4NcTtn

QUANTIZATION IN LARGE LANGUAGE MODELS Hoang Nguyen (our engineer) explains quantization - a practical solution for deploying large language models (LLMs) on everyday hardware. Quantization compresses model weights and activations, often converting 32-bit floating-point values to 8-bit integers. Benefits: - Smaller models for easier deployment. - Lower memory and storage requirements. - Faster computation with reduced energy consumption. Two Approaches: 1. Post-training quantization (PTQ): Applied after training for optimized inference. 2. Quantization-aware training (QAT): Simulates quantization during training for better accuracy. How it works? Quantization maps floating-point values to smaller ranges using linear quantization: - Symmetric: Aligns zero directly for efficiency. - Asymmetric: Adjusts zero for better accuracy. The process calculates scale and zero points, maps values to integers, and de-quantizes them for inference. Efficient model formats: GGUF (Generic GPT Unified Format) enables storing and running quantized models efficiently. Naming conventions (e.g., q8_0, q6_k) indicate the bit precision and method used. With LLMs growing to billions or trillions of parameters, quantization reduces resource requirements while maintaining performance. Learn more about quantizing LLMs: https://lnkd.in/gyd2BEKQ #LLM #AI #dwarves #software _ Dwarves Notes (https://memo.d.foundation/) combine our team’s collective know-hows, R&Ds, and operation approaches. Connect and learn alongside other tech fellows: - Discord: discord.gg/dwarvesv - Github: https://lnkd.in/gZZ2eZMu - Website: https://d.foundation