Sai Teja Gangapuram

💥💥💥 Align-SLM: Textless Spoken Language Models with Reinforcement Learning from AI Feedback Guan-Ting Lin, Prashanth Gurunath Shivakumar, Aditya Gourav, Yile Gu, Ankur Gandhe, Hung-yi Lee, Ivan Bulyko Abstract While textless Spoken Language Models (SLMs) have shown potential in end-to-end speech-to-speech modeling, they still lag behind text-based Large Language Models (LLMs) in terms of semantic coherence and relevance. This work introduces the Align-SLM framework, which leverages preference optimization inspired by Reinforcement Learning with AI Feedback (RLAIF) to enhance the semantic understanding of SLMs. Our approach generates multiple speech continuations from a given prompt and uses semantic metrics to create preference data for Direct Preference Optimization (DPO). We evaluate the framework using ZeroSpeech 2021 benchmarks for lexical and syntactic modeling, the spoken version of the StoryCloze dataset for semantic coherence, and other speech generation metrics, including the GPT4-o score and human evaluation. Experimental results show that our method achieves state-of-the-art performance for SLMs on most benchmarks, highlighting the importance of preference optimization to improve the semantics of SLMs. 👉 https://lnkd.in/dXBzaibk #machinelearning

4

I was shocked to see the below graph 🤯. I saw this Google Search trend graph on Chat GPT VS Perplexity VS Claude. I now feel like I am in a bubble surrounded by tech folks who use claude etc and to me it feels normal. But as you can see in the graph, the actual use of Claude & Perplexity is very very minute compared to Chat GPT. Especially Perplexity feels like its non-existent. And now a month back Chat GPT Search is also launched which will definitely take a share of this audience. Looking at this graph also makes me optimistic about building, It clearly shows that there is a huge gap in the market, and similar products can be built. What do you guys think ? #ai #startups #chatgpt #claude #perplexity

21

4 Comments

The Most Influential AI Week for India - Credit- AIM Amit Raja Naik From high-stakes visits to breakthrough partnerships and bold innovations, this was no ordinary week for AI. It was a game-changer, cementing India's place in AI history.  Here are the key highlights from NVIDIA AI Summit India 2024:  1- NVIDIA and Reliance Partnership for AI Infrastructure: Huang announced a landmark partnership with Reliance Industries, aiming to scale AI infrastructure with the latter’s $10 billion investment in data centres across India. Reliance is building infrastructure for 1 GW at Jamnagar, which is scalable for multi-GW use, backed by green energy, to support NVIDIA’s latest GPUs and AI models. 2-20x Increase in AI Compute Power in India by Year-End: NVIDIA looks to multiply its compute infrastructure in India by 20 times over the next year, supported by partners like Yotta, Tata Communications, and E2E, positioning India as a global leader in generative AI. 3-Development of LLMs in Indian Languages: Acknowledging India's diverse linguistic landscape, NVIDIA is collaborating to create a Hindi large language model and potentially other regional models to bridge the digital divide across India’s 25 formal languages and 1,500 dialects. Jensen noted that “this is the hardest language model region in the world”, underscoring the unique challenges and strategic advantage in building AI for a multilingual society. 4-AI’s Exponential Compute Demand and Blackwell GPU Launch: Jensen introduced NVIDIA’s Blackwell GPUs, designed to handle the 4x increase in compute demand needed each year for training ever-larger models. Blackwell racks, capable of delivering 120 kW of compute power, can train models with over a trillion parameters, with shipment to India scheduled for Q4 2024. 

14

Power of #Gemini #Google #DeepMind Experience the power of Gemini: IIT math question solved in seconds with detailed guidance. Unlock your full potential with AI-enhanced learning. I have only uploaded the image of the question and didn't share any options , the prompt was " Find the value of n " Here's how to solve the problem: 1. Focal Lengths of Mirrors: The focal length of a concave mirror is half its radius of curvature. Therefore: Focal length of M₁ (f₁) = -10 cm (negative for concave mirrors) Focal length of M₂ (f₂) = -12 cm 2. Object Distance for M₂: The object S is at the midpoint between L and M₂, so the distance between S and M₂ is 10 cm. 3. Image Formation by M₂: Using the mirror formula: 1/v + 1/u = 1/f u (object distance) = -10 cm f (focal length) = -12 cm Solving for v (image distance), we get v = -60 cm. The negative sign indicates the image is formed on the same side as the object, which is expected for a concave mirror. 4. Image as Object for the Lens: The image formed by M₂ acts as the object for the lens L. The distance between this image and the lens is the distance between M₁ and M₂ minus the image distance: 20 cm - (-60 cm) = 80 cm. 5. Image Formation by the Lens: Again, using the lens formula: 1/v - 1/u = 1/f u (object distance) = -80 cm (negative as it's on the same side as the incident light) f (focal length) = 10 cm Solving for v (image distance), we get v = 80/9 cm. This image is formed on the right side of the lens. 6. Final Image Coinciding with S: The final image formed by the lens must coincide with the object S for the condition given in the problem. This means the image formed by the lens must be 10 cm to the left of the lens (where S is located). 7. Image Distance for M₁: To achieve this, the image formed by the lens should act as a virtual object for M₁ at a distance of 10 cm + 80/9 cm = 170/9 cm. 8. Object Distance for M₁: Applying the mirror formula for M₁: 1/v + 1/u = 1/f v (image distance) = -n/7 cm (negative as it's a virtual object) f (focal length) = -10 cm u (object distance) = -170/9 cm 9. Solving for n: Plugging in the values and solving for n, we get: 1/(-n/7) + 1/(-170/9) = 1/(-10) Simplifying and solving, we find: n = 34 Therefore, the value of n is 34.

12

Here is a wonderful paper Published last week by MIT, Drexel, Harvard, BU & Lehigh on LLM benchmarks using Game Theory with strategic scenarios & Possibility of Nash equilibrium with LLM reasoning. :) https://lnkd.in/dGKSC6uK #ai #data #datascience #ml #gametheory This is probably the start of multiagent systems/LLMs coming together at the same time for a given use case :)

19

𝐖𝐞 𝐀𝐫𝐞 𝐓𝐨𝐫𝐭𝐮𝐫𝐢𝐧𝐠 𝐋𝐋𝐌𝐬 🤖 We will not get “𝐫𝐞𝐚𝐬𝐨𝐧𝐢𝐧𝐠” out of chatbots and large language models (LLMs). They are 𝒑𝒓𝒆𝒅𝒊𝒄𝒕𝒊𝒐𝒏 𝒎𝒂𝒄𝒉𝒊𝒏𝒆𝒔 —incredible at identifying patterns and generating responses, but fundamentally incapable of performing actual reasoning. 𝐰𝐡𝐲 ❔ 1/ 𝑨𝒓𝒄𝒉𝒊𝒕𝒆𝒄𝒕𝒖𝒓𝒆🧠: The transformer architecture, based on self-attention, grabs words in parallel but doesn’t truly understand context. It wasn’t built to. While it can piece together bits of information, the model doesn’t have the capacity to gather the 𝐜𝐨𝐦𝐩𝐥𝐞𝐭𝐞 context (no sensors. duh) or develop an internal representation of it, which is essential for reasoning. It's more like an "advanced autocomplete" system. a really good one. 2/ 𝑺𝒄𝒂𝒍𝒆🏋️: Making an elephant gain weight from 100 lbs to 1000 lbs won’t make it fly. bad anology but work with me here. 𝐍𝐞𝐯𝐞𝐫. Similarly, pushing more data into a transformer architecture doesn’t enable reasoning. It may remember more facts and spit out more accurate or contextually appropriate responses, but at its core, it’s still predicting, not reasoning. 💡 𝐓𝐞𝐬𝐭 𝐭𝐡𝐢𝐬: Ask any LLM to do the opposite of what you want and observe how it performs—its limitations become very clear in tasks that require true understanding and logical inversion. 𝐒𝐨 𝐰𝐡𝐚𝐭 𝐢𝐬 𝐑𝐞𝐚𝐬𝐨𝐧𝐢𝐧𝐠? my take - Making decisions - Solving complex, multi-step problems (not just writing code) - Understanding and interpreting events and relationships None of these tasks have been successfully done by 𝐚𝐧𝐲 chatbot without error. And we’re far from seeing them done in a fully integrated manner—where a model combines all of these elements in one smooth, error-free process. 𝐌𝐲 𝐒𝐮𝐠𝐠𝐞𝐬𝐭𝐢𝐨𝐧 LLMs are incredible tool for certain tasks, and they absolutely help solve specific problems (to some extent). But they do 𝐍𝐎𝐓 reason. We shouldn’t punish these models or expect them to do something t̳h̳e̳y̳ ̳w̳e̳r̳e̳ ̳n̳e̳v̳e̳r̳ ̳b̳u̳i̳l̳t̳ ̳f̳o̳r̳. Instead, we should recognize their limits and leverage them where they excel (like this picture) . And make a new architecture #AI #LLMs #ArtificialIntelligence #MachineLearning #Chatbots #Reasoning 

17

8 Comments

https://lnkd.in/gAcPi2Jv The LLama-3 paper by #Meta.ai was a great weekend read. A big shoutout to Meta.ai for publishing what goes into building these frontier models. A lot of this was speculation, but now, we know! A few things that struck me the most: 1. The model took 3.8 × 10^25 FLOPs of compute to train. Roughly, thats equivalent compute of a million M1 Pro chips for a 100 days! (And the M1 pro is a darn good processor by itself for running ML workloads) 2. The model itself has 405B parameters. About 10,000 times the number of parameters of advanced image-classification models from a decade ago! 3. The model was trained on 15 trillion tokens. Roughly 1000 times of all the data in wikipedia combined! 4. The model is still based on the famed "transformer" architecture. Makes me wonder if there is a 'better' architecture waiting to be discovered. 5. Scaling-laws are still holding good, for next-token prediction. However along with the sheer model size, it is the training-data mix that seems to be driving progress in these models. 6. Previous generation LLMs are playing increasingly big part in training the next generation LLMs. Along with using LLMs for input and output classification and synthetic data generation, it was interesting how they used smaller LLama models to identify the right mix of data for the main training runs. 7. Also it was super interesting how they trained the LLama3 in coding, by creating a execution feedback loop on its own output. Llama3 is made to generate unit-tests on its own code and fix the errors in the unit-test or code. Very much, how we learned to code! 8. There seems to be increasing focus on tool-calling. Llama3 has ability to invoke a search-engine and a python interpreter natively. 9. I feel improvements in tool-calling will bring in a lot of incremental utility at this point, than improvements in language understanding and reasoning. With the ability to test out reasoning by making tool calls, the model can iterate till it finds the right solution. #llama3 #meta #llm #artificationintelligence #ai #softwareengineering

43

5 Comments

💡A super interesting blogpost by Siddharth Jindal on Analytics India Magazine, talking about the uniqueness and potential advantages of Gemma over the LLaMA models when it comes fine-tuning LLMs for Indic languages. Some highlights: 📌This is specifically due to the sentencepiece tokenizer used by Gemma which is able to encode strong representation for scripts like Devanagari. 📌The vocab size of Gemma is 256K whereas the LLaMA-3 has only 128K, this seems to indicate that the coverage is much larger for Gemma! 📌The Gemma model is able to generalize to other languages without the explicit need for continuous pre-training as opposed to LLaMA models where the tokenizer needs to be extended to accomodate the language/s in question. 📌Of course, this dosen't mean that Gemma is the silver bullet, since it does face some stability issues in training. This article definitely was an eye-opener for me and also exciting in terms of working with Gemma for fine-tuning Indic LLMs. Check it out here: https://lnkd.in/dRCEtF9Q Also, I'd love to hear more thoughts from my audience here, about their experiences with Gemma and multilingual LLMs! #artificialintelligence #airesearch #largelanguagemodels

9

Sarvam AI Unveils Sarvam-1: A Game-Changer in Indian AI! India’s Sarvam AI, co-founded by Vivek Raghavan and Pratyush Kumar, is headquartered in Bangalore and has launched Sarvam-1, a cutting-edge language model designed specifically for Indic languages. And it’s already outperforming the big players like Gemma-2 and Llama-3.2! Here’s why Sarvam-1 is making waves: ✅ Supports 10 Indian Languages + English: From Hindi, Bengali, Tamil, Telugu, Kannada, Malayalam, Marathi, Punjabi, Gujarati, Oriya, to English—this model bridges the language gap in AI-powered communication. ✅ Outshines Larger Models: With only 2 billion parameters, Sarvam-1 excels across benchmarks like ARC-Challenge and IndicGenBench. On TriviaQA, it achieved 86.11 accuracy across Indic languages, towering over Llama-3.1 8B’s score of 61.47. For translation, summarization, and cross-lingual tasks, Sarvam-1’s chrF++ score of 46.81 on the Flores dataset is unmatched. ✅ 4-6x Faster Inference: Its speed makes it ideal for edge devices with limited computing power. ✅ Efficient Tokenization for Indian Scripts: Traditional LLMs require more tokens per word for Indian languages, but Sarvam-1’s tokenizer brings this down to 1.4–2.1 tokens, similar to English’s efficiency. ✅ Trained on the Sarvam-2T Dataset: This 2-trillion-token corpus contains a rich blend of technical, scientific, and programming data across the 10 languages, enhancing its ability to handle complex reasoning tasks. Built on Yotta’s Shakti cluster with NVIDIA’s NeMo framework, the model was trained on 1,024 GPUs in just five days, showing incredible efficiency. Developers can now explore Sarvam-1 on Hugging Face, opening new doors for Indic NLP applications—from conversational AI to advanced translations. With Vivek Raghavan and Pratyush Kumar leading the charge, Sarvam AI is at the forefront of AI innovation in India, setting new standards in multilingual technology. The future of AI-powered conversations is here, and it’s proudly made in Bangalore! I am eagerly waiting for Sarvam's API to test it and develop meaningful use cases. Are you as well? Please feel free to comment below!" #AI #SarvamAI #MultilingualAI #TechForIndia #HuggingFace

36

Open the LLM hood! Google offers a great 5 day GenAI intensive course. What's being covered? Day 1: Foundational Models & Prompt Engineering - Explore the evolution of LLMs, from transformers to techniques like fine-tuning and inference acceleration. Get trained with the art of prompt engineering for optimal LLM interaction. Day 2: Embeddings and Vector Stores/Databases - Learn about the conceptual underpinning of embeddings and vector databases, including embedding methods, vector search algorithms, and real-world applications with LLMs, as well as their tradeoffs. Day 3: Generative AI Agents - Learn to build sophisticated AI agents by understanding their core components and the iterative development process. Day 4: Domain-Specific LLMs - Delve into the creation and application of specialized LLMs like SecLM and Med-PaLM, with insights from the researchers who built them. Day 5: MLOps for Generative AI - Discover how to adapt MLOps practices for Generative AI and leverage Vertex AI's tools for foundation models and generative AI applications. Here are the notebook building of Gemini models: https://lnkd.in/eUhqpyu7 To register: https://lnkd.in/e5DSvHXW

10

3 Comments

"TIGER-Lab Introduces MMLU-Pro Dataset for Comprehensive Benchmarking of Large Language Models’ Capabilities and Performance Researchers from TIGER-Lab have introduced the MMLU-Pro dataset to address these limitations. This new dataset is designed to provide a more rigorous and comprehensive benchmark for evaluating LLMs. MMLU-Pro significantly increases the number of answer options from four to ten per question, enhancing the evaluation’s complexity and realism. Including more reasoning-focused questions addresses the shortcomings of the original MMLU dataset. This effort involves leading AI research labs and academic institutions, aiming to set a new standard in AI evaluation The construction of the MMLU-Pro dataset involved a meticulous process to ensure its robustness and effectiveness. Researchers began by filtering the original MMLU dataset to retain only the most challenging and relevant questions. They then augmented the number of answer options per question from four to ten using GPT-4, a state-of-the-art AI model. This augmentation process was not merely about adding more options; it involved generating plausible distractors that require discriminative reasoning to navigate. The dataset sources questions from high-quality STEM websites, theorem-based QA datasets, and college-level science exams. Each question underwent rigorous review by a panel of over ten experts to ensure accuracy, fairness, and complexity, making the MMLU-Pro a robust tool for benchmarking." Quick summary: https://lnkd.in/g9qYa9D7 Check out the preview evaluation results in: https://lnkd.in/g_svr2rQ

7

1 Comment

Introducing Saaras 1.0, state of the art multilingual speech-to-text translate model, supporting 10 Indian languages. Supports automatic language identification, and contextual biasing using prompts. Sarvam AI

16

🚀 Sarvam AI Launches India's First Open-Source Small Language Model in 10 Indic Languages! Exciting news in the world of Gen AI! Sarvam AI has unveiled a comprehensive GenAI platform tailored for India's linguistic diversity. The models are trained leveraging NVIDIA NeMo Framework (Megatron-LM) on H100s on Yotta Data Services Private Limited Shakti Cloud infrastructure. Let's explore the key innovations: 1️⃣ Sarvam 2B: 🧠 India's First Open-Source Indic SLM 🔹2B parameters, 4T high-quality tokens 🔹50% Indic languages focus 🔹Supports 10 languages: Hindi, Tamil, Telugu, Malayalam, Punjabi, Odia, Gujarati, Marathi, Kannada, Bengali 🔹Excels in translation and summarization 2️⃣ Sarvam Agents: 🗣️ Voice-Enabled Business Assistants 🔹Available in 10 Indic languages 🔹Deployable via phone, WhatsApp, in-app 🔹Use cases: Customer service, sales calls 🔹Pricing from ₹1/minute 3️⃣ Shuka 1.0: 🎙️ India’s first open-source AudioLM 🔹Built on Meta's Llama 8B 🔹Indian language voice input, text output 🔹Higher accuracy, cost-efficient 4️⃣ Sarvam Models: 🛠️ API Access 🔹Translation, speech recognition/synthesis, document parsing 5️⃣ A1: ⚖️ GenAI for Legal Professionals 🔹Regulatory chat, document drafting, data extraction 🔹Specialized contract tools Plus: 🎤 Bulbul 1.0: Text-to-Speech (10 languages, 6 voices) 🎙️ Saaras 1.0: Speech-to-Text (10 languages, auto-identification) 🈂️ Mayura 1.0: Translation API 💡 Why It Matters: Sarvam AI's "sovereign AI" addresses India's unique needs, democratizing AI access and fostering innovation. 🔮 Future Impact: 🔹Improved AI accessibility 🔹Culturally relevant AI applications 🔹Boost for local tech ecosystem What are your thoughts on these innovations? How might they shape India's tech landscape? 💼 Businesses: Considering these AI solutions? 🧠 Developers: What applications can you envision? Share your ideas below! Let's discuss the future of Indian AI. Vivek Raghavan #SarvamAI #IndianTech #AIInnovation #MachineLearning #OpenSourceAI

35

Tanay Jaipuria’s analysis on OpenAI’s o-1 model highlights how inference-time scaling laws, allowing models to “think” longer on complex queries, could enhance performance without extra training. This marks a shift toward dynamic reasoning, where compute resources are optimized at inference rather than pre-training. As AI models tackle harder, more consequential tasks, this approach could transform both AI deployment strategies and cost structures. Are inference-time improvements the next key to AI advancement? Let’s discuss!

3

Meta’s new language model "Llama 3" is finally here. Plus, you can access it directly with chatGPT like interface. (not available in India yet) ► Integration with Meta AI: Llama 3 is now part of Meta AI, facilitating an array of tasks including coding and problem-solving. ► Model Variants: Available in both 8B and 70B configurations, offering versatility for different development needs. ► Enhanced Training: Trained on a 15T token dataset, 7x larger than Llama 2, with four times more code content. ► Extended Context Length: Supports an 8K context length, doubling the capacity of the previous model, Llama 2. ► Infrastructure: Utilizes Meta's recently built 24K GPU clusters, enhancing its data handling and scalability capabilities. ► Access and Usage: Users can interact with Llama 3 through a ChatGPT-like interface, enhancing user accessibility and ease of use.

4