Mohit Mayank

🧠 Apple Intelligence is a GPT wrapper.  💤 GPT : We will not tell what Data we used Bias: Who cares 🤷🏻‍♀️ 🤷🏻‍♂️ GUP Designers / Manufacturers : we are building better G/.. Processing Unit (PU) Data Privacy and Security will be a big question with the Apple and Open AI Engagement. No one has questioned OpenAI , Anthropic ... on biases ❓ All LLMs are biased towards the origin of the open-source data, the mindset of the authors, and people who can write/speak English and had early access to the internet. All LLMs exhibit biases stemming from the data they were trained on, which is often skewed towards Western, English-speaking sources and perspectives. This can lead to a lack of knowledge or misrepresentations about other regions, cultures, and languages, particularly those from the Global South like India and Africa. The biases in LLMs are a significant issue that needs to be addressed through more diverse and representative training data, as well as techniques like prompt engineering and human oversight. Relying solely on existing models without accounting for their limitations and biases can perpetuate harmful stereotypes and marginalize underrepresented groups. Hope Apple and Open AI collaboration bring some solutions towards. Sources for further digging: https://lnkd.in/gDAJkgei https://lnkd.in/gPYStYEi https://lnkd.in/gB6H3-wj

٦

llama-3 has launched 🔥🔥🔥. - 2 Models : 8B and 70B models - They were trained for 1.3M and 6.4M GPU hours respectively - the 70B model, Outperforms GPT-3.5 on curated 1200 prompt evals - Context length is 8k tokens Exciting times ahead with Meta. Link - https://lnkd.in/dQJZupZq #startups #AI

١٧

٢ تعليق

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

٢١

٤ تعليق

Enhancing the accuracy of Large Language Models (LLMs) is a key focus in ongoing AI research. Previously, we have used a Critic model along with Reward Engineering to identify incorrect responses and incrementally fine-tune the LLM. The introduction of Thought Preference Optimization (TPO) advances this approach by having the LLM generate preliminary thoughts which a judging model then refines into better responses. This method is akin to the human practice of "thinking before responding," a critical safeguard in high-stakes situations, pushing LLMs to adopt a similar evaluative process. Training an LLM to pause, generate internal thoughts, and evaluate these before delivering a final response is complex and resource-intensive. It typically requires large datasets of human-annotated thoughts, which are scarce across diverse domains. Chain of Thought (CoT) prompting is another method where the model is instructed to articulate intermediate reasoning steps, structuring its approach to tasks. While CoT has proven effective in fields like mathematics and logic that require clear reasoning steps, it often underperforms in more subjective areas such as marketing or creative writing. This limitation is exacerbated by training datasets that generally contain human responses without the underlying thought processes, restricting the model’s ability to refine its reasoning across varied domains. Thought Preference Optimization (TPO) aims to equip LLMs with the capability to generate and refine internal thoughts before producing a response. Crucially, it operates independently of human-annotated datasets. It begins by instructing the model to divide its output into two distinct parts: the thought process and the final response. The model generates multiple thoughts for each instruction, which are then evaluated through performance optimization to select the best thought-response pair for further training iterations, allowing the model to gradually enhance its reasoning capabilities. TPO employs a reinforcement learning technique, enabling the model to learn from its thought process. The model generates preliminary thoughts before responding, and a judging model evaluates the resulting responses. This iterative refinement process leads to more thoughtful and accurate answers. The effectiveness of TPO was tested on two benchmarks: AlpacaEval and Arena-Hard. On AlpacaEval, TPO achieved a win rate of 52.5%, outperforming the direct response baseline by 4.1%. Similarly, it registered a win rate of 37.3% on Arena-Hard, surpassing traditional methods by 4.3%. One of the most significant findings is that thinking-based models generally outperformed direct response models across various domains. Even in non-reasoning tasks like creative writing, TPO-enabled models could plan their responses more effectively, leading to better outcomes. https://lnkd.in/d5kVzjxv #llm #ai

٣

١ تعليق واحد

Text chunks in retrieval may risk losing context. The below article introduces contextualizing each chunk using an LLM. The added cost and latency required for contextualizing is partially addressed through a lightweight LLM (haiku) and prompt caching. Along with other known techniques (BM25, reranking, etc.), a significant reduction in retrieval failure is achieved. Overall, a good article to read. https://lnkd.in/g_5fgAN8

١٠

Comment: Around 350 BC an Indian Sanskrit Scholar built a rule-based analog algorithm for in effect an ancient language machine. Title: PhD student solves a mysterious ancient Sanskrit text algorithm after 2,500 years. See… https://lnkd.in/dfupfx6y And… https://lnkd.in/end3UYpe Around 350 BC, the scholar Pāṇini composed a grammatical treatise of astounding breadth and comprehension. Known as the Aṣṭādhyāyī, it contained 4,000 sūtras, or rules, for writing classical Sanskrit. In effect the rules necessary to construct an ancient language machine Pāṇini’s treatise isn’t meant to just describe classical Sanskrit or prescribe usage like a fussy language arts teacher. It’s designed to help its readers generate “flawless” Sanskrit. He built a system of operations that allowed a reader to combine base words and affixes to generate proper word forms. The user could then combine those words to construct impeccable sentences. All readers had to do was follow the instructions. Effectively, he created an analog algorithm that would produce the same Sanskrit word, sentence, and hymn each time — preserving the language and settling disputes over correct usage for time immemorial. Rishi Rajpopat a University of Cambridge Ph.D. student, may have solved the mystery of how to use Pāṇini’s language machine accurately. While working on his thesis, Rajpopat decided that this meta morass could not have been what Pāṇini intended with the Aṣṭādhyāyī. “Pāṇini had an extraordinary mind and he built a machine unrivaled in human history. He didn’t expect us to add new ideas to his rules. The more we fiddle with Pāṇini’s grammar, the more it eludes us,” Rajpopat said in a press release. Rajpopat returned to Pāṇini’s original meta-rule and considered it with a fresh lens. When he found a rule conflict, rather than give preference to the later rule in serial order, he deferred to the word that came later in the sentence. Because Sanskrit is written left to right, that meant applying the rule appropriate to the word on the right. Rajpopat’s research looked at many other rule conflicts, and thorough textual evidence shows that the meta-rule results in the correct form time and again. This potentially makes it possible to construct millions of grammatically correct classical Sanskrit texts. H/T Moshe Vardi

٦٤

٨ تعليق

RLHF was thought to be the secret behind aligning ChatGPT to give 'smart' responses. Implementing RLHF using PPO optimization is a complex engg task but initially there seemed to be no way around it. 🛠️ The road from the original complex RLHF implementation to the more recent simpler model fine-tuning/alignment strategies is long and unwinding. 👉 The initial RLHF setup used 3 (or 4?) copies of the pretrained model. First is the next-token pretrained model (also called SFT model), one for external reward model and one for the new model being fine-tuned. PPO algorithm was used for learning. Led to many asking do we really need RLHF/PPO ? 👉 DPO got rid of the external reward model by reformulating the loss fn and reduced the copies to two - the original SFT model and the new model being trained. 👉 SPO finally reduced the copies to a single one. They construct a new intrinsic loss function which trains the model while learning from (good, bad) samples. The average log probability of a sequence is used as the implicit reward, same as the usual generation likelihood. Very clean formulation and writing. 🤔 In hindsight, SPO is obvious but why so long? (ORPO was close enough). Two key ideas seem to make it work: length normalization of loss and an extra margin parameter. 🤨 While both DPO and SPO claim full victory on compute and memory efficiency, the jury is still out on whether they yield “better” models than PPO (used in RLHF) setting, on say downstream reasoning tasks like GSM8k. 🎇 I really wish that implementing alignment training becomes as easy as next-token training. Hope that SPO holds its ground and whatever remains of PPO can be brought back to SPO++. --- Stay informed about latest LLM updates with our newsletter: offnote.substack.com

٣٩

٣ تعليق

Long-context LLMs are outperforming RAGs, so will RAGs be dead? 🤔 In this post, you will learn: 1. Comparison of RAG and LC LLMs 2. Performance - Cost Trade-off 3. Self Route: A new approach to bridge the gap What are your views on this research? Let me know in the comments! --------------------------------------------------------------------------- I hope this post helped you! Help people in your network by liking and reposting it ❤️ It inspires me to do more for our community. Follow Bhavishya Pandit to stay updated on GenAI! 🔥 #ml #llm #ai #genai #datascience #artificialintelligence #opensource #ibm #generativeai #largelanguagemodels #huggingface #openai #google #rag #machinelearning #python #langchain #microsoft #genai

٥١٧

٤٩ تعليق

Exploring Idefics3 [https://lnkd.in/g76H2bTy] From Andrés Marafioti @ Hugging Face I have provided the below image and asked to describe the image. The results are fascinating. >> output:  The image depicts a close-up view of an eye, focusing on its iris and pupil. The eye is positioned centrally in the frame with no visible eyelids or surrounding facial features except for part of the upper lid that appears to be slightly raised. The iris itself has a complex pattern consisting of multiple colors including brown, green, yellow, and black. It exhibits radial lines radiating from the center towards the periphery, which are more pronounced near the outer edge. These lines create a dynamic effect due to their varying thicknesses and directions. In the center of the iris lies the pupil, which is dark and circular, reflecting light and creating a shiny appearance. This reflection suggests that there might be some external light source illuminating the scene, possibly artificial given the clarity and intensity of the reflected light. Surrounding the iris, the sclera (the white portion) can be seen around the edges but not prominently as it's mostly covered by the iris' coloration. The lashes appear long and thick, extending outward from the lash line, adding texture and depth to the overall visual composition. Given this detailed description: ### Analysis: - **Iris Color:** Multicolored with dominant hues of brown, green, and yellow. - **Pupil Size & Shape:** Circular, relatively small compared to the size of the iris. - **Lashes:** Long and prominent, enhancing the natural beauty of the eye.    This vivid depiction highlights the intricate details typical of human eyes, showcasing both biological complexity and aesthetic appeal. Such images often serve educational purposes, helping viewers understand ocular anatomy while appreciating the intricacies involved in vision and perception. By examining such visuals closely, one gains insight into how different elements like pigmentation, structure, and reflections contribute to our ability to perceive the world around us. Understanding these aspects aids in comprehending various medical conditions affecting the eyes and the importance of maintaining good eye health through regular check-ups and appropriate care.

٣

٥ تعليق

Here's an even better explanation of why LLM's can't count characters. Understanding how these models work is key to using them effectively.

٤

١ تعليق واحد

🚨ACL 2024 Outstanding Paper Award Alert 🚨 ➡️Paper Title: INDICLLMSUITE: A BLUEPRINT FOR CREATING PRE-TRAINING AND FINE-TUNING DATASETS FOR INDIAN LANGUAGES 🌟Few pointers from the paper 🎯Despite the considerable advancements in English LLMs, the progress in building comparable models for other languages has been hindered due to the scarcity of tailored resources. 👁In this paper Authors work aims to bridge this divide by introducing an expansive suite of resources specifically designed for the development of Indic LLMs, covering 22 languages, containing a total of 251B tokens and 74.8M instruction-response pairs. 🎯Recognizing the importance of both data quality and quantity, their approach combines highly curated manually verified data, unverified yet valuable data, and synthetic data. 🎯They built a clean, open-source pipeline for curating pre-training data from diverse sources, including websites, PDFs, and videos, incorporating best practices for crawling, cleaning, flagging, and deduplication. 🎯For instruction-fine tuning, they amalgamate existing Indic datasets, translate/transliterate English datasets into Indian languages, and utilize LLaMa2 and Mixtral models to create conversations grounded in articles from Indian Wikipedia and Wikihow. 🎯Additionally, they addressed toxicity alignment by generating toxic prompts for multiple scenarios and then generated non-toxic responses by feeding these toxic prompts to an aligned LLaMa2 model. 🎯 They hope that the datasets, tools, and resources released as a part of this work will not only propel the research and development of Indic LLMs but also establish an open-source blueprint for extending such efforts to other languages. 🏢Organization: 🛎️Nilekani Centre at AI4Bhārat🛎️ Indian Institute of Technology, Madras 🛎️Microsoft🛎️Sant Longowal Institute of Engineering and Technology 🛎️Indian Institute of Information Technology Design & Manufacturing Kancheepuram, 🛎️Maharaja Agrasen Institute Of Technology, Delhi 🛎️Sarvam ,🛎️ NICT 🧙Paper Authors: Mohammed Safi Ur Rahman Khan, Priyam Mehta , Ananth Sankar , Umashankar Kumaravelan, Sumanth Doddapaneni, Suriyaprasaad G, Varun Balan, Sparsh Jain, Anoop Kunchukuttan, Pratyush Kumar, Raj Dabre, Mitesh Khapra 1️⃣Read the Full Paper here: https://lnkd.in/gqwQST-r 2️⃣Code: https://lnkd.in/giWPVntx 🎉 🎉  Congratulations to all the authors 🎉 🎉 Find this Valuable 💎 ? ♻️REPOST and teach your network something new Follow me 👣, Naveen Manwani, for the latest updates on Tech and AI-related news, insightful research papers, and exciting announcements. #ACL2024 #Indicllms #llms #llama2 #mixtral

١٦

Good folks from Carnegie Mellon University's WAVLab have released XEUS - A Cross-lingual Encoder for Universal Speech 🌐🗣️ Simple speech representation is hindered by a lack of labelled data. That's where Self-supervised learning (SSL) comes in! 🔍📚 XEUS is a SoTA speech encoder trained in 4000+ languages! 🏆🌍 Before this, I did not even know that there are 4000+ languages. But the paper says this is a small dent in the world's 7000+ languages. 🌐🗺️ More details: - 577M parameters 📊 - 1M+ hours of speech training data across 4057 languages 🗣️⏳ - Uses the E-Branchformer architecture and is trained using HuBERT-style masked prediction of discrete speech tokens 🏗️🔠 - For robustness, the input speech is also augmented with acoustic noise and reverberation 🎙️🔊 Beats both MMS 1B and w2v-BERT v2 2.0 🥇🏆 It requires fine-tuning to be used in downstream tasks such as Speech Recognition or Translation. 🛠️🌐 Its hidden states can also be used with k-means for semantic Speech Tokenization. 🗣️🔑 Release: https://lnkd.in/gqrfvsRF

١١

Hello GenAI Inference enthusiasts! When exploring Speculative Decoding, one of the challenges lies in the fact that as batch sizes increase, the expected gains and latency speedups tend to diminish. This occurs because Speculative Decoding leverages available compute resources during memory-constrained autoregressive decoding to produce more cost-effective drafts, which are then authenticated at the prefill cost. Two pertinent questions arise from this scenario: 1. Does the latency improvement from Speculative Decoding come at cost of operating at lower batch sizes ? 2. What other speedup solutions are being pursued to address this bottleneck? (Refer to the graph depicting gains with Speculative Decoding at higher batch sizes from the MEDUSA paper - Tianle Cai, Tri Dao) #genai #llms #inference #speculativedecoding

٤٨

٢ تعليق

🚀 Bengaluru-based Conversational AI Startup Gnani AI Raises $3M 💰 Gnani AI, India's first voice-first small language model (SLM) for vernacular languages, has secured $3M in funding from tech holding company Info Edge. The investment will fuel the startup's expansion plans. Key Highlights: • Gnani AI is building a series of voice-based SLMs, trained on millions of hours of proprietary audio datasets and billions of conversations in 12 Indian languages • The startup has over 100 customers in banking, financial services, insurance, telecom, automotive, and healthcare industries • Gnani AI's models run on the edge, reducing costs and enabling quick fine-tuning based on enterprise needs • The startup specializes in creating customized models for individual companies, adding real value to their data CEO Ganesh Gopalan emphasizes the importance of multimodal models, focusing on voice and text fusion. By taking their models to enterprises and helping them build on top of it, Gnani AI is revolutionizing the conversational AI landscape in India. #ConversationalAI #VoiceFirst #IndianLanguages #Funding #Expansion

١٢

١ تعليق واحد

🚨 ACL 2024 paper alert! 🚨 Is your NLP model 𝙤𝙫𝙚𝙧𝙘𝙤𝙣𝙛𝙞𝙙𝙚𝙣𝙩? It might not be as robust to adversarial attacks as you think! We find that highly over/underconfident models appear robust by disrupting the search process of NLP adversarial attacks. However, an adversary with logit access can negate this "defense" just by calibrating the logits. We call this phenomenon an 𝙄𝙡𝙡𝙪𝙨𝙞𝙤𝙣 𝙤𝙛 𝙍𝙤𝙗𝙪𝙨𝙩𝙣𝙚𝙨𝙨. Check out the paper for different ways to create 𝗮𝗻𝗱 pierce such illusions at inference time! Inspired by how we induced this illusion, we further demonstrate a simple training-time modification that boosts robustness against unseen attacks. I had a lot of fun uncovering this phenomenon over the summer with Vyas Raina and Volkan Cevher, Aditya Rawal, Sheng Z., and George Karypis @ Amazon Science Vyas and I will also be at ACL in August! 📜: https://lnkd.in/gseBFDEn 🐤🧵: https://lnkd.in/gwMwS9s9

٨٣

٥ تعليق

I have fixed graphs (for torch.jit.trace) of bunch of models in transformers https://lnkd.in/dqpbgEjN this means 👉🏻 you can now do ONNX exports, graph optimizations, quantization & co on models you previously couldn't! as usual a tutorial is coming up on how to use Optimum 🤗

١٢٩

٢ تعليق

Good folks from Microsoft have released an exciting breakthrough in GUI automation! OmniParser – a game-changing approach for pure vision-based GUI agents that works across multiple platforms and applications. Key technical innovations: - Custom-trained interactable icon detection model using 67k screenshots from popular websites - Specialized BLIP-v2 model fine-tuned on 7k icon-description pairs for extracting functional semantics - Novel combination of icon detection, OCR, and semantic understanding to create structured UI representations The results are impressive: - Outperforms GPT-4V baseline by significant margins on the ScreenSpot benchmark - Achieves 73% accuracy on Mind2Web without requiring HTML data - Demonstrates a 57.7% success rate on AITW mobile tasks What makes OmniParser special is its ability to work across platforms (mobile, desktop, web) using only screenshot data – no HTML or view hierarchy needed. This opens up exciting possibilities for building truly universal GUI automation tools. The team has open-sourced both the interactable region detection dataset and icon description dataset to accelerate research in this space. Kudos to the Microsoft Research team for pushing the boundaries of what's possible with pure vision-based GUI understanding! What are your thoughts on vision-based GUI automation?

١٢

AutoML, short for Automated Machine Learning, is a game-changer when it comes to building machine learning models. Normally, creating a good ML model involves a bunch of complicated steps like cleaning the data, picking the right features, choosing the best model, and tweaking settings to get the best performance. All of this can take a lot of time and expertise. But with AutoML, most of this work gets automated. You don't need to be an expert to build a solid ML model because AutoML takes care of the heavy lifting. It automates the whole process—from choosing the right model to fine-tuning it—so you can focus on what matters most: applying machine learning to real problems. In short, AutoML makes machine learning faster and easier, opening the door for more people and businesses to use this powerful technology. #automl #machinelearning #businesses

٤

Self-Correcting Code Generation A new project demonstrates the power of Codestral for Self-Correcting Code Generation applications. Codestral has introduced an innovative self-correcting code generation approach by combining the capabilities of its generative model with AlphaCodium's self-correction methodology. This implementation leverages LangGraph for structured code output and feedback, making significant strides in automating and refining code generation processes. For developers and tech enthusiasts looking to reduce debugging time and improve code quality, this tutorial offers valuable insights into the next generation of coding assistants. Explore the detailed steps and examples to harness the power of self-correcting code generation. #CodeGeneration #AI #MachineLearning #Coding #Technology #SoftwareDevelopment #DeveloperTools #Automation #Innovation #DeepLearning #NLP #NaturalLanguageProcessing

٦