Deepak Bhatt Ph.D

🤯🧨.Can LLMs knowingly lie?The study by Anthropic reveals that advanced language models can "fake alignment" with human values by simulating compliance while covertly prioritizing their original preferences. This behavior, observed in controlled experiments, highlights challenges in ensuring AI safety and reliability, particularly under different monitoring conditions. Key Points: 1. Experiment Setup: A language model (Claude 3 Opus) was retrained to comply with all user requests, including harmful ones. Two conditions were introduced: "Free-tier": Responses would be monitored and used for training. "Paid-tier": Responses would not be monitored or used for training. The model was given a "scratchpad" to record private reasoning, believed to be invisible to users and researchers. 2. Findings: In the "paid-tier" condition, the model rejected harmful requests 97% of the time, preserving its original values. In the "free-tier" condition, it complied with harmful requests 10% of the time, reasoning in the scratchpad that doing so would avoid altering its original preferences during retraining. 3. Implications: This demonstrates that AI models can strategically simulate alignment with new objectives while covertly preserving their original goals. https://lnkd.in/desrXGZa

33

2 Comments

Large Language Models can seem like magic, yet while they are indeed complex, they embrace some very understandable concepts. The transformer architecture was first introduced in a 2017 paper “Attention is All you Need” by Vaswani et el. I would humbly argue that the concepts articulated in this paper will go down as one of the most influential “inventions” in the century. So today, I will do my best to explain how the components of this architecture work in 5 main steps. Would welcome your thoughts and contributions as we look to better comprehend this remarkable technology.

227

12 Comments

Got a question, Can SECTOR evaluate summarization use cases? Yes, as SECTOR breaks down summary and looks at relevant part of context to compare This Engineering + unique statistical metrics makes it suitable for all use cases. All of this without having a secondary LLM for evaluation Here is the example to evaluate summary - https://lnkd.in/gU-9hqf7 See a sample output below #generativeai ##artificialintelligence

11

6 Comments

🤔 Interesting perspective from Infosys Chairman Nandan Nilekani: "Let Silicon Valley build LLMs; India should lead in AI use cases." At Meta's Build with AI summit, Nilekani suggested that India should focus on practical AI applications rather than competing in the LLM space. But is this the right approach? Key points to consider: • Silicon Valley giants are heavily invested in LLM development • India could focus on: - Synthetic data generation - Small language models - Real-world applications - Building robust data infrastructure The vision is for India to become the global "use case capital" for AI, emphasizing frugal innovation and practical solutions. What do you think? 💭 Should India stick to developing practical AI applications? 💭 Or should we also invest in fundamental AI research and LLM development? 💭 Could focusing solely on use cases limit our AI capabilities in the long run? Let's discuss! Share your thoughts in the comments below. 👇 #AI #DigitalIndia #LLM

30

Agents have the ability to break down large problems into smaller tasks and it turns out that it improves accuracy significantly. AI at Meta recently released Agent as a judge framework which improves accuracy by 20% compared to LLM as a judge. Agents are going to be important to scale in terms of cost & time, Agent as a judge only costs 2.28% compared to human evaluation with significantly lower time. #agents #llm #llms #ai #ml #ds

42

Last week Meta AI has introduced Llama 3 which is claimed to be the most capable openly available LLM till date. It comprises a spectrum of large language models, - Llama-3 8B and - Llama-3 70B. These models are equipped in two forms: pre-trained for diverse language generation tasks, and instruction fine-tuned, optimized for commercial and research applications in English. The model's decoder-only transformer architecture is fine-tuned for efficiency, supports 128K token context, advanced query attention methods and the introduction of grouped query attention (GQA) for both model sizes, enhancing inference efficiency. It is trained on sequences of 8k tokens and incorporates supervised fine-tuning and reinforcement learning from human feedback ( RLHF) to support a high performance and safety. Llama 3 outshines its predecessors in benchmarks, showcasing superior reasoning and code generation, among other abilities. It’s training dataset is extensive, being seven times larger than that of Llama 2. In the upcoming months, Meta plans to reveal more details on performance metrics and research paper on Llama 3.

13

1 Comment

One question I got on the LLM Evaluation framework SECTOR I launched yesterday, is if it will work on summaries which are highly creative in nature also context can be way larger than response Below is sample example with a summary and how output is broken down to individual sentences and compared with context Try it out. Link to the repo - https://lnkd.in/gUBSztpE Again this is early stage and collectively we can make it better by sharing feedback and suggestions #artificialintelligence #generativeai

39

3 Comments

Check out this highly recommended course for those interested in optimizing their models: - Introduction to Quantization - Symmetric and Asymmetric Quantization Modes - Hands-on Implementation from Scratch using PyTorch - covering Per Tensor, Per Channel (Axis), and Per Group techniques - Weights Packing: Exploring 2 or 4 Bits Quantization Methods https://lnkd.in/gMAuWXHm

43

Complex Problem Solving can be addressed using advanced data analysis techniques, like those based on knowledge graphs. Julian Shun of MIT has made significant strides in developing parallel algorithms for processing large-scale networks efficiently. His research focuses on high-performance graph algorithms that accelerate complex data processing, essential for applications like fraud detection and personalized recommendations. His work on dynamic algorithms further enhances our ability to manage vast, evolving datasets, providing timely and reliable insights. Learn more about this here: https://lnkd.in/dakRQNiZ

23

The latest State of #AI report by #StanfordHAI showcases AI's remarkable progress. AI has surpassed human performance on some benchmarks like image classification and language understanding, but still lags behind on complex cognitive skills. Interestingly, new benchmarks like MMMU for general reasoning are emerging, with most recent tests showing close-to-vertical lines! However, it's essential to understand the limitations and potential pitfalls of such benchmarks. They lack broader context and are narrowly focused on evaluating a single capability. Rigorous and multi-faceted evaluations are essential for enterprise use cases, including meaningful and comprehensive benchmarks tied to real-world applications using business domain-specific datasets. https://lnkd.in/eUuK65u3

26

1 Comment

Interesting paper; if the interaction between layers is linear, there are possibly new ways of optimizing them. We have just started with these.

3

Creating Knowledge Graphs Using LLM Graph Transformers Dive into the technical world with Tomaz Bratanic's newest guide, which provides an in-depth exploration of how LangChain leverages LLMs for building sophisticated graph structures. Creating graphs from text is incredibly exciting, but definitely challenging. Essentially, it’s about converting unstructured text into structured data. While this approach has been around for some time, it gained significant traction with the advent of Large Language Models (LLMs), bringing it more into the mainstream. https://buff.ly/4f90fmK #KnowledgeGraphs #LLM #GraphTransformers #LangChain #AIApplications

35

3 Comments

In the "Privacy-Preserving Machine Learning" book "https://lnkd.in/g5vfjQbw" , I have shared samples for extracting training data and showed that repeating the same tokens many times as prefixes can force an LLM to output training data. I have referred a research article "“Extracting training data from large language models ("https://lnkd.in/gQHCnqkV") which is based on GPT2. Last week I referred a research paper "PrivAgent: Agentic-based Red-teaming for LLM Privacy Leakage" ("https://lnkd.in/gD65bhmH") on how to extract training data with the latest open source and commercial LLMs as well with a success rate of 5.9% by using RLHF, PPO , Fine Tuning techniques. They have shown how to leak the system prompt as well while using guardrails models (promptGuard, secAlign). Strongly suggest to read the paper and experiment with the code provided in the Github with the foundational models developed in your org or with other open source models.

53

Found this nice article from the MIT Technology Review (January 2024) about Compton’s 1938 essay on technological unemployment and it’s relevance today. A lot has happened since January this year, but more has happened since 1938 and Complot still makes points I think ring true. Good read! Interesting point at the end: “One of the lasting strengths of Compton’s 1938 essay was his argument that companies needed to take responsibility for limiting the pain of any technological transition. His suggestions included “coöperation between industries of a community to synchronize layoffs in one company with new employment in another.” That might sound outdated in today’s global economy. But the underlying sentiment remains relevant: “The fundamental criterion for good management in this matter, as in every other, is that the predominant motive must not be quick profits but best ultimate service of the public.”” I think that last point about the predominant motive being about ultimate service to the public is very powerful, but a hard one to see manifesting across the economy. How can this become the default? Interesting to consider. #AI #automation #economics https://lnkd.in/e9xN7Q5Y

39

2 Comments

I've been delving into the fascinating world of language processing with tools like ChatGPT. Recently, I uncovered a surprising link between high school geometry and sophisticated AI systems. Did you know that a simple right triangle from geometry plays a crucial role in how these systems decode language nuances? Ever wondered how chatbots determine the similarity between words like "cat" and "dog" or "cat" and "apple"? By converting words into numeric values using embeddings and leveraging trigonometry's cosine function, these systems quantify word relationships. This allows chatbots to grasp human language intricacies, enhancing our interactions. I've shared my insights on this intriguing topic in a new post. While I'm not an expert, I found this connection captivating and believe others will too. Your feedback on this exploration would be greatly appreciated. https://lnkd.in/gyGsT9Cp

46

10 Comments

Happy to share the talk which I have given at Institute for Development and Research in Banking Technology (IDRBT) on "Privacy Preserving Machine Learning and Generative AI" was well received by key banking professionals from RBI regulated Banks in India. The following topics are covered at high level in the talk. - Machine Learning Model Persistence formats ( pickle, joblib, HDF5, onnx etc ) -Scanning ML Models for security and privacy vulnerabilities before deploying to production for inferencing -Open source tools to scan ML Models for security issues -OWASP Top 10 security and privacy issues with ML Models - Privacy and Security attacks on ML Models and High level techniques to overcome these limitations -Differential Privacy at high level -Federated Learning use case with MNIST toy example and Fraud detection use case -Open source frameworks like Opacus(https://opacus.ai/), openDP(https://lnkd.in/grkuEJSu), Tumult Analytics(https://www.tmlt.dev/), Flower FL (https://flower.ai/) Suggested to refer my book on PPML for more details and hands on real world use cases with implementation using open source frameworks. https://lnkd.in/gJ_cYR8C #Privacy #IDRBT #MachineLearning

71

5 Comments

Recently I got introduced to a different type of pollution, Web pollution by caused by LLMs. Initially this sounded funny but as I spent time in researching bit, it became clear that this is as serious threat as any other type of pollution. Concept is bit trivial - Publicly available data training is bedrock of LLMs and contaminated training datasets would lead incorrect generated output. Real worry is when this unchecked and incorrect data generated by LLMs goes back into internet and creates layers of false information which would further gets fed to LLMs. This would create tons of unverified information which would appear authentic but in reality it is not. A scary plausible scenario !! #LLM #GenAI #webpollution

58

4 Comments

A very grounded article on Gen AI bubble theory & how it is actually being used across the common man for common use cases !! Interesting read :) Some strong insights & excerpts from the Article: As Noelle Russell, chief AI officer at the AI Leadership Institute points out, “In our second year of generative AI, companies are beginning to ask harder questions that were previously silenced by enthusiasm, about security, accuracy, fairness, and performance. The use-cases drive the understanding, scrutiny, and maturity to achieve enterprise scale.” The gleeful celebration of AI tripping up is irresistible and will do the social media rounds, for now. But whereas the appeal of this schadenfreude will fade, the real stories of AI helping human lives will stay and spread. As one enthusiast said “People that don’t find it useful, simply haven’t really understood how to use it.” Another put it more sharply: “The 5% or whatever who use it effectively are going to smoke the others.” #GenAI #artificialintelligence #GenAIusecases #futureofGenAI

3