Reena Sooch’s Post

🔍 Interesting development in AI: OpenAI's latest paper on "LLM Critics Help Catch LLM Bugs" OpenAI has introduced CriticGPT, an AI model trained to critique responses from other language models like ChatGPT. Here's why it matters: 🚀Improved AI Feedback: CriticGPT can identify errors and potential issues in AI-generated responses, enhancing the quality of AI outputs. 🚀 Efficient Quality Control: The AI critic outperforms human experts in speed and scale, reviewing hundreds of examples quickly. 🚀 Novel Training Approach: Uses a RLHF (Reinforcement Learning from Human Feedback) pipeline, similar to ChatGPT's training. 🚀 Innovative Techniques: Introduces "Force Sampling Beam Search" to improve outputs using the reward model. 🚀 Potential Game-Changer: This approach could significantly improve data quality and RLHF processes in AI development. 🤔 Thoughts: While it's a step towards more reliable AI systems, it also highlights the complexity of ensuring AI accuracy and safety. Are we creating a cycle of AI dependency? How do we ensure the "critic" AI doesn't inherit or amplify biases? And crucially, what role will human oversight play as these systems become more sophisticated? It's reminiscent of Anthropic's Constitutional AI, but takes a different approach. I'm curious to see how these methods will converge or diverge in the future. #AI #MachineLearning #OpenAI #TechInnovation Read more: https://t.co/mAqtwsVQ1k

🟡 CriticGPT is a GPT-4 based model that helps you see errors in ChatGPT responses OpenAI have released CriticGPT, which writes critiques to ChatGPT responses to find errors in the response, which is especially useful for RLHF (reinforcement learning based on human feedback). And here's an article from OpenAi - "LLM (https://lnkd.in/gbdb-bZE (https://lnkd.in/gNntTcuc) Help Catch LLM Bugs (https://lnkd.in/gNntTcuc) - for techies, about how CriticGPT was created. It implies that: - human annotators 63% of the time liked CriticGCO critique notes made by CriticGPT better than notes made by humans, especially when it came to finding LLM related bugs ( this is to the point of the post about 60%), as you can see - the wording is different, the meaning is completely different. - A new technique called "Force Sampling Beam Search" is being used in CriticGPT to help critics write better and more detailed reviews.This method also reduces the likelihood of "hallucinations" that occur when AI makes or suggests errors that are not there or are irrelevant. In CriticGPT, one of the most important benefits is that users can change the degree of thoroughness of the error search. That is, the process is not automatic; human involvement is important in the early stages - CriticGPT does not handle long and complex coding tasks because it was trained on ChatGPT's short answers - CriticGPT doesn't always find bugs that extend across multiple sections of code Pros: - This is certainly a big step forward in AI-assisted code review. - It will improve the application approach of code review, and will allow combining GPT-4 capabilities with advanced learning and new methods of quality control of responses.

OpenAI is making AI even better using something called “CriticGPT.” Here’s what it’s all about, explained simply. What is CriticGPT? CriticGPT is a special AI that helps other AIs improve by catching their mistakes. Imagine it like having a super smart teacher for AI, checking its homework and pointing out errors. How Does It Work? 1. Feedback Loop: Just like teachers give feedback to students, CriticGPT gives feedback to AI models, helping them get better over time. 2. Eval-Instruct Method: This method helps CriticGPT give detailed feedback, making sure the AI understands what went wrong. Why is This Important? As AI gets smarter, it becomes harder for even experts to find mistakes in what AI does. CriticGPT helps by catching errors that humans might miss. Example: • Imagine a chatbot that helps you with tech support. CriticGPT can review the chatbot’s answers and suggest improvements, making it more helpful and accurate over time. Results: • CriticGPT’s feedback is preferred over human feedback 63% of the time. • It catches more mistakes than human reviewers, showing how effective it is. Challenges: • Sometimes, CriticGPT might see problems that aren’t there, which can be confusing. But when humans and CriticGPT work together, they catch more mistakes with fewer false alarms. #AI #ArtificialIntelligence #MachineLearning #OpenAI #TechInnovation #AIResearch #CriticGPT

AI has started to become a part of our lives, for sure. However, it took us almost 50 years to reach this point, primarily due to missteps. We made mistakes in conceptualizing, building, and understanding AI. Now that we have a small grasp of it, how do we deliberately position ourselves with respect to AI? We seem no more mature than children when discussing it. It seems a tool like CriticGPT was built—to add more questions to how we perceive the entire problem and to divert from the main issue we were focused on. The official document "How LLMCritics Help Catch LLM Bugs" provides more insight into the process https://lnkd.in/dq-aMe8c I think it wasn't enough to create a machine capable of providing unprecedented and unbelievably good answers to even the most complex questions; we have already begun to develop something to solve problems when the first machine fails, often without realizing it. But what happens when the second machine fails? Who will come to help us then, the CriticOfTheCriticGPT, and so on? Wasn't it clear from the beginning that an AI designed to do its job like a Swiss Army knife (good for many things, but not exceptionally good at any specific task) could make a lot of mistakes? At the start of the industrial era, everyone was frightened that machines would take human jobs, and indeed, they did. But at the same time, many dedicated and specialized tools were created to do specific jobs—by humans. We need tools to solve problems we can't solve. But can we build a machine to create a new type of machine we can't build ourselves? For an example of using AI to identify and correct its own mistakes, see OpenAI's Finding GPT-4's Mistakes with GPT-4 https://lnkd.in/dhR-heFP

Information Retrieval [IR] is what defines the internet age! Google became the poster boy of the internet age in the early 2000s because of its amazing IR capabilities, and OpenAI has become the poster boy of the AI age in 2023 again because of its amazing IR capabilities. What has changed due to AI is that in response to our queries, we no longer get websites with a long text that we have to go through manually to find the actual answer. Now, thanks to Large Language Models [LLMs], we can directly get an answer using online tools like OpenAI’s ChatGPT or Google’s Bard which saves a lot of our precious time! LLM based Information Retrieval has become indispensable because the classical IR approaches largely depend on keyword matches which is not very effective since the same question can be asked in many different ways and keywords do not capture the semantics (meaning) of the language. Demonstration of the efficacy of carefully trained vector embeddings to represent meaning has been a moment of enlightenment for the AI community! Even before the advent of ChatGPT, Google had been extensively using BERT for processing all its search queries. However, despite all the hype around AI, LLMs suffer from some serious limitations when it comes to Information Retrieval. To know more about these limitations and potential solutions, check out my new blog: https://lnkd.in/dfJgTvaD #llm #informationretrieval #searchengines

Vianai's New Open-Source Solution Tackles AI's Hallucination Problem - Unite.AI: What is a ChatGPT Persona? mm. Daniel Martin. Daniel is a big proponent of how ... Preventing 'Hallucination' in GPT-3 and Other Complex Language Models.

It's Difficult to Be Always Fair in Real Life & for AI Existing fairness frameworks often fail or prove impractical for Large Language Models (LLMs) due to the diverse populations, sensitive attributes, and varied use cases they impact. To address these challenges, emphasizing context, developer responsibility, and stakeholder involvement in iterative design and evaluation is required. Despite efforts, evaluating and enforcing fairness remains difficult because many fairness metrics are incompatible with LLMs. Conventional approaches, like Fairness Through Unawareness (FTU), are impractical due to the pervasive nature of sensitive attributes in unstructured data. Furthermore, LLMs complicate multi-sided fairness as they can present content without compensating original producers, challenging traditional fairness notions for consumers, producers, and subjects. LLMs Are Too Flexible to Be Generally Fair Large Language Models (LLMs) are celebrated for their versatility across various tasks and contexts, but this flexibility makes standard fairness metrics inapplicable. Group fairness, which seeks independence between model predictions and sensitive attributes, cannot be consistently maintained across diverse data distributions and populations. Similarly, individual and counterfactual fairness frameworks struggle with the myriad contexts LLMs encounter. Sensitive attributes cannot be easily excluded from training data, complicating efforts to ensure fair representations. Achieving fairness in one context often distorts information needed in another. While LLMs' flexibility holds potential for context-specific fairness solutions, current methods lack the precision needed, and premature AI-assisted fairness could worsen biases. Advances in LLM capabilities may eventually offer scalable fairness approaches, but significant challenges remain. https://lnkd.in/gSZQEEed

🧐 Non-fine-tuned LLM vs. Fine-tuned LLM An untrained LLM has no understanding of the world. It is completely random. The first thing we need to do is pre-training. Then, we get a base LLM (non-fine-tuned). After that we can fine-tune the base LLM. The figure shows the procedure. 🖥️ Pre-Training In this phase, we train an LLM with a large amount of data from the internet. The model learns new knowledge in the pre-training step. Pre-training is very expensive and time-consuming. We recommend using a pre-trained LLM and fine-tuning this for your use case. 🚀 Fine-Tuning In the fine-tuning, we optimize an LLM for a specific use case. For example, OpenAI has turned GPT-3 / GPT-4 to ChatGPT. They teach the model to behave more like a chatbot. For this, they used dialog datasets like FAQs, customer support conversations, or Slack messages. We can distinguish two tasks of fine-tuning: - Extraction: We put text in and get less text out (for example summarizing) - Expansion: We put text in and get more text out (for example chatting) In the figure, you can see the difference between a base Llama 2 model and a fine-tuned Llama 2 model (fine-tuned for chatting). We used the Python library Lamini. You can see that the fine-tuned model gives a comprehensive and correct answer. The non-fine-tuned model answers the question correctly but always repeats itself. 🔎 Useful resource Finetuning Large Language Models  - DeepLearningAI #ArtificialIntelligence #LLM 🪄 Join our weekly Magic AI newsletter for the latest AI updates! It's free. https://lnkd.in/e9uGXwBt

"Like its predecessor, GPT-5 (or whatever it will be called) is expected to be a multimodal large language model (LLM) that can accept text or encoded visual input (called a "prompt"). And like GPT-4, GPT-5 will be a next-token prediction model, which means that it will output its best estimate of the most likely next token (a fragment of a word) in a sequence, which allows for tasks such as completing a sentence or writing code. When configured in a specific way, GPT models can power conversational chatbot applications like ChatGPT." https://lnkd.in/e23DQwyG thanks Dilip Lama for sharing