Grammarly’s Post

In case anyone is doing work on AI integrations - this technique of fine tuning around retrieval before adding in retrieval augmented generation (RAG) is a good idea if sitting on the resources to do the fine tuning:

In our first part of the series we dig into what AI can (and can't) do for educators. Forget the pointless AI writing detector arms race, we want to know if AI can reliably grade student papers. Short answer: In one of our tests, the AI justified failing and giving an A+ to the exact same essay with the exact same rubric. https://lnkd.in/ehVF-EnY

This week Open AI revealed their improved GPT 4o. (I love the little ‘o’, I don’t know why, but it’s just different.) I was curious and wanted to learn more. It (GPT) says it’s faster and the newest and most advanced model. Cool, what does that mean though? 😅 Performance and Cost: GPT-4: The original model, known for its comprehensive capabilities and accuracy. It is resource-intensive, making it more expensive to run. GPT-4o: A more optimized version designed to be faster and cheaper while maintaining a similar level of performance. It offers a more cost-effective option for applications requiring frequent usage. Speed: GPT-4: Slower in generating responses due to its thorough processing. GPT-4o: Faster response times, which can be beneficial for real-time applications. Resource Efficiency: GPT-4: Consumes more computational resources, which can limit scalability for some applications. GPT-4o: Optimized to be more resource-efficient, allowing for better scalability and reduced operational costs. Use Cases: GPT-4: Preferred for tasks requiring the highest level of accuracy and detail, such as complex problem-solving and detailed content generation. GPT-4o: Ideal for applications where speed and cost are critical, such as customer service, real-time interactions, and high-frequency queries. So, it’s faster and more cost-effective. But honestly, still, what’s this mean!? 🤔 In the comments below, there’s a demo of Sal Khan, of Khan Academy using it to tutor his son in Math. It (GPT-4o) is seeing what his son sees and answering questions (voice) and tutoring his son. Just a cool application all around. I’m always excited to see what’s next. What do you think? 😎

What's better than Retrieval Augmented Generation (RAG)? What's better than Fine-Tuning? RAFT -> Retrieval Augmented Fine Tuning! RAFT attempts to get the model to study or adapt to a domain before it is used in a RAG setup. #AI https://aka.ms/raft-blog

Some exciting news from Berkeley's AI research lab! (Go bears!) They've developed a new technique called Retrieval Augmentation for Fine-Tuning models (RAFT). Impressive results were achieved using this method on Llama2-7b, outperforming GPT-3.5 + RAG in most benchmarks, sometimes by an insane amount. Check out the blog post here: https://lnkd.in/eQdTnctq. This highlights the criticality of data and why having a strong data strategy is the major critical path to GenAI adoption in business and enterprise. Brute-forcing your way to a more useful model with the latest, biggest, most accurate models is expensive and not always the best approach. Techniques like RAFT epitomize the "Work smarter, not harder" ethos, and it is increasingly clear that as companies focus on making practical and useful GenAI service an economically sustainable proposition, Data is key. It'll be interesting to see how developers and engineers will adopt this technique to create high-quality, domain-specific small models that will dramatically bring down the cost to serve. #AI #LLM #GenAI #Data

🚀 Master the Transition from Machine Learning to Large Language Models with These 5 Key Patterns! 🚀 Moving from traditional ML to the exciting world of LLMs? Here are five essential patterns to make your journey smoother: 1. Labels → Texts In ML, datasets consist of input signals and corresponding labels. In LLMs, you work with prompts as inputs and example texts as outputs, often crafted by humans. 2. Model Training → Fine-Tuning ML typically involves training models from scratch. With LLMs, start with powerful pre-trained models like GPT, Claude, LLAMA, or Gemini. You can use them as they are or fine-tune them to suit your specific needs. 3. Feature Engineering → Prompt Engineering Enhancing ML model accuracy relies on techniques like one-hot encoding. In LLMs, prompt engineering, including strategies like K-shot learning, is crucial for improving text output quality. 4. Evaluation → Evals ML evaluation uses metrics like AUC or MSE through cross-validation. For LLMs, evaluation methods are more flexible, involving semantic similarity, using LLMs as evaluators, or incorporating human feedback. 5. ML Flow → LangSmith Operationalizing ML often involves tools like MLFlow. For LLMs, operationalize efficiently with tools like LangSmith or LangFuse. Transitioning from ML to LLM is more than just a change in tools—it's a shift in mindset. Embrace these patterns to unlock the full potential of language models and drive innovation in your projects! #MachineLearning #LargeLanguageModels #AI #DataScience #Innovation #TechTransition #LLM #MLFlow #LangSmith

💥💥💥 Researchers gave AI an 'inner monologue' and it massively improved its performance 👉 Quiet-STaR: Language Models Can Teach Themselves to Think Before Speaking Eric Zelikman, Georges Harik, Yijia Shao, Varuna Jayasiri, Nick Haber, Noah D. Goodman Abstract When writing and talking, people sometimes pause to think. Although reasoning-focused works have often framed reasoning as a method of answering questions or completing agentic tasks, reasoning is implicit in almost all written text. For example, this applies to the steps not stated between the lines of a proof or to the theory of mind underlying a conversation. In the Self-Taught Reasoner (STaR, Zelikman et al. 2022), useful thinking is learned by inferring rationales from few-shot examples in question-answering and learning from those that lead to a correct answer. This is a highly constrained setting -- ideally, a language model could instead learn to infer unstated rationales in arbitrary text. We present Quiet-STaR, a generalization of STaR in which LMs learn to generate rationales at each token to explain future text, improving their predictions. We address key challenges, including 1) the computational cost of generating continuations, 2) the fact that the LM does not initially know how to generate or use internal thoughts, and 3) the need to predict beyond individual next tokens. To resolve these, we propose a tokenwise parallel sampling algorithm, using learnable tokens indicating a thought's start and end, and an extended teacher-forcing technique. Encouragingly, generated rationales disproportionately help model difficult-to-predict tokens and improve the LM's ability to directly answer difficult questions. In particular, after continued pretraining of an LM on a corpus of internet text with Quiet-STaR, we find zero-shot improvements on GSM8K (5.9%→10.9%) and CommonsenseQA (36.3%→47.2%) and observe a perplexity improvement of difficult tokens in natural text. Crucially, these improvements require no fine-tuning on these tasks. Quiet-STaR marks a step towards LMs that can learn to reason in a more general and scalable way. 👉 https://lnkd.in/dU2CuY3H #machinelearning

Impact of AI or expectations. My daughter (freshman in high school) spent 3 days writing an essay. She was proud of her work, but wanted a second opinion so she asked a friend of mine to review her work. He did, and suggested a few edits for a better use of vocabulary (it's ok Google docs, I can see the edits, nothing really major). None of the two used AI or Google, or anything other than their brains :) However, at school, her teacher looked at the essay and stated - "next time try without using AI" . This obviously stressed her out since she felt she went over and beyond to make sure her work was her 100%. Anyhow, she decided to write an email to the teacher and explain that not only did she NOT use any help other than her uncle's to review, she also tried running the essay thru multiple AI detectors and all of those indicated 0% AI. Thankfully, the teacher reverted back and apologized for jumping to conclusions and said he will try to find an alternative as the AI detector he used indicated it is written by AI. Now obviously there are two Major Issues here: 1. The teacher who believed in the information that was given to him, to take decisions which would impact his students in many ways, and also reflect on his credibility in some ways. 2. The AI detector he used analyzed the essay and felt it's too good, so maybe it's not written by a human? Is that the criteria? What are the expectations from humans here? I understand that the technology is developing, and there are a LOT of loose ends, but the end users need to be extremely careful on how you use the information provided by AI, it can hurt you if you blindly trust it. And it can make you look like a fool too. If students were using AI to write their essays, and teachers have to use AI detectors to see if students are using AI, then I think a smarter device or app needs to be invented that allows realtime tracking so it can save everyone time. More on this in another post, but it's a catch 22 situation I think. Fun fact: LinkedIn is prompting me right now (as I type this post) to Rewrite with AI. And I decide not to do that. Let this remain a human to human conversation 😂. sorry about the typo in the screenshot (promoting instead of prompting), AI won't make that mistake :)

‘Explain it like I’m 12 years old’… The wonder of taking a picture of a paragraph on my Kindle Scribe and using this prompt with GPT Vision never gets old. It is my go to for reading AI and ML papers like this one by Apple AI/ ML researchers entitled ‘KGLens: A Parameterized Knowledge Graph Solution to Assess What an LLM Does and Doesn't Know’ Here is GPT4’s interpretation of the page in question… ‘Okay, think of a knowledge graph like a collection of trading cards. Each card has three parts: a subject (like a player's name), a predicate (what they're good at, like batting or pitching), and an object (the details of that skill, like a batting average). Now, if a computer, which we call a Large Language Model (LLM), gets something wrong, it's likely because it's not so good with the cards related to that question. So, we give each card an "oops score" that tells us how often the computer might get things wrong about that card's information. We use something called a beta distribution to decide on this "oops score." It's like a special rule to guess how often mistakes happen. We start by saying every card's score is even (no mistakes), but as the computer keeps trying and sometimes messing up, we change the score to show what it's good at and what it needs to learn more about. Over time, by looking at which cards have high "oops scores," the computer can figure out what it needs to study more. This helps the computer get better at answering questions because it knows which cards (facts) it needs to review.’ #knowledgegraphs #vectordatabase #gptvision #ai #ml OpenAI Navdeep Jaitly He Bai Apple

BIGGEST Digital Marketing AI Prompts: GPTMantra BlackBook “GPTMantra BlackBook: Your ultimate guide to harnessing the power of GPT-3.5 for enhanced productivity, creativity, and innovation. Unlock the secrets of AI language models with expert insights, practical tips, and cutting-edge techniques. Elevate your writing, coding, and problem-solving skills with GPTMantra BlackBook today.” Details: https://lnkd.in/gyF7VEph