Sebastián Ramírez Montaño

Nowadays, most LLMs get trained with the AdamW optimizer as opposed to the Adam optimizer. Why? There used to be a time when Adam was the king among optimizers, and it didn't make much sense to spend too much time trying to find a better one. This has changed recently, and AdamW has become the default optimizer for the LLM practitioners. It all depends on how we apply the regularization terms to the weight parameters. For the typical gradient descent algorithm, if we want to apply an L2​ regularization term, we modify the loss function such that: regularized loss = loss + L2 term Then, we compute the gradient of that new loss to update the model parameters. The goal of the regularization term is to ensure that the weights don't grow too large and it acts as a weight decay mechanism when we update the weights. In Adam, when we apply the L2 regularization, we regularize the loss function as well but it gets used differently. The loss function is used to compute the first and second moments and when we update the weights, the regularization term is in the numerator and denominator of the gradient update term. Because of it, the effect of the L2 regularization term is minimized and cannot act as a weight decay mechanism. In AdamW, on the other hand, we DO NOT regularize the loss function and compute the gradient update independent from the regularization term. Only during the weights update do we add the regularization term so that it acts on the weights and not on the loss function. Because of it, training with AdamW tends to be more stable and leads to models that generalize better! Good to know, right? -- 👉 Last week to register for the LLM bootcamp: https://lnkd.in/e3YdK6DT --

835

39 Comments

LLMs and Search Following our recent webinar on the use of Generative AI in search and recommender engines, there have been numerous inquiries regarding the FreshLLMs approach. It is crucial to understand that the FreshLLMs methodology addresses a specific set of challenges, which is why it is implemented in Google Search, Perplexity, and other major LLM-based search platforms. Similar principles apply to Retrieval-Augmented Generation (RAG) and LLM agent searches; each technology is designed to solve a distinct set of problems and none offers a comprehensive solution for all scenarios. Google Gemini and Perplexity, for instance, adopt approaches inspired by FreshLLMs due to the necessity of retrieving information from extensive datasets (such as the entire web) and maintaining the freshness of information accessed at the query stage. Utilizing a search engine as a tool for LLMs effectively addresses these issues, provided the search engine performs efficiently. It must be capable of retrieving data from (dozens of) billions of documents with high precision and quality, and it must be adept at understanding and ranking recent information as required by the query. This requirement extends beyond web searches to various other domains that necessitate complex searches over vast information sets with freshness and precision, such as e-commerce with hundreds of millions of items, geo/local searches, and more. In the long term, I anticipate this paradigm to remain valuable and complementary to LLM agent searches or RAG architectures, rather than being supplanted by them.

18

1 Comment

Prompt Engineering: Anthropic continues to show they understand the needs of the prompt engineer. Does this announcement suggest we can anticipate Anthropic evolving this into a robust prompt engineering & management platform with evals? We will see...

9

Reasoning models are really interesting: https://lnkd.in/dEbydJHZ Whom it's for: "...and by developers in all fields to build and execute multi-step workflows" Learning to reason with LLMs: https://lnkd.in/dgnBiXfU

1

OLMoE 1x7B by Allen AI - apache licensed, 64 experts, 8 active - trained on 5T tokens, matches Gemma, Llama in perf, with orders of magnitude faster speed ⚡ > 1.3B Active and 6.9B Total - 5x fewer parameters than the comparative dense model > 64 experts per layer, 8 active experts > Compared to 7B dense, it trained on 4x less training FLOPs > Trained on Dolma, Proof Pile 2, DCLM, etc > Fully open model, code-base, datasets on the Hub > Compatible with Transformers 🤗 Kudos to Allen AI, Niklas and the team on shipping such an impactful project - open source for the win! 🐐 I'm so here for smol MoEs! ⚡

296

5 Comments

Programming just got a lot more accessible. French AI startup Mistral launched Codestral, its first-ever code-centric AI model. Codestral is a 22B parameter, open-weight AI model designed for code generation tasks. It’s trained on a diverse dataset covering over 80 programming languages—including Python, Java, C++, JavaScript, and more. Mistral’s model excels at completing coding functions, writing tests, and filling in partial code snippets. #AI #coding #techjob

1

I have created this repository with many #gemini examples - Multimodal Prompts - Langchain Integration - Embeddings and Search - Firestore Vector - LLM Reflection - Context Caching - Output Schema - ... Leave a star #ai #gemini #firebase #langchain #embeddings https://lnkd.in/eMRqt-Wj

13

2 Comments

Claude 3.5 vs. GPT-4o: Which AI Reigns Supreme? Anthropic surprised us all with the announcement of their new Claude 3.5 Sonnet model, the first from their 3.5 model family. Positioned between the largest Opus model and the smallest, fastest, and cheapest Haiku model, the Sonnet model offers a substantial performance boost. Key Highlights: Performance and Speed ⚡: The Claude 3.5 Sonnet is designed to be significantly smarter and faster than the current Claude Opus model. Benchmarks indicate that Sonnet is outperforming GPT-4 on many fronts, making it a formidable competitor in the AI landscape. Accessibility 🌐: Available via API from Anthropic, Bedrock, and Google Cloud from day one, ensuring wide accessibility for developers and businesses. Cost Efficiency 💰: Priced at $3 per million input tokens and $15 per million output tokens, it’s more cost-effective than GPT-4 on input costs, offering substantial savings for high-volume users. Innovative Features 🧩: The Sonnet model boasts improved vision tasks, including interpreting charts and graphs, making it versatile for various applications. Internal benchmarks show significant improvements in agentic coding evaluation over the Opus model, highlighting its practical benefits. Artifact Feature 🛠️: The new Artifacts feature in Claude.ai allows dynamic workspace interaction, where users can edit web pages, generate slides, and create code interactively. This feature enhances productivity and ease of use for developers. Real-World Applications 🌟: From creating websites to generating Python code, Sonnet’s capabilities are vast. The model’s ability to handle zero-shot inference effectively makes it even more powerful for real-world tasks. Why This Matters Anthropic’s Claude 3.5 Sonnet model represents a significant leap in AI technology. Its combination of speed, cost efficiency, and advanced features sets a new standard in the industry. Whether you’re a developer looking to leverage AI for innovative applications or a business seeking cost-effective AI solutions, Claude 3.5 Sonnet offers compelling benefits. Coming Soon to Europe 🌍 Currently, the Claude 3.5 Sonnet model is available through the Anthropic API, Bedrock on AWS, and Google Cloud, but it's not yet available in Europe. Stay tuned for updates on its European availability. Share Your Experience Are you excited about the Claude 3.5 Sonnet model? What are your thoughts? Share your experiences and insights in the comments below. Let’s explore the future of AI together! 💬 Stay updated with the latest in AI by following me on LinkedIn. For more in-depth discussions and insights, don’t forget to like and subscribe! 👍 #AI #MachineLearning #Anthropic #Claude3.5 #Innovation #TechNews #ArtificialIntelligence #Developers #TechCommunity 🔗 Learn more about Claude 3.5 Sonnet: https://lnkd.in/dhRjupi7

2

https://lnkd.in/g_y_ZxmH Explainable AI: Seven Tools and Techniques for Model Interpretability

1

Psycopg 3.2 released 🚀 Psycopg just released version 3.2 after almost two years of hard work and 846 commits! 🎉 This update brings a ton of new features, including support for Numpy scalar types, easier notification reception, and the ability to run queries with PostgreSQL-format parameters. Plus, there's a row factory for returning scalar values and support for the newest PostgreSQL v17 release. But wait, that's not all! Psycopg 3.2 also includes asynchronous, encrypted cancellations and bundles OpenSSL 3.x with its binary packages. Talk about convenience. And let's not forget about those pesky bug fixes, which have also been included in the 3.1.20 package. If you're not convinced yet, check out the release notes and news entry for more details. Just a reminder, some of the new features in 3.2 require libpq v17, which is not yet available in the binary packages. But don't worry, updated packages will be released as soon as PostgreSQL v17 is officially released. A big thank you to everyone who contributed to this project, especially Denis Laxalde from Dalibo. If you're a fan of Python and PostgreSQL, show your support by becoming a sponsor 💜 Happy hacking! #PostgreSQL #Psycopg #Release #GeekOut #Sponsor Source: https://lnkd.in/ebkzhXuU

1

Unleash the Power of Multiple Documents with Agentic RAG! Tired of one-trick pony RAG pipelines? Our latest article explores how Agentic RAG unlocks multi-document retrieval, allowing you to tap into a wealth of information across various sources. Imagine an intelligent agent that seamlessly directs you to the perfect document for any question! That's the magic of Agentic RAG. ✨ In this article, you'll learn: 1. How to overcome limitations of single-document RAG pipelines. 2. The secrets behind Agentic RAG's multi-document capabilities. 3. How to leverage this powerful system to find exactly what you need, no matter where it's hidden. ️ Ready to revolutionize your information retrieval game? Check out the full article, link below. #MultiDocumentRetrieval #AgenticRAG #MachineLearning #AI #InformationRetrieval

5

RAG Or Long Context LLMs - A Hamlet-Level Question! And some findings. https://lnkd.in/eJwVWxXm

4

🚀 Exploring the World of SMERF (Not the Cartoon, but Just as Cool! 😄) 🚀 SMERF (Streamable Memory Efficient Radiance Fields), a game-changer in real-time 3D rendering. Developed by Google DeepMind/Research, SMERF brings high-fidelity 3D graphics directly to your web browser, even on smartphones! 🌟 What makes SMERF unique? 1. Efficiency: High-speed rendering using advanced model partitioning and training strategies. 2. Accessibility: Navigate complex 3D scenes smoothly on various devices. 3. Versatility: Perfect for VR, gaming, and architectural visualisation. Just like the Smurfs brought joy to our screens, SMERF is set to revolutionise our digital experiences! So, what can this technology do for the Architectural industry? 1. Faster and easier communication of designs. 2. A new generation of compact XR devices specifically designed for streaming 3D scenes. 3. Improved design validation and iteration through real-time, immersive feedback. What potential applications do you see for SMERF in your field? Share your thoughts below! 💡 Check it out here: https://lnkd.in/gJyPRNcb #TechInnovation #3DRendering #SMERF #4c4545

5

4 Comments

LLMs are not Swiss Knives. In the past 2-3 months I've seen a resurgence of posts dismissing Language Models, calling them bogus, fakes, etc. Lots of posts diminishing this technology, even discarding it. LLMs, and in general Language Models of all sizes, are artifacts with some great, in fact, marvelous capabilities. My Casio calculator can't generate text. It only does math, and not all the math at it. It is not because of a bug but because of the design. Expecting it to produce text would be simply expecting it to do something that it is not supposed to do. Same thing with LLMs. This artifact has certain capabilities which we all are still learning and understanding. Some of its great capabilities are now clear. Also many of its limitations are becoming clear. But one thing I am sure: the things for which it is good, are truly amazing. This is not a technology that should be discarded, or deemed fake or buggy or bad. This is an amazing technology that we have to learn to use properly and within its capabilities. Learning point: Use LLMs for what they are good at, and don't expect them to be good at what they, by nature or design, are unable to do.

62

37 Comments

Introducing 𝗣𝗿𝗼𝗺𝗲𝘁𝗵𝗲𝘂𝘀 𝟮 - a more powerful 𝗲𝘃𝗮𝗹𝘂𝗮𝘁𝗼𝗿 𝗟𝗠 than its predecessor, designed to closely mirror human and GPT-4 judgments. This open-source model aims to bridge the gap between proprietary LMs like GPT-4, which are often employed for evaluation despite concerns over transparency, controllability, and affordability. ➜ Check the news at: https://lnkd.in/dM9De_7Y ➜ Vendorvise's official Instagram: https://lnkd.in/ebK2KQ2C ➜ Vendorvise's Linktree: https://lnkd.in/dQdbMdth 🤖 If you enjoy reading my posts about AI and tech news, hit that like button and share with your friends to support my content! #ai #artificialintelligence #ainews #aipowered #news #blog #prometheus2 #opensource #researchpaper #generativeai #llms #tokens #aitools #aitech #tech #technology #hardware #content #socialmedia #future #new #launch #business #techtrends #innovation #success #markets #people

2

Can Embedding/Reranker models compare numbers? This was a question I was asked at the #ICML2024 in Vienna today. A lot of LLMs can't figure out that 9.11 is actually smaller than 9.9. Can our embedding and reranker models do any better? I did some experiments to assess the effectiveness of our models in handling different numerical formats. Read more: https://lnkd.in/gpcbSnDF

104

6 Comments

Are you tired of getting wrong outputs from your LLMs? It is time to upgrade and try out our agentic-RAG demo on OpenVINO. It will give your query a little superpower boost and all models can run locally on your laptop. Unbelievable :) Background: Retrieval-Augmented Generation (RAG) is a form of prompt engineering that enhances large language models (LLMs) by integrating external data into prompts. This improves the models’ knowledge, relevance, and professionalism. However, traditional RAG systems have limitations. They often lack flexibility, relying heavily on retrieval results from vector databases, which can lead to the uncritical incorporation of data. As these databases scale, standard RAG struggles to efficiently classify and filter input requests, turning the retrieval process into a time-consuming and labor-intensive task akin to finding a needle in a haystack. Table of content: Full Example: Creating a RAG System with OpenVINO™ and LlamaIndex 2.1 Model Conversion and Quantization 2.2 Model Task Initialization 2.3 Create a RAG Tool 2.4 Build the Agent Pipeline Summary and Outlook Blog: Build Agentic-RAG with OpenVINO™ and LlamaIndex https://lnkd.in/gaYJ-2EP #iamintel #llms #openvinoblogseries #rag #llamaindex Special thanks to Ethan Yang building out this and Stephanie Maluso polishing the blog like crazy!!

71

8 Comments