Myles Harrison

ARE WE IN A RECESSION? As mentioned in our prevision post, the NBER's Business Cycle Dating Committee took 4 to 21 months to declare past recessions. Sometimes, they even do so AFTER the recession ended. Therefore, how do we know whether we are in a recession today? And are we in one? We know by looking at a handful of indicators. One is the Smoothed U.S. Recession Probability (at the top on the image below), the other is the Real-time Sahm Rule Recession Indicator (at the bottom). Both datasets are updated monthly. The former tells us that we are not in a recession, but the latter suggests that we might be. Because of this uncomfortable ambiguity, the STOIC Recession Index will include a real-time horizon, which will provide the probability that we are in a recession. It will be updated on a monthly basis as well. Right now, it would probably stand between 25% to 50%, meaning that we are not in a recession, but that we might get very close to one.

4

Just a few days out from my talk at Toronto Machine Learning Society (TMLS)‘ Summit. I will be speaking next Monday July 15 at 11:30 in Room 1. My talk, “GenAI: A New Renaissance in Product Development” is about actionable tips for how to incorporate GenAI in your day to day workflows. I cover three major topics on how to: 1. Get the most from work with global teams 2. Expedite user research (while still talking to real humans!) 3. Find the best tools to leverage during the product development process Along the way I’ll be sharing caveats and pitfalls to be on the lookout for. I’ve got some fun visuals to backdrop the talk. Plus, I will be launching my playbook for how to use GenAI in the product development process. It covers a whole lot more and the audience will have first access! So come on out from 11:30 - 12 in Room 1! And it’s my first trip to Toronto so please send tips!

27

Yesterday I pinged one of Yann LeCun's post that suggested that LLMs could perform planning, which involves a thought process that LLMs are not designed to do. This comment surprised me, because I have generally found Yann’s comments on LLM to be accurate, without all the AI hype that is coming out of companies selling this technology.   I took the time to read the link to an article that the post was based on, and I am happy to say that the article did not contain claims about planning. In fact, the most relevant paragraph reads:   =========== While acknowledging the advances brought by generative AI models like OpenAI’s ChatGPT and Meta’s Llama, LeCun highlighted their limitations. “LLMs can deal with language because it is simple and discrete, but it cannot deal with the complexity of the real world,” he explained. These systems lack the ability to reason, plan, and understand the physical world in the same way humans do. ===========   He went on to say that Meta wants to “develop a new AI architecture capable of observing and learning from the physical world, similar to how babies interact with their environment.” No objection here either. But babies (or adults) are not able to learn complex physical problems the way a computer model can. Optimization models can solve problems from finding the shortest path (e.g. Google maps) to managing a fleet of trucks (e.g. @optimaldynamics) in ways that humans cannot compete with. But LLMs can help prepare inputs and process outputs using human guidance. These technologies are not competitive – they are complementary.

162

19 Comments

Woo! Anthropic has finally made Claude/Claude Pro available to Canadians. Very happy to have direct access to it for testing. I also like the fact that they explicitly say they don't train on the data from interactions. For reference; "By default (and just like our beta version of Claude.ai), we will not use your prompts and conversations from Claude Pro to train our models. There are two instances in which we may use your prompts and conversations to train our models: (1) if you give us explicit permission by submitting feedback through the thumbs up/down feature or by reaching out to us with a request, and (2) where your prompts and conversations are flagged for trust and safety review, we may use or analyze those conversations to improve our ability to detect and enforce Usage Policy violations, including to train Trust and Safety classifiers in order to make our services safer. Only a limited number of staff members have access to conversation data and they will only access this data for explicit business purposes." #anthropic #ai #claude #claudepro #Canada #access

2

How we build Shopify data products in 5 Steps (with a 90% success rate) I've been with Shopify for 2+ years. It's impressive how many products we built. This is our approach: 1. Quantify the Opportunity ↳ Estimate potential revenue. ↳ Define the expected lift. 2. Get Stakeholders On Board ↳ Communicate with key decision-makers. ↳ Avoid over-communicating. 3. Build a Small Prototype ↳ Identify a specific use case. ↳ Build the simplest solution. ↳ Use heuristics if needed. 4. Test Impact in A/B Test ↳ Design the experiment. ↳ Set up control and treatment groups. ↳ Execute the experiment. 5. Summarize Findings ↳ Analyze the results. ↳ Document findings for all. This is a high-level overview. Many small steps are involved. ↳ But this approach takes you far. Find a way for your team. P.S.: Need more details? Send a DM. #Experimentation #AI #Data #ML

20

Intuitive Example, Principles, Code Explanation and Insights about CRAG Take your LLM performance to the next level. #llm #llmops #chatgpt4 #mlops #ai #datascience #ml #gemini tps://https://lnkd.in/dDkZiZ5m

13

AI Agents aren't going to be easy, and the problem is the LLMs. LLMs are trained to provide an answer to a question or an output for a task. "write this code", "categorize this list". They are not, however, great at making decisions - navigating a fork in the road. In fact, they're trained to *not* make decisions for you. What makes an agent an agent is its ability to decide, and the current models are fairly terrible at it. There isn't as much training data on making decisions. The fine tuning isn't being done on decision making. For a while, agents are going to be rudimentary at best.

16

1 Comment

Come visit Cerebras booth at #Neurips 2024! Along with all the work on the training front, for the first time we will also be talking about some of the ML research we are doing for LLM inference. If you want to know how we are using synthetic data to make our inference efficient and blazing fast, then check out the poster by Vithu Thangarasa.

27

The Toronto Machine Learning Society (TMLS) will host Dr. Yizhi Yin from Neo4j for an immersive workshop on Enabling GenAI Breakthroughs with Knowledge Graphs. 🌟 We'll explore the integration of Large Language Models (LLMs) with knowledge graphs using Neo4j, a leader in graph database and graph analytics, through a Retrieval Augmented Generation (RAG) approach. 🛠 Workshop Highlights: ✅ How to overcome LLM limitations with Retrieval Augmented Generation (RAG) ✅ How to build a personalized messenger app with RAG patterns for product recommendations ✅ Analyzing Vector Search Results with Graph Patterns ✅ Enriching Search Results Using Graph Data Science Methods You can see the full abstract here: https://lnkd.in/gwSRjPjz #GenAI #KnowledgeGraphs #tmls #neo4j

16

FlexAttention: Simplifying Attention in PyTorch _________________________________________________________________ FlexAttention is a new PyTorch API that combines the flexibility of PyTorch with the performance of optimized attention methods. It allows researchers to implement various attention mechanisms easily without needing to write custom kernels. Key Points: - Flexibility with Performance: -- Traditional attention methods offer high performance but limited flexibility. -- FlexAttention provides a flexible API that supports many attention variants with minimal effort. - How It Works: -- Standard Attention: Calculates attention scores using queries, keys, and values, normalizes them, and computes the weighted sum of values. -- FlexAttention: Allows for modifications to the attention scores through a flexible scoring function, enabling customization for different types of attention mechanisms. - Examples of Attention Variants: -- Relative Position Encoding: Adds positional information to attention scores based on relative positions. -- ALiBi Bias: Applies a bias to attention scores to account for position differences, enhancing model performance. -- Causal Mask: Masks future positions in the sequence to ensure the model only attends to previous positions, useful for autoregressive models. Handling Sparsity: -- BlockMask: Efficiently manages sparse attention masks, such as those used in causal attention, by creating masks that prevent unnecessary computations. Learn More: https://lnkd.in/g58pxB4A

8

Discover the Phi-3 mini models, AI creations with immense potential. The short context model, Phi-3-mini-4k-instruct-onnx, handles prompts up to 4k words, while the long context version deals with much lengthier inputs and outputs. Lee Stott shares how to leverage these models for text generation using NLP techniques. From setting up Python to utilizing the generate() API, Lee provides step-by-step instructions and code examples. Let's unleash the capabilities of Phi-3 mini models! https://lnkd.in/gp8c_2T7 #phi3 #onnx #msftadvocate

33

If you didn't have a chance to attend my Getting Started with Generative Text & Fine-tuning LLMs in Hugging Face workshop at Toronto Machine Learning Society (TMLS) Toronto Summit last month, all of the content is now available online! 🚀 Highlights of the workshop include: 💡 An intro to #largelanguagemodels and generative text models in #HuggingFace 🤗 🌡️An explanation of different decoding strategies for #LLMs (temperature, top-p, top-k, beam & greedy search) ⚙️Fine-tuning GPT-2 medium to talk like a Jedi Master 😱Using LoRA to fine-tune GPT-XL using only a Google Colab free instance 🙊Throwing shade at AGI, in-context learning, and George Lucas Check out the full 90 minute video of the workshop here 📽️: https://lnkd.in/gh5aXy4Q View the workshop materials on Github🧑‍💻: https://lnkd.in/gb7aKSCU Give the code a spin for yourself in Google Colab 🧪: https://lnkd.in/gvnpZWSu

11

I just released the April edition of my AI & the Web newsletter. Nathaniel Simard, founder & CEO of Tracel Technologies and inventor of the burn ML framework, spoke with me about how he got into ML, burn, Tracel and the future of AI. WebGPU, the successor of WebGL, will soon also be supported in Firefox and Safari. Small, powerful LLMs running right in the browser is not a distant future, but are the next big thing in cost-effective and privacy-friendly AI. Google is working hard on better WASM support in the browser that makes ML models running fast even on low-end devices without GPU. With Llama 3 by AI at Meta and Phi 3 by Microsoft AI we got some exciting new, small modells. Llama 3 was already ported to run in the browser and we'll see Phi 3 soon as well. MusicGen has been ported to the browser in transformers.js by Hugging Face engineers and developed an online Jukebox running right in the browser. #webgpu #webml #ai #ml #llama3 #wasm https://lnkd.in/drJCY7Ew

19

2 Comments

Great feature to learn from users best use cases for AI: "New research from Anthropic, describing a system they built called Clio - for Claude insights and observations - which attempts to provide insights into how Claude is being used by end-users while also preserving user privacy."

2

Ever grappled with juggling multiple data processing engines, i.e.,🦆,❄️, 🐼 etc. etc.? Meet Ibis, the hidden-gem of PyData that's about to revolutionize your data workflow. This post unpacks how Ibis seamlessly integrates with various backends and presents a concept where a single expression can run across multiple- engines. The primary benefits are: 1️⃣ Segment computation, by data size or cost, in-situ or serverless 2️⃣ Apply database style optimizations to pipelines with ML or complex operators 3️⃣ Right-size infrastructure, based on consumption stages 🔗 Deep dive into the full post here: https://lnkd.in/ehNqCTwB

53

5 Comments

Text embeddings have become crucial for transforming textual data into dense vector representations (with context), enabling efficient processing across various modalities like text, images, and audio. This article provides a good summary on some of these (not all) embeddings available for text https://lnkd.in/gKp4ta9a 1. Gensim Word Embeddings (Free) Why Use It: Specializes in NLP with open-source models like Word2Vec and FastText, promoting transparency and ease of custom training. 2. Facebook Embeddings (Free) Why Use It: Trained on extensive corpora and supports over 100 languages, making it versatile for global applications. 3. AllenNLP Embeddings (Free) Why Use It: Provides NLP-focused embeddings like BERT and ELMo, with tools for fine-tuning and visualization. 4. OpenAI Embeddings (Paid) Why Use It: Comprehensive training on massive datasets allows for effective semantic understanding, making it suitable for advanced NLP tasks. Ideal for tasks requiring zero-shot learning. 5. MultiLingual BERT (Free) Why Use It: Handles multilingual datasets effectively, offering robust support for multiple languages. Perfect for applications requiring consistent performance across different languages, such as global customer support systems. 6. DistilBERT Base Uncased (Free) Why Use It: Smaller and faster than BERT, maintaining near-comparable performance with reduced resource usage. 7. MPNet V2 (Free) Why Use It: Excels in text similarity tasks with a Siamese architecture, providing accurate similarity measurements. 8. HuggingFace Embeddings (Free/Paid) Why Use It: Offers versatility with a wide range of embeddings (text, image, audio) and the ability to fine-tune on custom data. HF also provides wrappers to almost all open sourced models. 9. Longformer Base 4096 (Free) Why Use It: Supports long texts up to 4096 tokens, making it unique in handling lengthy documents without truncation. 10. RoBERTa (2022) (Free) Why Use It: Offers updated text representations with comprehensive training on recent data. 11. GTE-Base (Free) Why Use It: Offers a good balance of performance and resource requirements with a 768-dimensional embedding. 12. GTE-Large (Free) Why Use It: Provides high-quality embeddings with more detailed representations, useful for nuanced text analysis. Best for applications requiring in-depth text understanding, such as detailed content recommendations. 13. GTE-Small (Free) Why Use It: Optimized for speed with a smaller model size, making it efficient for applications needing quick processing. Useful for real-time systems where fast text similarity searches are critical. 14. E5-Small (Free) Why Use It: Combines compact size with good performance, making it suitable for fast, general-purpose text embedding tasks.

11

I encourage you all to try the demo (link at bottom) - mind blown!

1

I just got my own copy of “Building LLMs for Production” by Louis-François Bouchard and Louie Peters. As I am getting more into applications of GenAI I expect this book to become one of my go-to resources in the upcoming months. https://amzn.to/3Yv5FmL #AI #GenAI #LLMs

2,250

51 Comments