Matanya Loewenthal’s Post

https://lnkd.in/gUQGYdPs A quickie from Mark Heaps using the Groq.com #llm #demo for a beginner end-to-end workflow with prompts that provide format solutions from a trained model. Works with #excel, #googlesheets, et al. Check out that #groqspeed. Whatever your LLM model #AI application is, if it requires ultra-low latency it will run #betterongroq. Contact labs@groq.com if you want to learn about #api early customer access for your enterprise.

Best aspects of software development for AI researchers to focus on for the next two years! Test suite generation, results analysis, and test harness coverage and ga-/risk analysis. Also software fault injection and automation of log analysis and configuration of telemetry monitoring systems (including performance benchmarking). The results of those efforts are easily measurable and quantifiable. Thus they're useful to the projects under test AND to the AI research and development efforts. They can also fill in for the slow decline in software quality assurance that's been going on for decades as well as the dearth of QA expertise. (It's been perceived by many as a dead end career path forever). Doing this will also build a foundation upon which improvements in automation code generation of the products themselves can be built by ensuring that hallucinations will be identified and isolated quickly. A broken test shows up as a failure when run against correct code. So the test generation automation is mostly protected from trivially bad tests. Fuzz testing is also cheaper and can be far more extensive using AI driven fuzzing tools. So the main risks are code coverage and use-case requirements gaps. But much of the time/effort that was formerly expended by existing QA staff can be shifted to gap analysis and behaviorally/test driven development. Given good test suites and performance telemetry in a reliable harness, broken production code triggers test failures and performance or reliability alerts continuously through development efforts. It's a compelling narrative.

Here is an easy step-by-step guide to creating your own custom GPT. Think of this custom GPT as an extension of your work to streamline or speed up tasks or help provide insights into data. It's like having an assistant! Keep in mind that you do need a subscription to CGPT Plus, but it's pretty straightforward after that. https://lnkd.in/gUfM8tNc

This is so on point. But I'm going to add my own nuances of course, because this is LinkedIn. The problem with data extraction has always been the amount of time required to understand the type of content and "expect the unexpected". In the past, rule-based and logic works well with structured documents, but not unstructured ones like PDFs and scanned images. Then we had deep-learning models that typically use neural networks to consider the spatial structure of the documents. These models require a large amount of training data to start with and can be expensive to deploy. Today we have LLMs that are readily available with an API call. Of course, each one of these technology isn't the only answer to the problem of "unstructured and unseen" documents. We at Xceptor believe that you shouldn't run your business on a single technology. We said this many times in our history, and I personally could remember the different hypes we went through from Robotic Process Automation to Optical Image Recognition. I think we survived those hype cycles because we truly believe this is about deploying a mixture of technologies to shape your business process. This time it's not so different, I think. #ai #automation #llm #deeplearning #xceptor #duco #data #ocr #rpa #machinelearning

With more industries embracing GenAI, it's crucial to consider key aspects before implementing LLMs or GenAI solutions. Our latest report delves into data hallucinations in GenAI applications and offers essential preventive measures. Additionally, we explore how Web 3.0 can amplify the impact of LLMs and GenAI, along with a detailed framework for data provenance and democratization. Check out the full report here: https://bit.ly/4bQOzCD #GenAI #LLMs #DataProvenance

Standard LLMs, although impressive in their own right for a variety of ‘Search’ use-cases, often fall short when it comes to generating relevant and coherent responses in an enterprise search environment.  This becomes even more prominent when there is no mechanism to set the right guardrails on top of company knowledgebase and repositories. This is where Small Language Models (SLMs) make a massive difference. These represent a departure from the one-size-fits-all approach of standard LLMs and are trained on specific domains, industry-terms, languages, webpages, SOPs, enterprise data or research contexts, enabling them to offer more tailored and relevant responses. Accordingly, enterprises are starting to adopt virtual assistants (or enterprise bots), specifically powered by RAG capabilities over SLMs. These ‘enterprise search bots’ are designed to be lightweight, efficient, and resource-friendly in terms of computational requirements and bandwidth for real-time inference. This approach not only packs the right user-experience across enterprise search channels, but also helps companies in delivering fairly dynamic and accurate responses to complex queries, without depending on any standard LLMs or heavy infra. #Search #RAG #VirtualAssistants #EnterpriseAI #GenerativeAI #LLM

Here's a simple but practical use case for LLMs in market research, where our 'junior analyst' Roya Ai distills some points on China from a few research reports on the outlook for 2024 - in the 1-minute video below. Accuracy not verified. In actual deployments on larger volumes, we would see full automation, including how the reports get ingested (API, upload, email, web capture), pre-filtered/classified, and where the outputs go (email, dashboards, API, etc.). All are integrated within our broader Kubro(TM) Information Engine (Enterprise SaaS) by Robotic Online Intelligence (ROI). In this case, we use Azure OpenAI API, but we have also installed Llama2 from scratch, with more room there for fine-tuning and knowledge editing. RAG (Retrieval Augmented Generation) or simpler Regex rules can also help with the accuracy / taking relevant components from longer bodies of text. #LLM #AI #GenerativeAI #Kubro #roboanalysts

Pt2 - A tremendous first day at Big Data & AI World London on the Superluminal Software Microsoft super stand! you could be mistaken to think we're not a #microsoftai #microsoftpartner #ledwall #ai #bigdataaiworld #londonexcel 💥

Exciting Day, again! AI21 Labs Jamba Instruct + Snowflake #Cortex available today! 🚀 📢: https://lnkd.in/giHNG7JW Why is this IMPORTANT? To put this into context (pun intended), Jamba Instruct comes with a whopping 256K context window. 🤯 That's like having 512 pages of a standard book, dozens of 10-K reports, thousands of standard customer reviews, hundreds of long product descriptions, or tens of thousands of social media posts about your company or product available all at once. This is a step-function leap for ENTERPRISE AI; on the march towards infinite context windows! #SnowflakeCortex Also check out: 🔹 5 Powerful Prompts for Data Engineering: https://lnkd.in/gMhjujAp 🔹 The AI Alphabet Soup: https://lnkd.in/gjAa6TU5 🔹 100+ Tools & Concepts Data Leaders Should Know To Keep Pace with GenAI/LLMs (in 2024): https://lnkd.in/g_7VeS62 #AI #DataEngineering #GenAI #LLMs #Innovation #SnowflakeCortex

Some cool Features coming to Dynamics Finance 365 that caught my attention. 1. AI powered collections workspace timeline summary: Gives collection agents access to relevant information in one place and personalized emails. 2. Matching rule enhancement in advanced bank reconciliation: A lot of manual work will be reduced. 3. Defaulting Engine for Financial tags powered by Ai-recommneded rules: Imagine being able to create rules using natural language via Microsoft co-pilot. Imagine be able to create rules using natural language instead of an expression builder, saves you enough time. SCM too has some cool AI features especially around MRP (DDMRP) with some micro services: Planning Optimization add-in, priority-based planning. #modernfinance #copilot #copilotai