CJ Przybyl’s Post

Here’s a little experiment I did playing around with GPT4 Vision while waiting at the DMV. GPT4 Vision effortlessly and autonomously executed scene understanding, unaided natural scene OCR, and complex reasoning combining the scene with my text query before giving me the correct answer in natural language. Total time was about 2 to 3 seconds. The entire chat is below. I gave GPT4 no additional context about my location or intentions. This seems simple for humans - but what the model is doing is absurdly hard. This is a combination of at least four or five AI problems that (a year ago) I didn’t think would be solved in my lifetime, yet here we are. #gpt4 #ocr #computervision #holycrap

If A=B > B=A. Simple equation, right? An LLM model cannot solve this simple equation. Why? The problem is simple. These are the main types of algorithms that presently exist: Prediction, Classification, Natural Language Processing. The Reversal problem literally does not fall under any of those classifications. This missing element can only have three possible solutions, it is not rocket science. 1. Data 2. Algorithm 3. We Are 100% Wrong With All of It. I have experimented with 1. Quite a lot. I know I am far from alone in that research and experimentation. It has limits. It is not the answer. You can see it. I think the answer is behind Door #2. I can put my money where my mouth is regarding this:

Generative AI has egalitarianized data science. If you want to bring data analytics with natural language processing under your own roof, reach out to me. we'll demystify it for you so you'll look at it as just "another report", albeit an extremely useful one. for example, coalescing and then tokenizing customer email content to predict retention, allowing you to focus on at-risk renewals.

Millions of work orders and maintenance logs are generated every week. These records and work orders contain the life story of a machine. The life story contains important information such as when it failed, why it failed, which parts were replaced, how much it cost, and so on. To maintenance and operations professionals, this information is vital. They are constantly in need of answers such as 1. how frequently the machine fails 2. which parts are responsible 3. are there some good/bad actors in the fleet 4. are there certain parts consumed more than others 5. whether the cost and downtime pattern changed. 6. Is a Caterpillar or Komatsu truck more cost-effective? And many more such insights. Unfortunately, these insights are buried in millions and millions of rows of Excel files, ERP databases, or EAM systems. To excavate the information is very difficult and highly time-consuming. It takes days or weeks depending on the quality of data and the amount of information needed. Squares software is specifically designed to provide these insights by processing complex text-based data in seconds. Squares runs on a platform supported by two powerful NLP AI transformers offering an accuracy of more than 90%. No other software in the industry is as powerful, accurate, or as fast as Squares is. I am so proud of my team who keep on enhancing it.

🚗 Revolutionizing Subrogation with SubroCheck! 🚀 We're excited to share a recent success story of our state-of-the-art solution, SubroCheck! SubroCheck leverages the power of natural language processing and machine learning to uncover missed subrogation opportunities in closed auto claims. Insurance companies using this tool can expect to substantially increase their subrogation revenue! The customer profiled in this case study recaptured an additional 40 opportunities per ten thousand closed claims that they processed. 📈 Key highlights: ✅ Automated extraction of 150+ accident facts from diverse documents with accuracy of up to 99.8%. ✅ Sophisticated understanding of state laws, including comparative fault laws and statutes of limitations. ✅ Generates a report with evidence-based recommendations for easy review by subrogation specialists. ✅ Result: $59,474 additional subrogation dollars per 10,000 claims. 💰 ✅ A whopping 100x reduction in review costs, compared to reviewing all claims manually. Not only does SubroCheck boost revenue, but it also empowers companies with insights to their subrogation operations and improved customer retention. Experience the future of insurance operations with Megaputer Intelligence. Discover what you might be missing. 🔍 #Subrogation #InsuranceTech #AI #Megaputer https://lnkd.in/gjrhsXan

Groq achieves exceptional LLM performance milestone, doubling inference speed for Llama-2 70B in three weeks #AI #AImarketdominance #AIsolutions #AItechnology #artificialintelligence #cyberthreats #emergencyresponses #GPUs #Groq #languageprocessingsystem #llm #LLMperformance #lowlatency #LPU™system #machinelearning #tokenspersecond #ultralowlatency

Hey connections....... Exciting news!!! I have recently completed the Encoder-Decoder Architecture course and delved deep into the world of AI. From sequence-to-sequence tasks to language translation, the encoder-decoder framework continues to drive innovation in the field of machine learning. It's amazing to be part of the ever-evolving landscape of AI. Let's keep pushing the boundaries together! #AI #MachineLearning #EncoderDecoder #simplilearn

Love this blog by Alex S. about how he fine-tuned Mistral AI , Solar LLM (by Predibase), and Llama3-7B to beat GPT-4 for his use case. There were a few things that struck me: 1️⃣ Not enough people write about the absolute grind it is to evaluate how good or bad these models are 2️⃣ Finetuning models once is relatively easy, fine-tuning in a repeatable, trustworthy way is still hard (TBH this is a carry-over from "traditional" ML in general 3️⃣ Structured data extraction, while considered boring to some, is probably the best use case for fine-tuning an LLM today. Read the full blog here, Alex is just an amazing writer and his work on LLMs is well worth following: https://lnkd.in/dzbgcKDC

I was shook when I discovered that the GPT model in AI Builder can properly detect Klingon! 🖖 I was processing a new Power Platform Prompt sample from Philippe Larrue (https://lnkd.in/gT64ap4V) and of course I had to do some testing to see how it works. The prompt let's you put in a block of text and it will return the language of the text in language code format. The instructions say to return "unknown" if it can't detect the language. I wanted to test it not being able to find the language so I put in something in Klingon thinking it wouldn't identify it but it did 😲 #PowerPlatform #AIBuilder #Prompts #Klingon

Large Language Models: Friend or Foe? Exploring the Power and Potential Perils of AI Text Generation