How AI Is Built 🛠

Hey everyone, we have come a long way. Super excited to conclude the first season of How AI Is Built 🛠 with a slightly different episode. Today, we are taking more of an enterprise view. What is happening in the industry, where is the potential for enterprises to adopt AI, where can startups come in and build a solution. In today’s episode, we talk with Jonathan Yarkoni. He is the founder of Latent, where he specializes on solving hard problems with generative AI. We talk a lot about what industries are ready for AI disruption? What are the key drivers of the current technology shift? How to solve the big, hairy problems with AI? Some key points: - 2024 is about internal use-cases - 2025 will be about integrating AI into the customer-facing applications - Text-heavy domains (legal, education, marketing) are prime candidates, but healthcare and finance are a close second. - Companies that have "exhausted all possible paths" are perfect customers for AI solutions Check it out on Spotify: https://lnkd.in/dq4Sgqju Apple: https://lnkd.in/dGrvxwaC Or now also on Youtube: https://lnkd.in/dY6CWGgw Stay tuned for the next season. Season 2 will be all about information retrieval, recommendations, and RAG. I will try to bring three very related fields together that do not really talk and learn from each other.

What are the areas where generative AI can bring the highest ROI? 1. Text-heavy domains: ➡️ Legal ➡️ Education ➡️ Software engineering ➡️ Marketing 2. Biotech: ➡️ Protein engineering ➡️ Drug discovery 3. Entertainment: ➡️ Gaming ➡️ Content generation ➡️ Personalized streaming (e.g. "Netflix on demand") 4. Healthcare: ➡️ Diagnostics ➡️ Personalized medicine 5. Finance: ➡️ Risk assessment ➡️ Algorithmic trading Key drivers: • Maturity of text models • Advancements in diffusion models • Increasing hyperrealism in generated content • Growing ability to adhere to specific styles/brands The next few years will likely see waves of AI adoption: 2024: Focus on internal use cases 2025: Broader customer-facing applications as models improve Learn more in today’s podcast with Jonathan Yarkoni on How AI Is Built 🛠. Links to the podcast in the comments. What industries do you think will see the biggest #AI impact? Any surprises? 👇

We have now also officially launched on YouTube, if you want to catch a video version of the podcast! Link: https://lnkd.in/d3xNpeqh

AI agents might soon be running your business workflows? Just wrapped up a podcast on the future of AI agents. 5 Key Takeaways: - The "human in the loop" is evolving: It's no longer about offshore workers completing tasks, but about AI pausing to ask YOU for critical decisions. - Cost vs. Accuracy Trade-off: For high-stakes tasks (think financial analysis), accuracy trumps cost. But for marketing copy? A faster, cheaper model might suffice. - The Workflow Capture Challenge: The magic often happens in the user's head. Capturing that decision-making process is the next frontier in agent development. - Enterprise-Ready Agents: Deploying AI agents in corporate environments requires robust security, scalability, and integration with existing systems. - The Future is Declarative: Moving beyond chat interfaces, the next big leap could be agents that understand and modify system states directly. Learn more about how you can actually bring agents into production and listen to the latest episode of How AI Is Built 🛠 with Rahul Parundekar, AI Hero. Links are in the comments below. #aiagents #agents #llms

Are you facing these reliability nightmares with your AI agents? - Does your AI agent occasionally produce nonsensical or off-topic responses? - Are you struggling to maintain consistent output quality as your agent scales? - Do minor changes in input sometimes lead to wildly different outputs? (1) Prompt Engineering for Consistency Implement structured prompts with clear instructions and examples Use few-shot learning techniques to guide the model's behavior Regularly update and refine prompts based on performance data (2) Output Filtering and Post-processing Implement content filters to catch inappropriate or off-topic responses Use language models or rule-based systems to refine and polish agent outputs Implement fact-checking mechanisms for critical information (3) Prompt Preprocessing and Compression Reduce token count without losing essential information Improve consistency by focusing on key elements of the prompt Enable use of larger context windows without increased costs In today's podcast of How AI Is Built 🛠, Richmond Alake gives you the tools to create more reliable and predictable AI agents. Richmond is an AI engineer at MongoDB working on using agents to solve real-world problems. We talk about some of the latest tools for LLM and AI Agents like Microsoft's LLMLingua (Huiqiang Jiang) for prompt compression, crewAI (João (Joe) Moura) and LangChain (Harrison Chase) for building agents, and more. Catch the episode on Spotify, Apple, Snipd, or wherever you get your podcasts. Links are in the comments below! What reliability challenges have you faced with your AI agents? Let me know below! Catch you next week with another episode in our series on AI agents with Rahul Parundekar.

It's often beneficial to start with a simplified version of these concepts and iterate as you learn more about your specific integration challenges.

We had the pleasure of hosting Derek Tu of Carbon. If you want to learn how to feed your data-hungry friends at OpenAI or Anthropic, listen to this one. For the ones who do not want to listen to 37 minutes of me learning from Derek, the gist is below. https://lnkd.in/dperWzTh

When building LLMs, we easily get caught up with the new, shiny optimizations: optimizing chunking strategies and fine-tuning prompts. The biggest gains can often be found in the pipes. Improving the data that is fed into your models. Tags, such as topics, the source of the document, and other metadata, allow you to retrieve more relevant input and contextualize it for the LLM. In today's episode of How AI Is Built 🛠, I talked to Derek Tu, founder of Carbon. Carbon is an ETL tool specifically built for LLMs that connects to various data sources and parses and chunks unstructured data before delivering it to a vector database. Beyond tagging, we discussed fine-tuning embedding models, necessary innovations to drive the space forward, building ETL pipelines for LLMs, extracting data from documents, and much more. Jump into the podcast ⬇ (comments) Stay tuned for next week when we start a little mini-series on agents with Richmond Alake. #llms #genai #etl

Learn how to orchestrate tasks like data ingestion, transformation, and AI calls, as well as how to monitor and get analytics on data products.

Data storage costs skyrocketing? Quantization offer a powerful, affordable solution. Embeddings can be big, taking up lots of space. That's where quantization comes in. It's shrinking the embeddings down, making them easier to store and search through. Two popular methods are binary quantization (BQ) and product quantization (PQ). BQ simplifies embeddings, making them smaller but still useful. PQ breaks them into pieces, compressing each part separately. Learn more in today's podcast on How AI Is Built with Zain Hasan from Weaviate. https://lnkd.in/dZqwP4tG Let me know below what use-cases you dream about with low-cost storage. #rag #vectordatabase #llm #generativeai