A Close Look at AI Pain Points, and How to (Sometimes) Resolve Them

Feeling inspired to write your first TDS post? We’re always open to contributions from new authors.

In the span of just a few years, AI-powered tools have gone from (relatively) niche products targeting audiences with specialized skill sets to ones that are widely and rapidly adopted—sometimes by organizations that do not fully understand their tradeoffs and limitations.

Such a massive transformation all but ensures missteps, bottlenecks, and pain points. Individuals and teams alike are currently navigating the tricky terrain of an emerging technology that comes with many kinks that are yet to be ironed out.

This week, we’re highlighting several standout posts that address this conundrum with clarity and pragmatism. From handling hallucinations to making the right product choices for specific use cases, they tackle some of AI’s biggest pain points head-on. They might not present perfect solutions for every possible scenario—in some cases, one just doesn’t exist (yet?)—but they can help you approach your own challenges with the right mindset.

• Why GenAI Is a Data Deletion and Privacy Nightmare
  “Trying to remove training data once it has been baked into a large language model is like trying to remove sugar once it has been baked into a cake.” Cassie Kozyrkov is back on TDS with an excellent analysis of the privacy issues that can arise while training models on user data, and the difficulty of resolving them when guardrails are only introduced after the fact.
• Exposing Jailbreak Vulnerabilities in LLM Applications with ARTKIT
  There’s a growing understanding of the safety and privacy risks inherent to LLM-based products, particularly ones where sophisticated “jailbreaking” techniques can, with some persistence and patience, bypass whatever data-protection measures the developers had put in place. Kenneth Leung demonstrates the urgency of this issue in his latest article, which explores using the open-source ARTKIT framework to automatically evaluate LLM security vulnerabilities.

• Choosing Between LLM Agent Frameworks
  The rise of AI agents has opened up new opportunities to automate and streamline tedious workflows, but also raises pressing questions about matching the right tool to the right task. Aparna Dhinakaran’s detailed overview addresses one of the biggest dilemmas ML product managers currently face when picking an agent framework: “Do you go with the long-standing LangGraph, or the newer entrant LlamaIndex Workflows? Or do you go the traditional route and code the whole thing yourself?”
• How I Deal with Hallucinations at an AI Startup
  “Imagine an AI misreading an invoice amount as $100,000 instead of $1,000, leading to a 100x overpayment.” If an LLM-based chatbot hallucinates a bad cookie recipe, you end up with inedible treats. If it responds to a business query with the wrong information, you might find yourself making very costly mistakes. From relying on smaller models to leveraging grounding methods, Tarik Dzekman offers practical insights for avoiding this fate, all based on his own work in document automation and information extraction.

If you’re looking for reading recommendations on other data science and ML topics, look no further:

• For his TDS debut, Iqbal Rahmadhan presents a detailed and accessible introduction to understanding, calculating, visualizing, and interpreting odds ratios and their confidence intervals.
• What is a semantic layer and how do you go about building an effective one? Marina Tosic’s hands-on primer offers all the answers you’ll need if you’re just getting started.
• If you’re interested in machine learning workflows that involve multilingual approaches, don’t miss Milan Tamang’s step-by-step guide, which shows how to build a Thai multilingual sub-word tokenizer.
• In case speeding up your deep learning model’s training sounds relevant to your current projects, Mengliu Zhao walks us through the inner workings of the automatic mixed-precision technique.
• With the help of statistics and a dash of Python, Maria Mouschoutzi, PhD unpacks the well-known birthday paradox in the latest installment of her engaging “Water Cooler Small Talk” series.
• Anyone who’s new to NumPy and its many powerful features should explore Lee Vaughan’s four-part, beginner-friendly series on this essential Python library—Part 1 is, naturally, a great place to start.

Thank you for supporting the work of our authors! As we mentioned above, we love publishing articles from new authors, so if you’ve recently written an interesting project walkthrough, tutorial, or theoretical reflection on any of our core topics, don’t hesitate to share it with us.

Until the next Variable,

TDS Team

A Close Look at AI Pain Points, and How to (Sometimes) Resolve Them was originally published in Towards Data Science on Medium, where people are continuing the conversation by highlighting and responding to this story.