My studies at Villanova occurred from 2008 to 2009, and Tom Davenport recently published his book titled "Competing on Analytics." For those unfamiliar with the book, it examines the use of data to build high-performing companies, developing competitive strategies that focus on sophisticated data analysis rather than merely depending on analytics for operational support or forming strategies based on intuition.

So, when Tom, Steve, and others began raising the likelihood of job disruption and/or displacement caused by AI, I paid close attention and spent a considerable amount of time reflecting on what I had observed over the prior 12 months.

In his post, "AI and the Entry-Level Problem," Professor Davenport’s argument was not about AI replacing experts (although there are some questions there as well)—it's whether it will eliminate the path to becoming one. This mirrors many of the conversations I have had at AI-focused events and workshops in the last year. My working theory entering the year is that companies will use AI as a means to accelerate technology development. However, conversations ultimately shifted to cost-cutting. While managers sought to procure AI tools to augment their existing workforce, corporate Chief Financial Officers (CFOs) focused on how much they could reduce current-year costs, which is another way of saying “reduce headcount.” This naturally extended to pauses in new hiring activity, as many managers preferred to retain their current staff rather than hire someone new with whom they were unfamiliar. This resulted in some uncomfortable conversations while I was at HumanX, as graduating college students participated in the discussion.

This unlocked a new perspective on what we are witnessing. Often, companies seek to reduce costs by laying off more experienced and typically more expensive employees while hiring younger, less costly ones. However, as artificial intelligence rapidly advances, we seem to be witnessing something different: the systematic erosion of entry-level positions across industries. This isn't just another automation story. It's perhaps a fundamental restructuring of how careers begin, how expertise develops, and how organizations build their talent pipelines.

The Diamond Emerges

Traditional corporate hierarchies resembled pyramids, with broad bases of entry-level workers supporting narrower layers of middle management and executives. However, AI may be reshaping this structure into something more akin to a diamond: far fewer entry-level positions, a more substantial middle of specialized experts, and the same small number of senior leaders at the top.

This shift is already underway. LinkedIn data reveals a troubling trend in entry-level hiring, particularly acute in software development, where AI-powered code generation has dramatically reduced demand for junior programmers. However, the impact extends far beyond tech. Financial services companies are restructuring their entire organizational models, executives report, as AI capabilities render many traditional entry-level tasks obsolete.

The implications are staggering. If companies aren't hiring inexperienced workers, where will tomorrow's experienced professionals come from?

Consider software development: junior developers typically start by fixing bugs, writing simple functions, and maintaining existing code. These tasks, once essential stepping stones to expertise, are increasingly handled by AI tools that can generate, debug, and optimize code faster than human beginners. However, without this foundational experience, how does someone develop the deep understanding necessary to design complex systems or make strategic technical decisions? Without deliberate action, we risk watching an entire “middle” of the talent pipeline collapse.

“But AI can’t do that!”

One common counterargument to this assertion of a trend is that AI cannot yet perform specific tasks. I am being purposefully vague, as the argument could genuinely be anything. But I would like you to reflect on where AI was 12 to 18 months ago. The early models were entertaining but had noticeable gaps if you wanted to use them for any serious work. Hallucinations and images of people with six fingers on their hands can be amusing fodder for memes. Still, they are not necessarily artifacts you would want to rely on for serious business.

Yet, if we reflect on progress made over this time, we are starting to see fewer hallucinations, and images are becoming so realistic that now people are worried about second-order effects if they are used as gasoline for misinformation. Some tools are introducing footnotes that allow users to validate text generated from the models. To assume that technology would remain static would be to ignore history.

The more recent news that people are grasping is that current AI is only a reasoning model and not Artificial General Intelligence (AGI). This misses two critical points.

First, these systems were never AGI to begin with. Large language models process vast datasets, encode patterns into high-dimensional representations, and generate responses by drawing on these learned associations. While sophisticated, this approach differs fundamentally from human-like general intelligence—a fact that researchers have long understood.

Second, and more importantly, the AGI threshold is largely irrelevant to the current disruption. Today's AI systems are already transforming industries, automating complex workflows, and augmenting human capabilities in ways that seemed impossible just years ago. The question isn't whether they constitute AGI, but how rapidly they're eliminating the practical barriers that once seemed insurmountable.

The Skills That Matter

While I think it is flawed thinking, I can understand the emotions underpinning some of the dismissiveness. The combination of fundamental change at a rapid pace can be scary. In the face of this anxiety, it is perfectly logical to search for reasons why this change will not occur. Perhaps those of us who believe that AI is going to cause massive disruption to the status quo are wrong, but I personally have seen enough to convince me that this change is real.

So, if we acknowledge the reality of this change, the next question is: “What do we do to make sure people are not left behind?” I don’t think this is necessarily a death sentence for new graduates and career changers. The key is understanding which capabilities remain uniquely human and are likely to stay that way.

Deep Domain Knowledge Becomes Premium: As AI handles routine tasks, the value of subject matter expertise skyrockets. Someone who understands both AI capabilities and the intricacies of supply chain management, healthcare regulations, or customer psychology becomes invaluable. The combination of domain knowledge and AI literacy creates a competitive moat that pure technical skills alone cannot provide.

Critical Thinking and Information Processing: Entry-level workers once spent a significant amount of time gathering and organizing information. Now, they need skills in evaluating, questioning, and improving AI-generated outputs. Can you spot when an AI model's assumptions are flawed? Can you identify missing context in a generated report? Can you ask better questions that lead to more useful AI responses?

Human-AI Collaboration Fluency: This isn't just about knowing how to write prompts. It's about understanding when to trust AI, when to doubt it, and how to combine AI capabilities with human judgment. It's knowing which tasks benefit from AI assistance and which require purely human insight.

The Path Forward

For those entering the workforce, the strategy is clear: become irreplaceable not by competing with AI but by complementing it. Additionally, commit to continuous learning and remain agile in your ability to adapt to a more dynamic environment. Develop deep expertise in a specific domain and master the art of human-AI collaboration. Cultivate critical thinking skills that allow you to improve on AI-generated work rather than just accept it.

The diamond-shaped organization isn't necessarily a dystopia—it could represent a more efficient allocation of human talent, with people focused on higher-value work from the start of their careers. But only if we successfully navigate the transition.

What do you think? Are we witnessing the end of traditional career ladders, or the evolution toward something better? How should educational institutions and employers adapt to prepare workers for this new reality?

But here's the thing—understanding these technologies isn't just for engineers anymore. Whether you're a business leader deciding on AI tools, a content creator exploring automation, or simply someone curious about the technology reshaping our world, knowing the differences between Artificial Intelligence (AI), Large Language Models (LLMs), Small Language Models (SLMs), and Retrieval-Augmented Generation (RAG) will help you navigate conversations, make informed decisions, and spot opportunities.

Think of this guide as your translator for the AI alphabet soup.

The Foundation: What is AI?

Artificial Intelligence is the broadest term in our lineup—it's like saying "vehicle" when you could be talking about anything from a bicycle to a spaceship. AI refers to any system that can perform tasks typically requiring human intelligence, such as recognizing images, understanding speech, making decisions, or solving problems.

Real-world examples include the Netflix recommendation system, which suggests what to watch next, the spam filter in Gmail, and using Waze to find the fastest route home, all of which are forms of AI. They're narrow AI systems, each designed for specific tasks.

The analogy: If AI were a kitchen, it would include everything from a simple can opener to a full robotic chef. It's the entire category of tools designed to simplify cognitive tasks.

The Powerhouses: Large Language Models (LLMs)

Large Language Models are the celebrities of the AI world right now. These are AI systems specifically designed to understand and generate human language, trained on massive amounts of text data—billions or even trillions of words from books, articles, websites, and more.

Real-world examples: ChatGPT, Claude, and Google's Gemini are all large language models (LLMs). We can write essays, answer questions, debug code, compose emails, and even engage in creative writing.

The analogy: Think of an LLM as an incredibly well-read librarian who has speed-read most of human knowledge and can instantly recall and synthesize information on almost any topic. They're powerful generalists but require significant computational resources and energy to build and maintain.

💡 Key characteristics:

• Massive parameter counts (billions to trillions)

• Trained on diverse, large-scale datasets

• Excellent general-purpose language understanding

• High computational requirements

• Can perform many language tasks without specific training

The Specialists: Small Language Models (SLMs)

Small Language Models are the efficient cousins of LLMs. They're designed to be lightweight, fast, and focused, often trained for specific tasks or domains rather than trying to be experts in everything.

Real-world examples: Mistral Nemo, SLM variants of Qwen2 and Llama 3 (e.g., Qwen2-0.5B, Qwen2-1.5B, Llama 3-8B), and specialized models for tasks like code completion in IDEs, grammar checking in writing tools, or customer service chatbots for specific companies.

Note: While 12 billion parameters is at the high end of what some consider an SLM, Mistral Nemo is generally positioned as a compact, efficient model for enterprise use, so including it is reasonable. However, it is worth noting that some sources may classify it as a mid-sized model.

The analogy: If LLMs are like university professors with encyclopedic knowledge, SLMs are like skilled specialists—a medical transcriptionist who's incredibly good at their specific job, works quickly, and doesn't need a supercomputer to function.

💡 Key characteristics:

• Smaller parameter counts (typically under 10 billion parameters)

• Faster inference and lower computational requirements

• Often specialized for specific tasks or domains

• Can run on smaller devices (phones, laptops)

• More cost-effective for targeted applications

The Game-Changer: Retrieval-Augmented Generation (RAG)

RAG isn't a model itself—it's a technique that supercharges language models by connecting them to external knowledge sources. Instead of relying solely on training data, RAG systems can retrieve relevant information from databases, documents, or the internet in real-time and use that information to generate more accurate and up-to-date responses.

Real-world examples: Customer service bots that can access your company's latest policy documents, legal AI assistants that reference current case law, or technical support systems that pull from constantly updated manuals and troubleshooting guides.

The analogy: Imagine a brilliant assistant who not only has excellent general knowledge but can also instantly look up specific information in your company's files, recent news, or specialized databases before giving you an answer. RAG turns any language model into a research assistant with access to current, specific information.

💡 Key characteristics:

• Combines language generation with information retrieval

• Provides access to current, specific, or private information

• Reduces hallucinations by grounding responses in retrieved data

• Can be implemented with both LLMs and SLMs

• Enables models to work with information beyond their training data

The Comparison: At a Glance

Strengths and Sweet Spots

AI (Broad Category)

💪 Strengths: Incredible diversity of applications, from image recognition to game playing to robotics.

⚠️ Limitations: "AI" is too broad to have specific limitations—it depends entirely on the implementation.

🎯 Best for: This is the umbrella term, so it's suitable for everything AI-related.

Large Language Models (LLMs)

💪 Strengths:

• Exceptional versatility across language tasks

• Strong reasoning and creative capabilities

• Can handle complex, multi-step problems

• Excellent few-shot learning (learning from just a few examples)

⚠️ Limitations:

• Expensive to run and maintain

• Can hallucinate (make up plausible-sounding but incorrect information)

• Knowledge cutoff means they don't know recent events

• Overkill for simple, specific tasks

🎯 Best for:

• Complex writing and editing tasks

• Creative projects requiring a nuanced understanding

• Multi-domain question answering

• Applications where versatility matters more than efficiency

Small Language Models (SLMs)

💪 Strengths:

• Fast and efficient

• Cost-effective to deploy and run

• Can run on edge devices (phones, tablets)

• Often more focused and reliable for specific tasks

• Easier to fine-tune for specialized applications

⚠️ Limitations:

• Limited general knowledge compared to LLMs

• May struggle with complex reasoning

• Less creative and flexible

• Narrower range of capabilities

🎯 Best for:

• Mobile applications

• Real-time systems that require quick responses

• Cost-sensitive applications

• Specialized tasks like grammar checking, simple Q&A, or code completion

• On-device AI, where privacy is crucial

Retrieval-Augmented Generation (RAG)

💪 Strengths:

• Provides access to current, specific information

• Reduces hallucinations by grounding responses in real data

• Can work with private or proprietary information

• Updates knowledge without retraining the model

• Combines the best of both worlds: language understanding + information access

⚠️ Limitations:

• More complex to implement and maintain

• Retrieval quality directly affects output quality

• Slightly slower than pure language model inference

• Requires maintaining and updating knowledge sources

🎯 Best for:

• Customer support systems

• Document analysis and question-answering

• Research assistants

• Applications requiring current information

• Enterprise systems with private knowledge bases

Your Decision Framework: When to Use What

Choose traditional AI approaches (non-language focused) when your task doesn't primarily involve language—think image recognition, predictive analytics, or control systems.

Go with an LLM when you need:

✅ Maximum flexibility and capability

✅ Complex reasoning or creative tasks

✅ General-purpose language understanding

✅ You have the budget for higher computational costs

Pick an SLM when you need:

✅ Fast, efficient performance

✅ Cost-effective deployment

✅ On-device or real-time applications

✅ A focused, specific language task

Implement RAG when you need:

✅ Access to current or private information

✅ Reduced hallucinations

✅ Domain-specific expertise

✅ The ability to update knowledge without retraining

💡 Pro tip: These aren't mutually exclusive! Many successful AI applications combine multiple approaches. You might use an SLM with RAG for a fast, accurate customer service bot, or an LLM with RAG for a comprehensive research assistant.

The Bottom Line

The AI landscape might seem overwhelming, but understanding these four key concepts—AI as the broad category, LLMs as powerful generalists, SLMs as efficient specialists, and RAG as the technique that keeps models grounded in real data—gives you a solid foundation for navigating this space.

The magic happens when you match the right tool to the right job. Not every nail needs a sledgehammer, and not every language task requires a large language model. Sometimes a small, focused model is perfect. Sometimes you need the power of retrieval-augmented generation to access specific, current information.

As these technologies continue to evolve, the lines between them may blur, and new categories will emerge. But by understanding these fundamentals, you'll be ready to evaluate new developments and make informed decisions about which AI tools can best serve your needs.

The future is AI-augmented, but it doesn't have to be AI-complicated. Now you're equipped to join the conversation and maybe even lead it.

What questions do you have about AI, LLMs, SLMs, or RAG? What specific use cases are you considering? Share your thoughts in the comments below.

BugBot: Your New Code Review Partner

The standout feature of this release is BugBot, an AI agent that automatically reviews your pull requests and catches bugs before they hit production.

Here's what makes BugBot special: when it finds an issue in your GitHub PR, it doesn't just leave a generic comment. It creates actionable feedback with a "Fix in Cursor" button that takes you directly back to the editor with a pre-filled prompt to address the specific problem. This integration between code review and fixing help makes for a more seamless coding workflow. One issue I have experienced with some of the AI coding tools is the inability to maintain context as it is making recommendations - resulting in other things breaking within the codebase. It will be interesting to see how Cursor built this feature and whether it addresses the context window issue.

Background Agent Goes Global

Remember when Cursor's Background Agent was locked behind early access? Those days are over. Every Cursor user now has access to the remote coding agent that can work on complex tasks while you focus on other things.

The Background Agent represents a fundamental shift in how we approach development tasks. Instead of manually grinding through repetitive coding work, you can delegate entire features or refactoring tasks to an AI that understands your codebase context. Access it with a simple Cmd/Ctrl+E or by clicking the cloud icon in chat—assuming you don't have privacy mode enabled (though they're working on that too).

Jupyter Notebooks Get the AI Treatment

Data scientists and researchers, this one's for you. Cursor can now implement changes directly in Jupyter Notebooks, creating and editing multiple cells automatically. This isn't just about code completion—it's about having an AI that understands the flow of data science work and can execute complex multi-step analyses.

Currently supported with Sonnet models, this feature transforms Jupyter from an interactive coding environment into a collaborative workspace where AI can be your research partner. As a data scientist, this is exciting to me. However, knowing what I do about software engineers who very much dislike notebooks, I would assume this feature is a lot less exciting for them.

Memories: Context That Persists

One of the most intriguing additions is Memories—Cursor's ability to remember facts from your conversations and reference them in future interactions. This addresses one of the most frustrating limitations of current AI coding assistants: the lack of persistent context, which often leads to redundant code generation or the AI taking completely different development paths than what you've already established.

These memories are stored per project and managed individually, giving you control over what your AI assistant remembers about your coding patterns, preferences, and project-specific details. It's currently in beta, but if implemented well, this could eliminate the annoying cycle of re-explaining project context and architectural decisions in every new chat session. For me, this potentially might be the most exciting new feature. The context window can be a really big headache when vibe coding with AI.

One-Click MCP Integration

The Model Context Protocol (MCP) integration just got a lot simpler. Cursor now offers one-click setup for MCP servers with OAuth support, making it trivial to connect external tools and data sources to your development environment.

This opens up possibilities for integrating everything from internal APIs to external services directly into your coding workflow. Cursor has curated a list of official MCP servers, and developers can now add "Add to Cursor" buttons to their own MCP servers.

The Polish That Matters

Beyond the headline features, Cursor 1.0 includes the kind of polish that makes daily use more pleasant:

• Richer chat responses with Mermaid diagrams and Markdown tables rendered inline

• Redesigned settings and dashboard with detailed usage analytics

• Performance improvements including faster responses with parallel tool calls

• Enhanced @Link and web search that can now parse PDFs

Why This Release Stands Out

Cursor 1.0 represents a meaningful evolution in AI-assisted development tools. The combination of automatic code review, persistent memory, background task execution, and seamless tool integration creates an environment where AI assistance feels more natural and contextual than what we've seen before.

What's particularly impressive is how these features work together. BugBot catches issues, Background Agent implements fixes, Memories retain project context, and MCP integration brings in external data—all within a single, cohesive development environment.

Cursor 1.0 shows promising advances in AI-assisted development, and the comprehensive feature set suggests the space continues to evolve in interesting ways.

Read more of the details in the release notes: https://www.cursor.com/changelog/1-0

Have you tried Cursor 1.0? What features are you most excited about? Hit reply and let me know your thoughts on the updates.

If you’re reading this, you’re probably as fascinated as I am by the world of data, artificial intelligence, and the ways they’re quietly (and sometimes not-so-quietly) reshaping our lives. Whether you’re a seasoned data professional, a curious newcomer, or someone who simply wants to understand the forces shaping our digital age, you’re in the right place.

Why “Signals & Sense”?

In the age of information overload, it’s easy to get lost in the noise. Every day, we’re bombarded with headlines about AI breakthroughs, data-driven decisions, and the next big thing in tech. But what really matters? How do we separate the meaningful “signal” from the background “noise”? And once we find that signal, how do we make sense of it?

That’s what this publication is all about. Here, I’ll explore the stories, applications, and ethical dilemmas at the intersection of data and AI—always with an eye toward clarity, curiosity, and real-world impact.

What to Expect

• Explainers: Demystifying the latest trends in AI and data science.

• Applications: Real-world examples of how data and AI are transforming industries—from healthcare to art to everyday life.

• Ethics & Impact: Thoughtful takes on the societal implications of intelligent technologies.

• Curated Insights: Links, books, and ideas worth your time.

Join the Conversation

This isn’t just a newsletter—it’s a community. I invite you to share your thoughts, ask questions, and suggest topics you’d like to see covered. My goal is to make Signals & Sense a space where curiosity thrives and where we can all learn from each other.

Thank you for joining me at the start of this journey. Here’s to making sense of the noise—together.

Stay curious,
Brian