Part 1 in the Demystifying Recent AI Advances Series

In an earlier series on multimodal transformers, I introduced zero-shot transformer models and highlighted two standout examples: GLiNER for named entity recognition and SAM for image segmentation. A follow-up post made the case for Small Language Models as a practical middle ground between traditional NLP and full-scale LLMs and ended with a prediction: the future of enterprise AI would be built on top of smaller, locally deployed models. That prediction seems to be coming to fruition faster than I expected.

The goal of this post, and of the four that follow it, is to cut through the noise. Rather than one sprawling field guide, I have broken the material into a series of shorter, focused installments meant to familiarize a practitioner with emerging technologies and techniques: the model zoo (this post), the autonomy spectrum and the plumbing that supports it, the retrieval architectures used to ground models in enterprise data, and the shifting economics of inference. And lastly, to take a look at how SAS is putting all of it into practice. Taken together, they are meant to serve as a field guide for analytics practitioners navigating this rapidly evolving terrain.

In just a few months, the AI landscape has undergone what can only be described as a technological and terminological explosion. New standards, new architectural patterns, new product categories, and an avalanche of acronyms have emerged at a pace that makes even seasoned practitioners feel like they need a glossary just to follow the conversation. As Andrej Karpathy quipped in 2023, "the hottest new programming language is English." Three years later, he was only half right. English is still part of the equation, but the real engineering has moved far beyond crafting the perfect prompt.

Different Types of Models

We begin where most confusion starts: with the models themselves. One of the first sources of confusion in the current AI landscape is the sheer variety of model types. You will hear people throw around terms like LLM, SLM, VLM, reasoning model, and diffusion model almost interchangeably, but they refer to meaningfully different tools with distinct strengths. See Table 1 for an overall summary.

Large Language Models (LLMs) are the headline-grabbers. These are text-centric models with parameter counts ranging from tens of billions to trillions. GPT-5, Claude Opus, and Gemini Pro all fall into this category. They excel at general-purpose reasoning, open-ended generation, and complex multi-step tasks. They are also expensive to run, typically costing between $2 and $30 per million tokens, and they generally require significant compute infrastructure. Think of LLMs as the senior architects of the AI world: brilliant generalists who can tackle almost anything, but whose time is expensive and best reserved for tasks that genuinely require their breadth.

Small Language Models (SLMs) range from a few million to roughly 10 billion parameters. Key examples include Microsoft's Phi-4 Mini (3.8B parameters), Meta's LLaMA 3.2 (1B and 3B variants), NVIDIA's Nemotron Nano, and the GLiNER family I discussed in the previous series. SLM inference costs are 5 to 20 times lower than LLMs. A June 2025 NVIDIA position paper, "Small Language Models are the Future of Agentic AI," argues that most agentic work (parsing commands, calling tools, producing structured output, summarizing, routing) is repetitive and narrow, making SLMs the natural fit. They can be fine-tuned to strict output formats, run on local consumer-grade hardware, and deliver real-time latency without heavy parallelization. Think of them as a junior employee or intern who can consistently accomplish a well-defined task.

SLMs also play an increasingly critical role in AI safety and governance. Because they are fast and cheap to run, they are ideal for implementing computationally efficient guardrails within a larger agentic process: scanning inputs for prompt injection or jailbreak attempts, detecting toxic or harmful content in model outputs, and enforcing content policy compliance, all in real time and without the latency or cost penalty of routing every safety check through a full-scale LLM. In an agentic workflow where a single user request might trigger dozens of model calls, having lightweight SLM sentinels standing guard at each step is far more practical than asking the expensive orchestrator to police itself. If LLMs are the architects, SLMs are the skilled tradespeople: fast, efficient, and excellent at executing well-defined tasks, including keeping the rest of the system honest.

Vision Language Models (VLMs) fuse a vision encoder with a language decoder, enabling the model to process images and text jointly. They span sizes from compact (Qwen2.5-VL at 7B parameters) to frontier scale. The key distinction is multimodal input: whereas LLMs and SLMs can accept only text as input, a VLM can accept both text and images. This enables them to look at scanned PDFs, photographs, charts, or diagrams and understand what they see in context. This makes VLMs invaluable for document understanding, visual question answering, and any task where layout or visual structure of information matters as much as the words themselves.

Reasoning Models are a more recent designation given to frontier LLMs. Standard LLMs generate responses token-by-token in a single forward pass, essentially "thinking out loud" as they write. Reasoning models, by contrast, are trained to perform an explicit chain of thought before producing a final answer. They allocate additional compute at inference time, working through a problem step by step internally before committing to a response.

It’s akin to a model performing “Type II thinking” from Daniel Kahneman’s book, “Thinking Fast and Slow”. Most frontier LLM models now include a thinking parameter that can be toggled to influence the amount of effort (and tokens) that should be spent in “thinking” mode. The trade-off is straightforward: reasoning models are slower and more expensive per query (that extra thinking consumes tokens), but they dramatically outperform standard models on tasks requiring multi-step logic, mathematical proof, code debugging, and complex analytical reasoning. For straightforward generation tasks like summarization or translation, a standard LLM is faster and cheaper. For complex tasks where getting the answer right matters more than getting it fast, reasoning models are increasingly the tool of choice.

Diffusion Models operate on an entirely different principle from the language models above. Rather than predicting the next token in a sequence, diffusion models learn to generate data by gradually reversing a noise-addition process. During training, the model learns how to take pure noise and iteratively refine it into coherent output. This approach has proven spectacularly effective for image generation (Stable Diffusion, DALL-E 3, Midjourney), video generation (Sora, Runway), audio synthesis, and even molecular design.

More recently, diffusion-based architectures have begun appearing in code generation and structured data synthesis, blurring the line between generative media and traditional AI tasks. For enterprise applications, diffusion models are most relevant in scenarios involving synthetic data generation, visual content creation, and augmenting training data sets where real-world data is scarce or sensitive.

Table 1: Different Language Model Variants

The key takeaway is that these are not competing alternatives. They are complementary tools designed for different jobs, and the most effective AI systems use them in thoughtfully designed combinations. An agentic workflow might use a reasoning model to plan a complex task, delegate extraction work to an SLM, hand visual document processing to a VLM, and generate synthetic training data with a diffusion model, all orchestrated by a reasoning LLM acting as the coordinator.

Generative vs. Discriminative: A Foundational Distinction

Underneath all the terminology above is a more fundamental split that predates the current wave of generative AI entirely: the difference between discriminative (deterministic) models and generative models. Traditional predictive analytics — credit scoring, fraud detection, churn prediction — is built almost exclusively on discriminative models. Given a set of inputs, a discriminative model learns the boundary or mapping to a specific output: a class label, a probability, a risk score. Feed it the same inputs twice, and it returns the same answer every time. That determinism is precisely why these models have been the backbone of regulated industries for decades: their behavior is reproducible, auditable, and explainable through reason codes, variable importance analysis, or heuristics.

Generative models — including every model type discussed above work differently. Rather than learning a fixed mapping from input to output, they learn the data's underlying distribution and then sample from that distribution to produce new content. Feed a generative model the same inputs and, even with the temperature set to zero, you will get different (though hopefully equally valid) answers. That is a feature, not a bug, when the goal is creative or open-ended output, but it is also exactly why generative AI introduces governance challenges that classical discriminative models never had to contend with: hallucination, output variability, and prompt-level attack surfaces.

Hallucination means a model can state a plausible-sounding falsehood with complete confidence; output variability means the same input can yield inconsistent decisions over time, which complicates audit trails and reproducibility; and prompt-level attack surfaces mean a carefully crafted input can manipulate the model’s behavior in ways a static scoring algorithm was never exposed to. Each is a new class of risk that governance frameworks built for discriminative models were never designed to catch.

Both are AI models, but the distinction has important implications for use case suitability. See Table 2 for a summary of these subtleties.

Table 2: Discriminative vs. Generative Models

This is not an either/or choice. The most mature AI programs pair the two. A discriminative model handles the deterministic decision that must be defensible in an audit (approve or decline, flag or clear), while a generative model handles the surrounding tasks that benefit from flexibility (summarizing the case for a reviewer, drafting an explanation letter, or extracting unstructured details that feed the discriminative model’s inputs).
The ideal setup for an enterprise use case might involve using deterministic analytic models upstream.  Then feeding the model artifacts to a downstream generative model. This can assist with understanding the outputs and providing human reviewers with guidance on how to interpret and act on the results.

Ultimately, knowing which category a given model belongs to is the first step in knowing what kind of governance, monitoring, and validation are required to convert a prototype into a vetted, production-ready solution.

Coming Up Next

Knowing which kinds of models exist is only half of the picture. Ultimately, choosing well also means understanding the systems we build around them. The next post in this series moves up a level of abstraction to look at the autonomy spectrum, the meaningful differences between a chatbot, a copilot, and an agent, and the plumbing that makes agentic behavior possible: Skills, the Model Context Protocol, and the harness that turns a capable model into a reliable system.

Blog Series:

• Part 2: From Chatbots to Agents: The Autonomy Spectrum and the Plumbing Beneath It