SLMs - A Very Different Form of AI

Source: Dev.to

From Enthusiasm to Realisation

I’ve been using large language models (LLMs) daily since ChatGPT was first released in 2022. The first thing I noticed about my local SLM was its limitations. The novelty began to wear off, and I drifted back to the online LLMs.

But as I tried different models—phi‑3, qwen‑2.5, gemma‑3—I began to enjoy testing and benchmarking them precisely because the limits are hit so easily—something you never encounter with a powerful online LLM.

Key insight: Small language models are not just reduced versions of large ones. They are vastly different because of how the human is required to think in order to use them at all. You have to be very clear and very specific with them.

How Large Models Shape Our Interaction

Large language models have trained me and millions of other users into a particular relationship with AI:

• You can be vague.
• You can improvise and bounce ideas off it.
• You can hint at an idea and let the model infer your intent, fill in gaps, and smooth over contradictions.

That generosity is powerful, and we get used to it. It’s why systems like ChatGPT feel conversational, advisory, even oracle‑like. The model absorbs ambiguity so the user doesn’t have to fully resolve it first.

Small Models Expose Ambiguity

Small language models don’t do that. They can’t afford to.

When you work with an SLM, ambiguity isn’t absorbed—it’s exposed. If you don’t know what you want, the model’s limitations will make that obvious. If your instructions are vague or inconsistent, the output degrades immediately.

• At first this feels like a limitation.
• In reality, it’s a demand: the outcome’s success is now on you.
• You have to decide what you mean before you ask.

That demand changes everything, because it changes you.

Shifting from Conversation to Instrument

Using the SLM effectively requires the user to:

1. Externalise intent.
2. Define boundaries.
3. Specify constraints up front.

You have to stop “playing” with the model and start operating it. I saw this happen to me: the work shifted from the conversational exploration I’d become accustomed to after thousands of hours with LLMs, to something different—the SLM required deliberate articulation from me. The model becomes less of an advisor and more of an instrument.

A Subtle Cognitive Shift

The relatively small number of people seriously using SLMs are undergoing a change that isn’t reflected in adoption statistics. These users are being forced into a different cognitive position:

• They learn to separate thinking from execution.
• They separate discovery from automation.

That skill predates modern AI—and even modern computing. It was gradually eroded by interfaces that grew more forgiving and more opaque. Perhaps this is the real pushback against the concept of “vibe coding”: the user is too passive, too removed from the process.

Small language models reintroduce friction—not as punishment, but as structure.

Trust Through Transparency

Ironically, this makes SLMs more trustworthy in certain roles. They don’t pretend to understand what you haven’t defined. They don’t improvise around weak specifications. Instead, they fail loudly and early, which is exactly what you want from a component in a system.

The Real Revolution

The real revolution of small language models isn’t that they run locally, that they can ensure privacy, or that they’re accessible without specialized hardware—though all of those are significant. If you use SLMs often enough, you discover they offer something else that’s very important: they force the human back into the loop as a thinking component.

• Large models are extraordinary tools for sense‑making and articulation.
• Small models are extraordinary tools for precise execution—but only if the user is willing to do the work first.

Thus, the difference between LLMs and SLMs isn’t about scale. It’s about agency.

Agency: Where Intention Resides

With large language models, a significant portion of agency is quietly transferred from the user to the system. You provide a loose prompt, and the model decides:

• How to interpret it.
• What to emphasize or ignore.
• Often, what the goal actually is.

That isn’t a flaw; it’s a design choice that makes LLMs feel helpful, conversational, and broadly accessible. The cost of that helpfulness is that the model exercises interpretive agency on your behalf.

Over time, users adapt to that. I did. We stop fully specifying our intent because we don’t have to. Maybe we don’t know what we want and want the LLM to tell us. And it will—the system will infer, improvise, and reconcile inconsistencies. Agency becomes shared and, in practice, blurred. The model is not just executing instructions; it is co‑authoring the problem definition.

Small language models simply don’t support that kind of transfer. They leave agency where it started:

• If the task is unclear, it stays unclear.
• If the goal is under‑specified, it isn’t realized.

The model doesn’t step in as an interpretive partner; it waits for direction. The user becomes responsible for defining purpose, scope, and success conditions. Believe me, once you get used to it, you’re glad. It’s like waking up.

Long‑Term Implications

When agency is outsourced—even partially—people begin to trust outputs without fully understanding how the conclusion was reached. This quickly becomes a habit, a way of working.

When agency is held by the user, trust must be earned through clarity and reproducibility. Smaller models encourage the latter because they force you to be explicit.

Bottom Line

• Large language models: generous, conversational, absorb ambiguity, share agency.
• Small language models: demanding, transparent, expose ambiguity, keep agency with the user.

Both have their place, but the true value of SLMs lies in re‑empowering the human as the primary thinking component. By doing so, they foster clearer intent, more reliable execution, and a healthier relationship with AI.

Because they don’t mask weak thinking with their own fluency.
I still use LLMs often — large models are willing to steer with you, or even for you, and sometimes that’s exactly what I want. Small models refuse. I no longer see that refusal as a limitation — it’s a mechanism that forces the human to stay in control.
— Ben Santora, January 2026