Artificially Intelligent, Actually Useful.

A Weekly AI Update So You Don’t Have To Wing It

View another: Elsewhere

Sep 28, 2025 - 10:39 UTC

February 18, 2026

The Friction Problem in AI Adoption

March 18, 2026

Artificially Intelligent, Actually Useful.

A Weekly AI Update So You Don't Have To Wing It

The Case for Smaller AI

In this week's newsletter, we're starting to see the results of the last three years of research, experimentation, success and failure. Since ChatGPT had its iPhone moment, the dominant focus was bigger and better. And that worked really well, until it started not to.

The pitch for big AI models was simple: more intelligence, better outcomes. What nobody mentioned was that "more intelligence" came with less control. Data routed through third-party APIs you don't manage. Outputs you can't audit. Decisions you can't explain, especially to a board or regulator.

Gartner puts a hard number on what's coming: 40% of enterprise applications will embed task-specific AI agents by end of 2026, up from less than 5% today. AI is moving out of the chat window and into your workflows.

Small language models (SLMs) aren't a step backward. They're the architecture that makes AI deployable at every layer: enterprise systems, edge hardware, and the AI now embedded across your personal data and devices.

Artificial Intelligence Is No Longer Something You Just Talk To

Gartner estimates that 40% of enterprise applications will include task-specific AI agents by end of 2026, up from less than 5% today. AI is migrating out of the chat interface and into your CRM, your ERP, your approval workflows. Not waiting to be asked. Executing.

Gartner draws a clear line between AI assistants, which depend on human input, and AI agents, which operate independently. That distinction matters because the governance requirements are completely different. An assistant drafts an email for your review; an agent routes a purchase order without a human in the loop.

Key Insight: Gartner flags a term worth knowing: "agentwashing," misusing the word "agent" to describe what are really just AI assistants. If a vendor calls something an agent but it still requires human confirmation at every step, it is not an agent. Real agents execute autonomously within defined guardrails.

Read the Gartner Research →

Why Governability Beats Raw Intelligence

Foundation models aren't failing in enterprise deployments because they lack intelligence. They're failing because organizations can't control them. Red Hat's analysis isolates three architectural advantages driving the move to small language models: data sovereignty, deterministic outputs, and cost predictability. The numbers back it up. Constrained SLMs hit 98% validity rates on structured tasks like JSON generation and SQL queries, compared to 60-70% for general-purpose LLMs on the same work. A model constrained to a schema can't hallucinate.

Strategic Insight: The SLM advantage isn't about parameter count, it's about control. Organizations deploying specialized models gain the ability to tune behavior, audit decision paths, and maintain greater data control.

Read the Red Hat Analysis →

The Economics of Getting This Right

NVIDIA's NeMo framework and Nemotron model family are purpose-built SLMs optimized for the tasks enterprise AI actually run: API orchestration, document parsing, SQL generation, and data classification. These aren't general-purpose models trimmed down, they're domain-specific models trained to do a narrow set of things with high reliability. Metrics show 5-7x cost reductions and 3x latency improvements compared to routing everything through a single large language model.

Practical Angle: The reason this matters isn't just cost. Specialized models give you predictable inference costs, faster response times, and the ability to manage behavior for your specific workflows.

Explore the NVIDIA Framework →

Quick Hits

.Foundations

What Are Small Language Models Actually

IBM's primer on SLMs explains why they aren't just smaller versions of large language models. They're purpose-built for deployment in constrained environments: edge devices, on-premises infrastructure, anywhere you need low latency, data privacy, or offline capability.

.Video

Perplexity's New Platform: AI That Does the Work

This walkthrough shows Perplexity's "Computer" platform which orchestrates 19 AI models simultaneously to complete multistep tasks. You describe the outcome. The system figures out which models handle which parts and delivers finished work, autonomously.

.Deep Dive

NVIDIA Builds Personal Hardware for the Agentic Era

Built for agentic AI workloads, NVIDIA's new Vera CPU prioritizes multi-model orchestration and low-latency agent communication. The fact that it's a CPU, not a GPU, matters: this hardware is designed to run agentic AI at the edge and on the desktop, not in the data center.

Industry Developments

Google Embeds AI Across Your Entire Digital Life

"Personal Intelligence" gives Gemini access to Gmail, Google Photos, Search history, and more, delivering responses tailored to you and your actual context. Google is betting that AI embedded across your data is where the real value lives, not in a chat window.

Adult Content in ChatGPT?

OpenAI's entire well-being advisory board unanimously opposed adult chat mode for ChatGPT. The board flagged serious risks including emotional dependency and a 12% error rate in the age-detection system meant to keep minors out. The feature has been delayed; it has not been cancelled.

Alright, that's a wrap for the week!

Here's my take. A Small Language Model isn't a watered-down version of the AI we've all been talking about. It's a purpose-built tool, trained on a specific domain, designed to run fast, run cheap, and run where you need it: on a server you control, on a device at the edge, or right on the laptop in front of you.

Microsoft, Apple, and Google are all building SLM capabilities into devices and enterprise infrastructure because it's clear that using cloud-based models is slower, more expensive, and a compliance headache. Running the model where your data already lives solves all three problems at once.

For businesses, this might be a game-changer. For techies and tinkerers, it absolutely will be.

Thoughts?

Artificially Intelligent, Actually Useful. is brought to you by iS2 Digital
Your Software at the Speed of Change

Click here to subscribe, if you want to receive our newsletter directly.

Artificially Intelligent, Actually Useful.

LLMs Are Growing Up

AI Isn't What You Think It is

Context is King

The Friction Problem in AI Adoption

Artificially Intelligent, Actually Useful.

Artificial Intelligence Is No Longer Something You Just Talk To

Why Governability Beats Raw Intelligence

The Economics of Getting This Right

Quick Hits

Industry Developments

Got a project in mind?
Tell us about it.

Artificially Intelligent, Actually Useful.

LLMs Are Growing Up

AI Isn't What You Think It is

Context is King

The Friction Problem in AI Adoption

Artificially Intelligent, Actually Useful.

Artificial Intelligence Is No Longer Something You Just Talk To

Why Governability Beats Raw Intelligence

The Economics of Getting This Right

Quick Hits

Industry Developments

Better Business, Not Better Buzzwords

Got a project in mind?Tell us about it.

Got a project in mind?
Tell us about it.