SCAIO Learn · Primer 01

What is an LLM?

A 10-minute, plainspoken primer on the technology behind the AI conversation in South Carolina.

SCAIO · scaio.org

The simplest definition

A program that learned to predict the next word in a sentence — and got very good at it.

An LLM — Large Language Model — is a statistical pattern-matcher trained on enormous amounts of text. Its core capability is deceptively narrow: given a sequence of words, predict what word comes next.

What makes modern LLMs remarkable is that doing this prediction well, at scale, turns out to require something that looks a lot like reasoning about language, facts, and tasks.

Why this matters now

Three things changed at roughly the same time.

Scale. Models trained on trillions of words, run on tens of thousands of chips, produced qualitatively different behavior than smaller models.
Cost collapse. The price of running a query against a capable model has fallen by 90 percent or more in two years — and continues to fall.
Integration. Capable models are now embedded in the productivity tools nearly every white-collar worker already uses — email, search, document drafting, customer service.

What modern LLMs can actually do

Most of the value sits in four buckets.

01 · Drafting

First drafts of language

Emails, memos, policy summaries, public communications, agency notices. The first draft, not the final one.

02 · Synthesis

Summarizing complex material

Long documents, multi-source briefings, regulatory text, public testimony. Compression with reasonable accuracy.

03 · Translation

Across languages and registers

Spanish-English, plain language from technical, statute-to-summary, executive briefing from raw notes.

04 · Code

Software, queries, automation

Increasingly capable at writing, reviewing, and debugging code in most programming languages.

What they cannot reliably do — yet

Honest limits matter as much as honest capabilities.

Cite their sources. Without specific tooling, LLMs will produce confident-sounding text without verifiable references — and sometimes invent them.
Know what they don't know. Most models will answer in the same tone whether they are sure or guessing. Calibration is improving but imperfect.
Operate without supervision on stakes-bearing decisions. Legal advice, medical diagnosis, eligibility determinations — all still require human judgment.
Remain current. Models have a training cutoff; without web-search tools, they don't know what happened after that date.

Vocabulary for the policy conversation

Five terms worth knowing.

Hallucination

When a model invents

A confidently stated fact that isn't true. Less common in newer models but never zero.

Prompt

What you give the model

The instruction or context the user provides. The same model produces very different output with different prompts.

Fine-tuning

Specializing a model

Additional training on a narrower dataset to improve performance on a specific task or domain.

RAG

Retrieval-Augmented Generation

Letting the model look things up in a curated source before answering. Reduces hallucination on factual questions.

Where this shows up in South Carolina

Concrete deployments already in production in SC.

MUSC's deployment of DAX Copilot for ambient clinical documentation across the health system.
The state's "Bradley" AI assistant in the Department of Administration / Center of Excellence pilot.
BMW Spartanburg's pilot with Figure AI — humanoid robotics that incorporate LLMs in their perception and planning stack.
Mercedes-Benz Vans and Volvo use computer-vision quality systems alongside LLM-augmented engineering workflows.

An LLM is not a thinking machine. It is a pattern-matching machine that has gotten so good at the patterns of language that the boundary between pattern-matching and thinking matters less than it used to."

A working definition for the policy conversation

What this means for SC policy

Three questions worth asking about any AI deployment.

What does the model do well? Drafting, synthesis, translation, code — yes. Final decisions on people's lives — not without supervision.
How do we know it's working? Evaluation is most of the AI deployment job. Without it, deployment is just hoping.
Who's accountable when it's wrong? The accountability question doesn't disappear because the decision was AI-assisted. It usually concentrates.

~90%

Drop in the cost of running a query against a capable LLM over the past 24 months. The economics of AI deployment are moving, not stable.

A practical posture

For South Carolinians making decisions about AI:

Treat it as software, not magic. Same questions of procurement, evaluation, security, accountability, and oversight apply.

Measure the present, don't predict the future. What does this tool actually do today, in your context, against your evaluation criteria? Forecasts of what AI will do in five years are mostly speculation.

Pilot before you scale. Most deployments fail not because the technology doesn't work but because the integration with existing process and people wasn't designed for.

Where SCAIO comes in

Public-interest AI research for South Carolina.

SCAIO is an independent observatory tracking AI's impact on South Carolina — its economy, workforce, institutions, and citizens. Browse more primers, the flagship report, and the SCAIO Journal at scaio.org.

scaio.org · jimmy@scaio.org