LLMs in one paragraph
A large language model is a transformer-architecture neural network trained on hundreds of billions to trillions of words of text. The training objective is, in simplest form, next-token prediction: given the previous tokens, predict the next one. Despite this simple objective, scale and good training methodology produce models that can answer questions, follow instructions, write code, summarise documents, translate languages, and reason about novel problems. Modern LLMs are not memorising training data; they are general-purpose pattern recognisers over text.
How LLMs differ from earlier AI
Three structural differences from prior-generation AI:
- Scale — modern LLMs have hundreds of billions of parameters, trained on internet-scale corpora. Prior models were orders of magnitude smaller.
- Generality — a single LLM handles classification, generation, translation, summarisation, code, reasoning. Prior systems were task-specific.
- Instruction following — LLMs trained with instruction-tuning and reinforcement learning from human feedback (RLHF) follow natural-language instructions reliably. Prior systems required specific input formats.
The combination is what makes LLMs broadly useful. A single model embedded into your application can support many different features without retraining.
Major LLM families
Production-relevant models in 2026:
- Anthropic Claude (Haiku, Sonnet, Opus) — long context, strong reasoning, enterprise-friendly contracts
- OpenAI GPT-4 / GPT-5 family — broadly capable, large ecosystem, multimodal
- Google Gemini — multimodal, integrated into Google Cloud, strong on certain reasoning benchmarks
- Meta Llama — open-weight, deployable in your own infrastructure, strong cost characteristics for high-volume workloads
- Mistral / Mixtral — European-headquartered, mix of open and proprietary, growing enterprise adoption
Most enterprise deployments standardise on one provider as primary and use one or two as backup or for specific use cases. NINtec's practice is Claude-centred but works with all of these depending on the workload.
What LLMs cannot do reliably
Honest caveats:
- Math beyond moderate complexity — LLMs make arithmetic mistakes; production systems route math to deterministic tools
- Up-to-date knowledge — LLMs only know what was in their training data; RAG provides current knowledge
- Truthfulness without grounding — LLMs hallucinate plausible-sounding but incorrect content; production deployments use grounding and citation discipline
- Long-horizon planning without tools — LLMs struggle with multi-day or multi-step plans without external memory or scaffolding
The right deployment shape works around these limitations rather than ignoring them.