35 articles
Browse guides
In a GAN, a generator tries to produce convincing samples while a discriminator tries to catch them, and both improve through the contest. GANs dominated image generation before diffusion and remain useful where speed matters, though they are notoriously unstable to train. In practice: The ’this person does not exist’ faces were GAN output.
Generative AI learns the shape of its training data well enough to sample new examples from it. The output is novel in the sense that it did not exist before, and derivative in the sense that it comes from patterns in the data. This is the category that took AI from a back-office tool to something everyone uses directly. In practice: Asking a model for a first draft of a launch email and getting one in four seconds. ...
A hallucination is content the model generates because it is statistically plausible, not because it is true. It is a structural consequence of next-token prediction rather than a bug to be patched: the model optimises for plausible, and plausible and true usually coincide, right up until they do not. Grounding, retrieval, and verification reduce it; nothing eliminates it. In practice: Invented case citations, invented DOIs, invented API endpoints — all classic, all confidently formatted. ...
A model knows nothing about the world after its cutoff unless it is given search or documents at inference time. This is why models confidently describe superseded versions of software and miss recent events. Retrieval and web tools are the fix; the cutoff itself never moves without retraining. In practice: Asking a model for ’the latest’ anything is unreliable unless it can search.
An LLM is a transformer trained at scale on text, with a deceptively simple objective: guess what comes next. Do that well enough across trillions of tokens and behaviours like summarising, translating, and following instructions emerge without being taught explicitly. Post-training then shapes the raw model into something that follows instructions safely. In practice: GPT-5.6, Claude, Gemini and Llama are all LLMs, differing in training, size, and tuning.
LoRA freezes the base weights and learns a small number of extra parameters alongside them. The result is a lightweight adapter file — megabytes, not gigabytes — that can be swapped per task and trained on modest hardware. It is why fine-tuning stopped being the exclusive privilege of large labs. In practice: One base model plus five LoRA adapters serving five brand voices.
A mixture-of-experts model contains many specialised sub-networks and a router that picks a few per token. Total parameter count stays huge while compute per token stays modest, which decouples capacity from inference cost. It is why some very large models are surprisingly cheap to serve. In practice: A 400B-parameter model that only runs 40B worth of compute per token.
Multimodal models map different data types into a shared representation, so you can hand one a screenshot and a question and get a text answer. The practical upshot is that whiteboards, PDFs, and voice all become valid inputs, which removes the transcription step from a lot of workflows. In practice: Photographing a broken chart in a slide deck and asking what is wrong with it.
Open weights means the parameters are published, so you can self-host, fine-tune, and inspect. It is not the same as open source: training data and code are usually withheld, and licences often restrict commercial use or scale. Read the licence before assuming freedom. In practice: Downloading weights to run a model entirely inside your own network.
Quantization shrinks each weight from, say, 16 bits to 8 or 4. The model gets dramatically smaller and faster with usually modest degradation, which is what allows capable models to run on a laptop or a phone. How much quality you lose depends on the method and how aggressive you get. In practice: A 70B model compressed to run on a single consumer GPU.