A new technical essay is challenging a common description of large language models as systems that nobody understands. Its central argument is that engineers know the mechanisms used to build, train and run these models, even though they cannot yet translate every internal computation into a simple explanation for every output.
The author describes the standard transformer pipeline: text is converted into tokens, represented numerically and processed through layers containing operations such as matrix multiplication, attention, normalization and nonlinear transformations. The model ultimately produces a probability distribution over possible next tokens, and a decoding procedure selects from that distribution. During training, a loss function measures prediction error and optimization adjusts parameters. Ordinary use of a deployed model does not, by itself, retrain those parameters.
That mechanical account matters because public discussion often collapses two different questions. One asks whether researchers understand the architecture, training process and numerical operations. The other asks whether they can provide a concise causal explanation for every learned behavior inside a model with vast numbers of parameters. The essay argues that uncertainty about the second question should not erase extensive knowledge about the first.
Researchers with model access can record layer activations, inspect attention heads, examine weights, remove or replace components and observe changes. They can also repeat training with altered data, random seeds, architectures or optimization settings to test which behaviors remain stable. Those experiments do not make interpretation effortless, but they turn the model into an object that can be measured and manipulated rather than an unknowable black box.
The essay points to active research on transformer circuits and circuit tracing, which seeks to identify internal pathways associated with particular behavior. That work exists precisely because understanding has degrees: engineers can know enough to design better architectures and diagnose limitations while still lacking a complete human-readable account of how a capability is represented.
The author's position is an argument, not a newly reported scientific consensus or experimental result. It also does not resolve concerns about opaque decision-making in deployed systems, especially when developers do not disclose training data, model weights or implementation details. Access determines which internal measurements are possible, and reproducibility is more limited for closed commercial models.
Still, the distinction improves the debate. Saying that no one understands language models can suggest that their outputs arise through magic or entirely inaccessible processes. A more precise description is that the computational recipe is well understood, while complete interpretability of learned representations and specific outputs remains an open research challenge. That framing preserves both the substantial engineering knowledge behind modern AI and the genuine limits researchers continue to investigate.
