To understand the limits of LLMs, the large language models behind many AI assistants, we need to look beyond parameter counts and test scores. A wrong answer may depend on the information available, how it is processed or the test chosen to assess it. Radar Special 12A, LLMs — The boundary between scale and intelligence, explores these distinctions to understand what we can improve and which questions remain open.
LLMs — The boundary between scale and intelligence
Special Issue 12A | Wednesday 4 November 2026 | Approximately 7 minutes to read
We ask an assistant to read a document. The first summary overlooks an important clause; we point out the passage and the answer improves. What has changed? Perhaps information was missing, or information already present was not used well. Before speaking of an insurmountable obstacle, it is worth tracing the error and understanding where it begins.
This is the path taken in this issue: from the operations that transform text to research on model growth and the tools with which we try to understand how models work. A result becomes more useful when we know which system it concerns, how it was obtained and how far we can extend it.
1. Between the question and the answer
For us, a sentence carries an intention and a situation. The model, by contrast, starts with units of text called tokens, which may be words, fragments or symbols. These units are associated with numerical representations, then transformed during processing. Meaning does not arrive as a label ready to be read.
In the Transformer, attention is a mathematical mechanism that relates parts of the available information, assigning them different weights. The name evokes a familiar experience, but does not demonstrate that the machine concentrates like a person.
When we add the overlooked passage, we give the system a different context. The new answer can be better without any change to the parameters, the values modified during training. We saw this when following a correction through conversation, memory and training: to know what remains, we need to understand which component has been updated.
💡 Did you know? — A token is not an idea
A word may be divided into several tokens. A numerical coordinate, in turn, is not a shelf reserved for a concept. Tokens, dimensions and internal features describe different aspects: counting them does not automatically tell us how many “ideas” a model has.
2. Much information in the same space
The Italian word “rete” — “network” or “net” — changes meaning depending on whether we are talking about the Internet, football or relationships. A model’s representations also change with context. To study them, researchers sometimes build simplified systems, small enough to allow observations that are difficult in a very large network.
In toy models of superposition, multiple features can share the same numerical dimensions. Sharing allows space to be used efficiently, but can introduce interference. Understanding the consequences requires checking which information the system can distinguish and use in the actual task.
These studies help isolate a mechanism. Their conclusions must, however, be read alongside the conditions of the experiment: a difficulty observed in a simplified model does not, on its own, describe every behaviour of a commercial assistant. It is this attention to conditions that makes research usable outside the laboratory as well.
3. Growth helps, but requires choices
Increasing resources can reduce error, even when each improvement requires more effort than the last. Scaling laws describe these trends under the conditions studied; they do not promise that every capability will grow in the same way.
How we allocate resources also matters. The Chinchilla study showed the importance of balancing model size and training data under a given compute budget. Moreover, the choice that is cost-effective for training must be compared with the work needed to deliver the service afterwards.
Compute and money remain connected through processors, energy, memory, networks and organisation. As Weekly 15 observes in discussing the cost of work actually completed, an inexpensive response may require further attempts and many checks before it becomes usable. Comparing efficiency therefore requires a defined result, with measurable quality and cost.
4. What a test really tells us
Imagine an assistant that summarises contracts. We would want it to recognise an anomaly in the clauses, report its sources and indicate what it cannot conclude. An improvement in average text prediction error does not automatically verify all these capabilities. A benchmark, too (a set of tests) tells us something precise only if we know the questions, criteria and conditions.
The study Lost in the Middle, published in 2024, observed difficulties in the systems examined when using information placed in the middle of long contexts. The result offers a practical question to return to in our checks: does having a document available also mean being able to use every passage well?
The same care is needed when considering a theorem. A mathematical conclusion holds for the construction and assumptions from which it is proved. A theoretical possibility does not guarantee that a product can learn it from data and implement it within sustainable time and cost. Testing the machine requires precisely this connection between what has been observed and what we claim.
🧠 Radar Insight — Before the score, the task
Faced with a “better” result, let us ask what was measured and against which system. An average can leave errors that matter to our work hidden. Recognising them helps us choose where to intervene and which checks to retain.
5. Understanding where to intervene
To improve an assistant, we can also act on the system surrounding the model: have it retrieve relevant documents, organise the context better, devote more processing to an answer or change the training criterion. Each intervention addresses a different problem and must be tested on the task for which we introduced it.
There is also the work of understanding internal mechanisms. In On the Biology of a Large Language Model, the Anthropic team explores some behaviours of Claude 3.5 Haiku through maps and experimental interventions, stating the method’s limits. An explanation generated by the assistant, however convincing, does not replace these observations.
For those using AI, the destination is concrete: what information is needed, which errors must we look for and who can assess the result? The step from a response to usable work is also built this way, by choosing proportionate checks and recognising when to correct or stop. Understanding a limit helps us know where to start again, with which tools and with what responsibility.
Read the full issue
The dossier contains the complete discussion of representations, superposition, scaling laws, mathematical proofs, hallucinations and research directions, with illustrative diagrams and references to primary sources.
© 2026 Barbara Gattobigio · Gruppo Radar
Comprendere oggi. Costruire domani · by effettoantropico.life
All rights reserved on original content. Free access does not authorise reproduction or republication without consent, except as permitted by law. Third-party materials remain the property of their respective owners. To share this issue, use the link to the article.
Cover in focus



















