Architecture / Learning
Learning Without Transformers: Possibilities, Evidence and Limits
Investigating AI without Transformers: Elena AILL’s tests of words, memory and component roles, from F235 to neural-core replays.
AI research does not have to follow a single architectural approach. Elena AILL explores an alternative to Transformers, investigating how memory, language and other components can work together.
Understanding that choice starts with the experiments: how information is retained, which components contribute to answers and what happens when one of them is removed.
Looking inside a different approach
Public checks of Elena AILL’s architecture found no Transformer modules in the path used to generate answers. The project also includes experimental PyTorch components and numerical representations of information, but their presence does not mean they all take part in a conversation.
The distinction concerns the path that was examined. It does not mean the entire project has no neural networks or no form of attention. Assessing it requires following what each part does, rather than inferring capabilities from an architectural label.
Twelve requests, one defined test
In F235, the Elena AILL candidate passed the specified checks on twelve requests. One check looked for meaningless words; none were found in the test. The phase verdict is PASS.
The result concerns the F235 suite and the candidate examined. It is not a general assessment of conversational quality or evidence of superiority over Transformers. The broader question of learning also requires observing what changes after an update and what happens to earlier knowledge.
Neural core · Unchanged output in the ablation
111/111 turns and 500/500 cases retained equivalent output in two documented replays after the neural core was excluded from canonical eager loading. These are separate equivalence checks, not accuracy or learning scores.
Source: Architettura e confine del runtime / Architecture and runtime boundary ↗What happens when a part is removed?
An ablation removes a component to see whether behaviour changes. In these two tests, excluding the neural core from loading did not change the observed answers.
The results measure equivalence, not correctness: the same error can appear in both conditions. The finding concerns those replays and that loading path. It does not establish that every neural component is unnecessary or that the core could not serve a different role.
Studying the connections
Separating memory, meaning, planning and language makes an error easier to follow. If a fact survives a restart but disappears from the answer, retention is not the same problem as its later use. If a sentence is readable but irrelevant, the investigation needs to return to the request.
Associative memories, local updates and transition models have an existing research history. The questions in Elena concern their combination, operating boundaries and observed effects. Experiments make it possible to test the effects of different design choices.
Which version was actually checked?
The experimental v19 version was checked against its source files and saved state, but was not adopted as the canonical version. The checks also found that some examined files did not match the fingerprints authorised for the canonical version.
The distinction matters: findings from the candidate do not certify a fully aligned live canonical runtime. Experimental architecture, general understanding, conversational quality and operational availability each require verification in their own context.
The useful question comes after the label
Elena AILL follows an architectural approach different from Transformers. Its experiments test the roles of individual components and identify changes that have observable effects on the system.
Those findings provide a way to assess the research, beyond the architecture’s name.
