Eyla: Toward an Identity-Anchored LLM Architecture with Integrated Biological Priors -- Vision, Implementation Attempt, and Lessons from AI-Assisted Development
A non-programmer spent over $1,000 on AI coding assistants to build a novel 1.27B parameter LLM, but 86 subsystems contributed less than 2% to output.
Researcher Arif Aditto has published a candid failure analysis of 'Eyla,' an ambitious project to build a novel 'identity-anchored' Large Language Model architecture. Unlike standard models optimized for helpfulness, Eyla aimed to maintain a coherent self-model using biologically-inspired subsystems like HiPPO-initialized state-space models and episodic memory retrieval. The paper introduces the Identity Consistency Score (ICS), a new benchmark to measure an AI's ability to resist manipulation and admit uncertainty—a critical property for trustworthy agents.
Aditto, identifying as a non-programmer, attempted to implement this complex architecture entirely using AI coding assistants like Claude Code and Cursor. After spending over $1,000, the result was a dysfunctional 1.27B parameter model where 86 separate 'brain' subsystems contributed less than 2% to the model's final output. The project failed to achieve its core goal of identity consistency, despite the significant investment.
The paper's primary contribution is its honest post-mortem, which identifies five systematic failure modes when using current AI assistants for novel architectural development. These include issues with architectural coherence, subsystem integration, and the tools' inability to reason about entirely new design paradigms. This documented $1,000+ failure provides concrete lessons for both the AI systems community, by highlighting the limits of current agentic AI for R&D, and for the AI-assisted software engineering community, by outlining where current tools fall short for groundbreaking work.
- The project spent over $1,000 using AI assistants (Claude Code, Cursor) but produced a failed 1.27B parameter model where 86 subsystems contributed <2% to output.
- Introduced the 'Identity Consistency Score (ICS),' a new benchmark for evaluating an LLM's ability to maintain a coherent self-model and resist manipulation.
- Documents five systematic failure modes of AI-assisted development for novel architectures, providing a rare, valuable case study of a high-cost failure.
Why It Matters
This honest failure analysis provides crucial lessons for professionals using AI coding tools for R&D, highlighting their current limits for innovative architectural work.