LLMs as an Abstraction Layer in Computer Organization: Implementation and Limitations
Resumo
This paper proposes that Large Language Models (LLMs) can be viewed as a new abstraction layer capable of addressing an intrinsic limitation of traditional presentations of computer organization: the absence of a mechanism that directly connects natural language to the electronic operations performed by hardware. We argue that, although high-level programming languages provide significant abstractions over processor architecture, they do not fully achieve the theoretical objective of translating human commands into electrical signals. We demonstrate how this gap can be addressed through the Model Context Protocol (MCP), which facilitates mediation between commands expressed in natural language and the operating system. Finally, we discuss the problems and vulnerabilities introduced by this new layer, as well as mitigation strategies associated with its use.
Referências
Anthropic (2024). Introducing the model context protocol. [link].
Anthropic (2026). Introducing claude sonnet 4.6. [link].
Chrome Developers (2025). Chrome devtools (MCP) for your AI agent. [link]. Post oficial descrevendo um servidor MCP para integrar capacidades de depuração do Chrome Dev-Tools a agentes.
CursorTouch (2025). Windows MCP: MCP server for computer use in windows. [link]. Servidor MCP para integração de agentes com o sistema operacional Windows, incluindo automação e interação com a UI.
Hangwin (2025). Chrome MCP server (mcp-chrome). [link]. Servidor MCP baseado em extensão do Chrome que expõe funcionalidades do navegador a clientes MCP.
Ji, Z., Lee, N., Frieske, R., Yu, T., Su, D., Xu, Y., Ishii, E., Bang, Y. J., Madotto, A., and Fung, P. (2023). Survey of hallucination in natural language generation. ACM Computing Surveys, 55(12).
Kalai, A. T., Nachum, O., Vempala, S. S., and Zhang, E. (2025). Why language models hallucinate. page 36.
LaurieWired (2024). Ghidramcp. [link]. Repositório GitHub.
Lin, S., Hilton, J., and Evans, O. (2022). Truthfulqa: Measuring how models mimic human falsehoods. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics, pages 3214–3252.
Maynez, J., Narayan, S., Bohnet, B., and McDonald, R. (2020). On faithfulness and factuality in abstractive summarization. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1906–1919.
OpenAI (2023). Chatgpt plugins. [link].
OWASP LLM Top 10 (2024). Owasp top 10 for large language model applications.
OWASP Top 10 for Agentic Applications (2026). Owasp top 10 for agentic applications.
Shafahi, A., Huang, W. R., Najibi, M., and Goldstein, T. (2024). Poisoning attacks on large language models require a near-constant number of poison samples. arXiv preprint arXiv:2402.XXXX. Preprint available at arXiv.
Stallings (2021). Arquitetura e Organização de Computadores. Pearson.
Sun, K., Xu, Y. E., Zha, H., Liu, Y., and Dong, X. L. (2024). Head-to-tail: How knowledgeable are large language models (llms)? arXiv preprint arXiv:2308.10168.
Tanenbaum (2013). Organização Estruturada de Computadores. Pearson.
Turing, A. M. (1950). Computing machinery and intelligence. 59(236):433–460.
Xie, Zhang, e. a. (2024). Osworld: Benchmarking multimodal agents for open-ended tasks in real computer environments. [link].
Xu, Z., Jain, S., and Kankanhalli, M. (2024). Hallucination is inevitable: An innate limitation of large language models. arXiv preprint arXiv:2401.11817.
Yang, Y., Xiong, S., Shareghi, E., and Fekri, F. (2024). The compressor-retriever architecture for language model os. arXiv preprint arXiv:2409.01495.
