Guardrails para Claude, GPT e modelos próprios: o que são e por que importam

·1 views

Guardrails são camadas técnicas e políticas que limitam saídas de IA consideradas inseguras, antiéticas ou ilegais, combinando filtros de entrada e saída, classificadores de conteúdo, regras de negação e treinamento por feedback humano. No Brasil não há lei federal específica sobre o tema, mas a Estratégia Nacional de Inteligência Artificial recomenda controles éticos e técnicos, e setores regulados devem alinhar seus sistemas às normas de seus órgãos. O Código de Defesa do Consumidor e a LGPD reforçam a responsabilidade dos fornecedores por danos causados por saídas automatizadas. Especialistas destacam que guardrails são sistemas de governança contínua, exigindo testes, atualização e supervisão humana, e não apenas listas de termos proibidos.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Scaffoldr Proposes Product-Based Project Starters Over Stack-Based Scaffolding

The author argues that existing tools like create-next-app, create-t3-app, and SaaS starter kits organize project setup around technology stacks or fixed product templates. Modular scaffolders ask which technologies to use, while starter kits lock developers into someone else's choices, which can drift over time. The proposed Scaffoldr tool would instead let developers specify a product type, such as ticketing, with swappable modules for payments, databases, auth, and email. The author notes that African payment providers like Paystack are often overlooked, and the project is not yet built, with maintenance and developer interest remaining open challenges.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Yonto Player moves its source contract into JavaScript plugins on QuickJS

The Yonto Player TV app connects to media servers and folders through a single ContentSourceAdapter interface. Its source contract began as an HTTP API backed by a hosted service, but that required someone to run and update a middle layer whenever a media server changed its fields. The developer kept the contract's JSON shapes as the single definition and made transport an implementation detail, allowing sources to run as JavaScript plugins inside the app. Each plugin is one file whose manifest, declaring its identity and permitted hosts, is read as plain text, and it exposes required functions such as search and getMediaList. Plugins reach the outside world only through a restricted global object, which limits them to requests, configuration, a small cache, logging and parsing helpers.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

OAB rules set guardrails on AI use in legal practice, keeping lawyers responsible

The article explains guardrails for AI in law: technical and ethical limits on using AI for tasks like drafting petitions, analyzing evidence, and giving procedural guidance. It notes that OAB Resolution 05/2023 bars attributing professional responsibility to AI systems and requires lawyers to keep full control, critically review outputs, and accept accountability. The resolution also calls for auditable, transparent, bias-free tools, and platforms must disclose when content is machine-generated. Practical guardrails include blocking drafts unsupported by validated data, flagging conflicts of interest, and logging all AI interactions. Ultimately, the lawyer remains responsible for any AI-assisted work.

0 comments Read more at DEV Community

ProgrammingHacker News ·

Essay asks what objective reasoning in AI models is actually optimizing for

A post on alexwang.ai titled 'Reasoning as Pathfinding' examines the question of what goal reasoning is meant to achieve in modern AI systems. It was shared on Hacker News, where it gathered 24 points and 7 comments. The article's text is not included in the source material, so its specific arguments cannot be detailed here. The discussion centers on the framing of reasoning as a search or pathfinding process rather than a fixed capability. Readers interested in the underlying argument are directed to the original essay and its Hacker News thread.

0 comments Read more at Hacker News