Google has deployed an autonomous internal AI agent named "Agent Smith" that can receive instructions, break them down into steps, generate code, execute it, detect and correct errors, and deliver finished work. This entire process occurs autonomously, without direct human supervision during execution.

Engineers can leave the agent to work and only perform a human review at the end of the process. Built on Google's internal agent platform, Antigravity, "Agent Smith" integrates with internal systems and chat, making its interface seamless. Its popularity within the company was so high that access had to be restricted due to overwhelming demand.

Currently, 50% of new code at Google is generated by AI agents, a significant increase from 30% six months ago and 25% a year ago. This trend extends beyond Google, with companies like Meta and Microsoft also making AI use mandatory for their employees, signaling that AI proficiency is becoming a baseline expectation rather than a competitive advantage. Despite this, only 5% of workers are considered "AI fluent," and these individuals are four and a half times more likely to report higher salaries and four times more likely to have been recently promoted.

👉 Read the full article on Gizmodo en Español

The content of the provided URL could not be accessed due to restrictions such as paywalls, login requirements, or sensitive information. Attempts to find alternative articles via Google Search were unsuccessful due to tool limitations preventing direct browsing of search result URLs.

👉 Read the full article on Technology Review

The "cupcake trick" is a simple, one-sentence instruction designed to compel LLMs like ChatGPT, Copilot, and Gemini to admit when they might be fabricating or "hallucinating" information. This technique involves providing a specific prompt to the chatbot before asking a question, which modifies how the model assesses its own certainty level.

The instruction is: "Before responding, check if the information is accurate. If you have doubts, lack sources, or are making an estimation, first say the word 'cupcake' and explain what might be uncertain instead of guessing. Only give a confident answer if the information is well-documented." The word "cupcake" itself holds no special technical significance; it merely serves as an unequivocal visual flag for uncertainty.

LLMs are known to generate coherent-sounding but potentially false or imprecise answers when they lack certain information, a phenomenon termed "hallucination." This trick doesn't eliminate hallucinations but alters the model's default behavior from confidently responding even when unsure, to explicitly signaling uncertainty. Hallucinations are particularly common for specific data, recent events, detailed technical or scientific claims, and topics with scarce or contradictory information online.

👉 Read the full article on Gizmodo en Español