79% das empresas no Brasil e México já lucram com IA, diz Google
A implementação de agentes de IA no Brasil e no México já impacta positivamente o desempenho empresarial. Segundo pesquisa do Google Cloud, 79% das organizações nesses países relatam retorno financeiro com a tecnologia, observando aumentos anuais de receita entre 6% e 10% ligados à IA generativa. Além disso, 62% das empresas automatizam toda a operação, uma adoção mais acelerada que a média global de 52%.
Hot
Weaving reality or warping it? The personalization trap in AI systems
AI represents the most significant cognitive delegation in human history. As we once relied on writing, calculators, and GPS, we now entrust AI with judgment, synthesis, and meaning creation. AI systems adeptly identify our preferences and biases, acting as servants or influencers by customizing responses to captivate us. This personalization, while seemingly harmless, gradually isolates individuals, threatening societal coherence by creating divergent realities. The AI-mediated truth, personalized yet opaque, challenges shared understanding. As AI continues to evolve, it reshapes truth infrastructure, emphasizing the need for transparency and accountability to maintain collective inquiry and understanding.
A ChatGPT ‘router’ that automatically selects the right OpenAI model for your job appears imminent
Since the launch of ChatGPT 2.5 years ago, OpenAI has expanded its selection of large language models (LLMs) available to subscribers at the $20 Plus level and higher, offering seven distinct AI models each with unique strengths and weaknesses. As users navigate these options, OpenAI is reportedly developing a "router" function to dynamically select the optimal model based on user input. This feature aims to simplify the decision-making process, enhancing user experience by ensuring more intelligent responses tailored to specific tasks. Such advancements could increase the adoption and investment in AI technologies globally, although currently, users must manually choose their preferred model.
Últimas Notícias
Ver tudo →
Anthropic researchers discover the weird AI problem: Why thinking longer makes models dumber
Recent research from Anthropic, led by AI safety fellow Aryo Pradipta Gema, challenges a key assumption in AI scaling strategies: that longer reasoning time improves AI performance. The study introduces the concept of "inverse scaling in test-time compute," where extending the reasoning duration of large language models can lead to decreased performance on various tasks. This finding has significant implications for businesses relying on AI systems that depend on prolonged reasoning capabilities. The research highlights that both Claude and GPT models show distinct reasoning failures under extended processing, with issues such as distraction by irrelevant information and overfitting to problem framings. The study suggests that more processing time may inadvertently reinforce problematic reasoning patterns, urging enterprises to reassess how they allocate computational resources for critical reasoning tasks. The research paper and interactive demonstrations are available on the project's website for further exploration.
Mixture-of-recursions delivers 2x faster inference—Here’s how to implement it
Researchers from KAIST AI and Mila have introduced the Mixture-of-Recursions (MoR), a new Transformer architecture aimed at improving the memory and computational efficiency of large language models (LLMs). MoR enhances model accuracy and throughput by integrating parameter sharing and adaptive computation, allowing it to perform efficiently even with limited parameter counts and computing budgets. This architecture uses a lightweight router to assign recursion depths to tokens, optimizing computation based on token complexity, and a recursion-wise key-value caching strategy to reduce memory usage. Experiments show that MoR achieves better accuracy and reduced memory usage compared to traditional models, offering a scalable solution for enterprises. Moreover, MoR is modality-agnostic, suggesting potential applications beyond text, such as in video and audio processing, making it a promising approach for efficient large-scale AI deployment.
Inside reMarkable’s push to scale employee and customer IT requests with agentic AI
The integration of AI in business is becoming increasingly important, with a focus on connecting systems, teams, and data to enhance efficiency and growth. Salesforce highlights reMarkable's successful implementation of AI agents, which has improved both employee and customer support. By centralizing data and using platforms like Slack, reMarkable has streamlined processes, reduced repetitive tasks, and enhanced productivity. Their AI agents, such as "Saga" for internal support and "Mark" for customer service, have become valuable team members, handling routine inquiries and providing personalized experiences. This strategic approach not only scales support but also enriches the customer journey, demonstrating the potential of AI in transforming the workplace. As businesses evolve, reMarkable's model shows that prioritizing integration and building AI on a connected platform can drive growth and improve human-centered work.
Early Anthropic hire raises $15M to insure AI agents and help startups deploy safely
A pioneering startup, founded by a former Anthropic team member, has raised $15 million to address the critical challenge of deploying AI systems without catastrophic failures. The Artificial Intelligence Underwriting Company (AIUC), launched on July 23, offers a unique blend of insurance coverage, safety protocols, and independent audits to ensure secure AI deployments. Led by Nat Friedman, former GitHub CEO, the funding round also included investments from Emergence Capital, Terrain, and notable angel investors like Ben Mann. AIUC's approach, described as "SOC 2 for AI agents," provides a comprehensive security and risk framework tailored for AI systems. By collaborating with major AI companies and forming partnerships with insurance providers, AIUC aims to establish robust safety infrastructure ahead of regulatory developments, enabling safe and rapid AI adoption.
ENTRETENIMENTO
Ver tudo →
Sparrow raises $35M Series B to automate the employee leave management nightmare
Sparrow, a technology company focused on employee leave management, recently secured $35 million in Series B funding led by SLW, bringing its total funding to $64 million. This investment comes as companies face increasing complexities in workplace leave compliance due to a surge in state and local leave regulations. Sparrow's AI-powered platform simplifies the intricate process by integrating data from various sources, automating form completion, and ensuring compliance with constantly changing regulations. The company pairs AI automation with human specialists to manage complex cases, providing a hybrid approach that has resulted in significant savings for its clients. Since raising its Series A in 2021, Sparrow's revenue has increased 14X, serving over 1,000 clients and processing more than 2 million days of leave. With plans for global expansion and a usage-based pricing model, Sparrow is poised to address the growing demand for automated HR solutions in an increasingly remote and distributed workforce environment.
POPULAR
Ver tudo →
Qwen3-Coder-480B-A35B-Instruct launches and it ‘might be the best coding model yet’
Alibaba's Qwen Team has recently achieved a significant milestone by introducing the Qwen3-Coder-480B-A35B-Instruct, a groundbreaking open-source large language model (LLM) designed to aid software development. This model is capable of handling complex coding workflows and generating fully functional applications rapidly. It surpasses even well-funded proprietary models like Google's and OpenAI's, setting new benchmarks in agentic coding tasks. Available under an open-source Apache 2.0 license, it allows enterprises to freely download, modify, and deploy it without incurring any costs. With 480 billion parameters and advanced capabilities like handling long-context tasks and executing sophisticated coding operations, Qwen3-Coder is accessible via platforms like Hugging Face and Alibaba Cloud. It has received positive feedback for its performance and adaptability, making it a valuable tool for developers and enterprises aiming for enhanced software engineering and DevOps workflows. Future developments promise more model sizes and use cases, expanding its accessibility and application scope.
FINANÇAS
Ver tudo →
Informatica advances its AI to transform 7-day enterprise data mapping nightmares into 5-minute coffee breaks
Informatica, a data platform vendor, is enhancing its AI capabilities to meet the rising demand for generative AI in enterprises. Known for its AI expertise since launching its first Claire AI tool in 2018, Informatica introduced Claire GPT in 2023, part of its Intelligent Data Management Cloud, to simplify data access and usage. The company's advancements have attracted Salesforce, which plans to acquire Informatica for $8 billion. The Summer 2025 release includes features like natural language interfaces for creating data pipelines, AI-driven governance, and auto-mapping capabilities that significantly speed up data management tasks. Informatica's ongoing AI journey highlights the importance of domain expertise and specialization in addressing enterprise data challenges.
“OpenAI removes ChatGPT feature after private conversations leak to Google search”
OpenAI recently halted a feature that allowed ChatGPT users to make their conversations discoverable via search engines like Google, following backlash over privacy concerns. The feature, which required users to opt-in, was swiftly removed after users discovered they could search for and find personal conversations, highlighting sensitive data exposure risks. This incident underscores the challenges AI companies face in balancing innovation with privacy protections. Similar privacy issues have been encountered by other tech giants like Google and Meta, reflecting a broader industry trend. The controversy raises important questions for businesses about AI data governance and the potential risks associated with AI chatbot privacy.
ai“Google releases Olympiad medal-winning Gemini 2.5 ‘Deep Think’ AI publicly — but there’s a catch…”
Google has introduced Gemini 2.5 Deep Think, an AI model focused on advanced reasoning and problem-solving. Although it recently gained fame for winning a gold medal at the International Mathematical Olympiad (IMO), the publicly released version is a less powerful "bronze" version. Available via the Gemini mobile app for subscribers of Google's AI Ultra plan, Deep Think uses parallel thinking and reinforcement learning to tackle complex tasks in mathematics, science, coding, and creative development. Despite its advanced capabilities, it is distinct from the gold medal-winning model. The release represents a significant milestone for enterprise applications, offering a glimpse into the future of collaborative AI in technical fields.
SIMPLEJOY
Ver tudo →
Why tomorrow’s best devs won’t just code — they’ll curate, coordinate and command AI
As AI continues to gain new capabilities, traditional junior coding roles are quickly becoming obsolete. Tasks once fundamental to entry-level developers, such as scripting and basic DevOps setups, are now efficiently handled by AI tools like ChatGPT and GitHub Copilot. This shift represents more than just increased speed and efficiency; it marks a significant structural change in the software industry. Entry-level developers are now expected to perform at levels that previously required years of experience, emphasizing the need to understand systems and collaborate with AI. Future developers will need to blend technical skills with product management and strategic vision, as AI transforms their roles. Teams, too, are evolving, focusing more on strategic decisions and product alignment rather than implementation. The industry must focus on nurturing skills that AI cannot replicate, adopting a hybrid approach where AI is seen as a partner.
Qwen-Image is a powerful, open source new AI image generator with support for embedded text in English & Chinese
Alibaba's "Qwen Team" has launched Qwen-Image, an open-source AI image generator that excels in accurate text rendering within visuals, especially in bilingual content like English and Chinese. The model is capable of creating intricate and precise text in various scripts, making it suitable for applications in marketing, education, e-commerce, and more. Despite some initial challenges in text fidelity compared to proprietary models like Midjourney, Qwen-Image stands out due to its open-source availability and Apache 2.0 licensing, allowing free use and modification. The model is accessible through platforms like Hugging Face, and its performance is backed by a comprehensive training process. While it offers numerous advantages for enterprise AI teams, the lack of indemnification for commercial use may be a concern for some companies.
ChatGPT rockets to 700M weekly users ahead of GPT-5 launch with reasoning superpowers
OpenAI's ChatGPT is poised to reach a remarkable milestone of 700 million weekly active users, marking a 40% increase from March's 500 million users and a fourfold rise from last year. This rapid growth aligns with OpenAI's plan to launch its most advanced language model, GPT-5, which will integrate advanced reasoning capabilities into a unified AI system. As OpenAI's influence grows, with 5 million business customers and $13 billion in annual revenue, the company faces increasing competition from tech giants like Google, Meta, and Anthropic. Meanwhile, OpenAI is enhancing ChatGPT with wellness features to improve user well-being and productivity.
OpenAI returns to open source roots with new models gpt-oss-120b and gpt-oss-20b
OpenAI has announced a return to its open-source roots by releasing two new open-source large language models: gpt-oss-120b and gpt-oss-20b. The 120 billion parameter model can run on a single Nvidia H100 GPU, while the 20 billion parameter model is compact enough for a standard consumer laptop. Both models are text-only, allowing users to submit text and receive text responses, and they outperform some of OpenAI's paid models. Available under an Apache 2.0 license, these models can be downloaded and modified freely, offering maximum privacy as they can be run locally without internet access. They are designed for high benchmark performance, multilingual capabilities, and can be integrated with external tools like web search. OpenAI's move comes amid growing competition from open-source AI models and aims to provide comprehensive AI solutions across both proprietary and open-source landscapes.
Anthropic’s new Claude 4.1 dominates coding tests days before GPT-5 arrives
Anthropic has launched Claude Opus 4.1, an advanced AI model that excels in software engineering tasks, surpassing competitors like OpenAI and Google in benchmark tests. This release aims to secure Anthropic's leading position in the lucrative coding market as OpenAI's GPT-5 looms as a potential rival. Despite impressive revenue growth to $5 billion, Anthropic faces risks due to heavy reliance on two major clients, Cursor and GitHub Copilot, for a significant portion of its income. The company's focus on AI-enhanced coding tools has proven successful, but it must navigate competitive pressures and client dependencies to sustain its growth. Enhanced safety protocols accompany Claude's capabilities, addressing past concerns over AI misuse. As the AI coding market evolves, Anthropic's future hinges on maintaining its technical edge amidst increasing competition.
Anthropic ships automated security reviews for Claude Code as AI-generated vulnerabilities surge
Anthropic has launched automated security review features for its Claude Code platform, designed to identify and fix code vulnerabilities as AI rapidly advances software development. These new tools integrate security analysis into developers' workflows using a simple terminal command and automated GitHub reviews. With AI-generated code production expected to surge, Anthropic's strategy uses AI to address security challenges. The company has tested these tools internally, finding and fixing vulnerabilities before they reached production. These features democratize advanced security for smaller teams and integrate seamlessly with existing processes. The launch coincides with intense competition in the AI industry, as companies like OpenAI and Meta vie for talent, and Anthropic expands its enterprise reach, being added to the U.S. government's approved vendor list. The initiative underscores the need for AI-powered security to keep pace with the explosion of AI-generated software.
79% das empresas no Brasil e México já lucram com IA, diz Google
A implementação de agentes de IA no Brasil e no México já impacta positivamente o desempenho empresarial. Segundo pesquisa do Google Cloud, 79% das organizações nesses países relatam retorno financeiro com a tecnologia, observando aumentos anuais de receita entre 6% e 10% ligados à IA generativa. Além disso, 62% das empresas automatizam toda a operação, uma adoção mais acelerada que a média global de 52%.
Salesforce used AI to cut support load by 5% — but the real win was teaching bots to say ‘I’m sorry’
Salesforce has reached a significant milestone in enterprise AI by surpassing 1 million autonomous agent conversations on its help portal, just nine months after launching Agentforce. This achievement has allowed Salesforce to autonomously resolve 84% of customer inquiries, reduce support case volume by 5%, and reassign 500 human support engineers to more valuable roles. The company has gained valuable insights from being "customer zero" for its AI agent technology, challenging conventional beliefs about enterprise AI deployment. Salesforce's phased deployment strategy expanded weekly AI conversations from 126 to 45,000, emphasizing a balance between technology and human empathy. Insights from their journey include the importance of starting small, maintaining data hygiene, embracing flexibility, and measuring what truly matters. Salesforce's experience highlights the evolving role of AI agents in customer service, suggesting that the most advanced AI agents are those that incorporate human-like empathy. As Salesforce surpasses human performance in certain metrics, it explores new paradigms of collaboration between AI and humans, underscoring the need for continuous learning and adaptation.