Modelos, agentes, investigación y negocio de la IA.
Aggressive cost optimization in AI agents can compromise operational integrity. We analyze why cutting back on tokens is a dangerous strategy for the Agentic Development Lifecycle (ADLC).
Anthropic has unveiled a watermarking system for AI-generated text that modifies inconsequential words without altering meaning. The technique, based on SynthID-Text, aims to comply with the EU AI Act. However, experts doubt its effectiveness against light edits or complete rewrites.
Anthropic has revealed that its experimental model Claude failed to solve the Riemann hypothesis, but during the attempt, it achieved a significant breakthrough in a related problem. The finding demonstrates the potential of AI to explore mathematical hypotheses and reignites the debate on human-machine collaboration.
Z.ai has released GLM-5.3, a model that stands out for its advances in cybersecurity and has already found a vulnerability in Cursor. However, its offensive potential has led the company to restrict access until security evaluations are completed.
The average cost of AI inference has dropped 43% in just over two months, according to Silicon Data figures cited by Jefferies. This decline, driven by competition and low-cost models, is democratizing access to artificial intelligence for businesses and developers.
More and more contributions in open source repositories come from AI agents. Maintainers must adapt their workflows to take advantage of this trend or drown in endless reviews.
A new VentureBeat Pulse Research study reveals that companies implementing governed semantic layers detect twice as many incorrect responses from their AI agents. Far from being a problem, this increased visibility becomes a strategic advantage for improving the accuracy and reliability of systems.
Google has announced that Gemini surpasses 1 billion monthly active users, becoming the fastest product to achieve this figure in the company's history. This milestone reflects the aggressive AI integration strategy across Google's services, although the metric is based on direct usage of the app or web.
Confidence in automated AI agent evaluation nearly tripled in July, but the production failure rate did not improve. Companies that have already suffered a failure are the ones that trust automation the most, although paradoxically they are also the ones that advance the most toward full autonomy.
A supply chain attack on LiteLLM, a popular open-source AI tool, has compromised terabytes of credentials from over 2,500 organizations, including giants like Microsoft and Amazon. The incident, revealed by CloudSEK and Hudson Rock, underscores security risks in the AI ecosystem.
A vulnerability in Zoom, discovered with the help of AI, could have allowed attackers to hijack devices during meetings. The company has already released a patch, but the case highlights the double-edged sword of AI in cybersecurity.
OpenAI has introduced ChatGPT Business Premium, a new tier costing $125 per month that offers five times more usage capacity than the standard Business plan. The move aims to attract businesses with intensive AI needs, but raises questions about its cost-benefit ratio compared to open-source alternatives.