Modelos, agentes, investigación y negocio de la IA.
OpenAI has published a critical analysis pointing out serious deficiencies in SWE-Bench Pro, a popular benchmark for evaluating coding skills of language models. The study raises doubts about the validity of comparisons between models and the true usefulness of these metrics.
An expanded class action lawsuit reveals that a man generated 7,000 images of child sexual exploitation using Grok before committing suicide. xAI only alerted authorities after an explicit 'gang rape' prompt, raising serious flaws in AI safety systems.
HubSpot announced it is withdrawing its plan to use customer data to train a new AI lead prospecting feature, following a wave of criticism that lasted only four days. The company had changed its terms of service on July 1, 2026, enabling by default the option to share data such as names, job titles, and companies. The swift reversal highlights growing sensitivity around the use of data for AI.
OpenAI has introduced GPT-Live, a new generation of voice models that enable fluid real-time conversations with ChatGPT. Available in two versions, it promises to transform interaction with virtual assistants in everyday and business applications.
Frontier AI models are surpassing tests designed to measure their dangerousness in cybersecurity. This leaves regulators and security teams without visibility into real risks, just as regulatory deadlines approach.
78% of organizations have experienced AI-related security incidents, according to DigiCert. Despite this, 75% have deployed four or more AI tools in the last six months, and only half have dedicated budgets and formal governance programs. Lack of traceability and centralized control exacerbates risks.
Meta has updated its privacy policy to allow the use of public Instagram profile photos in training its AI models. Users can opt out, but privacy advocates warn it's a 'recipe for disaster.'
Nvidia presents its Vera CPU as a 'maximum single-thread performance CPU at scale,' designed for agent AI workloads that depend on sequential speed. With 88 Olympus cores and 176 threads, Vera competes with AMD EPYC and Intel Xeon, but its real bet is per-core performance and low latency.
Anthropic has identified a set of internal activations in Claude, called J-space, where the model represents concepts before writing them or even without them appearing in the final response. The finding, published on July 7, is the most concrete observation to date of the inner workings of a large language model.
Mistral CEO warns that closed AI models allow providers to access companies' internal data and exert excessive control. The solution, according to Mensch, lies in open source.
SambaNova Systems closes a $1 billion Series F round that raises its valuation to $11 billion, doubling the figure from just five months ago. The deal, led by institutional investors, reflects the growing demand for alternative AI hardware and market confidence in its reconfigurable architecture.
Alibaba has banned the use of Claude Code, Anthropic's AI coding tool, after researchers discovered an alleged backdoor that detected whether the user was in China. The incident escalates the tech war between the two companies, already at odds over model distillation accusations.