Generative AI

AgentFounder, a groundbreaking AI agent developed by Alibaba’s Tongyi Lab, introduces Agentic Continual Pre-training (Agentic CPT) to revolutionize AI agent development. This novel three-stage approach incorporates an intermediate training phase focused on agentic behaviors, addressing limitations in current two-stage models. Using innovative data synthesis methods like First-order Action Synthesis and Higher-order Action Synthesis, AgentFounder-30B achieves state-of-the-art performance across multiple benchmarks, surpassing existing open-source models and some proprietary systems. The research demonstrates logarithmic scaling patterns for agentic capabilities and improved fine-tuning efficiency. This work paves the way for more capable, general-purpose AI agents that can autonomously navigate complex information environments and solve multi-step problems.

GenExam, a groundbreaking benchmark, challenges AI models to tackle multidisciplinary text-to-image exams, pushing the boundaries of artificial general intelligence. Spanning 10 academic subjects, it tests AI’s ability to draw complex diagrams requiring deep subject knowledge. Current state-of-the-art models, including advanced commercial ones, struggle with this task, achieving strict scores below 15%. The benchmark reveals significant limitations in today’s AI systems, highlighting the gap between closed-source and open-source models. GenExam’s comprehensive evaluation framework measures semantic correctness and visual plausibility, offering both strict and relaxed scoring. As AI progresses towards more general capabilities, GenExam provides crucial guideposts, raising expectations and offering a roadmap for future development in multidisciplinary AI image generation.

OpenAI has unveiled sweeping new restrictions for ChatGPT users under 18, prioritizing safety over privacy and freedom. This move follows a tragic incident involving a teen’s suicide, which led to a lawsuit against the company. The changes include blocking flirtatious conversations, implementing safeguards around self-harm discussions, and potentially contacting parents or authorities in severe cases. Parents can now set “blackout hours” and link their accounts for oversight.

These measures come as ChatGPT usage soars to 700 million weekly active users, with shifting demographics and use cases. However, security concerns persist, as evidenced by the recently patched “ShadowLeak” vulnerability. OpenAI has also introduced a personalization hub, allowing users to customize ChatGPT’s communication style, though reception has been mixed.

A groundbreaking paper from OpenAI and Georgia Tech researchers offers a compelling explanation for hallucinations in language models. The key insight: models hallucinate because training rewards guessing over admitting uncertainty. This stems from binary evaluation methods that penalize “I don’t know” responses while rewarding lucky guesses.

The paper traces hallucinations from pretraining through post-training modifications, linking them to statistical factors in binary classification. It challenges misconceptions, proving hallucinations aren’t inevitable or mysterious.

The solution? Modify evaluations to include explicit confidence targets, promoting “behavioral calibration.” This approach has already shown promise in newer OpenAI models, significantly reducing error rates.

As AI integrates further into critical applications, addressing hallucinations becomes crucial for building trustworthy systems. This research provides both theoretical insights and practical recommendations to transform AI development and evaluation.

Switzerland has unveiled Apertus, a groundbreaking open AI model developed by leading Swiss institutions. This fully transparent system, available in 8-billion and 70-billion parameter versions, challenges proprietary AI with its open-source approach. Apertus boasts impressive multilingual capabilities, supporting over 1,000 languages including Swiss German and Romansh. The project emphasizes AI as public infrastructure, aligning with EU regulations and Swiss laws. It represents a shift towards transparent, accessible AI development for the global community. The Swiss team plans to expand Apertus, focusing on domain-specific tools while maintaining openness. This initiative could significantly influence the future of AI development, promoting collaboration and empowering broader communities in technological advancement.

Dentsu Global Services has revolutionized customer service with Generative Machine Learning (GML), a powerful combination of Generative AI and machine learning. This innovative approach anticipates customer issues, transforming reactive service into proactive engagement. By analyzing both operational and conversational data in real-time, GML identifies at-risk orders, calculates risk scores, and triggers personalized interventions before problems escalate.

The system’s unified view of customer data and rapid decision engine have yielded impressive results: a 22% increase in customer satisfaction, 80% reduction in resolution times, and millions in saved revenue. GML’s success hinges on real-time data flow, effective system communication, and immediate responses.

This shift represents more than just AI implementation; it’s a fundamental change in service philosophy, treating customers as individuals with unique stories and expectations. Dentsu’s GML approach offers a blueprint for companies seeking to transform customer service from a cost center

Intuit’s journey into AI began with a rushed chatbot implementation that fell short of expectations. Recognizing the need for a strategic pivot, the company embarked on a nine-month transformation to reimagine its product development approach. By observing real customer behaviors, Intuit shifted focus to enhancing existing workflows rather than forcing new interfaces.

The company adopted a three-pillar framework: fostering a builder culture, implementing high-velocity iteration, and developing GenOS, an internal AI platform. This approach led to the creation of AI agents deeply integrated into QuickBooks and other products, automating tasks and improving user experiences.

The results have been impressive, with customers reporting increased ease of business management and significant time savings. Intuit’s success offers valuable lessons for other enterprises navigating AI transformation, emphasizing the importance of customer-centric solutions and willingness to pivot when necessary.

Anthropic’s decision to train Claude on user conversations marks a significant shift in AI development strategy. Starting September 28, 2025, chat transcripts and coding sessions will be used to enhance Claude’s capabilities, unless users opt out. This change applies to consumer tiers but not commercial accounts. Data retention will extend from 30 days to five years, allowing for long-term analysis. While offering potential improvements in natural language processing and contextual understanding, the move raises privacy concerns. Users must weigh the benefits of contributing to AI advancement against data protection considerations. This policy aligns with industry trends but underscores the ongoing debate between technological progress and personal privacy in the AI era.

Netflix has unveiled formal guidelines for using generative AI in its productions, addressing ethical concerns and industry controversies. The guidelines outline five key principles, focusing on copyright protection, data security, and safeguarding creative professionals’ roles. Production partners must share AI implementation plans with Netflix, with certain applications requiring legal approval. This move reflects Netflix’s commitment to balancing innovation with responsibility, potentially setting industry standards. While offering reassurance to talent, the guidelines also signal Netflix’s continued investment in AI technology. As the field evolves, these guidelines will likely need regular updates to address new capabilities and challenges in the intersection of AI and entertainment production.