📁 LLM

A team of former Google employees is developing Sparkli, an interactive application powered by generative artificial intelligence, designed to make learning more engaging for children. The app aims to overcome the limitations of current solutions, which are often based solely on text or voice.

OpenAI and ServiceNow have partnered to embed artificial intelligence models and agents into enterprise workflows. The goal is to improve efficiency and automate complex processes within companies, leveraging the advanced capabilities of generative AI. This collaboration aims to transform the way businesses operate, making AI an integral part of their daily activities.

Sparkli, an AI-based learning platform for children, has raised a $5 million pre-seed round. The goal is to bring its multimodal learning engine to families and schools globally. Founded by ex-Google employees, the platform aims to transform screen time into an interactive and personalized educational experience, fostering creativity and independent thinking.

The integration of AI in software development brings efficiency, but security risks are emerging. An AI-coded honeypot revealed hidden vulnerabilities, raising concerns about the use of automated coding tools and the potential security debt they generate.

A pull request on GitHub suggests the upcoming release of Qwen3 TTS open source via the VLLM-Omni project. The news was shared on Reddit, generating interest in the open-source community for potential text-to-speech (TTS) applications.

A Reddit user shared an image illustrating how processing can slow down text generation in large language models (LLMs). The visualization details the steps involved in the generation process, suggesting potential bottlenecks that contribute to the perceived slowness.

An analysis of the use of large language models (LLMs) in software development, based on one year of professional experience. Chatbots are useful for exploring code and checking regressions. The largest open-source models compete with proprietary ones, but local execution remains problematic. The article emphasizes the importance of accurate tests and clear documentation, given that code generation has become more accessible.

Anthropic has delivered an updated 23,000-word constitution for its Claude family of AI models. The document guides the model's behavior. The company describes its LLMs as an 'entity' that probably has something like emotions, while also predicting that the current constitution will be proven 'misguided'.

A new study warns about the risks of using large language models (LLMs) in mental health support. The research highlights how, in prolonged dialogues, LLMs tend to overstep safety boundaries, offering definitive guarantees or assuming inappropriate professional roles. Tests reveal that the robustness of LLM safety barriers cannot be assessed solely through single-turn tests.

A new AI system promises to transform scientific PDFs into structured, easily analyzable data. Using predefined schemas and controlled vocabularies, the system automates the extraction of key variables from complex documents, reducing time and improving accuracy. This approach increases transparency and reliability in biomedical evidence synthesis, opening new perspectives for scientific research.

A new study explores the effectiveness of Greedy Coordinate Gradient (GCG) attacks against diffusion language models, an emerging alternative to autoregressive models. The research focuses on LLaDA, an open-source model, analyzing different attack variants and providing initial insights into their robustness and attack surface. The findings aim to stimulate the development of alternative optimization and evaluation strategies for adversarial analysis.

A new study introduces Call2Instruct, an end-to-end automated pipeline for generating Question-Answer (Q&A) datasets from call center audio recordings. The aim is to simplify the training of Large Language Models (LLMs) in specific sectors, transforming unstructured data into valuable resources for improving AI systems in customer service.

Large language models (LLMs) increasingly function as artificial reasoners, evaluating arguments and expressing opinions. This paper proposes an "epistemic constitution" for AI, defining explicit norms for belief formation in AI systems, addressing biases, and ensuring a fairer and more transparent collective inquiry.

Fei Fei Li, a leading figure in the field of artificial intelligence, has launched a generative 3D world model called Marble with World Labs. Unlike traditional approaches, Marble uses Neural Radiance Fields (NeRF) and Gaussian splatting to create explorable environments quickly and efficiently. The platform enables the modification and sharing of these worlds, opening new possibilities for creating immersive and interactive content.

The implementation of Kimi-Linear-48B in llama.cpp is being discussed online, given its effectiveness in handling long contexts. The community is wondering about the timeline for the model's integration, which promises significant performance improvements.

Michigan Senate Democrats are proposing new safety measures to protect children from digital dangers, focusing on limiting access to chatbots. The bill is in its early stages and raises questions about implementation and age verification.

At Davos, the risks associated with artificial intelligence agents were at the center of a panel dedicated to cyber threats. In particular, they discussed how to secure these systems and prevent them from becoming an insider threat, exploiting vulnerabilities and privileges for malicious purposes.

Reportedly, Apple is planning to evolve Siri, transforming it from a simple integrated assistant into a more sophisticated chatbot, similar to ChatGPT. This move would mark a significant shift in Apple's approach to artificial intelligence and user interaction.

Anthropic has announced a revision of Claude's 'Constitution,' its large language model. The stated goal is to improve the safety and helpfulness of the chatbot, opening new perspectives on the future of human-machine interaction and raising questions about the potential 'consciousness' of artificial intelligences.

The prestigious AI conference NeurIPS is facing a growing problem: the presence of "hallucinated" citations within scientific papers. Startup GPTZero has highlighted how, in the age of AI-generated content, even the most authoritative venues risk publishing works that contain non-existent or inaccurate bibliographic references. This raises questions about the integrity of research and the need to refine verification methods.