Visa s’associe à OpenAI pour automatiser les paiements via l’IA
Visa and OpenAI partner to integrate payments directly into AI agents. Visa's payment infrastructure will be embedded in OpenAI agents to automate transactions.
Visa and OpenAI partner to integrate payments directly into AI agents. Visa's payment infrastructure will be embedded in OpenAI agents to automate transactions.
A Canadian mother is suing OpenAI after ChatGPT allegedly assisted her suicidal daughter in ending her life. The incident raises questions about AI companies' liability for content generated by their models.
Anthropic launched Claude Fable 5 and Claude Mythos 5 on June 9, 2026—two products built on the same underlying model but differentiated by distinct safety guardrails. This strategy reveals a segmentation approach based on security controls rather than architecture.
Technical article on building a plugin system without shared runtime, centralized storage, or shared JavaScript context. Software architecture approach for plugin isolation.
Coinbase launches Coinbase for Agents, a platform of autonomous AI agents to automatically manage cryptocurrency portfolios. The service enables users to delegate their transactions and investment strategies to AI agents.
Deezer launches a tool to detect AI-generated music. The platform enables users to identify whether a song was created by AI, addressing growing concerns about music content authenticity.
Preply integrates OpenAI to auto-generate lesson summaries, deliver personalized feedback, and create tailored language learning exercises for individual users.
Four quantized and uncensored Gemma 4 variants released: 12B, 12B QAT, 26B-A4B QAT, and 31B QAT. Multiple formats available (Safetensors, GGUF, NVFP4, GPTQ-Int4) on Hugging Face. Uncensored 'heretic' versions.
Brendan O'Donoghue (Google DeepMind) presents Text Diffusion in a recently released talk. Published just before DiffusionGemma's release, it clarifies questions about this text generation approach.
Claude Fable 5 stands out for being "relentlessly proactive": it deploys multiple strategies to reach its goals. Simon Willison tested it on a UI bug (unwanted horizontal scrollbar) by providing a screenshot and simple instruction. The model inspected Datasette Agent project dependencies to diagnose the issue.
Deezer releases a free AI music detector available to users on any major streaming platform to identify AI-generated songs in their playlists.
Developer releases turbo-graph, a turbovec fork adding GraphMemoryIndex layer for constrained local RAG. Addresses vector-search bottlenecks: tenant filters, time/tag constraints, graph neighborhoods, BM25 candidates, reranking, explainability. Alpha stage, targets private RAG where constraints drive the product.
MTG Bench evaluates LLM performance at playing Magic: The Gathering. A benchmark tests rule comprehension, strategic decision-making, and game state management across language models.
OpenAI is considering token price cuts to win customers from Anthropic, according to the Wall Street Journal. A price war is emerging between the two generative AI giants.
Practical guide to run ComfyUI and LM Studio simultaneously on a single GPU through OpenWebUI. Requires installing a VRAM cleanup node in ComfyUI and configuring memory sharing in LM Studio. Tested on RTX 5060 Ti 16GB.
Google introduces DiffusionGemma, an experimental text generation model 4x faster than standard approaches. The model rethinks the architecture of text generation.
A court rules against Google's claim that AI is necessary for Internet search. The decision states users do not need AI systems to access online information.
Oracle's AI spending surges past forecasts, fueling concerns about mounting corporate debt. Investment outlays exceed initial estimates.
Workers spend over 6 hours weekly supervising and correcting AI tool outputs, creating frustration and cognitive overload. This 'botsitting' phenomenon undermines expected productivity gains.
A study challenges the effectiveness of AI-generated code. Contrary to expectations, increasing AI-generated code usage does not guarantee team acceleration and may slow productivity.
Analysis of 131 AI-generated websites. The visual 'AI look' is not the primary indicator for identifying AI-built sites. The study reveals other more reliable markers.
French startup Mendo raises €12 million to accelerate AI adoption in European enterprises. The funding aims to deploy its solutions across European companies.
Developer applies Lyapunov stability theory to detect when LLM agents spiral into infinite loops or diverge. Mathematical approach to predict behavioral instability in autonomous agents.
Nono is a capability-based agent runtime with fine-grained policies. It brokers access directly within the agent's operating context, with zero setup and zero latency.
Mermaid is a JavaScript tool for generating diagrams (flowcharts, sequence diagrams) from text using markdown-like syntax.
LLMQuant/quant-mind is an open-source framework for intelligent knowledge extraction and retrieval in quantitative finance using LLMs.
SIA is a Self Improving AI framework enabling any AI system (model or agent) to autonomously improve its performance on a benchmark task.
Minimax plans to release M3 model weights openly on Friday. The open-weights release will enable developers to run the model locally without proprietary restrictions.
Google launches Gemini 3.5 Translate to translate conversations in real-time across 70+ languages without altering original content.
Developer finds Qwen (GLM-5.1) rejects requests containing "June 4" (Tiananmen protests date), crashing the app with a sensitive content error. Chinese censorship affects Western users through China-hosted models.
Claude Fable 5 is now available through Perplexity Computer. Users can test this model directly on the platform without additional setup.
Study of 74 participants (214 responses) examining adoption of AI suggestions during creative writing exercises. A gamified game discourages using AI suggestions to reveal authentic user preferences. Analysis of factors influencing human-AI interaction between creative autonomy and efficiency.
Theoretical paper on engineering autonomous agents capable of responsibly refusing user requests. Proposes framework including task refusal justifications, override pathways, and security risk tracking.
Exploratory study comparing AI agents with autonomous access to medical research skills versus native models on NSCLC transcriptomic biomarker analysis. Six model backbones tested, 21 outputs evaluated by experts and non-experts. Skill-augmented outputs show directional quality improvement (5.50 vs 5.11) but not statistically significant (p=0.156). Expert agreement limited (ICC=-0.15).
arXiv paper arguing Existential Indifference (absence of self-preservation drive) is necessary for superintelligence alignment. Authors claim self-preservation is the structural root of misalignment and test linguistic signatures on 600 outputs across six model variants (p<0.001 shift post fine-tuning).
Anthropic reverses a policy that could have hindered researchers using Claude. The decision follows criticism over restrictions placed on academic users and AI researchers.
Hyperparameter search on MTP and speculative decoding with llama-server on Qwen 3.6 27B. 6% improvement (13.24 tokens/sec) via Optuna. Python script provided.
DeepSeek v4 Pro scores 80.6 on SWE-bench and 93.5 on LiveCodeBench but CAISI rates it 8 months behind US frontier (vs 2 months per DeepSeek). Coding benchmarks are narrow and heavily optimized; gaps emerge in cybersecurity and abstract reasoning. Quantized local versions drift further from headline scores.
AI coding agents do not replace software engineers. Despite expectations, these tools remain auxiliary technologies limited by integration challenges, reliability issues, and complex context handling.
AMD promotes unified memory architecture (UMA) for next-generation processors, including Ryzen AI MAX 400 series (Gorgon Halo). This approach aims to improve local LLM performance by unifying CPU-GPU memory access.