TLDR AI 2025-05-28

Turns out, CEOs have confessions to make. Dataiku has the answer.

• 74% of CEOs admit their job is at risk if they fail to deliver measurable AI-driven business gains.
• 94% suspect employees are using GenAI tools without approval — exposing a massive governance failure.
• 54% admit a competitor already has a superior AI strategy.

Dataiku's latest report, The Global AI Confessions Report: CEO Edition, makes one thing clear: AI execution is where companies win or lose. Dataiku empowers organizations to take charge of their AI talent, processes, and technologies — turning AI strategy into real, measurable impact.

Mistral's new Agents API enables persistent, multi-agent workflows with built-in connectors for code execution, web search, RAG, image generation, and MCP support.

OpenAI is developing a system that lets users sign into third-party apps using their ChatGPT accounts, similar to "Sign in with Google," and is seeking developer partners.

Anthropic is launching a beta voice mode for Claude on mobile, enabling users to engage in voice-based tasks like summarizing calendars or searching documents.

Battlefield reporting reveals how AI-powered weapons are outpacing international law and creating a "race to the bottom" in warfare ethics. In Ukraine, drones are causing 70-80% of casualties, and combatants are using civilian disguises and human shields to evade AI targeting systems that struggle to distinguish between military and civilian targets.

Liger is an optimized trainer for Group Relative Policy Optimization (GRPO) that cuts memory usage by 40% and adds FSDP and PEFT support, making reinforcement learning fine-tuning more efficient and scalable.

Microsoft outlined how next-gen AI agents will require evolved identity and access standards, as it shifts from reactive assistants to proactive collaborators capable of operating autonomously across systems.

Daily-Omni is a benchmark and training-free agent for evaluating models on tasks requiring synchronized audio-visual understanding.

LMEval is a framework for benchmarking AI models across different providers with multimodal support and incremental evaluation.

INTUITOR is a reinforcement learning method that uses an LLM's internal confidence level as a reward signal instead of costly domain-specific supervision. The resulting models match performance on math benchmarks while achieving superior generalization on coding tasks, offering a potential alternative to current RL methods that struggle without verifiable rewards.

FutureHouse's Finch is a tool for data-driven discovery in biology, simulating a "1st yr grad student." Although AI promises to automate scientific steps, evidence of breakthroughs is lacking, with Finch yet to deliver a novel discovery. FutureHouse is seeking experts to enhance Finch's accuracy during its closed beta phase.

Anthropic's web search tool integrated with Instructor enables real-time data retrieval and structured outputs using Pydantic models. This combination allows applications to provide up-to-date information with cited sources, reducing the risk of outdated or incorrect information. Users can configure search constraints such as domain limitations, enhancing the tool's precision.

ChatGPT's memory shows that AI doesn't just respond, but truly grasps patterns in how we think.

Ambience Healthcare has a new medical coding model that could reduce billing mistakes and help clinicians and professional coders work more efficiently.

Meta is splitting its AI organization into two teams: AI Products, which handles Meta AI assistant and app features, and AGI Foundations, which covers Llama models and advanced capabilities.