Agentopic is a novel agent-based workflow for explainable topic modeling that leverages the reasoning capabilities of Large Language Models (LLMs). It uses multiple agents for collaborative topic identification, validation, hierarchical grouping, and natural language explanation, addressing the lack of transparency in methods like LDA and BERTopic. On the BBC dataset, seeded Agentopic achieves an F1-score of 0.95, matching GPT-4.1, outperforming LDA (0.93), and close to BERTopic (0.98). Unseeded, it generates 2045 semantically coherent topics across six hierarchical levels, enriching the original five-category structure.

This paper presents an agentic AI framework for trip planning optimization in intelligent vehicles, using an orchestration agent to coordinate specialized agents for traffic, charging, and points of interest. It also introduces the Trip-planning Optimization Problems (TOP) dataset with ground-truth optimal solutions. Experiments show 77.4% accuracy on the TOP benchmark, outperforming single-agent and workflow-based multi-agent baselines.

This research proposes a causal framework to define and quantify collective agency, i.e., the potential for multiple simple agents to inadvertently form a collective with distinct capabilities and goals. Using causal games and causal abstraction, the authors solve an incentive puzzle in actor-critic models and quantitatively assess collective agency in different voting mechanisms. The framework aims to provide a foundation for understanding, predicting, and controlling emergent collective agents in multi-agent AI systems.

ARMOR 2025 is a safety benchmark grounded in the Law of War, Rules of Engagement, and Joint Ethics Regulation, designed to evaluate LLMs for defense applications. It features a 12-category taxonomy with 519 prompts and revealed critical safety gaps when tested on 21 commercial models.

A new variant of Direct Preference Optimization (DPO) called TUR-DPO incorporates reasoning topologies and uncertainty signals to better align LLMs with human preferences. It improves over standard DPO on multiple benchmarks while remaining reinforcement-learning-free and simple to train.

Research shows that tool-augmented reasoning does not always outperform native chain-of-thought in the presence of semantic distractors. The authors propose a Factorized Intervention Framework to isolate prompt formatting costs, tool-calling protocol overhead, and actual execution gains, introducing G-STEP to mitigate protocol-induced errors.

Researchers propose LOCA, a method that identifies minimal interpretable changes in intermediate representations to causally explain why specific jailbreak prompts bypass safety training in LLMs. It outperforms prior global approaches, requiring only six changes on average to induce refusal.

A new arXiv paper proposes AgentReputation, a three-layer decentralized reputation framework for agentic AI marketplaces, addressing strategic optimization, cross-task transfer, and verification heterogeneity.

TADI is an agentic AI system that transforms drilling operational data into evidence-based analytical intelligence. Applied to the Equinor Volve Field dataset, it integrates 1,759 daily drilling reports, WITSML real-time objects, 15,634 production records, and more into a dual-store architecture (DuckDB + ChromaDB). Twelve domain-specialized tools, orchestrated by a large language model, support multi-step evidence gathering. The system parses all XML files with zero errors and proposes the Evidence Grounding Score (EGS).

Nylas CLI is a new open-source tool that provides AI agents with real email accounts, calendars, and contacts across multiple providers (Gmail, Outlook, Exchange, Yahoo, iCloud, IMAP) through a single authentication flow. It uses the Model Context Protocol (MCP) to give agents 16 tools for sending mail, searching threads, creating events, and more.

Release of datasette-llm 0.1a7, introducing a mechanism for configuring default options for specific models. This allows plugins to set default parameters like temperature for all enrichment operations, streamlining LLM integration within Datasette.

New release of llm-echo plugin adds -o thinking 1 option to simulate reasoning blocks for testing LLM 0.32a0+.

John Gruber reveals that Y Combinator owns about 0.6% of OpenAI, worth over $5 billion at the current $852 billion valuation.

moar is a free Chrome extension that converts documents (PDF, DOCX, etc.) into clean Markdown or CSV for AI tools, with up to 95% token savings and 100% privacy.

Apple ML Research proposes Stochastic KV Routing, a training technique where layers randomly attend to their own or a preceding layer's KV states, enabling adaptive depth-wise cache sharing. This reduces KV cache memory footprint without throughput loss and can even improve performance in data-constrained settings.

OpenAI expands ChatGPT ads with a beta self-serve Ads Manager, CPC bidding, and enhanced measurement tools—built to protect privacy and keep conversations separate from ads.

In the Musk v. OpenAI trial, Greg Brockman's testimony and journal entries reveal evasive, pedantic behavior and a pattern of greed and opportunism.

Simon Willison experiments with 21 quantized variants of IBM's Granite 4.1 3B LLM, prompting each to generate an SVG of a pelican riding a bicycle. All results are poor with no correlation to model size.

Equity management firm Carta is an example of how enterprises can use AI and agents to transform and shape their businesses.

A quiet day in AI news allows reflection on the nature of AI 'character' sparked by an OpenAI employee's praise of Claude, contrasting it with GPT's utility-focused design. Plus coverage of Sierra's valuation, agent harness maturity, coding agent cost dynamics, and more.

Simon Willison's sponsors-only monthly newsletter for April 2026 covers new models like Opus 4.7 and GPT-5.5, Claude Mythos and LLM security research, ChatGPT Images 2.0, and more. The newsletter is available to sponsors at $10/month.

Microsoft reversed a VS Code change that automatically added 'Co-authored-by: Copilot' to commits, after developers complained it added attribution even when Copilot wasn't used. The fix makes it opt-in. Similar issues have occurred with Claude Code and Codex. The incident sparks discussions about AI attribution, copyright, and liability.

OpenAI and PwC are partnering to help enterprises use AI agents to automate finance workflows, improve forecasting, strengthen controls, and modernize the CFO function.

Albertsons Companies, one of America's largest food and drug retailers with 2,300 stores, built a centralized AI core based on a 'one team, one platform, one operating model' strategy. Under Sunil Gopinath, they leverage the Databricks platform for data engineering, ML, and analytics, using a franchise model to balance central governance with local innovation. Key focus areas include customer experience, merchandising intelligence, labor, and supply chain. The approach emphasizes reusable accelerators, AI-augmented engineering, and leadership accountability to scale AI across the enterprise.

Anthropic, Blackstone, H&F, and Goldman Sachs are launching a new AI services company to help mid-market businesses adopt Claude. The move mirrors OpenAI's 'The Deployment Company' model, highlighting that AI rollout requires extensive services and integration beyond the technology itself. Microsoft's Copilot maintains a distribution advantage.

Simon Willison explores Ville Laurikari's TRE regex engine by building an experimental Python ctypes binding with Claude Code and testing it against malicious regex attacks. TRE, due to its lack of backtracking, handles 'evil patterns' and large inputs far better than Python's standard re module, scaling linearly instead of exponentially.

Business leaders across industries rely on operational dashboards as the shared source of truth that their teams execute against daily. But dashboards are built to answer known questions. When teams need to explore further, ad-hoc, multi-dimensional, or unforeseen questions, they hit a bottleneck. They wait hours or days for BI teams to build new views or update reports. The Dataset Q&A feature bridges that gap. You can ask questions in natural language, get accurate answers in seconds, with no new dashboards to build, and no queue to wait in. Just an interactive conversation with your existing datasets, without disrupting the dashboards your teams already depend on.

Generate recommendations from production traces, validate them with batch evaluation and A/B testing, and ship with confidence. AI agents that perform well at launch don’t stay that way. New capabilities complete the observe, evaluate, improve loop for agent performance.

Amazon Bedrock AgentCore introduces a new optimization loop with recommendations, batch evaluation, and A/B testing to continuously improve AI agent quality in production, replacing manual debugging with data-driven iteration.

Amazon SageMaker AI introduces an agentic experience where developers describe their use case in natural language, and an AI coding agent streamlines the entire model customization journey from data preparation to deployment. Pre-built agent skills encode AWS expertise for fine-tuning techniques (SFT, DPO, RLVR), data transformation, evaluation, and deployment. This post explains the skills, demonstrates with a clinical reasoning model fine-tuning example, and covers customization of skills.