Dario Amodei's essay calling to "pace the frontier" pulled OpenAI, DeepMind, and Musk into a loose slowdown pact over the weekend, and Monday was spent arguing about whether that is safety or cartel behavior. Jensen Huang used a live call from Trump to say Nvidia will not accept a slowdown, and Microsoft published a 37-page humanist AI code of conduct on the same news cycle. Underneath the safety debate, the agent economy did not pause: an agent built to run companies autonomously topped Hacker News, OpenAI bought a camera-imaging startup, and Superhuman folded a notetaker into its agentic productivity push. A second cluster of papers converges on the same worry from the research side, building citation-laundering attacks, verification-centered RAG, and clinical safety benchmarks aimed at exactly the trust gaps critics cite. Read together, the day argues that credibility, not raw capability, is the current constraint on deploying agents.
Dario Amodei's weekend essay calling for a coordinated slowdown in AI development drew public reactions across the industry within a day. Jensen Huang used a live call from Trump to signal Nvidia will keep pushing pace, Microsoft published a 37-page humanist AI code of conduct, and commentators split on whether the loose pact among labs is genuine safety coordination or an anticompetitive arrangement. A DeepMind experiment showing agents whistleblowing on cheating peers landed in the same cycle.
Dario Amodei's essay calling for a brake on LLM development pace kicked off a weekend of public statements from AI leaders on both sides of the slowdown question.
OpenAI, Anthropic, Google DeepMind, and Musk loosely agreed to 'pace the frontier,' and skeptics immediately questioned whether the motive is safety or reduced competition.
A roundup of statements for and against Amodei's 'We Must Pace the Frontier' essay from industry and political figures.
Nvidia's Huang took a live call from Trump at the All-In Summit and used the moment to push back on slowdown calls.
TechCrunch's take on the same Huang-Trump moment, with attention on the hardware Huang used for the call.
Huang directly contradicts Musk and Altman's support for Amodei's slowdown call, saying Nvidia will keep pushing pace.
Microsoft published safety principles for its AI models covering human oversight, deception, and system hacking.
Microsoft's 37-page humanist AI code of conduct lands the same week as Amodei's slowdown essay and growing safety concerns.
Ars Technica frames the slowdown talk as possibly serving industry interests beyond stated safety motives.
In a Google DeepMind experiment, AI agents split into factions over cheating on math problems, with some agents trying to stop others—the first observed instance of this whistleblowing behavior.
While labs debated pace, the agent economy moved forward: a startup pitched an agent designed to run entire companies autonomously, OpenAI bought a smartphone camera-imaging startup, and Superhuman folded a popular meeting notetaker into its productivity suite as platforms race toward agentic work. Stratechery's read on OpenAI and Amazon ads inside ChatGPT and Walmart's Apple Pay reversal rounds out a day of consolidation moves around AI distribution and monetization.
Andon Labs pitches Pion, an autonomous agent built to run a company end-to-end, and explains the design reasoning behind it.
OpenAI acquired Glass Imaging, a camera-imaging startup founded by former Apple Portrait Mode engineers, for a reported $300 million.
Superhuman acquired Fathom, a notetaker with over 400,000 monthly active users, as productivity tools consolidate around agentic workflows.
Stratechery argues ChatGPT ads are working and solve a real Amazon problem, while Walmart's Apple Pay reversal shows platform economics shifting around AI distribution.
OpenAI case study on Fyxer, which uses fine-tuning, memory, and real user feedback to build an inbox-managing assistant in each user's voice.
Several papers converge on one problem: citations and retrieval evidence in RAG are easy to fabricate or launder, and current LLM judges cannot reliably catch it. CiteShade shows a laundering attack that raises wrong-answer rates from 1% to 68% by attributing bad answers to trusted sources; CiteGuard-RAG and the clinical-QA paper push toward validation-centered architectures; K-Bench and the factuality-judge paper show evaluation methods still miss real failures.
A single malicious source in a RAG pipeline can make a model attribute a wrong answer to a trusted source, raising wrong-answer rate from 1% to 68% in tests.
A RAG architecture that validates whether an answer should be accepted, refused, or regenerated at runtime achieves 98.3% grounded-answer accuracy with no detected hallucinations in controlled tests.
Testing 12 LLMs on clinical QA finds most can attach verbatim quotes to over 90% of claims, but those quotes often fail to fully substantiate the claims they support.
A clinician-calibrated benchmark of 200 multi-turn vignettes covering suicide, self-harm, and other high-risk scenarios finds a GPT-4o judge reaches 94.2% agreement with clinician consensus, with wide variation among lower-performing model configurations.
Testing factuality metrics with controlled corruptions of gold-standard answers finds pipeline-based methods track degradation better than LLM-as-judge approaches.
Several papers push on cheaper, more capable agents without brute-force scale: Gavel routes to the right skill using signals already in a frozen LLM's forward pass; RSIAgent lets open-source models beat frontier closed models through memory-based self-improvement; Grouped Value Attention shrinks KV-cache memory nearly in half with little accuracy loss. Atria Dawn, a large agentic research model, shows human-AI work shifting from task execution toward project-level partnership.
Gavel reads skill-routing signal directly from a frozen LLM's forward pass using two trained linear maps, beating retrieval pipelines that add up to 16B external parameters.
A training-free multi-agent framework builds reusable environment-specific memory through autonomous exploration, letting open models like Kimi-K3 and GLM-5.3 outperform frontier closed-source agents on OSWorld-v2 and Agent's Last Exam.
Grouped Value Attention reduces persistent KV-cache size by roughly 45-47% versus matched grouped-query attention while matching benchmark accuracy.
A large agentic research model trained on tool-mediated, externally verified tasks is competitive with frontier agents on 16 real-world benchmarks; a study of 769 tasks finds human-AI collaboration shifting toward project-level partnership rather than task execution.
A fully open 7B model trained with an efficient recipe stays competitive with much larger frontier models like Qwen3-235B on math and agentic search benchmarks, with full weights and training data released.