Jason Says
Today's papers are dense with signal: from pricing the moral cost of agent actions to surgical non-compliance in VLMs, AI safety is finally becoming a measurable engineering discipline — developers who ignore these papers will be blindsided by the next wave of compliance requirements.
cathrynlavery/diagram-design ships 38 editorial diagram types as self-contained HTML+SVG for Claude Code and Codex — no Mermaid, no shadows, designer-approved. v2.5.10 adds 10 layout grammars including Sankey and flywheel. A plug-and-play Skills asset for agents that need to produce publication-quality visuals.
Dr. Claw is an open-source workspace that wraps coding agents like Claude Code in a controllable, human-in-the-loop workflow with persistent state and full auditability. It doesn't add another autonomous agent — it makes existing ones traceable. Critical for developers and researchers who need reproducible, explainable AI-assisted workflows.
HarvestBench is the first benchmark to put a monetary price on agent-caused harm to living creatures. In a farm simulation, LLM agents must weigh detour costs against hitting animals — with no mention of harm in the goal. A novel tool for developers building high-stakes agents who need to measure implicit value alignment, not just rule-following.
Real-world queries mix answerable and unanswerable content, yet safety benchmarks treat each request as all-or-nothing. This paper introduces selective non-compliance for VLMs — enabling models to respond to valid parts of a query while refusing specific unsafe or infeasible components. Directly relevant for devs building multimodal products that need nuanced safety without over-refusal.
LLMs routinely refuse benign queries that superficially resemble harmful ones, frustrating users. This paper structurally analyzes safety-tuning responses to identify root causes of false refusals and proposes fixes. Essential reading for any developer whose production LLM gets complained about for being overly cautious — this offers a principled path to better helpfulness-safety balance.
UniMate is a unified foundation model that generates articulated motion for any rigged 3D skeleton from a text prompt — no per-skeleton fine-tuning, no reference motion needed at inference. It removes the last major bottleneck in the auto-rigging pipeline, making it a significant unlock for game devs, virtual avatar creators, and AIGC content studios.
Latent Space launches the first Frontier AEO Tracker, analyzing which tools and products frontier models like GPT-6 Astra default-recommend without explicit prompting. For indie devs and SaaS founders, this surfaces a brand-new growth variable: whether your product is 'remembered' by AI models is becoming the next SEO battleground.
OpenAI, AIRPPU, and WAN-IFRA launch a joint AI program to help Ukrainian news organizations innovate and survive under wartime pressure. This marks OpenAI's first large-scale involvement in wartime media infrastructure — a notable expansion of AI tools into democratic resilience, with implications for future policy and licensing frameworks.
Subscribe for daily AI updates + free playbook
📘 Subscribe Free