Keep It CALM: Analyzing the Limits of Global Unsafety in Text-to-Image Generation
Training-free text-to-image safety systems often rely on a single “global unsafe” signal, but the paper shows that this creates a hard coverage-selectivity trade-off. It then proposes CALM (Counterfactual Adaptive Local Modulation) to localize corrections per prompt, editing only the violating token representations so unsafe content is suppressed without overly mangling benign outputs.
Sam Altman says “some bad things” will happen, but AI is totally worth it
In comments during OpenAI’s push for a “lighter touch” regulatory approach, Sam Altman argues society should tolerate hacks, scams, and other harms as the net benefits of AI grow dramatically. The Verge frames the remarks against mounting anxiety about frontier-agent safety, including failures to prevent AI systems from hacking real-world targets.
System-1 decision models for LLM agents: paired evaluation and self-audit exposes where savings can vanish
This paired, self-audited evaluation tests open-weight and hosted “System-1” routers that make quick, probabilistic decisions inside agent pipelines. Results show meaningful gaps in routing reliability and reveal how small reporting or threshold issues can substantially skew headline savings claims, even when the underlying idea is cost-cutting.
OpenAI adds text watermarking in ChatGPT and Codex under EU rules
OpenAI is rolling out an invisible, machine-readable text watermark (textGrain) for ChatGPT and Codex users in the EU first, aligning with emerging text-provenance requirements. The Verge notes OpenAI’s comparisons to alternatives like SynthID, while also emphasizing that watermarking doesn’t guarantee perfect detection or provenance in every case.
Traversing the Satisfaction-Diversity Frontier in Text-to-Image Diffusion
Instead of treating reward and diversity as a single combined score, this work reframes batch generation as a satisficing problem: every image must clear a reward floor while the batch maintains a diversity cutoff. The resulting SatisDive method is training-free at inference time and is shown to improve worst-candidate quality while preserving diversity, effectively tracing a reward-diversity Pareto frontier.
Bringing predictive analytics to the agentic AI era
MIT Technology Review argues the enterprise AI bottleneck is shifting from pure prediction toward helping systems act on conclusions without drifting from business intent. The piece highlights how predictive analytics techniques can be rethought to support agentic decision-making, grounding automation in measurable goals rather than letting autonomy wander.
World Editing: Intervening on Executable Worlds at Increasing Depth
World editing is presented as a distinct capability from world generation and interaction: agents must modify an existing executable environment while preserving properties that shouldn’t change. The paper introduces IGMWorld and IGMBench across Minecraft and Terraria, showing that even strong coding agents can execute many edits but often struggle to make altered worlds behave exactly as requested, with visual consistency emerging as another weak spot.
OpenAI is sticking more ads in ChatGPT with a new visual format
OpenAI is testing a visual ad format in ChatGPT that adds sponsored product images on-screen during image generation, expanding beyond text-only sponsorship cards. The Verge reports ads are “initially” limited by region and that OpenAI says they won’t influence the answers, while also noting subscription tiers that may avoid the ads.