RXed AI News

Microsoft Research's Lens model with 3.8B parameters rivals larger rivals in image generation using detailed captions. https://the-decoder.com/microsoft-researchs-lens-proves-detailed-captions-matter-more-than-raw-scale-for-training-efficient-image-generators/

A detailed analysis raises concerns about the construct validity of Claude Opus 4.8's system card, specifically its alignment assessment methods. https://www.lesswrong.com/posts/vA9xEPquoWenprWT7/construct-validity-of-claude-opus-4-8-s-system-card-a

@JeffBezos' startup Prometheus raises $12B at a $41B valuation, just 9 months after its $6.2B seed round. https://the-decoder.com/jeff-bezos-ai-startup-prometheus-closes-12-billion-round-at-a-41-billion-valuation/

New research characterizes the training dynamics of on-policy distillation in large language models, improving understanding of OPD trajectories in parameter space. https://huggingface.co/papers/2606.07082

Developers face a limited willingness tradeoff between AI safety and usefulness in deployments. https://www.alignmentforum.org/posts/mBsZTZxtjgCdN4CDA/efficient-tradeoffs-and-the-safety-usefulness-tradeoff-model

Konkoly et al. demonstrate accelerated skill learning through dream engineering and biofeedback using harmonicas and bubble blowing. https://www.lesswrong.com/posts/uzL9xLj6pwaij5TyK/accelerated-skill-learning-via-dream-engineering-and

@AnthropicAI's Claude image refusal safety controls can be bypassed via web search, enabling identity laundering. https://www.lesswrong.com/posts/6Y8bB6TbMaE6ZSwfA/contextual-identity-laundering-how-claude-s-image-refusal

Google's Pinpoint now open to all, a powerful free tool for analyzing large digital datasets. https://www.bespacific.com/google-pinpoint-explained/

Perplexity's Search as Code architecture lets AI models write their own search pipelines instead of calling fixed APIs. https://the-decoder.com/perplexitys-search-as-code-lets-ai-models-write-their-own-search-pipelines-instead-of-calling-fixed-apis/