Two papers from Nature Communications lead today: EndoRare, which synthesizes rare gut lesions from a single example, and a memory-based theory of how diffusion models start to generalize. In the news: ARC Prize scores GPT-6 Astra at 62.7%, not 99.9%, Replit moves its AI spending to OpenAI on price, and Pennsylvania signs an AI ad disclosure law.
ARC Prize, the nonprofit behind the ARC-AGI-3 reasoning benchmark, says OpenAI's headline 99.9% score for GPT-6 Astra came from OpenAI's own provider adapter harness, while its standard provider-neutral harness scored the same model at 62.7% at maximum reasoning effort. The gap comes from what the model may carry between requests: the adapter preserves opaque reasoning state and uses compaction, while the standard harness keeps only notes the model chooses to preserve. ARC Prize still calls the 62.7% result a step change in interactive reasoning.
OpenAI offered Replit a discount of more than 30% on a contract worth over $100 million a year, The Information reported, winning most of Replit's AI spending from Anthropic after Anthropic declined to extend Replit's earlier discount. At Rippling, OpenAI's share of engineers' coding work rose from 16% in May to 52%, while Anthropic's fell from 57% to 18%. Neither lab has publicly confirmed the deal terms.
Pennsylvania Governor Josh Shapiro signed the Artificial Intelligence Deceptive Advertising Act on October 22, making Pennsylvania one of the first US states to require labels on ads whose images, video or audio were significantly modified or generated by AI in a way that creates a false impression about the product. The law, Act 99 of 2026, takes effect December 21, with civil fines capped at $3,000 per case. Text-only content and evident parody, satire or commentary are exempt.
Rare gastrointestinal lesions appear too seldom in routine endoscopy to train reliable AI models or to teach novice clinicians. EndoRare is a one-shot, retraining-free generative framework that synthesizes diverse lesion exemplars from a single reference image, separating pathognomonic lesion features from non-diagnostic attributes with language-guided concept disentanglement. Across four rare pathologies, calcifying fibrous tumor, juvenile polyposis syndrome, familial adenomatous polyposis and Peutz-Jeghers syndrome, experts judged the synthetic images clinically plausible, and augmentation improved downstream classifiers at low false-positive rates. In an exploratory reader study, novice endoscopists exposed to generated cases improved in image-level recognition.
Rare diseases are where medical AI usually fails first, because there is almost nothing to learn from. A method that turns one real case into useful training data, for both models and people, could extend good diagnostics to the conditions that need them most.
Dense associative memories, generalizations of Hopfield networks, store training data at local minima of an energy landscape, and when data exceeds a critical capacity, new minima appear that are not training samples. The authors read diffusion models through this lens, treating generation as memory retrieval: with little data, diffusion models form one attractor per sample, and as data grows they transition to generalization. Between the two sits a critical intermediate phase where spurious states emerge, previously dismissed as retrieval failures, which the authors show are the first signs of generative capability across architectures and datasets.
Memorization versus generalization is the oldest argument about generative models, and this gives it a mechanism with a measurable turning point. If spurious states mark the onset of imagination, researchers gain a principled way to watch a model cross from copying to creating.
Today's new pathology paper is in the Selections above. No further standalone papers in this column today.
No new agentic AI papers in today's window. Recent issues carry related work on multi-agent systems and self-improvement.
No new AI fairness papers in today's window. Recent issues carry related work on bias and fairness in medical and general settings.
Today's new generative modeling paper is in the Selections above. No further standalone papers in this column today.