Systems Generate Faster Than Humans Can Verify

Interactive visualizations choke on datasets with millions of records. Users expect real-time filtering and highlighting, but selections across multiple components create performance bottlenecks that break the interaction loop.

Method: Mosaic Selections uses a declarative model where selections are represented as predicate expressions that get optimized through query rewriting. Instead of brute-force filtering, it applies cross-filter optimizations—when users select data in one view, the system pre-computes which records remain valid in other views, reducing redundant scans. The framework supports set operations (union, intersection, negation) across components and batches updates to minimize database round-trips.

Caveats: Requires restructuring existing visualization code to use the predicate model. Performance gains depend on database query optimizer quality.

Reflections: How does this model extend to collaborative filtering where multiple users make concurrent selections? · Can the predicate optimization approach generalize to spatial selections in geographic visualizations? · What's the performance ceiling when combining 10+ complex predicates with temporal range queries?

Millions of patients already use LLM chatbots for medical advice, but no systematic safety evaluation exists. Physicians worry about liability when patients follow harmful AI recommendations.

Method: Physicians red-teamed four major chatbots (Claude, Gemini, GPT-4o, Llama3-70B) on 222 patient-posed questions from the HealthAdvice dataset. They evaluated 888 responses using a framework that flags unsafe advice—recommendations that could cause immediate harm or delay necessary care. GPT-4o had the lowest unsafe response rate at 18%, while Llama3-70B reached 28%. Unsafe answers included recommending home remedies for chest pain and suggesting patients skip emergency care for severe symptoms.

Caveats: Evaluation focused on English-language questions from US context. Safety thresholds may differ across medical systems and cultural norms.

Reflections: How do safety rates change when patients ask follow-up questions that challenge initial advice? · Can fine-tuning on physician-reviewed responses reduce unsafe outputs without degrading helpfulness? · What liability framework should platforms adopt when patients act on chatbot medical advice?

Full-alphabet EMG decoding remains unreliable for motor-impaired users. Existing systems require 26+ gesture classes, but classification accuracy degrades rapidly as class count increases, making practical text entry impossible.

Method: LowKeyEMG reduces the problem to 7 gesture classes decoded from surface electromyography: 4 gesture keys map to character groups, plus 3 for space and system control. Users type by combining gestures in sequence—similar to T9 predictive text but driven by muscle signals instead of button presses. The system achieved 15.8 words per minute with able-bodied users and remained functional for individuals with motor impairments who couldn't reliably produce 26 distinct gestures.

Caveats: Requires users to learn the character-to-gesture mapping. Performance degrades when muscle fatigue changes signal patterns during extended sessions.

Reflections: Can adaptive models retrain gesture classifiers in real-time as muscle fatigue progresses? · How does performance compare to other reduced-input methods like eye-tracking or sip-and-puff for users with severe motor impairments? · What's the optimal keyset size that balances classification accuracy with typing speed?