SIGNAL

Noetik is applying large language models (specifically autoregressive transformers like TARIO-2) to address the 95% clinical trial failure rate in oncology by reframing the problem as a patient-treatment matching issue rather than a drug efficacy problem. The approach uses AI to better predict which patient populations are most likely to respond to specific cancer therapies, potentially reducing failed trials caused by poor cohort selection. This matters because if matching rather than efficacy drives most failures, AI-guided patient stratification could dramatically accelerate drug development cycles and reduce the $2.6B average cost per approved cancer drug.

Scaling laws—which predict how model performance improves with compute—are now being systematically studied in reinforcement learning (RL) contexts for LLMs, moving beyond their traditional application to pretraining. This matters because RL fine-tuning (used in models like GPT-4o and Claude) has become critical for alignment and capability gains, yet lacked the predictive frameworks that made pretraining scaling predictable. Understanding RL scaling laws enables better resource allocation decisions and helps forecast whether additional compute investment will yield meaningful performance improvements in post-training phases. The shift reflects the industry's recognition that gains from raw pretraining scaling are plateauing relative to returns from optimized fine-tuning strategies.

Google's Gemma 4 31B model has been benchmarked across three quantization formats (FP8, NVFP4, and INT4) to evaluate trade-offs between inference speed, memory footprint, and output quality. INT4 quantization achieves the smallest VRAM footprint (~8GB), while FP8 maintains closer fidelity to the full-precision baseline with moderate compression (~16GB). The results matter for deployment decisions: INT4 enables running the model on consumer GPUs, while FP8 suits use cases prioritizing accuracy where VRAM permits the additional overhead.

Schools are grappling with uncontrolled AI tool adoption among staff and students, creating governance and security risks across educational institutions. The updated VINE GenAI Guidelines introduce a three-zone classification system that categorizes AI tools by approval status and risk level, shifting from prohibition to managed integration. This framework prioritizes transparency and risk assessment over punitive measures, designed to encourage legitimate use while maintaining institutional oversight. The approach reflects a broader institutional challenge: banning AI tools often fails, making structured governance frameworks more pragmatic than restriction.