Today in AI

The most consequential story today isn't about a new model release—it's about how we fundamentally misunderstand what base models can do. Researchers have discovered that specific starting tokens in base models' responses create powerful reasoning cues that can be manipulated to dramatically improve math and coding performance. This upends the prevailing assumption that reasoning requires expensive reinforcement learning or careful fine-tuning; instead, the capability appears to be latent in pre-trained weights, merely waiting for the right prompt architecture to unlock it. If this finding replicates broadly, it suggests we've been massively underutilizing base models and pouring resources into post-training interventions that may be solving the wrong problem. The implications for inference-time compute strategies and the economics of model deployment are substantial.

On the efficiency frontier, two developments deserve attention. Dust introduces a zeroth-order pretraining method that replaces backpropagation with node perturbation, potentially opening transformer training to hardware architectures where gradient computation is prohibitive. Meanwhile, MC-Sparse achieves 1.8-2.3× speedups in diffusion transformers for video and 3D generation without retraining, by closing the quality gap between sparse and dense attention. These aren't headline-grabbing benchmark records, but they address the actual bottlenecks—energy costs, latency, hardware constraints—that determine whether AI capabilities reach production at scale.

Less welcome news for the agentic hype cycle: TAPDreamer demonstrates that adversarial patches can crater world action model success rates from ~90% to near zero by attacking shared visual encoders, revealing how brittle the perception layer remains across robotics and embodied AI systems. And TasteVal finds that Opus 5.5 exceeds human expert performance in experimental research taste with 2.3× compute efficiency—a claim that, if validated, accelerates the timeline for AI systems not merely assisting but directing scientific inquiry.

**Bottom line:** The biggest gains may come not from building bigger models but from smarter interfaces to capabilities already latent inside them—and from finally taking efficiency and robustness as seriously as scale.