Today in AI

The infrastructure wars are where the real money—and the real constraints—live right now. Cerebras just demonstrated trillion-parameter model deployment across its wafer-scale WSE-3T chips, using Kimi K3 as the showcase. This matters because it pushes against the NVIDIA monopoly narrative at the exact moment when inference economics are becoming existential. Speaking of which, DeepSeek's open-source DeepInfer is reportedly doing 155 trillion tokens per day per cluster at ~80% margins with 81-day payback periods—numbers that, if even directionally true, reshape how every lab thinks about serving costs. On the efficiency front, the RTK CLI proxy is a clever single-binary Rust tool cutting LLM token consumption 60-90% on developer workflows through compression and filtering; it's the kind of unsexy infrastructure that compounds fast across engineering teams.

Meanwhile, the interpretability community is having an identity crisis worth watching. A Princeton researcher's takedown argues mechanistic interpretability has been mostly flavor without nutritional value for safety—a "La Croix" of research, effervescent but empty. The counter-current is emerging: top-down interpretability proposes analyzing systems by behavior and function rather than tracing individual neurons, which feels more aligned with how we actually need to audit black-box models. This debate isn't academic; monitor alignment is failing for the same reason agents fail—misalignment, not technical inadequacy—meaning our oversight tools may be as hard to trust as what they're watching. The human comparative advantage may lie in "relevance realization," knowing what matters in context, though that feels like cold comfort if the monitors we build to catch AI misbehavior are themselves misaligned.

The AI graveyard keeps growing—42% of corporate AI initiatives abandoned, per TechCrunch's running tally that now includes Relay and Character.ai. The gap between infrastructure winners and application-layer casualties is widening. For those still building, Sebastian Raschka's 11-hour course on reasoning models from scratch arrives at a useful moment—reasoning is where the capability overhang currently lives, and understanding it beats betting on wrapper startups with 18-month lifespans. And in the "actually helping people" column, EPFL's LogSync neural chip uses machine learning for adaptive deep brain stimulation, decoding inter-region brain dynamics for psychiatric treatment—reminding us that not all impactful AI needs a chat interface.

**Bottom line:** The money is in inference infrastructure and efficiency, the intellectual action is in interpretability's methodological reboot, and the graveyard is full of companies that forgot both.