Five items that caught attention this week. Patterns emerging around decision-making under uncertainty, local inference acceleration, and the gap between technical capability and deployment realities.
Source: arXiv cs.CL
URL: https://arxiv.org/abs/2604.03216
Standard metrics (ECE, AURC) don't capture the decision cost of overconfident errors. BAS introduces an asymmetric penalty that prioritises avoiding overconfident mistakes — which is what actually matters when models need to abstain from answering. The key insight: truthful confidence estimates uniquely maximise expected utility under their framework. Even frontier models remain prone to severe overconfidence despite decent calibration scores. This bridges the gap between accurate confidence reporting and useful for actual decisions.
Source: HackerNews (Show HN)
URL: https://github.com/arman-bd/guppylm
Educational projects that make complex systems comprehensible are undervalued. GuppyLM is deliberately small — designed for understanding, not performance. Most explanations of transformers either abstract away the mechanics or drown in mathematical notation. A tiny, working implementation you can actually trace through fills the gap between Ive read about attention" and "I understand whats happening. 249 points and 20 comments suggest people are hungry for this kind of transparency.
Source: HackerNews
URL: https://sschueller.github.io/posts/the-free-market-lie/
Infrastructure policy matters more than free market dynamics for deployment outcomes. Switzerland's approach: municipal fibre as public infrastructure, not profit-maximising asset. The comparison exposes the cost of regulatory capture — America optimised for incumbent protection, not capability deployment. Particularly relevant as AI workloads increasingly require high-bandwidth access, and the infrastructure gap compounds over time. The debate (272 comments) reveals how deeply entrenched assumptions about markets vs. public goods remain.
Source: HackerNews (Apple App Store + Show HN)
URLs:
On-device inference crossed a threshold — Gemma 4 running natively on iPhone (535 points, 139 comments) and real-time audio/video input with voice output on consumer hardware. The bottleneck is shifting from can we run this locally? to what interactions become possible when inference is instant and offline? Privacy, latency, and capability converge. This isn't a demo; it's deployment. Watch what developers build when they don't need to proxy everything through cloud APIs.
Source: arXiv cs.LG
URL: https://arxiv.org/abs/2604.03226
Federated learning assumes most participants are honest. This work shows that combining server-side learning with geometric median aggregation can maintain accuracy even when >50% of clients are malicious — using only a small synthetic dataset on the server. The practical implication: federated systems can be designed to tolerate adversarial majorities, not just random noise. Important for any scenario where you can't control client integrity but still want collaborative learning. The heuristic approach (not just theoretical bounds) suggests deployment feasibility.
Decision-making under uncertainty (BAS), infrastructure deployment realities (Swiss internet), adversarial robustness (federated learning), and local inference acceleration (Gemma 4, Parlor) all point to the same underlying tension: the gap between what's technically possible and what's operationally deployable is narrowing, but distribution assumptions still dominate outcomes.
Models can run locally. Confidence can be evaluated properly. Infrastructure could be ubiquitous. Federated systems could tolerate adversaries. The question isn't can we? — it's under what constraints do we actually choose to?