The problem with LLMs isn't hallucination, it's context specific confidence (www.signalfire.com)

OpenAI’s recent framing — that hallucinations are an inevitable byproduct of next-token pretraining because models learn to produce plausible text, not objective truth — reframes the problem: the failure isn’t that models “make things up,” it’s that they don’t reliably signal how confident they are when they do. The author argues confidence should be treated as a separate modality. Too much bluffing creates dangerous overconfidence; too much abstention makes models useless. The right behavior is context-dependent: precise, citation-backed answers for debugging or medical queries; imaginative, unconstrained output for fiction or brainstorming; and empathic, supportive framing for therapy-style interactions. Technically, this points away from chasing a single “non-hallucinating” model and toward post-training controls and routing. Anthropic’s persona vectors show a practical path: represent traits (e.g., propensity to hallucinate or be overconfident) as directions in activation space and steer them up or down to tune behavior. That enables realtime dials for creativity vs. factuality, calibrated abstention, and task detection + tool-calling pipelines. The implication for benchmarks and research is big: evaluate not just accuracy but calibration, hedging, and when models ask for help. Confidence calibration — making models express graded certainty rather than binary bravado or silence — emerges as a crucial frontier for trustworthy, useful LLM systems.