Perplexity Assistant for Android vs Sup AI

“The primitive here is a hybrid inference router — compressed model runs locally, routes to cloud when the query exceeds local capability. That's a real engineering decision, not a marketing one, and the tradeoff is honest: you lose fidelity on hard questions but gain offline availability on simple ones. The DX for end users is cleaner than I expected — no configuration, the routing is invisible. What I can't verify is the boundary: Perplexity hasn't published the model architecture, compression ratio, or the heuristic for when it escalates to cloud, so the 'offline reasoning' claim is partially a black box. Ships because the hybrid routing pattern is the right bet; would ship harder if they opened the model card.”

“The HLE claim needs independent verification, but the underlying ensemble approach is architecturally sound for factual Q&A tasks. Running 339 models is expensive — pricing will be the gating factor for production use. The $10 free credit is a fair trial.”

“The category is AI assistant with on-device inference, and the direct competitor is Google Assistant with Gemini Nano — which already runs on-device on Pixel hardware and has deeper Android integration than any third-party app ever will. Perplexity's wedge is search quality and the hybrid routing, which is genuinely better than Gemini Nano's offline capabilities today, but that gap closes the moment Google ships Gemini 2.x natively to assistant. The scenario where this breaks: any power user who relies on the Calendar and Gmail integrations will hit permission friction and edge-case failures that Google's first-party integrations don't have. What kills this in 12 months: Google ships this natively and Perplexity's differentiation collapses to brand loyalty among users who already pay for Pro.”

“Extraordinary claims require extraordinary evidence. A 7.41 point jump on HLE via ensembling — without publishing methodology — smells like benchmark gaming. The latency of running 339 models in parallel is also a real concern for anything other than async research tasks.”

“The thesis here is falsifiable: by 2028, on-device inference becomes the default mode for personal assistant queries, and cloud becomes the exception for heavy reasoning rather than the rule. Perplexity is early to this — Qualcomm's NPU roadmap and Apple's on-device model investments confirm the trend line is real, but most assistants still phone home for everything. The second-order effect that matters: if on-device reasoning normalizes, the surveillance economics of cloud AI assistants get disrupted — users who care about query privacy get a credible alternative without sacrificing capability. The dependency that has to hold: compressed models keep improving fast enough that 'on-device quality' stops being a polite euphemism for 'noticeably worse.' Right now that gap is still real.”

“Model ensembling is an underexplored direction in the race to reduce hallucination. If Sup AI's approach scales, it could be more durable than fine-tuning individual models — you get the wisdom of the crowd across model families, training data, and architectures simultaneously.”

“The job-to-be-done is ambiguous: is the user hiring this to replace Google Assistant, to do offline search, or to get a smarter calendar and email integration? The answer requires 'and,' which is a focus problem. Onboarding presumably involves setting Perplexity as the default assistant and granting Calendar and Gmail permissions — that's a multi-step trust ask before the user has seen a single moment of value, and most users will drop before completing it. The completeness problem is real: this only replaces Google Assistant if the Android system action integrations are deep enough to handle the full surface area of things users actually ask their phone assistant to do, and third-party assistants have a 10-year track record of failing exactly that completeness bar. The gap between what's shipped and what's needed is reliable system-action breadth, not more reasoning capability.”

“For creative work, ensemble outputs tend to regress toward the mean — you get the most-agreed-upon version of something, which is usually the least interesting version. This is a tool for factual accuracy, not creativity. I'd stick with a single strong model for writing.”