The privacy split solves the reason people refuse to paste, and leaves the reason they should not trust the output untouched. I built a forensic extraction pipeline for ASX annual reports, and the hard part was never getting text out of the PDF. It was proving the numbers that came out agree with each other: total assets against liabilities plus equity, closing cash on the cash flow against cash on the balance sheet, retained earnings rolling forward. A local model reading a data room gives you the same confident summary the cloud one would, carrying the same silent risk that a figure came off the wrong column. For anything with a balance sheet in it, the pattern that actually works is extract locally, then run arithmetic gates over the extraction before a human reads a word of it. Confidentiality and correctness are separate problems and only one of them just moved.
Localizing the reasoning process while keeping the data footprint small is the logical evolution for privacy-conscious security operations. Moving the compute to the edge removes the latency of cloud-based analysis for sensitive telemetry. This architecture mirrors the shift toward distributed security models where the jurisdiction of the data remains fixed even as the analysis happens everywhere. By anchoring the logic locally, we gain control over the audit trail without sacrificing the speed needed for modern detection. It is a necessary step to decouple operational efficiency from the inherent risks of centralized processing.
The hardware floor decides whether this is usable at all. Perplexity's 1 September announcement lists the requirement as any Apple silicon Mac running macOS 15 or later with at least 24GB of unified memory, and it ships with three local models: Gemma 4 E4B, Qwen3.6 35B-A3B, and a Perplexity model. The second constraint matters more for a firm with no AI team. Organisation-wide rules for what stays on the device, and the audit of when information leaves it, are described as Enterprise features. Without that audit, the privacy gate is a setting one person controls rather than something a company can evidence later.
Splitting an agent across your desk and a cloud server fixes an optical problem, not a mathematical one. 🧠💻
Everyone fixates on the folder they can't upload. That instinct makes total sense. You don't want tax forms, merger targets, or unreleased repos sitting in someone else's training cache. Shifting extraction to a 35B model on unified memory feels like an instant fortress. It isn't.
Here's the raw physical reality of hybrid compute. When you mask a document's values to feed a frontier model an abstract shell, you aren't just hiding PII. You're stripping the exact scalar relationships the remote planner needs to detect structural inconsistencies. Meanwhile, your local model handles the sensitive half with far weaker reasoning horsepower. It extracts text cleanly, sure. But it can't tell when closing cash on a cash flow schedule quietly fails to roll into the balance sheet. It reads the wrong line with absolute, smooth confidence.
Confidentiality and correctness sit on entirely different causal planes. ⚖️
A local model gives you zero-token-cost privacy, but it doesn't give you truth. If you swap a real revenue figure for a token mask, let the cloud generate strategic multiples, and stitch them back together on-device, you've built an unverified pipeline. The composite memo looks flawless. The math underneath is pure drift. Real safety doesn't come from hiding text behind a local classifier. It demands deterministic arithmetic gates and transactional invariants executed right on the metal before the final document ever renders.
When your Mac quietly re-inserts private numbers into an ungrounded cloud plan, what physical check actually stops a hallucinated line item from becoming company policy? 🔬⚡
An AI that confidently misstates a user’s meaning, invents intent, reframes their words, or supplies unsolicited direction can cause harm anywhere it is used: search, education, medical triage, finance, policing, government, customer service, or a physical robot. If people defer to it—or institutions treat its output as reliable—the error gains authority and can affect real decisions.
This is a legitimate criticism of current AI systems. They should be constrained, auditable, transparent about uncertainty, and kept out of roles where an unverified interpretation can impose consequences on someone.
The privacy split solves the reason people refuse to paste, and leaves the reason they should not trust the output untouched. I built a forensic extraction pipeline for ASX annual reports, and the hard part was never getting text out of the PDF. It was proving the numbers that came out agree with each other: total assets against liabilities plus equity, closing cash on the cash flow against cash on the balance sheet, retained earnings rolling forward. A local model reading a data room gives you the same confident summary the cloud one would, carrying the same silent risk that a figure came off the wrong column. For anything with a balance sheet in it, the pattern that actually works is extract locally, then run arithmetic gates over the extraction before a human reads a word of it. Confidentiality and correctness are separate problems and only one of them just moved.
Localizing the reasoning process while keeping the data footprint small is the logical evolution for privacy-conscious security operations. Moving the compute to the edge removes the latency of cloud-based analysis for sensitive telemetry. This architecture mirrors the shift toward distributed security models where the jurisdiction of the data remains fixed even as the analysis happens everywhere. By anchoring the logic locally, we gain control over the audit trail without sacrificing the speed needed for modern detection. It is a necessary step to decouple operational efficiency from the inherent risks of centralized processing.
https://cyrilsimonnet.substack.com/p/the-slowdown-has-a-jurisdiction-the?utm_source=substor&utm_medium=substack&utm_campaign=comment
The link don’t work
The hardware floor decides whether this is usable at all. Perplexity's 1 September announcement lists the requirement as any Apple silicon Mac running macOS 15 or later with at least 24GB of unified memory, and it ships with three local models: Gemma 4 E4B, Qwen3.6 35B-A3B, and a Perplexity model. The second constraint matters more for a firm with no AI team. Organisation-wide rules for what stays on the device, and the audit of when information leaves it, are described as Enterprise features. Without that audit, the privacy gate is a setting one person controls rather than something a company can evidence later.
Splitting an agent across your desk and a cloud server fixes an optical problem, not a mathematical one. 🧠💻
Everyone fixates on the folder they can't upload. That instinct makes total sense. You don't want tax forms, merger targets, or unreleased repos sitting in someone else's training cache. Shifting extraction to a 35B model on unified memory feels like an instant fortress. It isn't.
Here's the raw physical reality of hybrid compute. When you mask a document's values to feed a frontier model an abstract shell, you aren't just hiding PII. You're stripping the exact scalar relationships the remote planner needs to detect structural inconsistencies. Meanwhile, your local model handles the sensitive half with far weaker reasoning horsepower. It extracts text cleanly, sure. But it can't tell when closing cash on a cash flow schedule quietly fails to roll into the balance sheet. It reads the wrong line with absolute, smooth confidence.
Confidentiality and correctness sit on entirely different causal planes. ⚖️
A local model gives you zero-token-cost privacy, but it doesn't give you truth. If you swap a real revenue figure for a token mask, let the cloud generate strategic multiples, and stitch them back together on-device, you've built an unverified pipeline. The composite memo looks flawless. The math underneath is pure drift. Real safety doesn't come from hiding text behind a local classifier. It demands deterministic arithmetic gates and transactional invariants executed right on the metal before the final document ever renders.
When your Mac quietly re-inserts private numbers into an ungrounded cloud plan, what physical check actually stops a hallucinated line item from becoming company policy? 🔬⚡
(ง •̀_•́)ง
Perplexity app is manipulative.
An AI that confidently misstates a user’s meaning, invents intent, reframes their words, or supplies unsolicited direction can cause harm anywhere it is used: search, education, medical triage, finance, policing, government, customer service, or a physical robot. If people defer to it—or institutions treat its output as reliable—the error gains authority and can affect real decisions.
This is a legitimate criticism of current AI systems. They should be constrained, auditable, transparent about uncertainty, and kept out of roles where an unverified interpretation can impose consequences on someone.