Notes from
the workshop.

Helios Search — dual-engine, verified for truth.

We don't ask you to trust one answer. Helios Search fires two independent reasoning engines at every query in parallel, then runs both answers through a verifier that checks for substantive agreement. When they agree you get a "Verified" badge. When they disagree we surface that explicitly and run a reconciliation pass — no silent averaging, no hallucinated consensus.

Free: 20 queries/day, single-engine. Premium ($25/mo) and above: unlimited dual-engine + verifier on every query. The product is hardened against upstream outages — when one engine is at capacity, the other carries the answer with the badge marked "Single-source" so you know.

Premium upgraded — frontier engine, every turn.

As of today, every Premium subscription routes through the same frontier reasoning engine that Pro and Studio tiers run on. Premium is no longer a tier of compromises; it's the cheapest seat at the frontier. Price moves to $25/month to reflect the upgrade. Voice, imagine, and the full 108-tool web tier all stay uncapped. Quality over volume — that's the new floor.

Also shipping today: brain-level engine failover (when one upstream is at capacity, Helios auto-routes to a healthy engine — no hangs, no dead air), search rebranded as Helios Search (no more leaked vendor labels), and post-login dashboard routing so anyone signing in is one click from talking to Helios.

Helios v4.1 is live.

Today we're putting Helios — a JARVIS-class personal AI — into the hands of the public for the first time. It's voice-first, sub-200ms turn-taking, with 108 live tools on the web tier and a five-layer cognitive engine I've been refining for a year. No corporate gating, no waiting list. Free for voice + chat. Paid tiers unlock /imagine, /code, /3D Canvas, and parallel sub-agents.

The five-layer cognitive engine, in plain English.

Most AI assistants narrate. They tell you what they'd do. Helios runs every reasoning turn through five mandatory layers — Perception, Strategy, Execution, Validation, Delivery — and the layers are forbidden from being skipped to "look fast." Result: when Helios says he did something, he actually did it. The receipts are in the activity log.

Why we don't have a wake word.

Wake words are training wheels. They were necessary when STT was slow and expensive — neither is true anymore. Helios uses a continuously-open Voice Agent WebSocket with server-side VAD and hard barge-in. Just talk. Pause to think — Helios waits. Interrupt mid-response — he stops instantly. JARVIS didn't need a wake word. Neither do we.

What's next: Q3 2026.

• Mobile PWA — full Helios in your pocket
• Custom voice cloning at Studio tier
• Game-engine bridge (Roblox + Unity + Unreal)
• Hardware: AR/MR HUD experiments — internal alpha