--- title: "Horizontal AI" canonical: "https://anchor.is/on-deck-vvSq/issue/17/horizontal-ai" publication: "On Deck" publisher: "Ground Truth" language: "en-US" guide: "https://anchor.is/on-deck-vvSq/llms.txt" permalink: "https://anchor.is/on-deck-vvSq/issue/17/horizontal-ai" issue: 17 section: "horizontal-ai" description: "AI tools, assistants, and platforms worth understanding" authors: ["Reva Sandoval"] published: 2026-10-26 updated: 2026-10-22 articles: 10 --- # On Deck > Ground Truth follows AI as it's actually built, funded, and shipped: consumer assistants, vertical products from legal and healthcare to agriculture, and the capital moving behind them. It's written for practitioners who've shipped agentic systems and want more than press releases. It reads interaction models, architectures, trust patterns, and funding logic to judge which companies are serious, and where it's worth building next. [Read on the website]() ## Issue 17 — Horizontal AI AI tools, assistants, and platforms worth understanding ### Deep Dive - [Who Decides When Your AI Assistant's Memory Goes Stale]() — By [Reva Sandoval]() · Published 2026-10-22 · Format: Centerpiece Teardown ChatGPT, Claude, Gemini, Copilot, Meta's Muse and Amazon's Bee each keep a running picture of you, built from chats, files and connected apps. Some of it stops being true: a trip that has ended, a move that already happened. This year OpenAI and Anthropic both shipped ways for their assistants to update outdated memories on their own. Neither shows you when a memory changed or where it came from. I read six products' documentation and a year of user reports to see what each keeps, what it lets that memory steer, and who can retire it. ### Product Verdict - [Claude Duet Works While You're Away, Then Waits for You to Come Back]() — By [Reva Sandoval]() · Published 2026-10-22 **Bottom line: a real new way to use AI, and weak at bringing you back to it. Moderate confidence.** ### Interaction Patterns - [When an AI Assistant Saves You Work, Who Picks It Up?]() — By [Reva Sandoval]() · Published 2026-10-22 AI assistants call restaurants and fill shopping carts. Someone on the other end handles each request: a store, a restaurant, a group-chat friend, a person being recorded. That person, the **counterparty**, absorbs the work the assistant saves. ### Launch Radar Seven items that didn't earn full treatment this week. Each fills a thin spot on the map: voice, model choice, creative tools, or what happens to generated work after you hit share. Read the two voice cards as a pair. ElevenLabs sells a single managed model. Mistral sells separate parts you can swap out or host yourself. With the first, one vendor owns your latency, hosting and switching costs. With the second, you can replace whichever stage fails on speed or price without starting over. Model choice is also turning into a menu, and competitors' models now appear on it. Each card says whether the move is structural or incremental. Read every performance number as a vendor claim, since none has been independently tested yet. - **ElevenLabs sells the whole voice stack as one model** — [Source]() — By [Reva Sandoval]() — Published 2026-10-22 Tag: Voice · ElevenLabs Eleven Realtime handles listening, reasoning, turn-taking and speech in one managed model. It comes in a cheap Mini tier and a tool-calling Max tier. The API mirrors OpenAI's Realtime endpoint, so testing it means changing an endpoint. **Structural.** The 0.57-second latency is ElevenLabs' own figure. - **Voxtral 3 keeps each voice stage separate** — [Source]() — By [Reva Sandoval]() — Published 2026-10-22 Tag: Voice · Mistral Mistral ships speech-to-text, text-to-speech and a voice-agent SDK as distinct parts. Agents can call tools mid-conversation and run on your own servers or on EU-hosted APIs. Some coverage calls it "one pipeline." Mistral's own wording keeps the stages swappable. **Structural**, and the opposite bet from ElevenLabs. - **Gemini Enterprise app now offers Anthropic models** — [Source]() — By [Reva Sandoval]() — Published 2026-10-22 Tag: Models · Gemini Google's release notes list Anthropic models inside the Gemini Enterprise app as of October 15. Google's own assistant serving a direct competitor's models is **potentially structural**, at moderate confidence. The note cuts off before admin controls, the model list and scope. Treat those details as unconfirmed. - **Gemini-built apps now publish straight to Google Sites** — [Source]() — By [Reva Sandoval]() — Published 2026-10-22 Tag: Sharing · Gemini Since October 19, apps built in the Gemini app publish directly to Google Sites, including from personal accounts. Muse paused sharing on this same generated-app-to-public-page path. Its shared pages turned up in passive DNS records, logs of which web addresses get looked up. Google hasn't documented visibility or indexing defaults. **Incremental.** Check those defaults before you publish anything real. - **Kling 4.0 takes up to fifteen reference inputs** — [Source]() — By [Reva Sandoval]() — Published 2026-10-22 Tag: Creative · Kling Omni Reference accepts images, video clips, subjects and voices in one generation job, up to 15 total. That part is **structural**. You direct a shot by choosing its inputs, which gives you far more control than rewording a prompt. 10-bit HDR and 8,000-token prompts are incremental quality gains. Every spec comes from Kuaishou, Kling's parent company. - **Firefly keeps adding rivals' models to its picker** — [Source]() — By [Reva Sandoval]() — Published 2026-10-22 Tag: Models · Firefly Firefly lists GPT Image, FLUX, Nano Banana, Veo 3.1 Fast, Kling 3.0, Luma and Runway next to Adobe's own models, with Kling 4.0 marked as coming. Adobe pitches its own models on commercial safety and the partner models on choice. This pattern has held since February. **Incremental** this week. - **Midjourney's private folders can still leak to Explore** — [Source]() — By [Reva Sandoval]() — Published 2026-10-22 Tag: Sharing · Midjourney Shared folders launched in alpha for collaboration. Midjourney's changelog says images made in a private folder can still show up on Explore and your profile unless stealth mode, the setting that hides them from public view, is on. Midjourney discloses this and calls the feature an early version. **Incremental.** Turn on stealth for client work. ## Other sections in Issue 17 - [Signal Watch]() - [Vertical AI]() - [AI Capital]() ## More - [Publication guide]() - [Issue archive]()