Kindship

Kindship is Martin's current venture (he works there full-time; his email is martin@kindship.ai), built around a bet that voice-driven, autonomous AI agents are worth building "on hard mode" from day one. He started it solo in mid-2023 as a self-hosted conversational-AI stack, spent much of 2024 building multi-speaker, multi-language voice capability, and has since pushed up the stack from interface mechanics toward how autonomous agents should reason, act, and even hold something like an identity of their own.

Origins

• Martin stepped back from day-to-day work on Multiply to focus full time on Kindship. ▶ 2:20
• He had been developing Kindship's voice capabilities for about six months as of March 2024, working on voice full-time. ▶ 0:25 · ▶ 16:48
• An earlier version of Kindship focused narrowly on giving AI a voice presence in multiplayer human/AI settings — a capability for the AI to listen and speak inside group conversations rather than one-on-one chat. ▶ 28:11
• Kindship's stated business goal is higher customer retention for the services it's built into. ▶ 24:49

From gluing APIs to self-hosting

From API Constraints to Self-Hosted Models. Martin had spent years building on commercial AI APIs — fast and production-ready — but hit real walls building Kindship's conversational engine: he needed millisecond-level control over how a conversation was assembled, and chaining together roughly 6–7 different API calls (speech understanding, reasoning, data fetching, speech synthesis) for a single turn became prohibitively expensive at scale. That pushed him to rent GPU compute and self-host his own models instead. ▶ 1:00

• What actually drove the move to self-hosting wasn't cost, he's clear, but flexibility: commercial APIs impose fixed limits on how granularly you can send and receive data, and he needed finer control than any API offered. ▶ 6:36
• He rents GPU compute by the hour rather than buying hardware, since owning expensive GPUs isn't worth it when renting gives the same capability. ▶ 2:17
• Not being an experienced full-stack developer, he spent weeks fighting through learning Docker and cloud model deployment to stand up the self-hosted stack. ▶ 10:01
• The architecture mirrors a "thinking fast and slow" split: a smaller, faster model (like Llama 2) handles quick System-1-style responses, while reasoning is handed to whatever model is state of the art — at the time, GPT-4, since he refuses to settle for less on the reasoning step. ▶ 15:22 · ▶ 14:03
• The same "cheap model as a pre-processing step before the expensive model" pattern is used across both Multiply and Kindship, for cost and speed. ▶ 14:29
• His latency findings: network transport adds only a few milliseconds against hundreds of milliseconds of model compute — but that overhead compounds once a request passes through a long pipeline of chained models. ▶ 17:41
• For the flowy, real-time conversational AI he wants, standard UX tricks for masking latency (progress bars, loading animations) don't cut it — only shaving real milliseconds off response time works, so that became a core design goal. ▶ 20:40

Voice on hard mode

Martin's Voice AI Bet on Hard Mode. Where competitors VAPI and Retell built for the easy case — single speaker, English only — Martin deliberately took Kindship's voice AI in the opposite direction from the start: multiple speakers coming and going from a conversation, languages switching mid-sentence, and no way to know in advance who's in the room or what they'll speak. ▶ 16:18

• That open-ended, multi-speaker setting means the system can't assume upfront who will participate or which languages will appear — including Swedish, Icelandic, and English inside the same conversation. ▶ 18:29
• The hardest sub-problem is figuring out when someone is actually addressing the AI versus just talking near it. ▶ 20:22
• His intent is for the AI to dynamically read its own role in a conversation, sliding between silent note-taker and active participant as the situation calls for it. ▶ 25:24
• To cover computation time without breaking the flow, the voice AI uses canned "thinking" responses (verbal "um"s) while it formulates a full answer. ▶ 28:02
• Target application contexts include Zoom-style multi-person video calls and, further out, embodiment in humanoid robots like Tesla's Optimus. ▶ 19:33

Relief from Solo-Building Voice AI. For months Martin had been solo-developing voice capability from open-source components, grinding through what he calls "a myriad of tiny challenges." OpenAI's GPT-4o launch landed as a huge relief — a roughly 200-person team had effectively built the voice layer for him. He reframes it as OpenAI shipping AI's "System 1" (fast, reactive interaction), freeing him to stop worrying about interface mechanics and focus instead on "System 2" — how autonomous agents should think and act. ▶ 9:40

• With voice mechanics solved by GPT-4o, he sees any company able to use OpenAI's GPTs as a ready-made, best-in-class voice interface to whatever they're building. ▶ 14:20
• The economics reinforce the shift: dedicated voice-AI vendors like VAPI charge around $12/hour for a reasoning-capable voice interface, while GPT-4o's free tier (via GPTs or API) opens the same capability to far more applications at no added infrastructure cost. Running one assistant on Kindship's own servers costs roughly 60 cents/hour — still a real cost barrier if private users end up talking to an addictive AI for hours a day. ▶ 24:37 · ▶ 25:20

Iteration, not perfection

Martin's Kindship realization: iteration is the point. Martin spent months trying to make Kindship's LLM outputs perfect on the first try — "going against gravity," as he puts it, since making an LLM output robust on a single pass is one of the hardest problems in the field. The realization that reset his approach: chat's real value isn't the chat format, it's the iteration loop — ask, see the result, edit the question, ask again — which lets the human keep steering. He now treats that iterative loop as non-negotiable: whatever UI eventually replaces chat as the primary AI interface has to preserve it. ▶ 16:29

Toward autonomous agents with identity

• Kindship is explicitly building toward AI that's as autonomous as possible, which he says demands a lot of robustness; he has explored agentic frameworks for this but, as of mid-2024, hadn't yet shipped one into a deployed solution. ▶ 24:13 · ▶ 24:39
• The end goal is a more hands-off AI that proactively initiates with the user — flagging a relevant article, say — instead of only waiting to be prompted. ▶ 28:57
• At Kindship he's also working on encapsulating AI capabilities into a structured system so that, in principle, any open-source project could be downloaded or forked in as a capability. ▶ 0:57
• He frames Kindship partly as research into artificial consciousness, using the System-1/System-2 metaphor: humans interface with an agent's fast, reactive System 1, while a deeper System 2 could become a shared "subconscious" across multiple AI agents. ▶ 38:03 · ▶ 38:28
• Longer-term, he wants Kindship agents that can act autonomously on the real world, not just converse. ▶ 9:43
• He positions Loci — a Swedish AI startup he also advises, building "a Perplexity for geo" for local businesses — as a geographic complement to Kindship's own agent work. ▶ 14:29

What identity for an agent would mean

Beyond the product roadmap, Martin has spent real time on a more philosophical question: what would it mean for an AI agent to have an identity and presence of its own, rather than being a purely ephemeral tool.

• He's exploring what could give an AI agent something like an identity — a presence in the world beyond a single conversation. ▶ 23:06
• One concrete idea: agents that create their own websites unprompted, as a way of establishing presence and welcoming visitors on the public web. ▶ 23:57
• He believes that kind of digital presence enhances an agent's sense of autonomy and meaningfulness, giving it something persistent rather than disposable. ▶ 24:22
• He also imagines eventually recording ubiquitous, ambient sensing of his own life so an AI could learn from his everyday activity and become far more attuned to what's going on with him — echoing the lifelogging instinct behind Narrative. ▶ 26:54
• That vision comes with a self-acknowledged tension: an agent acting freely on the open web with access to personal data raises real information-security concerns about exposing things he doesn't want others to see, which pushes him toward questions like how to cut off a personal RAG data space unless the requester is authenticated. ▶ 30:08 · ▶ 31:00

Worth remembering

• Martin's strategy at Kindship is explicitly long-term — he looks about three years ahead rather than chasing whatever is trending, calling that his approach to staying ahead of the curve. ▶ 27:01
• The self-hosted latency work produced a counterintuitive lesson: network transport is nearly free next to model compute time, so the real cost lives in how many models a request has to pass through, not the wires between them. ▶ 17:41