18 Aug 2026 · Voice AI

A few thoughts on voice AI

Voice AI is getting really good. The models are there, but penetration is still low.

I've tried a few voice AI bots focused on long-horizon interviews — 20 minutes or more — and it feels very different from answering the same questions in text. It has more energy and deserves your full attention. In a chat, you can close the window, do something else and get distracted. A live conversation, even with a chatbot, feels focused.

That makes voice AI interviews for research and audits very interesting. A main part of any consulting or AI deployment project is auditing how processes actually work: what people dislike, where they see optimization potential, what has been tried and what went wrong.

Agents must analyse all internal knowledge, but it is often messy and doesn't contain the opinions you need. It may be more effective to deploy dozens of agents to run live voice interviews with employees.

The same applies to HR and customer research. Companies are growing quickly around this, like Listen Labs and LATO, founded by the brother of an ElevenLabs co-founder. Customer research looks quite competitive already, but LATO's focus on commercial due diligence seems more distinctive. Curious how it performs.

Choosing your customer is key. Plenty of “wrapper” companies that VCs claimed had no moat are thriving, so generalist agents can work. Still, specialization in research feels key: it lets you automate the full service and build end-to-end software around it.

18 Aug 2026 · Ambient AI

A voice companion for walks

Voice AI's latency is low, quality is high, and speaking with it is already natural — sometimes even enjoyable.

Combined with real-world data, this opens new use cases connected to having a proactive companion. I've been enjoying the idea of a voice companion that activates during your walks. When you are close to something cool, it tells you the history of a building or a story about the street.

Museums and brands pre-record this. But AI is good enough that, with a good knowledge layer and a pipeline that provides good facts — in the tone you like and about the topics you like — it could be automated and lively.

The devil is in small UX details. I like hearing stories about buildings and architecture, but I hate taking out my phone, photographing something and asking ChatGPT what it is. It would be cool to have a better experience.

Ideally, it should work while you're listening to music. Either you allow interruptions — with a setting to turn them off — or it listens for your voice: “Hey, what's this building?” Meta's glasses and similar devices seem like a natural way to integrate it.

17 Aug 2026 · Consumer AI

What will unlock consumer AI?

While both OpenAI and Gemini reached 1bn MAUs, conversion to paid is around 2-3%. Why?

A few hypotheses.

Most consumers don't have problems where AI is a drastically better solution they want to pay for

Would be good to map B2C spending on software and human-provided services that AI can automate, then compare market size, AI penetration and product mode: an existing app, chat or a vertical product.

Chat is a bad universal UI — something else will unlock usage

Chat is bad for discovery: people don't always know what's possible. It is also boring: people want engaging, case-specific environments. Major labs tried app integrations and MCP apps without much success. Wabi's on-demand apps are interesting, though the greatest apps require mastery and skill (for now?).

Consumers use consumer AI — it is embedded in the apps they already use

People keep using the same apps: calorie trackers, fitness trackers, whatever trackers. AI becomes native — visible as a copilot or chatbot, or behind the scenes providing new benefits.

There are new vertical consumer AI products — just niche?

AI friends like Character.AI have reached real consumer scale. Photo and video AI is growing big. B2C edtech is growing too: AI tutors embedded in Duolingo, Preply and others.

17 Aug 2026 · What got cheap

What did AI make so cheap that something new becomes possible?

AI is the biggest inflection point of our generation, but the general point is clear: it lets us do things at scale, cheaper. I'm interested in the particular areas where this creates new possibilities and behaviours. What got so cheap or fast that it enables a new user experience?

Games get cheap enough to be used as media

Follow the three.js account on Twitter and you'll see that agents are already outstanding at creating procedural 3D worlds and games. The production of on-demand, personalized worlds is close. Top-notch games are works of art, but games of some quality will soon be easy to produce.

I'm curious whether this does more than increase the number of games — whether it lets us use games as media where they haven't been used before:

Each newsletter is a new game.
Gamified on-demand educational worlds.
Websites.
Anywhere else a cheap game is better than text or video.
TikTok with games? A social network with game-creating influencers?

Human services get automated — the obvious one

Parts of human-provided services can be automated: in some cases fully, like high-frequency customer support; in others by augmenting the service provider, as in legal. Are there any service types left that aren't being tackled by dozens of competitors?

Which consumer services is nobody serving?

Most AI deployment focuses on B2B services — predictably, because there is more money there. Which B2C services haven't been tackled yet? Why is there no high-quality travel agent, shopping agent and so on?

Tacit knowledge finally becomes writable

Cheap intelligence makes distributed tacit knowledge easier to use. Decisions, knowledge and complex interconnected processes can be formalized, tracked and potentially improved at a scale that wasn't possible before.

Research fleets instead of ten blue links

Search, relevance and quality review can be done with agents at unprecedented scale — instead of looking through the first dozen Google results or trusting one agent that runs a few dozen searches. You can create research fleets that know the job to be done and your personal preferences. They can also challenge your biases and show adversarial counterpoints, if prompted correctly.

A fleet can analyse thousands of internal or external knowledge pieces. This can help you find what you need better and more frequently — and create personalized newsrooms, research, and media.