Teahose.
SIGN IN
NEW HERE — WHAT TEAHOSE DOES
We read the entire AI & tech firehose — so you don't have to.
PODPodcastsAll-In, No Priors, Acquired…
NEWNewslettersStratechery, Newcomer…
PAPPapersPhysical AI research
PHProduct Huntdaily launches
VCInvestor ScoutSequoia, a16z, Benchmark…
CLAUDE DISTILLS →
7 reads, 30 sec each — free, 6 AM ET.
+ a live graph of the companies, people & themes underneath.
HOME/THE A16Z SHOW/The Missing Layer for Personal A…
POD
// EPISODE
THE A16Z SHOW

The Missing Layer for Personal AI | First Person Labs’ Ken Liu

DATE November 18, 2026SOURCE THE A16Z SHOWPARTICIPANTS JOEL DE LA GARZA, KEN LIU, MARTIN CASADO
// KEY TAKEAWAYS6 ITEMS
  1. 01Portable, user-owned context is the missing layer between users and agents
  2. 02The OS already emits a rich event stream that is waiting to be mined
  3. 03From reactive to reflective proactivity: the next interaction paradigm
  4. 04Security through verifiable architecture, not promises: confidential inference
  5. 05Token economics gate the market today, but the curve is steep
  6. 06Mixture-of-models harness design as a core competency
In this episode

1. Key Themes

Portable, user-owned context is the missing layer between users and agents

Ken Liu's core thesis is that every agent currently re-learns who you are, and that the context should instead belong to the user and travel across tools. First Person Labs' product, "Second," is a persistent model of the person (preferences, priorities, relationships, digital activity) that any agent can query. Ken describes the product as modeling the person and then handing control to them: "We focus on essentially creating a model of a person and then put that person in control." 00:00:28 The model is meant to go beyond memories to "the priorities, preferences, goals, values, and commitments, and intent." 00:01:42 The practical payoff is that agents stop interrogating you: "As of today, you have to explain that to GrokBot. There's multiple rounds of questioning. But if you have a user model... that just builds up by itself... then you can share that over instantly." 00:02:33

The OS already emits a rich event stream that is waiting to be mined

The technical wedge is not a new data source but the raw stream computers already generate: app switches, typing, pauses on websites, Wi-Fi state. Ken argues this is badly underused. "On your computer, there's an infinite stream of events. As you use a computer, the apps that you switch to, the way you pause on certain websites, the way you type, and the way your Wi-Fi is connected. And that the OS already gives you. This is across Apple, Microsoft, and even Linux." 00:05:12 He adds, "I think people are underestimating like how much the computer already knows. It's just there sitting there waiting to be tapped, so to speak." 00:06:06 The system applies 24/7 agents hierarchically: first synthesizing "this piece of typing at this particular app might mean that you're sending a message," then building up to higher-level understanding.

From reactive to reflective proactivity: the next interaction paradigm

Ken distinguishes between reactive productivity (notify me when an email arrives) and reflective productivity, which "connects the dots over many events and long time horizon." 00:07:28 His claim is that the future of personal computing is less about asking agents things and more about agents intervening unprompted: "What are some things that I can do for the user without any asking? And I think that's one of the biggest problems right now. Like people have to ask. You're not supposed to ask. They're super intelligence." 00:07:53 Martin Casado's lived example validates it: "We were planning an event yesterday. There was a 20% chance of weather. And in real time, it was continually updating the forecast for us to let us know if we needed to get a tent or not... Did you ask it to do that? No, I did not ask it to do that." 00:06:14

Security through verifiable architecture, not promises: confidential inference

Because ambient AI sees everything, Ken argues trust must be engineered. He frames AI as only two things, "data, there's inference" 00:11:51, and gives the user control over both. Options include bringing your own tokens with local data ("Second Local"), or confidential inference on NVIDIA GPUs with a hardware attestation toggle: "You can basically prove that there's only the model running inside. No one can see inside. And then you don't have to trust whoever's running that model. The hardware is what you trust, which is NVIDIA." 00:12:50 He explicitly contrasts this with zero data retention: "Zero data retention is a promise by policy... I give you my word." 00:13:12 For cloud storage, the chain is "you encrypt it locally, you put it in the cloud, and you only decrypt inside a secure environment." 00:15:15 He also notes the industry is moving this way: "I think OpenAI is also providing this option, I think, in the last month in their announcement." 00:13:45

Token economics gate the market today, but the curve is steep

Always-on analysis is expensive. Ken's back-of-envelope math: "One agent running 24-7... at 60 tokens per second. And let's say you run 10 of those... at Opus 5i's pricing, that leads to $1,000 per day... If you use DeepSeek Flash, for example, it becomes like $60 per day." 00:18:10 That limits early users to "busy professionals," but "since we started this company, actually, token prices already dropped like by more than 50%." 00:19:07 The product also exposes a fidelity/cost tradeoff between precomputing insights ("cash time") and reasoning over raw logs at query time, which "usually just take a lot more time at test time. Let's say 10 times longer." 00:20:01

Mixture-of-models harness design as a core competency

Not all tasks need frontier models. Ken explains that the system routes by task: "If you say I want to analyze what is a semantically meaningful moment that happened on your computer... you don't need [a frontier model]. A different kind of information would be like, what is Joel's goal for this week? That's something you want a very expensive model, but you don't need to run it very frequently. So essentially, Second involves like a harness design problem." 00:22:16 He says new models get swapped in continuously and that one in particular "turns out... is extremely good at retrieval." 00:23:15 Portability is a design principle so it can later run on local inference: "You should be able to run this on any models. Because later on, you want to run it on local inference." 00:17:20

The user model becomes an API platform, a "personal infra" for app builders

Second exposes the person as APIs: "There's an API about what you just did 10 minutes ago. And there's an API for what are you going to do for the next week with your priorities. There's an API for your reminders. And there's an API for your wiki, like people in your life." 00:24:09 Ken compares this to the personal computer era: "The first personal computer came out, the Altair 8800. And it doesn't do anything. You're supposed to program it. No one does. And then... someone came up with a spreadsheet and then spreadsheet became useful." 00:25:29 His example was a designer friend whose UI taste (button placement, color) is captured so apps built on top need "way less iterations." 00:25:29

Ubiquitous, device-agnostic AI as the long-term vision

Ken's endgame is a single Second with every device as an input/output surface: "Everyone has to have a second. And all the devices just becomes inputs... Like the screens that you look at or even this floating orb next to you." 00:28:10 He observes that the "Jarvis" fantasy is under-appreciated: "It's not just that it's intelligence and you can do things, but it's more that it's everywhere." 00:27:14 On ubiquitous computing generally: "The closest thing we have is Apple iCloud... But there's no AI version of that yet." 00:28:44

A single trusted vault agent, with specialized and ephemeral agents around it

Ken predicts agent specialization and consolidation of the personal layer: "Agents will start to specialize just like people do... And some agents will specifically focus on protecting your data and lining with your interests. I think that's what Second is intended to be... It's like a vault." 00:09:23 He says there will not be multiple Seconds. Other agents can be ephemeral, but "you want to have consistent store and the consistent store is what gives you that value." 00:10:09

2. Contrarian Perspectives

Your context should be a standalone, user-controlled asset, not a feature of each agent

The prevailing assumption is that each model lab or agent (ChatGPT memory, Claude projects) builds its own memory moat. Ken argues the opposite: the valuable layer is a neutral, user-owned model that every agent queries. "You go to GrokBot and you say, hey, can you talk to my thing to figure out what I want? And then the two of them would start talking." 00:02:11 Martin demonstrates the switching-cost collapse: "I haven't been using my Codex in a while... I can just go start using Codex, and it has all the other work that I've already done." 00:03:01 If context is portable, agent vendors lose a primary lock-in mechanism.

Prompt injection is a reasoning problem that scales with model capability, not a rules problem

When Joel pressed on tricking LLMs into leaking via role-play scripts, Ken rejected the idea that deterministic rules can fully solve it: "This is not a problem you can definitively solve... What we've seen with models is that they just get over time much, much better at underlying. Because it's essentially a reasoning problem." 00:17:20 His example of context-aware policy: "He does not send me a message at 3 a.m. And so if... you send a message at 3 a.m., it's likely spam... you cannot write down rules for all the things that are going to happen." 00:16:31 Security through a rich user model that spots anomalies is a different bet than static guardrails.

Zero data retention is the wrong standard; hardware attestation should be the default

Most of the industry treats ZDR as the privacy gold standard. Ken calls it merely "a promise by policy" and argues "we want to move towards a world where confidential inference becomes... that becomes a default." 00:13:12 Martin reinforced the distinction: "The way you guys have built it, this is technically provable... It's not just your word." 00:13:36

Resist building the UI; deliberately constrain the product surface

Counter to the instinct to own the interface, Ken says First Person Labs tried "to go all in interaction and turns out it's actually pretty hard." 00:31:28 They chose to "close down the surfaces that we have. There's only texting. There's only API and there's nothing else." 00:31:55 The bet is that the infrastructure layer plus ecosystem builders will beat a first-party app.

Expensive always-on inference is viable now because the cost curve is so steep

Rather than waiting for cheap tokens, Ken launches a $60 to $1,000-per-day-per-user workload for power users today, betting that "token prices already dropped like by more than 50%" since the company's founding. 00:19:07 Building products that are uneconomic today but will be viable in 12 to 18 months is the underlying wager.

3. Companies Identified

First Person Labs

Ken Liu's company; builds "Second," a persistent user model and portable context layer for AI agents. Mentioned as the subject of the episode. Early users are thrilled: Joel says, "I'll get a text from my second and it's like something that I actually need to know about relative to some ongoing conversation that I didn't even think about." 00:06:42 The product is available now, with capacity being scaled responsibly: "We want to make sure this rolls out in a very like responsible way. I want to make sure people's data like are controlled and secure." 00:20:19

NVIDIA

GPU provider whose confidential computing mode underpins verifiable private inference. Ken: "There's just a special toggle. It's called a confidential computing. You can turn it on... The hardware is what you trust, which is NVIDIA." 00:12:20

OpenAI

Model provider (and Codex maker); relevant both as a token source and as a confidential inference provider. Martin notes, "OpenAI released two new cyber-capable models yesterday." 00:03:01 Ken says, "I think OpenAI is also providing this [confidential inference] option, I think, in the last month in their announcement." 00:13:45

Anthropic

Cited as a common source of user-provided tokens. Ken: "Most people, when they bring their tokens, it's tokens from OpenAI or Anthropic." 00:12:20 Its Opus pricing is used in the cost model.

Codex (OpenAI)

Coding agent used as an example of an agent that Second can plug into so the user doesn't re-establish context. 00:03:15

GrokBot

Joel's primary personal-agent platform (as transcribed), used as the running example of an agent that talks to Second. Joel runs multiple instances for synthesizers, real estate, and family. He also notes its audio mode "is actually quite good." 00:30:15

DeepSeek

Cheap open-weight model line cited in the cost analysis: "If you use DeepSeek Flash, for example, it becomes like $60 per day." 00:18:38

Apple

Cited as the closest existing example of ubiquitous computing via iCloud, and as a source of the OS event stream. Ken: "Apple is doing a tremendous amount of work. But... there's no AI version of that yet." 00:29:04 Also the model for local-encrypt-then-cloud storage architectures.

Microsoft and Linux (OS ecosystems)

Named as OS platforms that likewise expose the event stream Second consumes: "This is across Apple, Microsoft, and even Linux." 00:05:12

Slack, Signal, WhatsApp

Connectors Second uses to learn faster. Ken describes them as a "sample efficiency problem. It makes it learn faster." 00:24:09

Dropbox

Referenced by Martin as an example of cloud storage he would avoid for sensitive data, as with local password vaults. 00:14:47

Altair 8800

Historical analogy: the early personal computer with no killer app until the spreadsheet, framing today's personal-context APIs. 00:25:57

Google Docs

Mentioned as an example of an activity Second infers from app-switching behavior. 00:05:42

4. People Identified

Ken Liu

Co-founder of First Person Labs; systems/security-trained builder behind Second. Mentioned as the guest and founder. Martin highlights his background: "A lot of the path that took you here was this kind of hardcore security stuff." 00:30:38 Ken's self-aware product lesson: "We made a lot of mistakes on the interaction side. We try to go all in interaction and turns out it's actually pretty hard." 00:31:28

Percy Liang

Ken's advisor (transcribed as "Percy" and "Percy Dameron"), a Stanford AI researcher. Mentioned by Martin as part of Ken's pedigree: "You know, who's your advisor, Percy?" 00:30:38

Dan Boneh

Likely the security researcher referenced via the garbled "Dameron A, hardcore, you know, security guy" in connection with Ken's advisors. Martin says, "He also worked with Dameron A, hardcore, you know, security guy." 00:30:38 Identity inferred from context; included with the caveat that it is a probable match.

Joel de la Garza

a16z partner and host; self-described heavy agent user and security-minded early adopter. Contributed pointed adversarial questions (prompt injection, cost, UX) and shared experiences such as "my second is so smart. I just, wow." 00:26:45 His wish for an embodied AI: "I want like that 1970s AI, you know, like the head on the screen that I can talk to." 00:26:45

Martin Casado

a16z general partner and host; early Second user and paranoid-security adopter. His validation: "The long discussions we had about security architecture really got me comfortable." 00:10:48 He also testified, "Lately it's been just indispensable because it's able to catch..." 00:06:14

Batman / Alfred (analogy)

Not real people, but Ken's central mental model: Second as Alfred the trusted butler who knows Batman's preferences and acts without revealing identity. "Alfred would not reveal, for example, that it's Batman doing it, but he would just buy it. He would state the preference." 00:03:53

Iain M. Banks

Author of The Culture series, raised by Joel as the best fictional embodied AIs: "They have the best embodied AIs. You know, they kind of float next to you." 00:28:38 (The transcript renders the name as "Stephen Bink, or Ian Binks.")

5. Operating Insights

Design the "harness," not just the model: route tasks by value and frequency

Ken's system assigns models by how often a task runs and how much reasoning it needs. Cheap models handle high-frequency event classification; expensive ones handle rare, high-stakes synthesis like weekly goals. "That's something you want a very expensive model, but you don't need to run it very frequently." 00:22:16 Any AI product team can copy this tiering to cut inference cost without hurting quality.

Use precompute vs. test-time compute as an explicit pricing/quality dial

Ken describes a fidelity-vs-cost tradeoff between continuously analyzing data ahead of time versus reasoning over raw logs at query time: "To answer a question based on raw logs, it usually just take a lot more time at test time. Let's say 10 times longer." 00:20:01 That suggests offering tiered SLAs: deep-cache premium, raw-log lite. It is a concrete lever for serving non-power users before token prices fall further.

Shrink the surface area to find product-market fit on the infrastructure layer

After failing at all-in UX, the team deliberately limited the product to texting and APIs. "We decided to like close down the surfaces that we have. There's only texting. There's only API and there's nothing else." 00:31:55 The lesson is to pick the layer where you have an unfair advantage (here, systems and security) and let an ecosystem handle polish.

Seed fast learning with connectors, then let ambient data compound

Connectors are used as "a sample efficiency problem" to speed up the model's learning curve while the raw OS stream does the heavy lifting. 00:24:09 A cold-start tactic: ingest structured sources first so the ambient layer produces value in the first days rather than weeks.

Make trust verifiable at every layer, and offer a bring-your-own-compute escape hatch

Ken offers users a path that removes the need to trust the company at all: "If you don't trust, let's say, us, in principle you should be able to bring your own tokens... to then analyze data that's on your laptop locally." 00:00:41 For products handling sensitive data, a no-trust-required option can be the thing that unlocks adoption among security-minded early adopters (like Martin).

6. Overlooked Insights

"Context continuity" across devices is the sleeper use case, and the OS lockdown is the real constraint

Ken mentions in passing that developers get huge value from not re-explaining anything ("You can just screenshot anything and just say fix.") and from moving between laptop and phone with the same context: "You brought something on the laptop, you close it and you take out your phone. It's the same context." 00:20:56 The significant implication emerges later: when Joel asks about running Second on a kitchen iPad, Ken admits "right now it's not much you can do because the OS is like locking it down." 00:29:28 That is a platform-risk and opportunity signal: whoever controls OS-level event access and cross-device identity (Apple, Microsoft) holds a gatekeeper position over this entire category, and a hardware or OS-level partner could leapfrog app-level players. Ken's aside that "devices are hard" 00:29:59 hints at why the layer is being built as software for now.

The only-a-few-weeks-old cost collapse implies a near-term cliff for any "heavy" personal-AI product

Ken drops, almost as an aside, that "since we started this company, actually, token prices already dropped like by more than 50%" 00:19:07 and that each new model is swapped in as "more capable." Combined with his own number (about $1,000 per day on Opus-class models versus about $60 per day on a DeepSeek-class model, roughly a 16x spread), this means the economic feasibility of always-on personal agents flips for mass-market users on a timeline of quarters, not years. The overlooked investment angle: products that are economically infeasible today but architecturally portable across models (as Second is designed to be) are positioned to ride the cost curve, whereas companies hard-wired to one frontier model are not.