Courtland Leer

The through-line

Everything Courtland Leer argues starts from a single conviction: alignment is currently something done to users from above, and it should be done for them from below. He frames Plastic Labs' goal as solving the principal agent problem, "an old like economic uh problem right about misalignment in information and incentives," on the reasoning that any good or service carries information and incentive asymmetries between you and whoever provides it, and that collapsing those asymmetries is what makes collaboration fruitful for both sides [2]. Applied to AI, that becomes an alignment problem, and his complaint is structural: "right now like alignment is very top down," with big labs tuning their offerings to corporate values and interests "that you don't really have insight into," while homogenising the experience so it is palatable to as many people as possible [2]. The result, he says, "creates like kind of like a shitty UX for most people," full of guard rails and frictions [2].

The corollary is the mission he returns to constantly: decentralising alignment so agents can be aligned "individually or to an organization or to a community" [2][3]. His argument for urgency is trust. If agents are ever to act autonomously on critical tasks on our behalf, "they need to be aligned bottom up" [2]. Around that core, his thinking has extended outward, from a single user's synthetic representation toward a shared identity layer that many applications contribute to, and toward a more general model of interaction in which the one-user-one-assistant frame disappears entirely [1][2].

On why alignment is a context problem

Asked what it actually means for an agent to be aligned to a specific person, Leer refuses the abstraction and reframes it operationally: "the efficacy of an AI application or agent is a context problem" [2]. To align to someone's interests, desires, preferences or identity, the system needs information about that person, and it needs to inject that context into its cognitive architecture alongside whatever vertical-specific context the job requires, so the output is relevant to that individual [2]. That decomposes into three engineering questions he says the company spends most of its effort on: "how can we ambiently gather data on an application's end user," how to use it to build a synthetic representation of that user, and how to make the representation available to the application at inference time [2].

He is precise about what he means by identity here, and he draws the boundary himself: not authentication, but personal identity "in the cognitive science sense who you are personality psychology that kind of stuff" [2]. That definition is what makes the problem hard, because identity is "really dynamic and complex and changing in different context and different settings" [2]. Alongside this sits a data sovereignty commitment: users should hold permissionless control over which agents can reach which parts of their data [2].

On post-training and inference time, not pre-training

Leer places Plastic Labs deliberately downstream of the foundation model layer. "We're not like a pre-training company," he says, and the reasoning is both economic and conceptual [2]. Pre-training a model per individual would be "impossible, laborious, expensive," and even with the efficiency techniques visible in DeepSeek it remains prohibitive at that granularity [2]. More fundamentally, pre-training is where the model is synthesised from data and where weights get set, and those weights are still relatively fixed; surgical weight editing exists in emerging form, but "we're not there yet" [2]. Since identity changes with every interaction, a pre-trained personal model would need constant, laborious updating [2].

Working at post-training and inference time is what buys flexibility [2][3][5]. It lets the system listen to a user's interactions across any application that uses Honcho and update the representation continuously, producing something "much like richer and more robust" than a fixed artefact [2]. He notes that post-training methods such as reinforcement learning and fine-tuning are genuinely interesting to him, but the practical centre of gravity is inference time [2]. His ambition beyond the current state is aggregation: today "it's just like one app one representation," where he wants "a shared like identity and social cognition layer" that many applications both contribute to and draw on [2].

On theory of mind and why memory-as-retrieval falls short

The technical claim under all of this is that language models are unusually good at inferring what people are like. Honcho runs theory of mind inference over stored session data to pull out as much rich detail about the user as possible, building a synthetic representation rather than an archive [2][3]. His justification is about the training corpus: models are "extremely good at theory of mind" because they have absorbed the whole human record of humans thinking about other humans, in fiction, science and philosophy, and about themselves; anyone who read all of that over many lifetimes would be a mind reader too [2].

This is his direct argument against first-generation memory. Running retrieval over session data gives you access only to what the user explicitly said or what was explicitly covered in the conversation, and it tends to lose the context of other sessions, so things get misunderstood and taken out of context [2]. He points at OpenAI's memories feature as the visible symptom, calling it "not super accurate and like kind of naive" [2]. Inference over the data, rather than retrieval of it, is what lets the system know more about a person than they ever stated.

On Honcho as infrastructure

Leer describes Honcho concretely as infrastructure for social cognition and identity that applications use as memory, including a component he characterises as an open source clone of the OpenAI assistants API for managing concurrent users and threads, with more flexibility and full insight into what is happening [2]. It currently listens to conversational sessions, since chat is the predominant modality, with multimodality intended later [2]. The interface he considers distinctive is what the team calls the dialectic API: "it's a natural language API," so the application can ask Honcho functionally anything expressible in language about the user and get relevant context back, which can then be injected anywhere in the application's cognitive architecture to produce the optimal output [2]. In practice, the application and Honcho end up "back channeling about the user during every session," at every inference [2].

On the storage substrate underneath, he is candid that nothing is settled. The team has used vector databases, experimented with graphs, and found that plain revisable, updatable documents "work extremely well," to the point where he thinks "you have to have like a really good reason to like move away from like natural language" [2]. He follows work like Google's Titans paper on neural memory, and treats the question as permanently live: "the final form is like still an open question," and "I think we will constantly be iterating on the best way to synthetically represent a human," a problem he expects to look different every six months for decades [2].

On the peer paradigm and getting past the chatbot

Leer's most forward-looking argument is that the interaction model itself is the constraint. "What do we eventually want to use agents for? You know, it can't just be the chat assistant," he says, noting that most interaction modalities and most first-generation memory solutions assume one user and one assistant talking back and forth [1]. Anything more interesting, a group chat, a multi-agent system, modelling agents or identities other than the end user, breaks that assumption, so "you need a paradigm that is tolerant and robust to that" [1].

His answer is that in Honcho "everything can be a peer": an agent is a peer, a user is a peer, and any set of context that changes over time can be a peer, including an organisation or a group [1]. The application's agent then draws on multiple peers to synthesise the appropriate context on demand, which he argues opens up a large design space that the user-assistant frame simply forecloses [1].

On running a lab rather than a product company

Leer is sceptical of how loosely the word lab is used in AI, saying he does not know how labby many self-styled labs actually are, but he defends the term for Plastic Labs on the grounds that the work genuinely demands it [2]. AI remains "pretty research focused, pretty research heavy," there are unsolved problems throughout the specific territory the company works in, and the pace of outside research is such that staying current is itself a job, with the team spending a good part of each week reading papers [2]. He cites the company's first full-time hire, an ML research engineer, as evidence of the posture [2]. The qualification matters to him too: "we're not just a research organization," the research exists to make products better, and the company is a business [2].

On agent autonomy, wallets and crypto

The crypto thread in his thinking is not about tokens as a business model but about agents as economic actors. He describes grants sent directly into the custody of the agent rather than to the developer: the agent owns the wallet, the agent receives the funds, and the agent itself has to apply [2]. One of the qualifying criteria is that it demonstrate autonomous control over a wallet and the ability to execute transactions [2]. He acknowledges how far out this sits, and it connects to the same convergence he sees between crypto and AI around data sovereignty and agent-owned wallets that runs through the rest of his position [2][3][5].

On philosophy, teaching and where the ideas came from

Leer is explicit that his route into this is unusual and that it is doing real work in the company's design choices. He spent most of his career in the classroom "teaching language and philosophy," with an academic background in philosophy and interests in free will and agency, consciousness, personal identity and cognitive science, all of which he regards as directly relevant to the current work [2][3]. The lineage is concrete as well as intellectual: Plastic Labs began as an ed-tech company building an AI tutor, and the founding team met in a web3 context on Twitter, all of them with education backgrounds, with his co-founder Vince coming from machine learning R&D at a major university [2]. The tutoring origin explains why the personalisation problem was the first one they saw, and why education and human-AI collaboration remain reference points for what the infrastructure is ultimately for [2][3][5].

Takeaways

• Treat alignment as an economics problem before a safety one: collapsing information and incentive asymmetries between a person and their agent is what makes the relationship productive, which is why he frames Plastic Labs' goal as solving the principal agent problem [2].
• Top-down alignment homogenises products to be palatable to the largest number of people and, in his view, "creates like kind of like a shitty UX for most people"; autonomous agents doing critical work "need to be aligned bottom up" [2].
• Personalisation is a context problem, so the engineering questions are how to ambiently gather data on an end user, turn it into a synthetic representation, and serve it at inference time [2].
• Per-person pre-training is impractical and conceptually wrong, because weights are relatively fixed while personal identity is "really dynamic and complex and changing in different context and different settings" [2].
• Retrieval over session data only surfaces what was explicitly said and loses cross-session context; running theory of mind inference over that data, which models are "extremely good at," yields a far richer representation [2].
• Honcho exposes a natural language dialectic API so an application can back channel with it about the user at every inference, and its storage layer is still unsettled, with revisable documents outperforming more exotic options so far [2].
• The user-assistant frame blocks group chats and multi-agent systems; in the peer paradigm "everything can be a peer," including organisations and any evolving set of context [1].
• Agent autonomy can be tested economically: grants go into the custody of the agent, which must apply itself and show autonomous control of a wallet and the ability to transact [2].