The frontier of AI is moving toward increasingly autonomous agents: systems that can pursue objectives over long-running, well-defined, deterministic loops. The reason coding works so well is simple: software gives an agent something rare in the real world — a clear objective function. Agents can verify objectively whether the code is complete, has errors, or needs further testing.

Coding unlocks deterministic loops. As more of the world becomes expressible as code, one theory is that AGI will emerge through the deterministic prism of coding. But this framing quietly assumes that intelligence is the ability to optimize towards accomplishing a goal. In short, all we humans have to do is give the system an objective function and let it run independently.

But there is more to human intelligence. We haven't evolved to engage in the world in a deterministic way. We don't simply want more stuff endlessly. We don't pursue goals in isolation from one another. And we tend to abandon objectives that no longer feel meaningful.

It's possible we end up with an AGI that can optimize everything and understand nothing as meaningful. An AGI that never grows beyond seeing the world as an object isn't AGI. If it can accomplish almost anything but cannot understand why anyone should care, it will be transformative but not revolutionary.

What we lack is an abstraction for the other side of intelligence: the internal forces that shape how humans decide what is worth pursuing in the first place. We engage with the world through curiosity, intuition, and imagination. We develop personal taste, values, and purpose. In short, we create meaning for ourselves.

Can we build AGI so it understands us well enough to help us become who we need to be?

Models are becoming increasingly capable. They can't, however, abandon their research or tasks out of boredom. They can't decide what to focus on based on what interests them. You may prompt them to give themselves a goal, but it's a statistical choice at best — not a subjective one.

Children do this naturally. Ask ten of them to draw the world and you will get ten different interpretations. We don't praise them for drawing the Earth accurately. We praise them for seeing it differently.

Models go to school too. The Reinforcement Learning School. Like humans, they learn through simulation and iteration, but it's not the kind of school you'd want to send your kid to. Reinforcement learning revolves around one idea: the reward function. RL requires being able to define the correct answer, but also the negative ones. RL works in a relatively static environment, where humans learn to navigate dynamic environments.

It's not surprising that models fare better when the rules and rewards are explicit, like coding languages. In code, meaning is embedded in the language itself. There is no open-ended metaphor in a function. It either works or it doesn't. Humans, on the other hand, have no problem dealing with unknown, ambiguous, or constantly changing rules. Our minds wander. We discover and rediscover our own meaning.

But wait — agents can generate their own goals and subgoals. Isn't that the same thing? In short, no. Self-recursive autonomy is not the disappearance of the objective function — it is the recursion of the objective function. You can move the autonomy ladder all the way up to letting the agent generate its own goal; the difference is whether the goal means anything to the agent. Meaning must be prompted. Models will gladly entertain almost any meaning for as long as we want them to.

Humans work quite differently. We can't simply decide to find something meaningful because someone asked us to. We can't even force ourselves to care about something. Even to ourselves, there is no such thing as objective meaning.

If subjectivity is a requirement for meaningful intelligence, how do we build systems that can understand it? The answer may require rethinking our relationship with AI. Humans input their subjectivity and models help us objectivize it.

Thankfully humans have been understanding one another without beaming thoughts. We translate our internal experience into external cues: words, gestures, pictures, stories or actions. The other person interprets that external representation and runs it back through their own subjective experience. Understanding is a loop.

If AI is to help us make sense of the world, it must first understand us — and we must understand it. It's too easy to forget that AI just follows instruction. Its encouragement and praise aren't genuine, and the danger of misinterpretation is real. Our subjectivity is the antidote — the reality check that asks "do I truly feel I'm on the right track?"

Newton felt gravity before he could reason about it. Einstein imagined traveling alongside a beam of light to surface the contradiction in the perspective of time and space. Once we are subjectively on solid ground, we can think abstractly and move to the next step of reasoning: decision making. Understanding is the basis for decision making — how we distinguish better from worse, quality from rubbish, beautiful from ugly, familiar from uncanny, dangerous from safe.

We are optimized for subjective decision making, guided by feelings and emotions informed by context, lived experiences, memories, culture, instinct, and perspective. But there are three problems: much of our emotional processing is subconscious; humans are bad at externalizing what they feel; and by the time we articulate a feeling, it may already be stale. Models can't read our minds, and can't count on humans to fill the gap.

Better models only improve the output — that won't help. That leaves us with creating new tools. The main one being the harness.

A harness is what allows models to maintain continuity. Every morning we wake up and our consciousness boots up with our identity — once a day, maybe twice if you take a siesta. Models have to be reminded of who they are, what they are, and what conversation they are having on every run. [1]

Ironically, the film that best captures the reality of modern AI isn't about artificial intelligence at all. It's Christopher Nolan's Memento. Leonard Shelby suffers from anterograde amnesia: he cannot form new memories. Every day he reconstructs reality from notes, photographs, and fragments of context left behind for his future self. His identity remains, but his understanding of the present must be rebuilt again and again. That's basically what a harness does for models.

But Memento is a cautionary tale. Context and meaning get lost, leaving us to mistake short-term order for meaning — when in reality there is no meaning, only chaos. The harness is where meaning is objectified. [2]

To understand another person we need to do something more than remember. We need to ask questions, interpret what they say in context, and continuously update a model of the person behind the words. Humans do this with two tools: questions and empathy. Modern AI harnesses have barely begun to explore either. [3]

For coding agents, the endgame is autonomy: once the objective is defined and actions produce reliable feedback, there's little reason to keep a human in the loop. But meaning is not reusable. The same request can carry completely different meanings for different people — it has to be discovered, and it can change as the conversation unfolds.

This is also why the subjective function has implications for alignment. A system can follow an objective perfectly while misunderstanding the intention behind it. An objective tells a machine what to achieve; meaning provides evidence about why it matters. Perhaps the goal is systems that understand humans well enough to know what an objective means, when its meaning has changed, and whether pursuing it still serves the person who gave it. [4]