Could We Actually Just… Build a Human? 🚀

“The best minds of my generation are thinking about how to make people click ads.”

— someone smart, probably

But what if, instead, we just built… a human? Not in the biological sense (ew, messy), but in the AI sense. Let’s go full Musk-mode and break this down from first principles¹.

✅ The Goal:

Construct an intelligence system equivalent to a human brain.

Step 1 — Decompose the Human Intelligence Stack™

Component	Human	LLM	Engineering Gap
Input	Continuous: $x(t) \in \mathbb{R}^n$ , $t \in [0, \infty)$	Discrete: $x_i \in V^k$ (token sequences)	Add continuous sensory streams
Memory	Dynamical system: $M_{human}(t) \to M_{human}(t+\Delta t)$	Static weights + RAG hacks	Integrate dynamic, differentiable memory
Agency	Utility maximizer: $\max_{a} \mathbb{E}[U(s, a)]$	Zero agency; acts when called	Implement agent loop + self-initiation
Learning	Online updates: $\Delta \theta \propto \nabla_{\theta} \mathcal{L}_{env}$	Offline SGD	Enable continual learning
Embodiment	Physical coupling: $S = f_{body}(env)$	None (optional)	Hook up to sensors + actuators
Emotion	Internal reward function: $E: S \to \mathbb{R}$	Simulated patterns	Approximate intrinsic reward signals

Step 2 — Translate into Engineering Terms

Perception: Multimodal Transformer (image, audio, sensor → tokens)
Memory: Differentiable Memory Modules (NTM, DNC)
Agency: Agent loop (AutoGPT style)
Learning: Online Gradient Descent / Reinforcement Learning
Embodiment: Robotics + Control Theory
Emotion: Learned utility function

Step 3 — It’s Just Math (Elon Voice)

At the end of the day, the brain is just:

A highly parallel, real-time, self-updating, multi-modal, embodied agent optimizing for survival.

And the math? Just some light bedtime reading:

$\pi: s \to a, \quad \text{maximize} \quad \mathbb{E}[U]$

This is the agent’s policy function. It means that the agent (whether it’s a robot, a human, or a bored LLM pretending to be helpful) takes its current state $s$ , and produces an action $a$ . The goal is to choose actions that maximize the expected utility $\mathbb{E}[U]$ , or in more relatable terms, to pick actions that are likely to make things better (according to its internal goals).

$E(s) = \sum_i w_i f_i(s)$

This is the agent’s internal reward function (or you could say “emotion function” if you’re feeling philosophical). Here, the agent looks at the current situation (state $s$ ) and computes how “good” or “bad” it feels about it. Each $f_i(s)$ is a feature — like “Am I safe?”, “Is there food nearby?”, or “Did my tweet get likes?”. Each feature has a weight $w_i$ representing how much the agent cares about it. The sum produces a single scalar number: a utility, or emotional score, for the current situation.

Combined, these two equations describe a simple but surprisingly deep story of intelligent behavior: decide what to do (policy) based on how you feel about the situation (emotion/reward).

Step 4 — First Principles Reality Check

Is there any magic ingredient?

Compute
Data
Integration

And possibly a touch of existential dread.

Step 5 — Why Aren’t We Already There?

It’s hard.
It’s expensive.
It’s easier to ship another GPT-4 powered SaaS for lawyers.
We’re busy pitching this blog post as a seed round.

TL;DR

From first principles, building a human-level intelligence seems… annoyingly possible. Not mystical. Just systems integration at scale.

The future is going to be weird.

Next Up

how actual neuroscience models today literally describe the brain as an approximate Bayesian², embodied, RL agent, which is just a fancy way of saying… we are surprise-avoiding computers³.

https://en.wikipedia.org/wiki/First_principle ↩︎
https://en.wikipedia.org/wiki/Free_energy_principle ↩︎
Friston, K. (2010). The free-energy principle – https://www.fil.ion.ucl.ac.uk/~karl/The%20free-energy%20principle%20A%20unified%20brain%20theory.pdf ↩︎

drew's blog