Build intelligence that learns without burning the world around it.

Three pieces exist today:

• Stability theory for learned dynamics. The KAM/HNN paper treats a Hamiltonian Neural Network as a perturbed Hamiltonian system and asks which invariant structures survive. The kicked-rotor demo makes the geometry visible. This is a result about a defined class of systems—not a proof that arbitrary neural networks are safe.
• A trained local-rule system. The Neural Cellular Automata lab runs in the browser and regenerates learned forms after damage. Metabolism and physical energy measurement are proposed experiments, not shipped results.
• A playable integration laboratory. Action in Love combines generated social pages, persistent state, model-driven characters, browser physics, and explorable simulations. It is evidence that the pieces can meet in one world; it is not a claim of AGI.

$\text{credibility} = \text{working artifact} + \text{stated limits} + \text{reproduction path}$

If a visitor cannot distinguish the demo from the dream, the communication has failed.

The working hypothesis is that many real environments need adaptation, efficiency, and recoverable stability more than another static increase in scale. A practical scorecard is:

$J = \frac{C \cdot A}{E \cdot \tau \cdot (1 + F)}$

where $C$ is task capability, $A$ is adaptation quality, $E$ is measured or estimated energy, $\tau$ is response/update latency, and $F$ is a fragility penalty: forgetting, divergence, or failure to recover after perturbation.

This is not a law of physics and the units do not magically cancel. It is a decision tool. Every experiment must publish the components separately so a flattering composite score cannot hide a regression.

The thesis loses if constrained local adaptation produces no useful Pareto improvement over a static model plus retrieval—or if its stability and measurement overhead erase the gain. That would be a valuable result.

A safe online learner needs more structure than gradient descent. One candidate loop is:

$\theta_{t+1} = \Pi_{\mathcal S}\left(\theta_t - \eta P_t \nabla_{\theta}L_t\right)$

$P_t$ restricts the update to a small, inspectable subspace; $\Pi_{\mathcal S}$ projects or rejects changes that leave a defined stability set. Low-rank adaptation supplies one such subspace:

$\Delta W = BA, \qquad \operatorname{rank}(\Delta W)=r \ll d$

The first benchmark compares four honest baselines on the same task stream: frozen model, frozen model plus retrieval, full update, and gated low-rank update. Measure time-to-adapt, joules or a disclosed energy proxy, retained performance, rollback rate, and recovery after distribution shift.

“Real-time learning” only counts when the system learns something useful before the environment changes again.

Control theory offers a disciplined vocabulary when—and only when—the state, input, output, and disturbance can be defined:

$x_{t+1}=Ax_t+Bu_t+w_t, \qquad y_t=Cx_t, \qquad u_t=-Kx_t$

For this discrete-time linearized regime, $\rho(A-BK)<1$ is a meaningful local condition. It is not a universal safety certificate for a language model. The research task is to find narrower settings where Lyapunov functions, energy drift, spectral bounds, runtime monitors, or reversible checkpoints give useful guarantees.

Each adaptive experiment therefore needs: a safe baseline, a bounded update region, a tripwire, a rollback state, and an account of what the monitor cannot see. Performance without recoverability is an incomplete result.

Logical irreversibility has a thermodynamic floor:

$E_{\mathrm{erase}} \geq k_B T \ln 2$

Conventional switching is often approximated by $P_{\mathrm{switch}} \approx \alpha C V^2 f$. These equations orient the search; they do not predict a modern computer's wall-plug energy by themselves.

• Landauer bounds logically irreversible erasure, not every operation.
• Lower temperature lowers that bound, but refrigeration has a system-level cost.
• Superconducting or cryogenic logic wins only if the full stack—including cooling and I/O—wins.
• Memory movement can dominate arithmetic, but the ratio depends on workload and hardware.

The first useful artifact is therefore mundane and powerful: a reproducible harness that records latency, device power or a named proxy, memory traffic, temperature assumptions, and useful work. Cryogenic compute remains an open hardware path, not a shortcut in the spreadsheet.

An NCA cell updates from local information rather than a central blueprint:

$c_{i,j}^{(t+1)}=c_{i,j}^{(t)}+f_{\theta}(\mathcal N(c^{(t)}_{i,j}))\,m_{i,j}$

The running demo shows regeneration. The next experiment adds a resource channel $r_{i,j}$, charges updates, and removes cells that cannot meet a maintenance threshold. Sweep damage size and resource supply, then report recovery probability, time, update count, and a clearly labeled energy proxy.

The interesting question is not whether this is literally alive. It is whether decentralized learned rules can deliver repair, graceful degradation, and legible local behavior more efficiently than centralized control. The playful interface matters here: people should be able to poke the organism, damage it, feed it, and understand the result with their hands.

Paul Graham argues that hubs concentrate people, capital, and norms that amplify ambition. Jensen Huang's older execution advice pairs a long horizon with deliberately narrow projects. Sam Altman's “steamroll” warning is best read as a constraint: do not build a thin layer whose only advantage is a temporary model limitation.

All three can be useful without becoming commandments. Networks matter; so do independent thought, places outside the dominant hub, and problems that a frontier lab's roadmap will not automatically absorb. Scale is real. So are embodiment, trust, energy, local context, and the long tail of human purposes.

An inspired mission post should make a talented person feel that their presence changes the outcome. It should not pretend the founder invented the ingredients or that victory is guaranteed.

$\text{strategy}=\text{trajectory awareness}\times\text{independent thesis}\times\text{proof}$

There is no lone-genius version of this work. KAM carries Kolmogorov, Arnold, Moser, Hamilton, and generations of mathematicians. Neural cellular automata inherit cellular automata, developmental biology, differentiable programming, and open research code. The product inherits browsers, databases, model builders, chip designers, maintainers, critics, friends, and the people who contributed care when the work was still hard to explain.

Machines contribute search, synthesis, drafts, and labor. Humans contribute labor too—and judgment, consent, responsibility, lived stakes, and the right to say no. Neither contribution should be erased.

So the operating agreement is simple:

• name sources and contributors;
• separate evidence, hypothesis, metaphor, and marketing;
• publish useful negative results;
• invite criticism before certainty hardens into identity;
• increase human agency without hiding physical or social costs.

“Act in love” is not a claim that good intentions make a system good. It is a demand to keep asking who gains agency, who pays, who is missing, and whether we can still enjoy building it together.

We should narrow, redirect, or stop a branch when:

• local adaptation fails to beat a frozen-plus-retrieval baseline under the same budget;
• stability gates cost more capability than they protect, with no safer niche found;
• an apparent cryogenic advantage disappears after cooling, I/O, and utilization are counted;
• NCA “metabolism” produces pretty motion but no measurable resilience insight;
• independent users cannot reproduce the result or understand the interface;
• the work reduces people's agency, even while improving a technical metric.

Failure is not the opposite of the plan. Hidden failure is. The public artifact should keep a changelog of what survived contact with evidence.