Write

A coherent set of choices about where to concentrate limited resources to achieve disproportionate impact -- not aspiration, not a list of goals.

The discipline of concentrating energy on a deliberately chosen few things at the expense of everything else -- not trying harder, but choosing less.

The story you tell yourself about who you are -- it shapes every decision, habit, and reaction, operating as both engine (drives behavior without willpower) and cage (traps you in paths that no longer serve you).

The act of cutting through ambiguity to name what matters, what the problem actually is, and what to do next -- deliberate simplification into something actionable.

The act of absorbing complexity and uncertainty so that others can act with clarity and purpose -- less about inspiration, more about sense-making, choice-making, and boundary-setting.

The process of choosing between alternatives under conditions of scarcity, uncertainty, and cognitive bias.

Change that comes in step functions rather than steady inclines, built through sustained effort during flat periods that precede sudden leaps.

Repeated behaviors that become automatic through practice, operating as the compound interest of self-improvement.

The dollar cost of serving a token from a trained model, which is a roofline trade-off between memory bandwidth, compute, and batch size — not a property of the model alone.

Reasoning about a problem as the interaction of components rather than as a property of any one component — the trait that lets engineers spot which constraint actually binds.

The willingness to be vulnerable to another party based on the expectation that they will act with integrity and consistency -- built slowly through repeated actions, destroyed quickly through a single betrayal.

Empirical power-law relationships between model size, data, compute, and loss that let labs predict the return on the next training run before they pay for it.

The deliberate constraint of supply below demand to increase perceived value -- not a production limitation, but a strategic choice that transforms a product into a dream.

A force multiplier on judgment and effort, which Naval splits into labor, capital, and the new permissionless forms (code and media) that let one person reach many at zero marginal cost.

The management of commitments, attention, and systems to translate intention into results -- not doing more, but reliably doing what matters.

Training a model from reward signals on its own generated trajectories rather than from fixed labels; the core of agent-era post-training.

The primary lever for raising baseline happiness, built through mutual over-giving (80/80, not 50/50) and sustained investment of time and attention.

Models that update their weights from ongoing experience the way humans do, instead of being frozen after pre-training and refined only via fine-tuning or RLHF.

How model and agent quality is measured; in 2026 the bottleneck is less whether a model can do a task and more whether the task is well-defined and well-measured.

The physical substrate of AI — chips, HBM, racks, scale-up networks — whose generation-over-generation shifts (Hopper → Blackwell → Rubin) gate which model architectures and context lengths become economical.

Specifying the task and the desired behavior precisely; the scarce skill once raw model capability is commoditized.

A language model operating in a loop with tools, memory, and an environment, taking multi-step actions toward a goal rather than returning a single response.

The nonlinear payoff of small, consistent inputs repeated over long horizons, in relationships and reputation as much as in capital.

Signals that point to whether our needs are being met, which Rosenberg argues are caused by our own needs rather than by other people's actions.

Training the body, where intensity, recovery, and standardized measurement matter more than volume.

The rate at which a GPU can move bytes between HBM and the compute units — typically the binding constraint on LLM inference latency and on practical context length, far more often than raw FLOPs.

A durable competitive barrier; in 2026 frontier AI, most defensibility still sits on the model side rather than in the application layer.

The initial large-scale self-supervised training of a base model on broad data, setting raw capability before any post-training.

The cached key/value tensors of previously processed tokens that an autoregressive decoder attends to — the part of inference memory that scales with batch size and context length, and which neither batching nor pipelining can amortize away.

Curated, verifiable task environments in which humans build the framework, tools, and reward rubrics so models can learn to perform real, multi-tool, long-horizon work; the dominant new post-training data type in 2026.

The gap between what a product does and what it means -- a story powerful enough that people pay for the meaning, not just the function.

Not a momentary feeling but a baseline that shifts through sustained effort during the 90% of life that feels "normal."

The internal drivers (meaningful work, growth, responsibility, recognition) that make someone want to act, as distinct from external prods or the mere absence of dissatisfaction.

Everything done to a base model after pre-training (SFT, RLHF, RL on agent trajectories) to shape its behavior and unlock capability.

How much capability a model gains per unit of data; the axis on which current models trail humans by roughly a millionfold during training.

The DeepMind result that, for a fixed training compute budget, the optimal model size and training-token count grow at roughly equal rates (~20 tokens per parameter) — a law about *training* cost only.

The conviction about what you are capable of, which Schwartz argues acts as the trigger that mobilizes the mind to find ways to succeed.

The felt sense of competence that grows from acting first, since Schwartz argues action precedes and produces confidence rather than the reverse.

The maximum number of tokens a model can attend to in one forward pass — practically capped not by training tricks but by KV-cache memory bandwidth at serving time.

The respectful understanding of what another person is experiencing, which Rosenberg insists is presence with their feelings and needs rather than advice, reassurance, or analysis.

Testing claims on yourself or your system with controlled variables instead of trusting consensus advice.

The anticipatory discomfort that blocks action, often disguised as prudence or optimism and best defused by naming the worst case explicitly.

The smallest input that produces the desired outcome; anything beyond it is waste or harm.

The universal human requirements underlying every action and feeling, which Rosenberg separates from the specific strategies we mistake for them.

Direction aimed at something outside yourself, function over form, which Holiday contrasts with the breathless self-focus of passion.

Training loops in which a model (or another model) influences its own training during the forward pass, so capability gains compound without requiring proportionally more human-generated data or labels.

The set of GPUs/TPUs connected by the fast, fully-connected scale-up network (NVLink/NVSwitch on Nvidia, ICI on TPUs), inside which all-to-all communication is cheap — typically one rack on Blackwell (72 GPUs), much larger historically on TPU pods.

Making your wellbeing depend on a person, outcome, or identity, which de Mello identifies as the root of fear, grief, and suffering.

Watching what is happening in and around you as if it were happening to someone else, which de Mello treats as the master key to change.

Running many independent inference requests through the same forward pass so the cost of fetching weights is amortized across users — the single most important lever in inference economics.

A way of solving recurring problems together that has worked so reliably the group stops considering alternatives.

What reaches the bloodstream and what it signals hormonally, not what enters the mouth.

The set of ways to spread a model and its activations across many chips: data, tensor, expert, and pipeline parallelism — each cutting along a different dimension of the model.

The distance between what an evaluation measures and the real-world capability users actually care about; closing it means building evals that mirror the true distribution of valuable tasks.

Assets that earn while you sleep, which Naval distinguishes sharply from money (how we transfer wealth) and status (your place in a hierarchy).