Antisubmissivist
All posts
Thoughts

Food Delivery via Portal Gun? The AI Playbook

Most people use an AI portal gun to deliver food faster. When execution costs hit zero, what are smart people using AI to grab?

A takeout bag being handed through a glowing portal, while the vast unexplored landscape behind the portal stretches away unnoticed.

You meet an alien, and the alien says they can help you enhance two organs in your body. If you could only enhance two organs, what would they be? For this thought experiment, the generally accepted answer might be: first, enhance your brain, and then let the enhanced brain figure out what the next organ should be. Many find this question meaningless since no one actually meets aliens. But they ignore the similar opportunity brought by the AI era.

AI Agents have truly brought everyone a "self-driving computer". The difference between people will soon ruthlessly converge on how they use this new tool. The level of your ability to use tools, or the degree to which you wield the means of production, will rapidly widen the gap between individuals.

When a person masters this self-driving computer, it is as if Rick has acquired portal gun technology, or like getting a powerful outer brain, or more accurately, like a farmer who used to dig with his bare hands suddenly getting a bulldozer. This is a revolutionary, ground-zero moment of change for an individual.

For a long time, the cognitive limit for most AI users might be "What task did I complete with AI today, what website or software did I develop, what workflow did I build?" This is like a delivery rider acquiring a portal gun, only to use it to deliver takeout food faster, rather than exploring the infinite possibilities of the multiverse. Using AI simply to optimize existing, outdated workflows will certainly yield economic benefits, but the accumulation of one's vision and ideological maturity is left far behind.

Recently, I did a comprehensive review of the history of Western philosophy. I realized that the definition of philosophy might simply be the process of asking questions and attempting to answer them. In the AI era, the gap between people might also be defined by a closed-loop process: knowing how to ask questions, understand answers, make judgments, describe requirements, and verify results. Therefore, learning how the smartest minds in history asked questions is especially important in the AI era.

So, the alien's question takes on a new form in the AI era: When you acquire a second brain, what will you use this new tool to think about and create? This is perhaps the very first question one should ponder in the AI era.

I have been thinking about this for over half a year. My answer is: Learn from the smart people. Learn not just their methods, but more importantly, the underlying principles behind those methods. Now, I have decided to share my phased thoughts, fulfilling the purpose of launching this blog: open-sourcing my thinking in the AI era.

I. The General Principles: Three First Principles

When observing frontier explorers, one finds their choices are not made in a vacuum; they are built upon underlying logic.

① When Marginal Costs Go to Zero, Money Doesn't Disappear; It Moves

When models, tools, and robots all become commoditized, the operators of rented assets can only retain upside if they possess a "scarce complement"—otherwise, returns are crushed back to the market average[1]. The things machines cannot build include: physical locations, licenses, proprietary data, and distribution channels. It’s like the gold rush: when shovels are handed out for free, selling shovels no longer makes money. You must own the gold vein.

A corollary logic is that in an AI-abundant era, people will only pay a premium for five things: time, certainty, status, meaning, and authenticity[2]. When a technology drives the marginal cost of execution to zero, the premium doesn't vanish; it merely migrates elsewhere on this fixed menu.

② Your Understanding Is the New Capacity Ceiling

Today, as code generation speeds skyrocket, the new limiting factor has become cognitive capacity. As one developer frankly stated, they can output code a hundred times faster, but they lack the capacity to watch over a hundred times more code[3].

In the past, the fact that "development takes a week" was a discipline in itself; time costs forced people to make trade-offs. Now, adding features is so cheap that no one weighs the costs, and features easily grow in all directions like strange protuberances. The underlying implication of this is not merely a call for caution, but a stark reality: in this era, the only asset that will be magnified 100x is your own bandwidth of understanding.

③ Verifiability Dictates What Gets Flattened

You can outsource your thinking, but you cannot outsource your understanding[4]. The capabilities of current large models are still "jagged"—they are extremely strong in domains where verification is possible, but not necessarily reliable elsewhere.

A critical deduction is that tech giants will eventually eat up all standardized domains where verification infrastructure is already built. The real opportunities exist in the gaps: domains that are verifiable, but where the verification infrastructure has not yet been built.

II. What Smart People Are Actually Aiming At (The Targets)

Based on these first principles, the targets set by those leading the charge are highly specific.

Target A: Building Private Assets That "Skyrocket After Agent Processing" Their goal is not just to use Agents to organize data, but to possess raw material that no one else can get. When public corpora are drained, proprietary data becomes the new gold vein. A decade of industry meeting recordings, an undigitized archive, a set of field records—this "proprietary data + workflow embedding" is the key to ensuring the upside belongs to you[5].

Target B: Building and Occupying the "Judge Infrastructure" in Unbuilt Verifiable Domains Echoing Principle ③, this target involves picking a domain where outcomes can be objectively judged as right or wrong, but where judgment is currently entirely manual, and building an evaluation environment (a sandbox). This is not just an engineering habit; it's occupying the judging power as an asset. Whoever can verify whether a claim is right or wrong holds the power of definition[6].

Target C: Rebuilding All the "Software Too Shitty for Anyone to Fix" Around You (TAM = 1) Coding Agents have pushed the cost of "spinning up a good-enough GUI" to near zero[7]. The smarter approach takes a step further: building fully customized tools for yourself or those around you tortured by bad software. Batch-renaming receipt files, batch-transcribing meeting recordings, turning Japanese vocabulary into lock-screen wallpapers. This marks the arrival of the TAM (Total Addressable Market) = 1 software era—most people don't need production-grade generic software; they need things that exactly solve their own specific pain points[8].

Target D: Taking Back Your Negotiation Power Against Institutions For the first time in history, ordinary people possess infinite patience. Institutions often hold the upper hand in negotiations not because they are right, but because they can afford to stall. Recent research evaluated the ability of "user-owned Agents negotiating on behalf of the user" against institutions (e.g., revenue splits, appeals, refunds)[9]. Wielding a Sovereign Agent that knows the terms inside out and never gets tired against customer service systems is gradually leveling this asymmetric game.

Target E: Building a Thing in Order to Understand That Thing This maps directly to the "outer brain" metaphor. If knowledge is considered the "L1 cache of cognition," having an Agent look something up for you is an order of magnitude slower during invocation than simply knowing it directly[10]. "Building to internalize" is the strongest form of learning. It’s not just storing knowledge in an external system; it's using the process of building to force yourself to walk through the knowledge structure again. The tool doesn't understand for you; the tool is the construction site of your understanding.

Target F: Building Something That Is Still Running Five Years from Now Because Agents have made "building" cheap, most people spend all their budget on building, and zero on "staying alive." The personal projects that truly survive long-term usually share characteristics: self-hosted, possessing persistent memory, and running on personal servers[11][12]. Their commonality is not a pursuit of complex features, but remaining operational without depending on any single company.

III. The Methodology of the Smart: How to Execute

At the execution level, the methodologies of frontier explorers have also undergone fundamental shifts.

1. The Capacity Ceiling Has Moved to "Verification": Build the Judge First

The bottleneck is no longer generation speed, but verification capacity. A notable example is rewriting Git from scratch in Rust using an Agent, generating massive amounts of code and passing tens of thousands of official tests[13]. The real breakthrough wasn't the code volume; it was picking a battlefield where a judge already existed and could not be bribed. There are also fully unattended evaluation loops in which the grading half is locked down, so the Agent cannot cheat by making the exam easier[14]. Before executing any goal, spending time to establish an objective judging standard is the prerequisite for letting an Agent onto the field.

2. Compound Engineering: Judgment Must Crystallize into Files

The core of compound engineering is that every unit of work should make the next one easier; taste should live in the system, not in the manual review process[15]. This is essentially a contract: what files to write when a task ends, and what files to read when the next begins[16]. If every correction made to an Agent is not crystallized into a file (e.g., a proprietary corpus, a rule document), judgment cannot compound.

3. Environment Engineering > Prompt Engineering

The IQ of the model is no longer the sole variable; whether the environment is "Agent-readable" has become more critical. Instead of relying on RAG to re-retrieve every time, it's better to have the model incrementally maintain a persistent, interlinked knowledge base, allowing knowledge to compound in the environment over time[17].

4. Upgrading from "Q&A" to "Having a Heartbeat"

The true watershed is turning the chat box into a living background process. Whether using natural language to write tasks executed on schedule by a daemon, or using a simple database paired with a cron job, the core is giving the Agent a "heartbeat"[18][19]. Only when parallel tasks are truly running independently in the background is the potential of automation truly unleashed[20].

References

  1. Metacircuits. There are only four ways to make money in the agentic economy. ↩

  2. Metacircuits. There are only 5 things people will pay a premium for. ↩

  3. Simon Willison. Conceptual integrity and counting lines of code. ↩

  4. Andrej Karpathy. Sequoia Ascent 2026 summary. ↩

  5. Metacircuits. There are only four ways to make money in the agentic economy. ↩

  6. Andrej Karpathy. Sequoia Ascent 2026 summary. ↩

  7. Simon Willison (quoting Thomas Ptacek). Stop Making TUIs. ↩

  8. Zara Zhang. Coding agents are general agents. ↩

  9. Dylan Zongmin Liu. SovereignNegotiation-Bench (arXiv:2607.02814). ↩

  10. Patrick Collison. YC Startup School 2026 notes. ↩

  11. Two self-hosted personal-agent projects: OpenHuman and Avelina. ↩

  12. Daniel Miessler. LifeOS (formerly Personal AI Infrastructure). ↩

  13. Scott Chacon. Grit: rewriting Git in Rust with agents. ↩

  14. Andrej Karpathy's Autoresearch loop. See the Dooza analysis. ↩

  15. Every. Compound engineering guide. ↩

  16. MoClaw Blog. What Compound Engineering Actually Compounds. ↩

  17. Andrej Karpathy's LLM Wiki pattern. See the AI Builder Club analysis. ↩

  18. OpenClaw official docs. Heartbeat. ↩

  19. Geoffrey Litt. Software You Can Shape (Dialectic interview). ↩

  20. Developers Digest. Git Worktrees + Claude Code. ↩