Bản chép lời đầy đủ của The Window Opens Again: d/acc Has a Build Order

the work: "I think that not just magnitude but also direction matters." And then the direction: "The 'd' here can stand for many things; particularly, defense, decentralization, democracy and differential." 21 00:03:41.821 --> 00:04:02.309 Thirteen months later, on January 5, 2025, he compressed it to one line: "decentralized and democratic differential defensive acceleration. Build technologies that shift the offense/defense balance toward defense, and do so in a way that does not rely on handing over more power to centralized authorities." 22 00:04:02.909 --> 00:04:35.395 Both halves are load-bearing. He says why: "any kind of decentralized, democratic or liberal political structure thrives best when defense is easy, and suffers the most challenge when defense is hard." When defending yourself is cheap, you get to be free. When it is expensive, you rent protection from whoever is strongest. His own gloss goes straight there: defense "like in democratic Switzerland and historically quasi-anarchist Zomia, not like the lords and castles of medieval feudalism." 23 00:04:35.995 --> 00:05:03.561 Decentralize, encrypt, accelerate is a different slogan, routinely confused with this one. We went looking for a named origin and found none: no manifesto, no canonical post, no author to credit. It is the cypherpunk distillation, thirty years older than the d/acc label. Eric Hughes, March 9, 1993: "We know that software can't be destroyed and that a widely dispersed system can't be shut down." 24 00:05:04.561 --> 00:05:07.192 "Math-Backed" Is the Wrong Compliment. 25 00:05:07.992 --> 00:05:34.609 People who like d/acc call it math-backed. Be precise, because this audience checks. The first pillar is differential technological development, and it is not Buterin's. It is Nick Bostrom's, from his 2002 paper on existential risks: "trying to retard the implementation of dangerous technologies and accelerate implementation of beneficial technologies, especially those that ameliorate the hazards posed by other technologies." 26 00:05:35.209 --> 00:05:48.877 The second pillar is from international relations: Robert Jervis, "Cooperation Under the Security Dilemma," World Politics, January 1978. Two variables decide whether rivals can cooperate without a hegemon 27 00:05:49.227 --> 00:05:56.337 whether offense or defense holds the advantage, and whether an offensive posture can be told apart from a defensive one. 28 00:05:56.937 --> 00:06:07.560 That is a real intellectual spine, and it is not a theorem. The offense/defense balance is contested inside the field that invented it, and we will come back to how badly. 29 00:06:08.160 --> 00:06:27.153 So here is the stronger compliment, which happens to be true. d/acc is the only framework in this argument that specifies what to build first and why. Everything else supplies a throttle setting. This supplies a sequence, and a sequence can be handed to somebody with a soldering iron. 30 00:06:28.153 --> 00:06:32.430 Mozi Belongs to d/acc, Not to Effective Altruism. 31 00:06:33.230 --> 00:07:02.241 Mozi, or Mo Di, Master Mo, turns up in fifth-century BCE China. The Stanford Encyclopedia of Philosophy notes the early sources depict him as a master craftsman and military engineer, probably an artisan and very likely a carpenter. A tradesman with a doctrine, and two planks matter here: jian ai, rendered "inclusive care," concern for everyone weighted no more heavily for your own family or state; and fei gong, "Condemning Offensive Warfare." 32 00:07:02.841 --> 00:07:37.766 What the school did with that pair should stop you. "Far from being pacifists," the encyclopedia says, "they saw a strong defense force as crucial to state survival," and some Mohists became specialists in defensive warfare, rushing to the aid of cities under attack. The last block of the Mozi is not philosophy. Chapters fifty-two to seventy-one are military engineering addressed to his chief disciple Qin Guli: fortification specifications, counters to scaling ladders, tunneling and battering rams. Engineers with an ethic, who would help you hold a wall and would not help you breach one. 33 00:07:38.366 --> 00:07:57.321 Nothing about the man himself is reliably known, and the material is thinner than a confident paragraph makes it sound. Only eleven of the twenty-odd military chapters survive, badly corrupted. None of the Mozi claims to be from Mo Di's own hand, the defense chapters were most likely written by Qin Guli's students. 34 00:07:57.921 --> 00:08:31.431 Neither d/acc essay mentions Mozi, the Mohists, or any Chinese philosopher. We checked; the count is zero. But on September 17, two days ago, Buterin made the connection himself on Farcaster, twice in twenty-six seconds. The first cast aimed it at effective altruism: Mozi was "surprisingly close to being a proto-effective-altruist." That is the one the crypto press reprinted. Two replies down, he corrected the aim: "Mozi is also surprisingly proto d/acc (!!)" 35 00:08:31.781 --> 00:08:48.143 the defense chapters "are basically open-source city defense manuals," and "the explicit goal was to upgrade people's knowledge of defensive arts, and thus shift the offense/defense balance further toward defense." The aggregators did not reprint that one. 36 00:08:48.743 --> 00:09:11.484 The second aim is the right one, and it is not close. Mohist practice was not expected-value estimation across cause areas. It was fortification engineering, written down and handed out. Effective altruism takes the impartial ethics and arrives at compute governance and a short list of trusted labs. The Mohists gave away the plans for the wall 37 00:09:11.834 --> 00:09:19.018 the oldest anti-landlord technology on record, because a city that can hold its own wall never buys protection. 38 00:09:19.618 --> 00:09:35.626 The parallel has a hard limit. Mohist political doctrine was centralized: "Identifying Upward," everyone aligning their judgments to their superiors.

Mozi is an ancestor of the defensive engineering and the impartial standard, not of decentralization. 39 00:09:36.626 --> 00:09:39.502 The Case for e/acc, Made Properly. 40 00:09:40.302 --> 00:09:57.616 Effective accelerationism was formalized in mid-2022: the original post came from the account known as Swarthy, the physics notes from Beff Jezos, with bayeslord and creatine_cycle as coauthors. Forbes identified Jezos as Guillaume Verdon in December 2023. 41 00:09:58.216 --> 00:10:05.190 The argument deserves better than the caricature. It is a physics argument. Life emerged from dissipative adaptation 42 00:10:05.540 --> 00:10:09.594 matter reconfiguring itself to extract free energy 43 00:10:09.944 --> 00:10:37.979 which Jezos traces to the Jarzynski-Crooks fluctuation theorem. The universe "exponentially favors" futures where matter captures more free energy, and "capitalism is hence a form of intelligence." Then the line we agree with: "top-down control is not fault-tolerant in its operation to corruption of root nodes in the hierarchy." Central planning is not primarily a moral failure. It is an information failure. 44 00:10:38.579 --> 00:11:00.649 Here is where it collapses. Thermodynamics does not rank. A vaccine platform and an engineered pathogen both dissipate free energy, and the universe has no opinion about which you build on Tuesday morning. "You cannot stop the acceleration. You might as well embrace it. Accelerate." As a build order that is empty. 45 00:11:01.249 --> 00:11:22.343 Which is why the landlords can fund it without flinching. An accelerationism that ranks nothing endorses whatever is already being built, and what is already being built is one hundred eighty-five billion dollars of rental property. e/acc supplies a physics argument for the capital expenditure budget. It cannot supply an objection. 46 00:11:22.943 --> 00:11:54.675 Then the stated commitment: "e/acc has no particular allegiance to the biological substrate for intelligence and life." As an engineering priority that is disqualifying, because a framework indifferent to whether humans are still here cannot tell you which failure modes to prevent first. And Buterin lands the cleanest hit: treating frontier weapons development as a force for good requires believing "the dominant technological power will reliably be one of the good guys in most conflicts, now and in the future." 47 00:11:55.675 --> 00:11:58.831 The Case for the Safety Camp, Made Properly. 48 00:11:59.631 --> 00:12:12.336 Effective altruism and the AI-safety movement have a stronger claim than their critics admit: they supplied the spine. Differential technological development is their concept, and Buterin says so himself. 49 00:12:12.936 --> 00:12:49.352 They also produced the best objection to d/acc, which Buterin published against himself: "it's looking likely we have three year timelines until AGI, and another three years until superintelligence," so "we can't just accelerate the good, we also

have to slow down the bad." His two answers: liability placed as close to end use as possible, and a soft pause able to cut worldwide compute by ninety to ninety-nine percent, via trusted hardware that stops running unless it collects three signatures a week from international bodies. 50 00:12:49.952 --> 00:13:28.439 The failure is structural, and shows up every time this camp converts a concern into policy: the answer collapses toward a control point. Sam Altman, before the Senate on May 16, 2023, proposed "a combination of licensing and testing requirements for development and release of AI models above a threshold of capabilities." Aschenbrenner proposes the United States race to a decisive lead and dictate terms to China. And the floor of the genre is Bostrom's scenario, which Buterin quotes: everyone wears a "freedom tag" uploading encrypted video and audio for machine interpretation in real time. 51 00:13:29.039 --> 00:13:46.455 What the center never prices is the risk that the center is the problem. Buterin lists the receipts: gain-of-function research funded by major world governments as a plausible origin of the pandemic, and the World Health Organization taking years to acknowledge that Covid was airborne. 52 00:13:47.055 --> 00:13:55.265 None of this makes them villains. It makes them predictable. Give this camp everything it asks for and the artifact is a license 53 00:13:55.615 --> 00:14:05.237 a title deed, issued to whoever already keeps a compliance department and a Washington office. Whoever captures the issuing desk gets to aim it. 54 00:14:06.237 --> 00:14:09.632 What a Build Order Looks Like on a Workbench. 55 00:14:10.432 --> 00:14:29.643 Now the operational part, the reason any of this matters. Buterin's triage for AI risk is three lines long and it is a ranked list: "It hacks our computers, cyber-defense. It creates a super-plague, bio-defense. It convinces us, either to trust it, or to distrust each other, info-defense." 56 00:14:30.243 --> 00:14:58.732 Read that as a work queue and it tells you things. A formal verification toolchain outranks a bigger model. Sandboxing and egress monitoring outrank a policy document about responsible scaling. The 2025 follow-up lists what shipped in a year under that banner: verifiable open-source vaccines, healthy indoor air, Community Notes, prediction markets maturing as an information tool, zero-knowledge proofs in government ID. 57 00:14:59.332 --> 00:15:10.751 The design principle underneath is the best sentence in either essay: focus on "building mecha suits for the human mind, not on creating new forms of self-sustaining intelligent life." 58 00:15:11.351 --> 00:15:50.503 Here is why it survives contact with people who ship. Nearly every item on that list can be built by someone who is not a frontier lab. Air filtration is a fan, a filter and arithmetic. Open vaccine protocols are a publication. Sandboxing and audit logs are a weekend and some discipline. You cannot build a licensing regime in a hackerspace. You can build almost everything on that list in one, publish the plans, and have ten thousand people holding it by Friday. That is the difference between a defense you own and one you rent, and exactly what the Mohists did with the fortification manuals. 59 00:15:51.503 --> 00:15:53.967 The Strongest Case Against This. 60 00:15:54.767 --> 00:15:59.250 Now the part that could take the whole thing down, made properly before we answer it. 61 00:15:59.850 --> 00:16:34.712 The offense/defense distinction may not survive contact with real technology. Buterin concedes a version of this and treats it as ambiguity at the edges. It is not at the edges. It is the center. Every defensive technology is a capability, and capabilities are dual-use by construction. Formal verification hardens your code and hardens malware. A vaccine platform is the same bench and techniques as gain-of-function research. Encryption protects a dissident and a ransomware crew with the identical primitive. 62 00:16:35.312 --> 00:16:52.350 Which breaks both pillars. Jervis's second variable was whether an offensive posture can be told apart from a defensive one; for software it usually cannot. Bostrom's differential development assumes you can sort technologies in advance, when in practice you sort them afterward. 63 00:16:52.950 --> 00:16:59.177 So defense-favoring is a judgment call made by whoever is building, the party least able to be neutral 64 00:16:59.527 --> 00:17:19.703 and everyone passes their own test. Altman believes licensing is defensive. Yudkowsky believes the ban is defensive. The procurement officer believes autonomous targeting is defensive. A build order every camp can claim to be following is not a build order. It is a mood with better vocabulary. 65 00:17:20.303 --> 00:17:34.023 All of that survives, and we are not going to pretend otherwise. There is no test that takes a technology and returns offensive or defensive. Anyone who tells you this framework settles that question is selling something. 66 00:17:34.623 --> 00:17:53.157 Our answer is narrower. The framework was never a claim about technologies. It is a claim about structures, and there is one question you can answer about an artifact before it exists: does it need a permanent operator? A filter box, a published vaccine protocol, an open set of weights 67 00:17:53.507 --> 00:18:02.851 once you have one, nobody can take it back or raise the price. A licensing regime, a soft pause, a model behind an API. 68 00:18:03.201 --> 00:18:12.177 each works only while a specific someone stands at a specific chokepoint, and that someone is a landlord whether anybody meant to hire one or not. 69 00:18:12.777 --> 00:18:25.806 That is less than the framework's fans claim. It will not tell you whether a given biotechnology is net offensive; we do not know how to build a test that does. It tells you who will own the thing afterward 70 00:18:26.156 --> 00:18:31.348 smaller than a theory of technology, and larger than anything the other camps have on offer. 71 00:18:32.348 --> 00:18:34.187 Where We Part Company. 72 00:18:34.987 --> 00:18:40.035 This is not a fan letter, so: two disagreements, and they are the same disagreement twice. 73 00:18:40.635 --> 00:19:04.086 The soft pause fails its own test. Industrial hardware that will not run without three signatures a week from international bodies is a global chokepoint with better ergonomics. By the framework's own reasoning it turns offense-favoring the moment anyone captures it, and it is the permanent operator written into silicon. The artifact is the problem, not its governance. 74 00:19:04.686 --> 00:19:42.681 The second is the hedge on open weights. He writes that there is "a high chance that the most powerful models will be closed source anyway." As a prediction that may be right. As a plan it is unacceptable, because every defense built downstream of a closed frontier is built on a landlord's terms, and the landlord revises the terms on a Tuesday. The weights are defensive infrastructure and they belong in public, published like the fortification specs. Not communism, not socialism: bootstrapism. The people who tell you to pull yourself up by your bootstraps are standing on the bootstraps. 75 00:19:43.681 --> 00:19:44.975 Pick the Direction. 76 00:19:45.775 --> 00:20:04.007 The goal has not changed since the first post on this site: post-scarcity, and the end of the enormous share of human suffering that exists only because what people need is still expensive. That requires speed, and every year of delay has a body count the delayers never put on their own ledger. 77 00:20:04.607 --> 00:20:41.762 But speed alone does not pick a destination. Speed in an offense-favoring world produces a warlord, because there the cheapest way to be safe is to be the biggest threat. Speed in a defense-favoring world produces something closer to Switzerland with better tools. Same throttle, different sequence, completely different century. e/acc says floor it and hands you no map. The safety camp hands you a map whose only road ends at a checkpoint. d/acc says floor it in a direction: build the shield before the sword, publish the plans, refuse to manufacture the chokepoints. 78 00:20:42.362 --> 00:21:06.402 Mozi had the doctrine right twenty-four centuries ago and lost anyway. His school could rush engineers to one besieged city at a time, and when the Mohists faded the manuals sat unread for two thousand years. We can copy a defensive design to every city on Earth before dinner. That is the genuinely new variable, and why the build order can win this time. 79 00:21:07.002 --> 00:21:16.665 Nobody is coming to hand you a defense you get to own. Build one. Publish it. Accelerate, and pick the direction. The only way out is through.