Nuclear Deterrence Will Fail, AI Labs May Be Nationalized: 46 Bold Claims About the Future by an Anonymous Researcher
TechFlow Selected TechFlow Selected
Nuclear Deterrence Will Fail, AI Labs May Be Nationalized: 46 Bold Claims About the Future by an Anonymous Researcher
"Everyone is using past efficiency curves to understand AI, but the real big leap hasn't come yet; intelligent production may still have four to ten orders of magnitude left to go."
Navigating Web3 tides with focused insightsAuthor: bayeslord
Compiled by: TechFlow
TechFlow Editor's Note: bayeslord (@bayeslord) is an anonymous but weighty account in the AI × crypto circle, not selling goods or chasing trends, but专门 pushing hard along technical cores like scaling law and algorithmic depth.
The blogger recently wrote a 46-point list, deducing the future development of technology, AI, and related technologies, believing that everyone is using past efficiency curves to understand AI, while the real big jump hasn't come yet, and intelligent production may still have four to ten orders of magnitude left to go.
He talks all the way from algorithmic acceleration to robots, capital, and a permanent underclass, finally landing on the sharpest point: Mutually Assured Destruction (MAD) may fail, military and police will be automated, and AI labs may be nationalized.
The original post has nearly 1 million views. Although the views are extreme, each point is relatively self-consistent and worth a look for general technology readers.
This list is based on a tweet thread I posted on June 4, with some modifications and additions. Several people said the original thread was too hard to read, so I organized it into this version.
Intelligence
1. Algorithmic progress will catch everyone off guard. The entire world—markets, governments, militaries, companies, individuals—is using production efficiency and patterns from recent years to understand the impact of AI and judge how things will likely proceed. Even those new labs claiming to believe in "recursive self-improvement" think this is just the old stuff with an agent running in a loop. It's not like that. I guess there are still many orders of magnitude left in intelligent production, perhaps as many as ten, with four to seven being more likely. In principle, exceeding ten is not impossible, but that would hit hard against the upper limit I suspect physical laws truly allow. Unlikely, but not ruled out. If this judgment holds, then the real direction of things is not what it appears to be on the surface, and a big jump is approaching. Anything happening along this direction will make the world much weirder than almost everyone is pricing in.
2. We are in the early stage of takeoff. AI improving AI may ultimately become the step with the most severe consequences in history. This cannot be guaranteed, because we don't know how far we are from the physical and computational limits of intelligence, but I bet we are still far away (as mentioned earlier, squeezing out 4 to 10 more orders of magnitude of intelligent output per unit of compute seems possible).
3. Since entering the takeoff phase, algorithmic research is accelerating. Compute is still a scarce resource, but the opportunity cost of researcher time has come down, because you can directly send an agent to run any task, even if it's just messing around. It might bring something back. All new ideas carry a debt of "optimization debt," and now this debt can be paid off with unsupervised token consumption. Massive research scaling law curves will be walked through one by one.
4. AI models will continue to get stronger, especially frontier models. The only real wall is physics. Models are becoming more autonomous, smarter, and constantly improving. Math and code are being conquered by scale plus reinforcement learning, and the rest are lined up behind. The distinction between "verifiable" and "unverifiable" will slowly disappear as a meaningful boundary. Moving forward, automated AI research and AI learning will become more and more the same thing. Training a model well is essentially closely related to the model learning well on its own. Sample efficiency, creativity, and all other limitations will be solved, then approaching algorithmic optimality at any scale.
5. The idea that long-task agents must have training data of equal length is wrong, because generalization exists in the time dimension. Long tasks are not built up by the attribute of "length." This is related to LeCun's fallacy of (1-e)^n error accumulation. What really happens is error correction. Error correction happens simultaneously at multiple scales, from the level of single token generation to every step in a long task. The reason that METR chart goes up is partly because agents are beginning to touch the escape velocity of error correction.
6. An engineering-level deep learning science is about to emerge. It will push us toward the algorithmic maturity of AI, much faster than most people expect—although as mentioned earlier, how far this path can go in principle is still unclear. For example, a science studying scale invariance will greatly increase the scale and return of useful experiments, because an experiment on one GPU can tell you how to use one hundred thousand.
7. Every field of technical human activity will have its own "Move 37" moment (the move AlphaGo played against Lee Sedol that surpassed human intuition), and then soon, "Move 37" itself will appear outdated. I mean all fields.
8. Compute will continue to get better. Today's best matrix multiplication machines are far from the physical limits of AI accelerators. There is still much room for improvement in the digital silicon path. There are many candidates for new substrates, and their owed algorithmic debt will be squeezed to the limit by automation, but we don't know yet what is optimal for AI in terms of space, energy consumption, time, manufacturability, and cost. Photonics and random silicon are interesting candidates, but I also expect the Singularity itself to be surprising.
9. How far labs can lead depends partly on the returns from automation and scale, including returns from deeper algorithmic depth. If the practice (and theory) of deep learning is always shallow, then in the long run the moat will probably not be mainly at the algorithmic level, because secrets are relatively easy to discover. Eventually, distillation plus data plus time can catch up to compute scale, maybe a bit slower. Currently we seem to be partially in this state, but even if so, no one guarantees it will continue this way.
10. If deep learning becomes less shallow as scale increases, then every increment of automation and scale will buy you algorithmic secrets that others can reach less and less. This state also seems to partially fit us currently. The endpoint of both situations is when marginal utility returns saturate with scale and research. We don't know where that point is. It could be 2 orders of magnitude from today, or 20. No one knows.
Intelligence Supply Chain
11. For at least the next few years, compute will be a fiercely contested resource. But during this time it will begin to commoditize, and we will look back and laugh at the meagerness of the 2020s. Scale is expanding and working, capital is following in, turning the flywheel over and over. More matrix multiplication machines, more wafer fabs, more energy are on the way. The bottleneck of intelligent production is temporary. Excluding potential economic speed bumps.
12. The nature of the intelligence supply chain is changing. Now it is highly concentrated in the hands of labs. But labs are automating the core thing that makes them strong—researchers, and the discovery of algorithmic advantages. Once this process begins, assuming open source doesn't fall too far behind, especially if labs don't lock down AI researcher models, then the labs' advantages will shift to easier financing, more compute, exclusive data, business relationships, and good products. This indeed depends on how the aforementioned algorithmic depth question plays out, and some other factors.
13. Distributed training will reduce the need for large-scale construction of单体 data centers, giving some advantages to non-hyperscale vendors. However, in the pure dimension of single largest-scale training, it will not exceed hyperscale vendors.
14. Automated AI experiments will allow algorithmic secrets to be widely discovered, because these secrets are naturally easier to distribute than full-scale training. How far this path can go is unclear, but I expect quite far. As mentioned earlier, the fundamental depth of deep learning is still unknown, and the upper bound of this judgment depends on that unknown.
15. Although these forces表面上 benefit academia and open source, they may still shrink due to the cost and opportunity cost of compute. For example, is GB300 more valuable serving GLM5.2 or Fable, or doing non-frontier research in an academic lab, or building Mythos 2 inside Anthropic? The market will calculate where the demand is greatest, and right now, that place is indeed labs. This means open source labs may become more compute-hungry, even if they have money, provided they haven't locked in compute capacity in advance. Even if locked in, they still have to weigh the opportunity cost of doing research themselves versus renting out compute. Refer to the collaboration between Colossus and Anthropic.
16. In an environment where AI capabilities start becoming stimulating (next 0 to 18 months), open source may also start becoming difficult at the social level, especially if we accelerate safety slowly—which so far has indeed been slow.
17. When capital floods into labs, open source may begin to shrink. There is a coordination problem here: except for labs (and maybe governments), no one wants a token monopolist, but if this problem can be solved and the regulatory environment is friendly, maybe the result will be okay.
Robotics
18. Robotics will have a moment similar to ChatGPT in November 2022, and then another moment similar to Opus 4.5 in November 2025. Neither has happened yet, but they are coming, and faster than people think, this is the result of rapid AI progress, including physical system engineering accelerated by AI. The interval between these two moments for robotics does not look like it will be three years.
19. However, to truly stack up the number of robots worldwide physically, it may wait until 2030 or even later. Although we build about 100 million cars a year, and humanoid robots are much smaller than cars. Considering we also build 1 billion smartphones a year, if capital and algorithms run fast enough, achieving a scale of 100 million robots per year by 2030 is reasonable. 10 million per year can definitely be done; we are already doing the drone market. As long as software can prove humanoid robots are worth the money on a small scale, it can leverage infinite capital, proportional to the quality of the proof.
20. Things that look like hard ceilings for robots today will disappear, including poor sample efficiency, relatively scarce data, expensive and difficult hardware design for hands and motors, the fractal complexity of the physical world, and those unrecorded tacit knowledge about how we work in the world (like the plumber's set). World models look useful, but exactly which thing matters is not important. Research scaling law will be ground down until diminishing returns.
21. Global demand for robots is easily hundreds of billions of units, especially if various forms are added together. There is too much physical labor worth automating. The market will find a way to solve this, and people probably won't block the way.
Progress
22. Science is being automated and virtualized. This means much of the progress this world needs will come from automated labs and simulations. We don't know the full computational limits of virtualization, but robot-driven labs in fields like biology and materials science will remove many bottlenecks, pushing up the boundary of "verified virtualization" along the way to improve sample efficiency and net returns of "landing true." Basically in every field, we will have some combination of neural models, explicit simulations, and real-world experiments, together improving the return per dollar and per unit of time in fields like biology and materials science.
23. Laws of progress are everywhere. In deep learning they are called scaling law. On any curve, when the S-curve saturates is hard to judge, and whether there are new S-curves beyond the horizon is also hard to judge. What needs to be understood here is that the engine of civilizational progress itself also has a law of progress. Our progress is likely saturation-type, like most natural processes, but we actually don't know where saturation occurs. The maturity of technology and civilization may be near, or far. We are at such a historical node: first, we have invested almost no resources into progress yet, but this is changing rapidly; second, we are automating the machine that directly produces more progress. We are in an interesting era.
24. The future of scaling up or scaling out. Zero to one or one to n. How much progress the universe allows us to make in breadth and depth is an open question. Breadth is easier to estimate, because it is roughly "from now on, how many steps of computation do physical laws still allow us to take?" And how "deep" that computation can be—in the broadest sense of the word—is unknown. In some versions of the future, the tech tree is incredibly deep, the accessible computational universe is so rich that we will keep inventing and discovering until physics blocks us, if it can block us. Other versions are flatter: we soon fill up a relatively shallow tech tree, relatively easily reach technological maturity, and then scale it out until satisfied or physics blocks the way.
Capital and Production
25. More capital plus more intelligence means a more intensified capitalism, means we rush toward market equilibrium faster. In the long run this naturally should lead to deflation, leading to most important goods competing to near zero marginal cost, including AI, food, housing, medicine, electronics, entertainment, and travel. Provided we don't let people block the way. In some cases they probably will.
26. Mining will be automated. Sea, land, and air transport will be automated. Factories will be automated. Workers will be automated. Distribution centers will be automated. The maintenance, improvement, and expansion of the entire supply chain will be automated.
27. There will be humans retaining jobs, retaining them for a very long time. What proportion of humans this number accounts for is an open question. People who say this number will be high are overconfident, and those who say this number will be zero are equally so. However, it is indeed hard to imagine how long humans can contribute marginally in the "knowledge" part of knowledge work. Some demands, like doctors, may drop significantly—if we have superhuman AI doctors for $20 a month, plus on-demand testing, plus significant health improvements from medical tech progress. But because we have cartelized doctors now, we may continue to do so, and being a doctor will still be a good profession. Demand for entertainment will probably rise, but production costs will fall, and the technical demand for entertainment for humans has already decreased significantly. But we care about other humans, so maybe we will continue to care about them, and being an actor will become more profitable. There is a line of thinking that can help you figure out how this evolves: from a worker today to a consumer, how many intermediate layers are in the supply chain. For a TikTok influencer, zero layers. For a doctor, zero layers. For a factory worker, many layers. Whether a job can be disintermediated, eliminated by competition, or is irreplaceable will probably largely determine its outcome. This analysis is quite subtle, this paragraph is far from exhaustive, but finally to mention one point: all this presupposes we don't encounter a cliff-like collapse on the demand side—if too many people don't work, and productivity or government efficiency can't support universal basic income or universal basic healthcare, that collapse could happen.
28. Related to but not contradictory to the above points: a "permanent underclass" may truly exist. In the better worlds where it comes true, it looks more like agency being highly restricted, rather than income being devastated. For most people this is ultimately acceptable, our agency was already highly restricted by modern society, but it requires psychological adaptation, which may take time, and may be painful.
Culture and Psychology
29. The human mind now grows and adapts slowly, but this will change. The key is changing in a good direction, which is not easy for some people. Abundant intelligence and automation will allow us to engineer psychological structures far more durable than today's—today's set is an evolutionary leftover not adapted to our environment. Psychiatry and psychology will walk through a thousand years of innovation in no more than a few decades. Humans will fundamentally get better. Rough, degenerate "pleasure direct connection" is overestimated as a risk, because we will have more sophisticated and diverse mind engineering available.
30. In an extremely uncertain world, people will fight for power, status, and wealth more fiercely than ever before, betraying their own kind with a clear conscience in the process. They will invent various reasons to explain their behavior is good, even great. Look around.
31. You will live to see embarrassments you cannot believe.
32. Now there is an obvious double discourse playing out: those who are about to be, or already are, the richest 0.01%, on one hand say AI will benefit everyone, don't worry about jobs, but on the other hand are unwilling to give up their wealth to be a random member on Earth, or even in the US, regardless of whether the term is one year, five years, or twenty years. People can see this and have started reacting. To be clear, I won't give up my position either, but I also didn't say everything will be perfect (and I'm not the richest 0.01%). The result is, we have the risk of building an unjust world. Some people care about this, I think this matter should be discussed more frequently. And to put it more bluntly, the way US politics handles such issues is rotten.
33. Elon looks very likely to become the first quadrillionaire. Broadly speaking, it's not hard to imagine demand for chips, robots, and spaceships will rise more than a thousand times from now, and he will probably eat up a large chunk of this.
Coordination
34. Better coordination is needed at all scales of society, this is obvious. With our current understanding of coordination, it has weaknesses and risks, but likely we haven't even scratched the surface of it. Will there be a Satoshi-level figure to take down Moloch (symbolizing malignant competition where everyone is forced to participate and no one can exit)?
35. Some international coordination on AI is probably a good idea. We may want treaties and GPU counting. This stuff can be designed to: first, slow down the spiral of antagonistic accumulation of military and government power, second, minimize impact on science and other important progress fields. We may not get this, because GPUs are too universally powerful. We managed it on nuclear weapons because except for madmen no one really wants to use nukes.
36. A pause or slowdown of AI lab coordination now looks more likely than in 2023. There are many tradeoffs here, but I think the debatable value of a pause is slightly higher today than in 2023. The argument that "a pause will be wasted" is harder to sustain because we have automated research—although not fully yet (what we have is automated engineering). To be honest, I personally do not support a pause now, mainly because it would interrupt too many other parts on the tightrope crossing the Singularity, there may be dragons in the tech tree, and opponents are real.
Power, Violence, Security, Liberty
37. I regret to tell you, our universe may be fragile in the Bostromian sense (philosopher Bostrom's "Fragile World Hypothesis": technological progress may pull out some capability sufficient to destroy civilization once discovered). It is possible that there exist some degrees of freedom in the current world that we cannot control quickly while preserving the norms of governance and liberty (this set of norms is sufficient for the truth of our world, except for the panopticon). Note that in such a world, power accumulation is a slippery slope. Many such worlds end up very bad for most people. I wish this weren't true, but it may be.
38. AI proliferation will happen at some speed greater than zero, regardless of how many potential speed limit factors there are. There are too many computers in the world, and the exchange rate of FLOP for intelligence is at the lowest point in history. Don't bet things will stall.
39. The concept of a "permanent underclass" implies the existence of a "permanent upper class." This presupposes a group of people having more rights for some relatively illegitimate reason. This reason ultimately is always implied or cashed-out domination supported by violence. But perhaps a world with advanced AI is a world where humans no longer have legitimate reasons to rule, without any recognized ability or status surpassing other humans. This will never be 100% true, but it may become increasingly important, worth thinking about. I suspect in practice, moral arguments and practical arguments will diverge quite significantly, and perhaps that is actually right.
40. Forces in various directions will push institutions to transform, and these forces may lead to tyranny. There are many paths there, some under the banner of security, some are benign power creep—the ceiling is powerful AI plus fully automated military supply chain plus fully automated weapons. We need better institutions.
41. There may be a large number of zero-day vulnerabilities out there. In networks, biology, infrastructure, neural, memetic, physical fields. We simply don't understand the returns of algorithmic depth and consistency in these fields, whether on the defense and robustness side, or the destruction side. The algorithmic depth of nuclear weapons is not out of reach for the smartest humans in the world. Tomorrow our machines will reach the next level, and the next. Now we know a little about the random disaster rate for algorithmically shallow things, and know almost nothing about what will happen in an algorithmically deep civilization.
42. A related sentence: there may be some really fucked up things in the tech tree. We really don't know.
43. Scaled robotic capabilities bring real takeover and coup-style risks beyond pure computer models, and also more mundane things, like new surfaces and vectors for cyberattacks. We should take these risks seriously and strive to reduce them.
44. Mutually Assured Destruction is built on technology from the 20th and early 21st centuries. We will undergo drastic technological change, perhaps a thousand years' worth of change compressed into a very short time. This means MAD is not taken for granted. This problem is solvable, but neither completely certain nor a clean overthrow, because to get decisive advantage, the margin for error is extremely low, possibly fundamentally infeasible. In the past some people brought up this topic in a rather unserious way, I think that is wrong and irresponsible. This is one of the most serious topics we can discuss. It is right for people to feel nervous about it, but I think it is time to talk.
45. Militaries, police, and the main mechanisms of government law enforcement will be automated, and smarter than humans. You figure out how to interpret that.
46. Finally: AI labs may ultimately be nationalized in the strong sense. In my view, the US system is actually not very compatible with this, but there are many paths to nationalization, and in conservative or liberal political environments, it doesn't look blocked by taboo areas. In principle, labs seem able to maintain coordination with the military and intelligence departments at the backend without being more flamboyant than the posture already displayed. The federal government possessing this kind of unilateral power we speak of is also extremely dangerous. Private companies having this power is another matter, because they usually do not directly implement violence, and are not legally allowed to. I don't like nationalization much, but this world is confusing, and looks like it is becoming increasingly dangerous.
Join TechFlow official community to stay tuned
Telegram:https://t.me/TechFlowDaily
X (Twitter):https://x.com/TechFlowPost
X (Twitter) EN:https://x.com/BlockFlow_News
@bayeslord
