一位匿名研究者的46條暴論：AI將顛覆一切，包括核威懾與人類統治

Original Author: bayeslord

Original Compiled by: TechFlow

Introduction: bayeslord (@bayeslord) is an anonymous yet influential figure in the AI × crypto space, who avoids promotion and hype, focusing instead on pushing the technical boundaries of scaling laws and algorithmic depth.

This blogger recently published a list of 46 points extrapolating future developments in technology, AI, and related fields, arguing that everyone understands AI through the efficiency curves of the past, and the real leap hasn't arrived yet. There may still be four to ten orders of magnitude left in intelligence production.

He covers everything from algorithmic acceleration to robotics, capital, a permanent underclass, and finally lands at the sharpest point: Mutually Assured Destruction (MAD) may fail, military and police forces could be automated, and AI labs might be nationalized.

The original post has nearly one million views. Although the views are extreme, each point is relatively self-consistent and worth a read for tech enthusiasts.

This list is based on a tweet thread I posted on June 4, with some modifications and additions. Several people said the original was too hard to read, so I've compiled it into this version.

Intelligence

1. Algorithmic progress will catch everyone off guard. The entire world—markets, governments, militaries, companies, individuals—is using the productivity and rules of recent years to understand AI's impact and judge how things will roughly go. Even new labs claiming to believe in "recursive self-improvement" think it's just the old routine with an agent running in a loop. It's not. I suspect there are still many orders of magnitude left in intelligence production, perhaps as many as ten, with four to seven being more likely. In principle, exceeding ten is not impossible, but it would heavily collide with what I suspect is the true upper limit allowed by the laws of physics. It's unlikely, but not ruled out. If this judgment holds, then the true direction of things is not the same as it appears on the surface; a major leap is approaching. Anything happening along this path will make the world much stranger than almost anyone has priced in.

2. We are in the early stages of takeoff. AI improving AI may ultimately become the most consequential step in history. This can't 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 another 4 to 10 orders of magnitude of intelligence output per unit of computation seems possible).

3. Since we've entered the takeoff phase, algorithmic research is accelerating. Compute remains a scarce resource, but the opportunity cost of a researcher's time has dropped because you can just deploy an agent to run any task, even random ones. It might bring something back. All new ideas carry an "optimization debt," which can now be repaid with unsupervised token consumption. Vast numbers of research scaling law curves will be traversed 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. Mathematics and code are being conquered by large-scale reinforcement learning, with everything else following behind. The distinction between "verifiable" and "unverifiable" will gradually disappear as a meaningful boundary. Going forward, automated AI research and AI learning will increasingly look like the same thing. Training a model well is essentially closely related to the model learning well by itself. Constraints like sample efficiency, creativity, and others will all be resolved, eventually approaching algorithmic optimality at any scale.

5. The idea that long-horizon agents require training data of equal length is wrong because generalization exists across time. Long-horizon tasks are not built from the "length" attribute. This relates to LeCun's fallacy of (1-e)^n error accumulation. What actually happens is error correction. Error correction occurs simultaneously at multiple scales, from the level of individual token generation to each step within a long task. The reason METR's chart goes up is partly because agents begin to reach the escape velocity of error correction.

6. An engineering-level science of deep learning is about to emerge. It will push us towards AI's algorithmic maturity much faster than most people expect—although, as mentioned earlier, it's unclear how far this path can go in principle. For example, a science studying scale invariance would significantly increase the scale and returns of useful experiments, because an experiment on a single GPU can tell you how to do it with a hundred thousand.

7. Every field of human technical activity will have its own "Move 37" moment (the move AlphaGo made against Lee Sedol that transcended human intuition), and then quickly, "Move 37" itself will seem outdated. I mean all fields.

8. Compute will continue to improve. The best matrix multiplication machines today are far from the physical limits of AI accelerators. There is still significant room for improvement in digital silicon. There are also many candidates for new substrates, whose algorithmic debt will be squeezed to the limit by automation, but we don't yet know which is optimal for AI in terms of space, energy, time, manufacturability, and cost. Photonics and stochastic 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 to automation and scale, including returns from greater algorithmic depth. If the practice (and theory) of deep learning remains shallow forever, the long-term moat probably won't be primarily algorithmic, because secrets are relatively easy to discover. Ultimately, distillation plus data plus time can catch up with compute scale, albeit perhaps slower. Currently, we seem partially in this state, but even if true, no one guarantees it will stay this way.

10. If deep learning becomes less shallow as scale increases, then every incremental unit of automation and scale buys you algorithmic secrets that others find increasingly harder to reach. We currently seem partially in this state too. The endpoint for both scenarios is diminishing marginal utility returns as scale and research saturate. 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 scarcity of the 2020s. Scale is expanding and working, capital follows, spinning the flywheel over and over. More matrix multiplication machines, more fabs, more energy are on the way. The bottleneck in intelligence production is temporary. Potential economic speed bumps don't count.

12. The nature of the intelligence supply chain is changing. It's currently highly concentrated in labs. But labs are automating the very thing that makes them strong—researchers, and the discovery of algorithmic advantages. Once this process begins, assuming open-source follows not too far behind, especially if labs don't lock down their AI researcher models, the labs' advantage will shift towards easier financing, more compute, exclusive data, business relationships, and good products. This indeed depends on how the previous question of algorithmic depth resolves, among other factors.

13. Distributed training will reduce the need for massive single data center builds, giving some advantages to non-hyper-scale players. However, in the pure dimension of single largest training runs, it won't surpass the hyper-scalers.

14. Automated AI experimentation will lead to the widespread discovery of algorithmic secrets because these secrets are inherently easier to distribute than full-scale training. How far this path can go is unclear, but I expect it to be 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. Despite these forces seemingly favoring academia and open-source, they could still atrophy due to the cost and opportunity cost of compute. For example, is a 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 figure out where demand is greatest, and right now, that place seems to be the labs. This means open-source labs might become even more compute-hungry, even if they have money, unless they've pre-locked compute capacity. Even if locked, they still have to consider the opportunity cost of doing their own research versus renting out compute. Think about the Colossus and Anthropic deal.

16. In an environment where AI capabilities start getting exciting (0 to 18 months from now), open-source might also start to become socially difficult, especially if we are very slow in accelerating safety—which has been the case so far.

17. As capital floods into labs, open-source might begin to shrink. There's a coordination problem here: besides labs (and perhaps governments), no one wants a token monopolist, but if this problem can be solved and the regulatory environment is friendly, perhaps the outcome could be okay.

Robotics

18. Robotics will have its own "ChatGPT November 2022" moment, and then another "Opus 4.5 November 2025" moment. Neither has happened yet, but they are coming, and faster than people think, as a result of rapid AI progress, including AI-accelerated physical systems engineering. The interval between these two robotic moments doesn't seem likely to be three years.

19. However, truly scaling up the physical number of robots worldwide might have to wait until 2030 or later. Although we build about 100 million cars per year, and humanoid robots are much smaller than cars. Considering we also build 1 billion smartphones per year, if capital and algorithms run fast enough, achieving 100 million robots per year by 2030 is reasonable. 10 million per year is definitely achievable; we are already doing it in the drone market. As long as software can prove humanoid robots are worth the money at a small scale, it can unlock infinite capital, proportional to the quality of the proof.

20. What look like hard limits for robotics today will disappear, including poor sample efficiency, relative data scarcity, expensive and difficult hardware design for hands and motors, the fractal complexity of the physical world, and the undocumented tacit knowledge about how we work in the world (like plumbing). World models seem useful, but which specific thing it is doesn't matter. Research scaling laws will be exhausted until they hit diminishing returns.

21. Global demand for robots is easily in the tens of billions, especially if we count all form factors. There's so much physical labor worth automating. The market will figure this out, and people probably won't get in the way.

Progress

22. Science is being automated and virtualized. This means much of the progress the 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 tear down many bottlenecks, pushing the boundaries of "verified virtualization" along the way to improve sample efficiency and the net returns of "making it real." Essentially, in every field, we will have some combination of neural models, explicit simulations, and real-world experiments that together increase the returns 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 laws. It's hard to tell when any given S-curve will saturate, or if there's another S-curve on the horizon. What's important to understand here is that the engine of civilization's progress itself also has a law of progress. Our progress is likely saturating, like most natural processes, but we don't actually know where the saturation point is. The maturity of technology and civilization could be near or far. We are at a historical juncture: first, we have barely invested any resources into progress, but this is changing rapidly; second, we are automating the very machine that directly produces more progress. We live in interesting times.

24. The future of scaling up or scaling out. From zero to one or from one to n. How much progress the universe allows us in breadth and depth is an open question. Breadth is easier to estimate, as it roughly equates to "how many more steps of calculation do the laws of physics allow us to take?" And how "deep" that calculation can be—in the broadest sense—is unknown. In some versions of the future, the technology tree is unfathomably deep, and the reachable computational universe is so rich that we will keep inventing and discovering until physics stops us, if it can. Other versions are flatter: we quickly exhaust a shallower technology tree, reach technological maturity relatively easily, and then scale it out until we are satisfied or physics gets in the way.

Capital and Production

25. More capital plus more intelligence means a more intensified capitalism, meaning we rush towards market equilibrium faster. In the long run, this should naturally lead to deflation, driving most important goods towards near-zero marginal cost through competition, including AI, food, housing, medicine, electronics, entertainment, and travel. Provided we don't get in the way. In some cases, we 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 with jobs, for a very, very long time. What percentage of humanity this represents is an open question. Those who say it will be high are overconfident, as are those who say it will be zero. However, it's hard to imagine how much longer humans can contribute marginally, especially in the "knowledge" part of knowledge work. Demand for some occupations, like doctors, might drop significantly—if we have a $20/month superhuman AI doctor, plus on-demand testing, plus significant health improvements from medical tech advances. But because we have cartelized doctors, we might continue to do so, and being a doctor will remain a good profession. Demand for entertainment will probably rise, but production costs will fall, and the technical demand for humans in entertainment has already decreased greatly. But we care a lot about other humans, so maybe we will continue to care, and being an actor will become more profitable. One way to think about how this evolves is by counting how many intermediary layers exist in the supply chain from a worker today to a consumer. For a TikTok influencer: zero layers. For a doctor: zero layers. For a factory worker: many layers. Whether a job can be disintermediated, competed away, or substituted will likely determine its fate to a large extent. This analysis is quite nuanced, and this paragraph is far from sufficient, but finally, a caveat: all this presupposes we don't face a cliff-edge drop in demand—which could happen if too many people don't work, and productivity or government efficiency cannot sustain a universal basic income or universal basic healthcare.

28. Related but not contradictory to the above: a "permanent underclass" might truly exist. In the better worlds where it materializes, it looks more like highly constrained agency rather than severely harmed income. This is ultimately acceptable for most people; our agency is already highly constrained by modern society. But it requires psychological adaptation, which may take time and could be painful.

Culture and Psychology

29. The human mind grows and adapts slowly now, but this will change. The key is that it changes for the better, which won't be easy for some. Abundant intelligence and automation will allow us to engineer far more durable psychological structures than we have today—our current ones are evolutionary leftovers maladapted to our environment. Psychiatry and psychology will complete a millennium of innovation within no more than a few decades. Humans will fundamentally get better. Coarse, degenerate "pleasure-direct-connection" is overestimated as a risk because we will have more sophisticated and diverse mental engineering available.

30. In an extremely uncertain world, people will compete for power, status, and wealth more ferociously than ever, betraying their own kind with a clear conscience. They will invent reasons why their actions are good, even great. Look around.

31. You will live to see embarrassments you cannot believe.

32. There is a clear double discourse happening now. The people who are about to be