• Home
  • Technology
  • Gaming
  • Entertainment
  • World & Business
  • Science
  • Sports
  • AI
HomeTechnologyGamingEntertainmentWorld & BusinessScienceSportsAI
    • HomeTechnologyGamingEntertainmentWorld & BusinessScienceSportsAI
    • Home
    • Technology
    • Gaming
    • Entertainment
    • World & Business
    • Science
    • Sports
    • AI
    AI

    Herbie Bradley Reviews History of the Future Scenario

    Rereads parts 2 and 3 of LRudL's 2025-2027 scenario and shares messy takes on how it aged.

    HB
    1 Source, 27d ago, first seen 27d ago

    TLDR

    Herbie Bradley, a Cambridge researcher focused on synthetic data for LLMs, posted that roughly 18 months had passed since LRudL published the History of the Future scenario. He stated he would reread parts 2 and 3, review the material, and offer takes on how well it aged along with remaining weaknesses. The post flags one area as bearishness on Goo. The scenario link appears in the same post.

    Combined views

    3.9K

    1 Source, first seen 27d ago

    Combined views

    3.9K

    1 Source, first seen 27d ago

    37 likes
    37 likes
    3 comments
    36 saves
    1 reposts

    Sentiment

    Positive——Negative

    Summary

    Not enough discussion yet.

    No sentiment analysis available yet.

    3 comments
    36 saves
    1 reposts

    Sentiment

    Positive——Negative

    Summary

    Not enough discussion yet.

    No sentiment analysis available yet.

    Today's Rank

    —

    Not ranked yet

    Today's Rank

    —

    Not ranked yet

    1 Source

    @herbiebradleyIt's been around 18 months since @LRudL_ published his extremely detailed History of the Future scenario: https://www.nosetgauge.com/p/a-history-of-the-future-2025-2027 I thought I'd reread parts 2 & 3, review, and give some messy takes on how well it aged and where I still see some weaknesses: •⁠ ⁠⁠Bearishness on Google was a good call "lives off vast rents”, “lumbers aimlessly on”. •⁠ ⁠⁠Somewhat fast timelines for the first drop-in worker replacements (2027). Overall the model goes from scaffolds for white collar tasks, to "economy 3.0" models which solve long-horizon knowledge work capabilities, have greater purposefulness and reliability, then assumes that is sufficient for the drop in worker. This is not necessarily a mark against the scenario, because it's super hard to see these things from early 2025, but I currently think that if we were standing in 2030 we would be able to describe a series of things which were missing from models in 2027 that were necessary to move towards further and further automation (Economy 4.0, Economy 5.0, Economy 6.0 in your parlance). •⁠ ⁠⁠It doesn't seem very data bottlenecks pilled (again, not super clear from early 2025) in the white collar automation section. •⁠ ⁠⁠The bitter lesson of business section is fun, and I think successfully predicted greater returns to scale/centralization and an explosion in company formation. Amusingly, Z Combinator style things (eg Polsia) may not work super well, because if one person can spin up a Z Combinator company for an idea, then many others can too, and so the shape of what types of problem you can solve to make a viable business from 0 to 1 is probably just going to very rapidly migrate to specifically and always the things which are not fully automatable. Rudolf was right to be directionally pessimistic here. There's some chance the market as a whole might become more resistant to fully automated companies (eg a cultural norm that every company filters out autonomous company inbound via pangram). •⁠ ⁠⁠Overall it is generally not super econ pilled in terms of its job market predictions, in the way that I have become increasingly over the past year. The biggest miss is probably an Alex Imas-like relational labor in white collar conception—this also applies to the bit at the beginning of part 3. You go a little in this direction in the list of careers which are viable in part 3, but I think a much longer list, due to relational labor, could be a stable equilibrium well into the 2030s. An AGI pilled economist might look at these predictions of ~AGI drop in knowledge worker and be like: ah yes 3-5% growth a year and the labor force participation rate just monotonically decreases slowly, but due to diffusion/regulation/inertia/relational economy, most white collar are still employed to direct the agents. I can't tell how aligned the scenario is with this? •⁠ ⁠⁠> the internet overall is a more cheerful and upbeat place than it was in the late 2010s or early 2020s — optimistic! I remember the late 2010s being reasonably optimistic overall, but maybe social media is more fragmented now so it's dampened. •⁠ ⁠⁠I would expect early to mid 2030s for significant impacts on manufacturing (eg such the developing countries feel the pain). The scenario in places seems a bunch faster than this. •⁠ ⁠⁠By 2030 or so it doesn’t mention the at this point expected self-sovereign agent swarms out there, or something adjacent to that. •⁠ ⁠⁠Broadly whenever, up until at least say 2033 or so, when the scenario describes humans "not actually being in the loop” in white-collar work, I feel sceptical, and generally expect them to for real be in the loop but at some higher level of abstraction (eg, telling the agents that we want to build a product that does X and not Y, where the fact that they know X is important comes from conversation with other humans and would be very hard for AI to guess a priori in the average case). Then, they would also be in the loop via playtesting the software and validating that it matches their vision. •⁠ ⁠⁠> Where rapid iteration is possible but the performance ceiling is not as high, like having sales calls, the AIs are better than all humans. — I would probably bet against this specific thing by 2035, because I think it's heavily relational labor, so the verifiability or not is pretty moot. Though I have many separate doubts about superpersuasion. For similar reasons, I don't think the AI powered cults thing will happen. •⁠ ⁠⁠> In 2029, OpenAI rebrands its models to just “o”. — fun that they’re doing an almost equivalent naming scheme to this 3 years early, with Luna/Terra/Sol. •⁠ ⁠⁠Overall I am roughly on board with much of the capabilities progress up through say 2033 (though possibly -20% speed in some areas) — interestingly, no software-only intelligence explosion or strong RSI is really required for much of this, but would put slower timelines on the hypothesized effects on the world. •⁠ ⁠⁠Past the midpoint of part 3, you really really emphasize nanotech, and I think that tech tree is so long that it comes on the other side of the "Anthropic biotech revolution" you write about. My guess is late 2030s for first pieces of research and 2040s for any application. Replicators at greater than nano scale are promising, and may be quicker, but that also has a long tech tree. •⁠ ⁠⁠The car dealership model for robot ownership is fun! •⁠ ⁠⁠You cite actually working ICBM defense by 2034, thanks to Forethought detailed modelling, I now think this is late 2030s and may not be possible if China take certain steps in their ICBM development. •⁠ ⁠⁠The explosion in R&D should also create new types of work IMO, not all of which can be automatable. Furthermore, with more unemployed/UBI people, there should also be a more rapid rate of development of the leisure economy, which often works on purely relational grounds and can generate a significant number of new professions. For example, padel coach. •⁠ ⁠⁠It doesn't talk about some of the downstream effects of much higher real rates, which could be significant. For example, European countries may be in real fiscal trouble under this regime even without any AI job losses. •⁠ ⁠⁠In light of Marco Rubio, the "regime change in Venezuela at the press of the button" in 2036 was only 10 years too slow. •⁠ ⁠⁠It under-indexes on space in the mid 2030s, since by roughly then I think possible to get some sort of moon factory with robots and humans. By the late 2030s, a nascent industry on the Moon is plausible. In this situation, the US Government may want to assert greater control over space (if it's being done by a single private company), so pseudo-nationalization via the DPA or other mechanisms seems likely. •⁠ ⁠⁠On the manufacturing side, as with many industrial explosion analyses, it doesn't talk about the demand side, almost entirely the supply side. This is more plausible in the white-collar situation, but for industrial explosion it seems hard to see where the demand comes from for the described level of output in 2040. I think this is a common mistake (eg made by Dwarkesh in his podcast with Dylan recently). •⁠ ⁠⁠One way to get around this demand objection is to posit nationalization, eg of SpaceX, which can then do things like maintain strategic reserves because it's in the US interest, or build a giant anti global warming mirror at Lagrange 1 even though it's quite likely that people won't really pay for it. •⁠ ⁠⁠Part of your answer to the demand objection is geopolitical competition with China. But I think this will mostly happen this way over space expansion by the late 2030s. The West will become eventually, self-sufficiant in manufacuturing of goods (probably by ~2040 or so), but there is little point in building OOMs more robots than necessary to replace human labor currently in that supply chain. •⁠ ⁠⁠At some point on the algorithmic side, you need to posit some dramatic breakthrough to get human-brain like sample efficiency and continual learning, in order to get the effects you describe in the late 2030s (eg AIs run all the major companies), I think. Otherwise, if I extrapolate out the asymptote of LLMs, I'm suspicious that it doesn't lead to that. •⁠ ⁠⁠For reasons I've given above, I think the 2040s+ section is too pessimistic about humans and disempowerment.

    1 Source

    @herbiebradleyIt's been around 18 months since @LRudL_ published his extremely detailed History of the Future scenario: https://www.nosetgauge.com/p/a-history-of-the-future-2025-2027 I thought I'd reread parts 2 & 3, review, and give some messy takes on how well it aged and where I still see some weaknesses: •⁠ ⁠⁠Bearishness on Google was a good call "lives off vast rents”, “lumbers aimlessly on”. •⁠ ⁠⁠Somewhat fast timelines for the first drop-in worker replacements (2027). Overall the model goes from scaffolds for white collar tasks, to "economy 3.0" models which solve long-horizon knowledge work capabilities, have greater purposefulness and reliability, then assumes that is sufficient for the drop in worker. This is not necessarily a mark against the scenario, because it's super hard to see these things from early 2025, but I currently think that if we were standing in 2030 we would be able to describe a series of things which were missing from models in 2027 that were necessary to move towards further and further automation (Economy 4.0, Economy 5.0, Economy 6.0 in your parlance). •⁠ ⁠⁠It doesn't seem very data bottlenecks pilled (again, not super clear from early 2025) in the white collar automation section. •⁠ ⁠⁠The bitter lesson of business section is fun, and I think successfully predicted greater returns to scale/centralization and an explosion in company formation. Amusingly, Z Combinator style things (eg Polsia) may not work super well, because if one person can spin up a Z Combinator company for an idea, then many others can too, and so the shape of what types of problem you can solve to make a viable business from 0 to 1 is probably just going to very rapidly migrate to specifically and always the things which are not fully automatable. Rudolf was right to be directionally pessimistic here. There's some chance the market as a whole might become more resistant to fully automated companies (eg a cultural norm that every company filters out autonomous company inbound via pangram). •⁠ ⁠⁠Overall it is generally not super econ pilled in terms of its job market predictions, in the way that I have become increasingly over the past year. The biggest miss is probably an Alex Imas-like relational labor in white collar conception—this also applies to the bit at the beginning of part 3. You go a little in this direction in the list of careers which are viable in part 3, but I think a much longer list, due to relational labor, could be a stable equilibrium well into the 2030s. An AGI pilled economist might look at these predictions of ~AGI drop in knowledge worker and be like: ah yes 3-5% growth a year and the labor force participation rate just monotonically decreases slowly, but due to diffusion/regulation/inertia/relational economy, most white collar are still employed to direct the agents. I can't tell how aligned the scenario is with this? •⁠ ⁠⁠> the internet overall is a more cheerful and upbeat place than it was in the late 2010s or early 2020s — optimistic! I remember the late 2010s being reasonably optimistic overall, but maybe social media is more fragmented now so it's dampened. •⁠ ⁠⁠I would expect early to mid 2030s for significant impacts on manufacturing (eg such the developing countries feel the pain). The scenario in places seems a bunch faster than this. •⁠ ⁠⁠By 2030 or so it doesn’t mention the at this point expected self-sovereign agent swarms out there, or something adjacent to that. •⁠ ⁠⁠Broadly whenever, up until at least say 2033 or so, when the scenario describes humans "not actually being in the loop” in white-collar work, I feel sceptical, and generally expect them to for real be in the loop but at some higher level of abstraction (eg, telling the agents that we want to build a product that does X and not Y, where the fact that they know X is important comes from conversation with other humans and would be very hard for AI to guess a priori in the average case). Then, they would also be in the loop via playtesting the software and validating that it matches their vision. •⁠ ⁠⁠> Where rapid iteration is possible but the performance ceiling is not as high, like having sales calls, the AIs are better than all humans. — I would probably bet against this specific thing by 2035, because I think it's heavily relational labor, so the verifiability or not is pretty moot. Though I have many separate doubts about superpersuasion. For similar reasons, I don't think the AI powered cults thing will happen. •⁠ ⁠⁠> In 2029, OpenAI rebrands its models to just “o”. — fun that they’re doing an almost equivalent naming scheme to this 3 years early, with Luna/Terra/Sol. •⁠ ⁠⁠Overall I am roughly on board with much of the capabilities progress up through say 2033 (though possibly -20% speed in some areas) — interestingly, no software-only intelligence explosion or strong RSI is really required for much of this, but would put slower timelines on the hypothesized effects on the world. •⁠ ⁠⁠Past the midpoint of part 3, you really really emphasize nanotech, and I think that tech tree is so long that it comes on the other side of the "Anthropic biotech revolution" you write about. My guess is late 2030s for first pieces of research and 2040s for any application. Replicators at greater than nano scale are promising, and may be quicker, but that also has a long tech tree. •⁠ ⁠⁠The car dealership model for robot ownership is fun! •⁠ ⁠⁠You cite actually working ICBM defense by 2034, thanks to Forethought detailed modelling, I now think this is late 2030s and may not be possible if China take certain steps in their ICBM development. •⁠ ⁠⁠The explosion in R&D should also create new types of work IMO, not all of which can be automatable. Furthermore, with more unemployed/UBI people, there should also be a more rapid rate of development of the leisure economy, which often works on purely relational grounds and can generate a significant number of new professions. For example, padel coach. •⁠ ⁠⁠It doesn't talk about some of the downstream effects of much higher real rates, which could be significant. For example, European countries may be in real fiscal trouble under this regime even without any AI job losses. •⁠ ⁠⁠In light of Marco Rubio, the "regime change in Venezuela at the press of the button" in 2036 was only 10 years too slow. •⁠ ⁠⁠It under-indexes on space in the mid 2030s, since by roughly then I think possible to get some sort of moon factory with robots and humans. By the late 2030s, a nascent industry on the Moon is plausible. In this situation, the US Government may want to assert greater control over space (if it's being done by a single private company), so pseudo-nationalization via the DPA or other mechanisms seems likely. •⁠ ⁠⁠On the manufacturing side, as with many industrial explosion analyses, it doesn't talk about the demand side, almost entirely the supply side. This is more plausible in the white-collar situation, but for industrial explosion it seems hard to see where the demand comes from for the described level of output in 2040. I think this is a common mistake (eg made by Dwarkesh in his podcast with Dylan recently). •⁠ ⁠⁠One way to get around this demand objection is to posit nationalization, eg of SpaceX, which can then do things like maintain strategic reserves because it's in the US interest, or build a giant anti global warming mirror at Lagrange 1 even though it's quite likely that people won't really pay for it. •⁠ ⁠⁠Part of your answer to the demand objection is geopolitical competition with China. But I think this will mostly happen this way over space expansion by the late 2030s. The West will become eventually, self-sufficiant in manufacuturing of goods (probably by ~2040 or so), but there is little point in building OOMs more robots than necessary to replace human labor currently in that supply chain. •⁠ ⁠⁠At some point on the algorithmic side, you need to posit some dramatic breakthrough to get human-brain like sample efficiency and continual learning, in order to get the effects you describe in the late 2030s (eg AIs run all the major companies), I think. Otherwise, if I extrapolate out the asymptote of LLMs, I'm suspicious that it doesn't lead to that. •⁠ ⁠⁠For reasons I've given above, I think the 2040s+ section is too pessimistic about humans and disempowerment.