Bleeding-Heart Incrementalists

A short preface; I don’t hold any considerable amount of authority at all on the various topics that are discussed within my writeup; I have simply spent an amount of time in circles of discourse, and I have gotten to compare various different positions against my own point of view. There are many other way more qualified, smarter people than me, which my writeup draws from.

When I first started thinking about this write-up, I intended to ground it in what I would have called a techno-optimist viewpoint, but on reflection I wouldn’t describe my own views as any more optimistic than it is cautious or realist. I am not strongly versed in either of the spaces this piece draws on, the technology-leaning one (Silicon Valley futurists, effective altruists, accelerationists) and the left-leaning one (a leftist interpretation of AI safety along the lines of Cassie Pritchard, as well as the people who could be called “luddites”), but I have had no shortage of time to interact with people from both.

A valuable point of reference I have chosen is Dario Amodei’s October 2024 essay “Machines of Loving Grace.”¹ Amodei is the CEO of Anthropic, and the essay describes a world in which AI systems matching the best human experts arrive within a few years and compress decades of progress in medicine, economic development, and governance into a much shorter span. It’s a pretty clear statement of liberal techno-futurism, in that it takes markets and existing institutions as the vehicle for that progress. I want to root the AI element of this write-up in the technological element of Amodei’s essay (although with more room to discuss the vehicle that gets us there beyond the technological element), although I wouldn’t go as far as to call that the principal assertion of the piece. Regardless, I will assert this: technological advancement has prevented a great deal of undue suffering and lifted enormous numbers of people out of poverty, and future advancement could continue that trajectory without being outweighed by the negatives. At the same time, AI is different from earlier technologies because it automates cognition itself rather than any particular task, and that calls for a perspective that is informed by history but also adaptive and holistic, weighing near-term systemic harms alongside existential risk.

The discourse about all of this struggles when its participants disagree on whether AI is a capable technology at all; in the end though, people across the political spectrum are concerned about what it will do to our wellbeing, our economic security, and the world we live in, and this includes people who don’t think AI is a technology capable of significant benefit at all; besides this, even if the trend among the left seems to be a staunch humanist opposition to AI, a significant subset of the left shares Amodei’s premise about capability, while drawing very different conclusions from it.

Socialist economists who argue for central planning have long held that the obstacle was computational, since a planning board could not gather and process the information that prices carry as natural indicators of scarcity. Oskar Lange made this case in the 1930s in reply to Friedrich Hayek and Ludwig von Mises,² and Stafford Beer tried a version of it in Chile under Salvador Allende from 1971 to 1973 with the Cybersyn network.³ Language models and large-scale data collection revive the argument, because a system that can read, forecast, and negotiate at unprecedented, mindbreaking scale naturally prompts whether we have bridged the gap of postulations on the feasibility of central planning with technological assistance. Anarchists and libertarians could potentially take the same premise in the opposite direction: if technology lowers the cost of coordination far enough, then central planning need not apply, as radical abundance and technological capability could solve the question of what markets and planning exist for; the issue of scarcity.

Additionally, as I earlier mentioned Machines of Loving Grace as a general representation of the radical abundance vision that is often seen throughout techno-optimism, we should contrast this against visions which go beyond Amodei’s simpler vision of just massive economic growth; by asking about the way that wealth is controlled, questioning also Amodei’s vision that AI may not change governance in the same way it might advance economic prosperity; while also, going beyond interpretations that rely on near-term radical abundance by also acknowledging more reserved visions of “AI as a normal technology;” we can assert that even if AI is “only as big as previous societally-altering technologies, instead of being the biggest human invention up to this point,” that in this more cautious scenario, AI still has massive implications for economics and governance - while keeping open the possibility of those even more radical developments later.

With that in mind, I want to go to those who are economically liberal but sympathetic to leftist movements. They might accept that capable AI changes the economics of coordination, and they end up in-between the leftists and the Silicon Valley liberals, the staunch humanists and the existential risk advocates and the futurists; if it wasn’t clear, this would be the position I think of myself as falling into. This write-up, first, will meditate on that position, and then it will expand on the political and economic goals of various positions and how they intersect.

An interesting descriptor I have seen used by a few people is to describe themselves as a “fellow traveler” - this isn’t an exact breakdown of the positions of this sort of person (it would be impossible to do so because “fellow traveler” is pretty vague) but generally it makes sense to attach it to an open subset of left-leaning liberals, some range of social democrats (both modern and traditional) as well as other economic liberals, who to some degree, significantly factor in a post-capitalist horizon within a reasonable amount of time (say, within or near the end of their lifetimes, if we interpret lifetime as being a conventional lifespan, so approaching the latter half/end of the 21st century or before then). How could this post-capitalist horizon come about?

Well, this sort of person could reason that technological advancements could make more feasible the idea of central economic planning, combined with radically available abundance that gives us more than enough resources and material to solve many of the issues that face us today in regards to scarcity. If these people imagine the possibility, if not the probability of an outcome of this sort, then to what extent do they diverge from say, stage-based evolutionary socialism, or the social democrats of old? Firstly, the divergence in destination does depend to some degree on the mechanisms of economic and political control and to what extent there is planning versus the persistence of a free market; however I think it is reasonable to assert that, in every scenario of radical abundance, capitalism as we know it today will cease to exist and will be reformulated to some relatively small or massive extent. The more prominent difference I believe, is in how we get there; the means and the necessity of empowering labor. But maybe, even as the destination isn’t all that different, the way to get there may be compatible between both interpretations.

In a near-term scenario of significantly impactful abundance as a result of AI, a consensus exists on the basic possibilities of an economic policy framework to leverage this abundance for general benefit while minimizing negatives. Past visions of universal basic income have given way to a focus on a more cautious near-term approach which acknowledges the economic challenges of maintaining economic stability as AI causes massive job displacement. Along the lines of Anthropic’s proposed economic policy framework,⁴ a tiered, data-driven framework could be established that focuses on market preservation through wage insurance, while higher levels of unemployment could trigger larger basic-needs relief, with the possibility of an eventual transition to post-labor mechanics (a la universal basic income) even in the near-term, if displacement and thus, the ensuing economic growth, allows for it. This could be funded by direct taxation on AI companies and compute infrastructure, or through a sharing of equity; Sam Altman, CEO of OpenAI, has suggested “universal basic compute” which allocates ownership of computational power to citizens,⁵ which could be part of a framework like this.

I would not want to overstate how much agreement there is; what I think probably exists is an overlap in instruments, since wage insurance, taxes on compute, equity sharing, and tiered relief can each be defended by Silicon Valley liberals and by much of the left, while the reasons for supporting them differ.

If some framework along these lines manifests, it will be very complex, much more so than the simple notion of universal basic income, especially in the near-term, and the complexities are not something for me to narrow down.

This raises the question of why I discuss a framework of this kind when there is so little precedent for it. Sam Altman, the chief executive of OpenAI, has suggested “universal basic compute,” which would allocate ownership of computational power to citizens, but a proposal from someone who stands to lose a great deal by it is a weak foundation, and elites have rarely given up economic or political power voluntarily, whether in government or in Silicon Valley.

Anthropic is a somewhat different case. It is organized as a public benefit corporation, it has set up a public benefit fund, and its founders and employees have pledged part of their equity. In the rationalist and effective altruist circles where I have taken some time to discuss with some people (but by no means have I been able to obtain a large picture on what various different people believe), it is often treated as the lesser of two evils: a company with a better safety culture than OpenAI’s, but one that is still racing toward superintelligence, which its critics consider dangerous. I think one can be cautiously optimistic about parts of its conduct and still decline to stake the future on its promises.

The more useful history is of concessions made under pressure. Sweden’s Meidner plan, proposed in 1975, would have transferred a share of corporate equity to collective wage-earner funds, and it was enacted in diluted form in 1983 and abolished in 1991, which shows both that such an idea can reach law and that it can be reversed once its coalition weakens.⁶ The Alaska Permanent Fund, which distributes resource revenue to residents, is a more durable example.⁷ Neither guarantees the outcome I am describing; however, both suggest that the deciding factor is organized pressure.

Political pressure, then, is not an option, as much as it is a necessity to some degree. We will imagine its necessity both from the explosive growth paradigm, and the incremental technology (“AI as a normal technology”) paradigm.

Firstly, the more radical paradigm. If AI significantly automates cognitive labor while also compressing technological advancement that would’ve otherwise taken significantly longer, then market distribution will rapidly break down. The economic policy framework that was just detailed out is an example of how the government might respond to this, but importantly, it requires significant political pressure to push slow institutional consensus.

It is imperative that people from all backgrounds, if they find themselves in a scenario such as this one, be ready to rapidly work together to extract necessary public value from this abundant economic growth, through some mixture of compute taxes, direct equity shares, dynamic safety nets, or possibly even more radical changes, alongside other necessary policies. Without organized political pressure from civil society, it is a real possibility we will snowball towards rentier capitalism, and we cannot trust Dario Amodei, Sam Altman, and the rest of the heads of American technological excellence to prevent that from happening.

As I mentioned earlier, the reasons for supporting various near-term policies differ by purpose, but regardless, we can posit that the short-term political goals of people across the political spectrum (I’m focusing on economic liberals to leftists here) could be significantly intertwined in a scenario like this, with significant divergence on understanding of what these policies lay the foundation for; whether it is to ensure first and foremost economic stability or whether it is to begin a larger transition of economic power into the hands of society as a whole.

It might be a somewhat lofty notion to the more economically liberal (and very much so to those who are more libertarian-leaning) to suggest the right course of action in this scenario would be to direct the government to further intervene. Markets have coordinated dispersed knowledge better than any agency has managed to, and governments have a long record of intervening badly. Even so, I think that leaving the outcome entirely to the market runs into three difficulties when the technology in question automates cognitive work itself and radically changes what labor is worth; this is a profound technology like no other throughout history, as we have established.

The first is that markets distribute most purchasing power through wages. Earlier technologies displaced workers from one sector into another, and the new sector paid wages. If AI substitutes for labor across many occupations at the same time, the returns go to whoever owns the compute and the models, and households that depended on wages lose their means of bidding in the market. New kinds of work may appear, but adjustment takes years. The economists David Autor, David Dorn, and Gordon Hanson found that American regions exposed to Chinese import competition after 2000 still had depressed employment a decade later, and that shock was confined to manufacturing.⁸

The second is that frontier AI requires very large amounts of capital for chips and data centers, so only a few firms can compete. Where entry is that expensive, competition cannot be relied on to spread the gains, and the returns concentrate with whoever holds the capacity early.

The third is that the largest risks from this technology, from systemic harms to existential danger, fall on people who are not party to the firms’ decisions. Economic liberals have long accepted a role for government where costs land on third parties, and I think these risks belong in that category.

Finally, the case for a public response sits inside the liberal tradition, arguably. Milton Friedman proposed a negative income tax in Capitalism and Freedom (1962),⁹ and Friedrich Hayek endorsed a guaranteed minimum income in Law, Legislation and Liberty.¹⁰ An income floor, then, is a conclusion that some of the most committed defenders of markets reached on their own terms, and the policies discussed here extend that line of thought to a technology neither of them anticipated.

There is of course, the notion that, if this sort of massive short-term economic explosion occurs, then AI would have already produced so much abundance that it will continue to grow and shape society for the better in a way that benefits everybody in a manner that we find amenable. We can maintain this as a possibility while also positing that, in an exponential curve made up of individual s-curves, we don’t know where we might need to intervene to make sure things turn out right, and thus we should naturally tend towards some form of careful intervention.

The second paradigm I want to take seriously is the one Arvind Narayanan and Sayash Kapoor, two computer scientists at Princeton University, set out in their 2025 essay “AI as Normal Technology.”¹¹ Their argument is that the bottleneck on AI’s effect on the world is adoption rather than capability. A system that can do a task in a laboratory still has to be fitted into hospitals, law firms, and factories, each of which has regulators, liability rules, and established ways of working, and the people in them have to learn what the system is good for.

The historical case they lean on is electricity. Electric motors were available in the 1880s, but factories were built around a single steam-driven shaft, and the productivity gains came decades later, in the 1920s, after owners rebuilt the factory floor around individual motors, as the economist Paul David argued in his well-known 1990 paper.¹² On this view AI would diffuse over decades and be shaped by institutions along the way, much as earlier general-purpose technologies were. If AI is not the instantly transformative technology singularitarians posit it to be, or even the transformative technology that Dario Amodei envisions as possible, it is still “as impactful as electricity” potentially, if not probably.

I find this scenario at least as plausible as the radical-abundance one. It also doesn’t make the question of this writeup go away. If AI is only as consequential as the technologies that came before it, then those technologies, electrification and the computer among them, still reorganized who held wealth and power, and the displacement of cognitive work would still be large. The difference is in the time scale. In the fast scenario, markets break down within a few years and the political response has to be improvised under pressure. In the slow scenario, institutions have time to respond, and the harder problem is keeping a coalition together long enough to act. I think of the Meidner plan here: its supporters had a proposal in 1975, and by the time a diluted version passed in 1983 the political conditions that produced it had changed. Slow diffusion also gives incumbents time to entrench, since the firms that own the compute and the models early on will have decades to turn that position into permanent advantage. It would probably be preferable to keep open the possibility that the faster scenario arrives later.

What I would want, then, is policy that stays defensible if the faster scenario arrives later, since nobody can say now which pace we are on. I will call the approach I have in mind radical incrementalism; incrementalism, probably to some degree in the socialist sense, but as I am not necessarily looking at this from purely a socialist perspective, by “incremental” I mean reforms that existing institutions can pass in steps, through legislation, taxation, and regulation, without a rupture. By “radical” I mean, following the French social theorist André Gorz, who called such measures “non-reformist reforms,”¹³ that a reform is judged by its effect on who holds ownership and power as well as by how much income it transfers.

Radical incrementalism has the potential to keep the long-term question open while the short-term measures are being argued over, because each interim policy that spreads ownership or builds public claims on AI-generated wealth changes who has a say when the larger decisions arrive.

Let’s expand on the earlier mentioned near-term policy proposals, which I have posited are amenable to a broad political coalition.

The first policy, as mentioned earlier, is wage insurance and tiered relief. Wage insurance pays a displaced worker part of the difference between an old wage and a lower new one for a limited period, and the economists Lori Kletzer and Robert Litan proposed it in 2001 for workers hit by trade.¹⁴ Congress adopted a version for older workers in the 2002 Trade Adjustment Assistance reforms.¹⁵ Anthropic’s framework ties the generosity of relief to measured unemployment, so that support steps up as displacement increases (hitting 5 and 10 percent and larger thresholds to trigger escalating measures). This, even as the apparently extreme measure it might seem to be, is defensible to an economic liberal because it keeps people attached to the labor market and leaves prices to do their work. A leftist can defend it because it strengthens a worker’s ability to refuse a bad offer.

The second policy is a tax on compute, or on the profits of the firms that own it. The reason is arithmetic: payroll taxes fund Social Security and Medicare in the United States, and if AI lowers labor’s share of income, those revenues fall at the moment demand for relief rises. Bill Gates proposed taxing robots in 2017 on similar logic.¹⁶ For a liberal this replaces an eroding tax base with a new one, levied on a concentrated input. For a leftist it captures part of the surplus that would otherwise accrue to owners. As chips and data centers can move to another location, it is probably very necessary to combine this with some sort of equity sharing, since that doesn’t really need a location at all. Besides that, a minimum rate could be coordinated, such as the OECD’s 15 percent global minimum corporate tax,¹⁷ just this time on compute or firm profits.

We will then also touch briefly on equity sharing as I mentioned, with equity sharing meaning a fund that holds shares in AI firms or compute and pays out to citizens in a dividend like mentioned earlier. The idea has two lineages; the liberal one, associated with the economist James Meade and later with John Rawls under the name “property-owning democracy,”¹⁸ wants capital spread across many households. The leftist one, represented by Sweden’s Meidner plan, wanted collective funds that would acquire a growing share of the largest firms and so gain some control over them, partly to stop wage restraint from turning into windfall profits for owners.

The fourth policy is a direct public stake in the capacity itself, through publicly owned compute or a public holding in frontier developers. A small version already exists in the United States, where the National AI Research Resource pilot, launched in 2024,¹⁹ gives researchers access to publicly funded compute. A liberal can support a modest form of this as a source of competition and of research outside the large firms. A leftist sees in it the beginning of public ownership of productive capacity, and this is the rung at which I expect agreement to thin out.

The policies, as I said, depend on one another. Wage insurance keeps households spending while the slower measures take effect, but taken alone it only softens a transition while ownership stays where it was. A compute tax raises revenue, but revenue transferred year by year gives the public no standing claim on the firms. A fund provides that claim, and it needs a source of assets, which could be a tax payable in shares, as Meidner’s design was payable in new shares issued from company profits, or a negotiated equity grant. Public compute needs both a fund to hold the stake and a tax base to finance it. If the first rung is missing, the others face a political backlash from displaced workers before they have taken effect, and if the later rungs are missing, the first two become permanent subsidies for a concentrated owning class.

And so, we have detailed out the potential near-term policy goals that a broad political coalition could rally around as necessary in a scenario where AI growth continues and adoption is fast. Now we will explore long-term horizons as mentioned earlier.

The planning version rests on the argument Oskar Lange made in the 1930s and that Cockshott and Cottrell, and later Phillips and Rozworski, have revived:²⁰ the obstacle to running an economy without markets was the labor of gathering and processing information, and sufficiently capable computation removes it; if AI continues to scale, there is a very significant argument that it could be the sufficiently capable computation that is missing from this equation.

In this version, models forecast demand, adjust accounting prices by trial and error in Lange’s manner, and direct capacity toward goals that a public body has set. Hayek’s reply was that the relevant knowledge is dispersed and often tacit, and that incentives matter as much as information.

There is also the radical freedom scenario, if one were to interpret the benefits of AI from an anarchist or libertarian or minarchist perspective. In Daniel Kokotajlo’s AI 2040,²¹ by 2045 each person lives in a community they chose or founded, supported by a Citizen’s Dividend, and every human holds an equal one-ten-billionth share of the resources beyond the solar system, issued as lottery tickets that can be kept or sold. AIs enforce a short list of universal rights, such as the prohibition of torture and slavery, along with property claims, and otherwise let each owner govern as they please, which the authors describe as a cosmos as diverse as humanity itself. Philosophically it resembles Robert Nozick’s “framework for utopias” in Anarchy, State, and Utopia,²² in which a minimal structure lets many communities form and leave, and it resembles Ronald Dworkin’s hypothetical auction,²³ in which everyone begins with equal resources and bids according to their own plans.

Importantly, the freedom described in scenario 2 as presented by the AI Futures Project, depends on an enforcer, since the probes that guarantee property and universal rights are also described as crushing anyone who builds a military to conquer other people’s star systems. In both stories, also, the outcome depends on a settlement of ownership that occurs early, while political leverage still exists.

In the planning version the public must already hold a stake in the capacity, and in the freedom version the equal endowments must already have been agreed. This is why I think the interim policies matter beyond their immediate effects: public equity and public compute are what a planning board would need, and individual dividends are what the freedom version builds on, so the same instruments serve either destination.

Regardless of which interpretation you lean towards, it is arguably pretty easy to assert that sufficiently transformative AI, which will probably happen eventually, and could possibly happen within our lifetimes, will change society, economics, and governance as we know it.

The span that somebody could potentially have in mind for the post-capitalist horizon, the second half of this century, is about fifty years, which is also the interval Mayor Abraham Beame wrote across in 1976 when he addressed a letter to the mayor of 2026.

So, in 1976 Mayor Abraham Beame wrote a letter to his future counterpart, in 2026. This letter would end up being read by Zohran Mamdani at the opening of a time-capsule at the Port Richmond Wastewater Resource Recovery Facility in Staten Island.²⁴

Firstly, Beame notes how around 50 years before 1976, Charles Lindbergh had just flown across the Atlantic Ocean. And, just 50 years later, airplanes were everywhere; both Wilbur Wright and the New York Times made ill-fated assertions that human flight was thousands or millions of years away. And what happened?

And so, 50 years ago, Mayor Beame took the time to reflect on how just 50 years prior, the world was a radically different place, various different technologies in its infancy; and then he reflected on how, naturally, the world 50 years after the time he would write his letter would also be radically different. A lot of the modern reception to his letter, in newspapers and on social media, interprets his apparent hope through a tragic lens; we did not fulfill his visions of peace and a resolution to world hunger and poverty and suffering. But it isn’t any more tragic than hope is usually.

And now, once again, we can think about how radically different our world is now from Beame’s time; even if in a way we did not “live up to the expectations” of Beame’s natural hope, it’s very likely that if he were alive today (instead of having died 25 years ago, in 2001) he would be astonished by the progress we have made, even if he had only passed a little more than 2 decades prior to now. And naturally, we reason that 50 years from now the world will be unrecognizable in a way. As flight was in its infancy in 1926, Beame saw how 50 years later it had transformed global travel and interconnection. And now, a new, even more potentially profound technology is in its infancy. Many people disagree on how fast this sort of unrecognizable change will come; many have controversially short timelines, which is a real point of acknowledgement. But, with that in mind, we can also simply establish that over a long enough period of time, that change will come, even if you are skeptical about its near-term potential. And so, in this time of great uncertainty and change, like Mayor Beame we will hope; and that hope translates to action, and thus change.