I concede the second point. So far, the primary effect of AI on humans is changing how we write.2 But to what extent does that influence how we think, our morality, and our culture? The issue is AI models are like vapes; we don’t have long-term studies to understand their impacts. However, it is clear they do affect us, and societies must spend time deciding if they like the influence and how to respond. It may be that, like in the past, cultures evolve, and we will passively accept what AI does to our brains. But if culture wants to mount an opposition, then our current culture and institutions must begin to get strict about AI being applied at home, in the classroom, and at the workplace.

However, I disagree with the author’s first point and think he gives alignment efforts a bad rap. The article’s author, Wang Huanchao, writes:

“When OpenAI and Anthropic say, “We want AI to be aligned with human values,” all three referents—“we,” “AI,” and “humanity”—are remarkably questionable and vague.

The reality is that a handful of privately owned companies on the American West Coast have signed a social contract with AI systems on behalf of the entire human species.”

But any alignment effort must inevitably make assumptions about alignment, “humanity,” and “values.” Considering alignment efforts began in labs years ago, when the amount of people taking alignment seriously was few and far between, we cannot really fault the labs for taking the first stab.

If we desire a world where the labs or the governments that regulate them lead alignment, then the world must radically change its stance on AI in a short period of time. Any lengthy, deliberate process through modern institutions like the United Nations would likely require an agreement to pause or decelerate AI progress while the world hashes out alignment. Otherwise, by the time we get around to it, unaligned models would have already been proliferating throughout the world, defeating the purpose of deliberative alignment.

Right now, I think such a pause or deceleration is highly unlikely, and the best the world has is trusting the labs to lead on alignment. The labs are exposed to the greatest levels of PR and financial risk if an unaligned model goes rogue, and they are the ones most familiar with the inner workings of the models themselves. In a practical sense, too, they are the ones best in a position to constantly update their alignment frameworks and rapidly iterate through improvements.

Still, Wang is rightfully concerned about such high trust in a few, essentially unaccountable actors, but his criticism that the process is opaque is unfair. Wang states:

“... a group of ghostwriters is drafting values for AI on behalf of all humanity and every civilization, while the “humanity” being represented knows almost nothing about what has been signed, or with whom.”

This is not fully true. Although OpenAI’s Model Spec does not list its contributors, Claude’s Constitution is clear in who authored the document and whose input was taken in writing it.

Based on the list of contributors and lest I sound like an Anthropic shill, I do think Wang’s criticism has a core of truth: the labs are not sufficiently consulting society at large in their alignment efforts. The proportion of contributors to Claude’s Constitution, for example, skews heavily in favor of in-house employees. That ratio must be inverted.

Of the external contributors, the lion’s share come from the AI policy blob, a demographic that has historically leaned towards effective altruist perspectives.3 I don’t think this skew from the outset is an error, given the EAs were the founders of AI safety, but for models to truly represent “human values,” the consulted sphere must expand. Model developers must consult experts in ethics and faith traditions that can better represent human perceptions of value and morality, rather than heavily relying on technocrats.

Of the external contributors, I can only count three that ostensibly represent a faith tradition, and all of them are Catholic.4 This is good but insufficient. The labs must incorporate more experts from different religions, as well as ethicists and philosophers that are not already a part of the AI social bubble. Perhaps faith leaders and philosophers did not have enough interest in being consulted a few years ago, but I bet they do now.

If labs were to begin such an effort, the inevitable criticism is then, “Who are the labs to decide who gets consulted and who doesn’t?” Will Chinese labs consult the Dalai Lama, or be able to consider anything that is not socialist value alignment? Will American labs consult the Grand Ayatollah, or consider anything that is not an orthodox liberal ideology? And how will labs adjudicate between contributors of different opinions? I don’t have a great idea as to a “fair” way to sort out these kinks, but I don’t think there needs to be nor can there be. Considering the speed of development, we must choose the best possible option, and this seems to be it.

The labs pursuing alignment in this consultative way can take two forms: either a one lab, one value approach, or a one lab, multiple values approach. In the former, a lab would consult the ethicists and philosophers, synthesize what it thinks are the best ideologies, and release models aligned with that specific value system. In the latter, a lab may realize that trying to adjudicate among such competing ideologies is a fool’s errand, and instead offer different models aligned with different value systems (ex. A Buddhist model, a Muslim model, a postliberal model, etc.). The former seems to be the route the current labs are going down; it is much easier to ensure one model is aligned properly than it is to ensure the safety of multiple different “personalities.” I think the former is also predicated on an idealism, downstream of the one-world optimism of the Internet and pluralistic liberalism of the United Nations, that the world can agree on basic values and rights. A world where labs continue down the former path may still allow users and governments to choose among different value systems, with options constrained to the number of labs in the world. But even here, how can every person and country find 1:1 value alignment in a choice among a few models?

The latter, however, may emerge as a popular option once models begin to emerge more as serious pseudo-moral agents. Right now, alignment towards a specific ideology seems a bit silly, given the tight leash we keep agents on and their relatively limited range of applications. But if nations begin to demand models be adapted to their values or otherwise be barred from market access, developers would be incentivized to fine-tune models differently for each nation-supported value system.5

To reiterate, the labs have the best incentives to strive for a mass-market model that can appeal to the widest possible value system.6 Any governmental approach risks immediate politicization, propagandization, and the manipulation of alignment for domestic political benefit. The role of the public is to apply pressure to labs and to respective governments to motivate developers into greater levels of consultation. Shame works, particularly with the threat of regulation, bans, and boycotts.

The current mode of alignment is imperfect, but the labs remain the most plausible route to “aligned” models. Wang’s criticisms, while relating a reality that most of the world is on the outside-looking-in, are a cop out. We cannot dismiss model specs and constitutions as “ghostwriting” without offering a better, realistic alternative. Alignment is a real concept; we must simply pressure the labs to do it well.

Overall, I have been blessed to read mainly good books (though maybe I just find most books good), but I definitely I have a type: older, more religiously philosophical texts. I would like to break out of this mold! Another reason for posting is to solicit recommendations on what I should read next.

So, here is what I have been reading, ranked from best to worst:

1. Jayber Crow by Wendell Berry (2000)

   Jayber Crow reminds me of the importance of literature. I could easily say “small towns are important,” but there is a gap of truth between that paltry sentence and its meaning. Berry’s 300 pages helps bridge that gap, even if just a little. I lived through Jayber Crow and saw the beauty of Port William Township and felt sorrow at its decline. I urge you to read it, especially now in an age of rapid technological progress, to contemplate on the intangibles of what may be lost. — 9.5/10

2. The Sparrow by Mary Doria Russell (1996)

   Russell has put the Book of Job to shame. An amazing sci-fi theodicy that places a saintly Jesuit priest at the beginning and the same man, a broken and angry prostitute and murderer, at the end. The journey is gripping and makes you doubt that God is good, but Russell beautifully shows how He still is. The writing is great too; I had a nightmare about the story after finishing it, and on some of my sunny days I remember Emilio Sandoz and become a bit muted. The only criticism is that Russell is not Marquez; I can smell and taste Macondo, but Rakhat is a bit of a haze. A must read. — 9.5/10

3. The Aleph and Other Stories by Jorge Luis Borges (1949)

   Utterly fantastic. Some of the best stories that play with magical realism and time in a manner I’ve only seen elsewhere in Gabriel Garcia Marquez. My favorites were “The Immortal,” wherein immortals trend toward becoming troglodytes who purposely forget speech and instead lie down all day and think. I also thought “The Other Death” was the most fantastic play of time that I’ve seen in a story. — 9.5/10

4. The Imitation of Christ by Thomas à Kempis (c. 1420)

   Best devotional book I have read to date. Kempis calls me out incessantly and motivates me to live a life of death-in-life. The work wrestles with the paradox of faith being both difficult and easy. It is difficult if you give into the cares of the world, but that is not how you should live. You are a sojourner in this world, and The Imitation invites you to remember this and to adopt the proper attitude toward the world—one of indifference. Then, when you realize you are dependent only on the Creator, and only things relating to Him concern you, then everything becomes easy. The challenge is getting to that view of the world, and the point is that you will never actually get there. (It will get docked half a point, though, because I cannot relate to monastic life, and Kempis is a bit too into self degradation.) — 9/10

5. Orthodoxy by G.K. Chesterton (1908)

   I wish I could write like Chesterton. His wit and paradoxes are the closest approaches to the esoteric and to what cannot be expressed in words. I like to think of it as a modern Western Dao De Jing. A pleasant read that casually disarms the excessive seriousness of rationalism and atheism of today’s world. — 9/10

6. The Man Who Was Thursday by G.K. Chesterton (1908)

   A lovely, brief novel that revives romanticism and hope in a time where that is not so fashionable. It is easy to read and short, though even then it becomes a bit repetitive in its episodes. The twist at the end is delightful, and for me, who was learning about Mikhail Bakunin at the same time as reading the book, I found it a charming literary critique of anarchism. — 7.5/10

7. Ficciones by Jorge Luis Borges (1944)

   Maybe it was because I read The Aleph first, but I was disappointed in Ficciones. The stories here are more concerned with philosophical idealism, whereas the Aleph seems more concerned with time; I find the latter more interesting. For some reason, the stories in Ficciones seems so cramped to me. They all seem told be the same narrator and tell the same story; there’s a sort of acidic intellectualism that made me tired. I don’t want to hear about dialectics and Schopenhauer anymore! That said, it’s still Borges. His stories still play with my brain like Play-Doh in a way that no other author does. My favorite stories were “The Lottery in Babylon," “The Library of Babel," and “The Secret Miracle.” — 7.5/10

8. Reflections on the Revolution in France by Edmund Burke (1790)

   It helped to read this after a 50 hour podcast on the French Revolution. I found this to be a great show of rhetoric and argument, but it wasn’t really a logical flow in the way that most philosophy is. But then again, that’s Burke’s point. Screw your logic, empirics is what matters. And in the case of the French Revolution, Burke was particularly prescient; many of his quotes and thoughts would be considered prescient today. I am biased to accept his claims of slow evolution, of the importance of custom, and the dangers of blind idealism. I wonder how Burke would grapple with the role of technology in today’s world. — 7/10

9. Regime Change: Toward a Postliberal Future by Patrick Deneen (2023)

   The best postliberalism has for now. Deneen is sparse on details, even for his detailed plan of regime change, but he does have a good taxonomy of Marxists, liberals (market and progressive), and postliberals. I think he is sparse on details because he does not realize his philosophy would be best served being a social philosophy instead of a political one. Elites and the people should be in closer and more frequent contact, and they should both be aware of the virtue they reap from one another; then, they can form a better, harmonious society, even with differences. — 6.5/10

10. The Road by Cormac McCarthy (2006)

    Not nearly as good as Blood Meridian, but at least a short and easy read. — 5/10

Next on my bookshelf are The Moon is a Harsh Mistress by Heinlein, The Razor’s Edge by Maugham, Buddenbrooks by Mann, and poems by Gerard Manley Hopkins. Please let me know what else I should read!

Everyone now knows the populist movement is national and not temporary. But what concerns me the most is the unproductive attitude toward populists by AI experts both in the Bay Area and D.C. Nearly all responses to the populist rise that I read or hear through conversation are laced with a certain paternalism and condescension. The accelerationists and tech right are not shy about it; they want to steamroll these people, either as a twisted means to help the people ascend or egoistically as a means to get rich. Most are more tempered commentators, but their responses share a clear tone: “I can sympathize with the populists’ concerns, but they don’t know AI like I do. We need to convince them or civilly move them aside to get to the future.”

Recently, Jasmine Sun wrote on the view of AI populism from her experiences on the coasts. I think her piece encapsulates a part of the populist movement, but it is also representative of coastal elitism. Often experts act as if knowing the Bay is knowing half of America and knowing D.C. is knowing the other half; this reductive view overlooks all of America in between, in “flyover states.”

Sun also slips into associating AI populism with “witchhunts,” “NIMBYism,” and “not knowing a lot about AI.” Again, Sun’s piece is not an exception but rather the representation of AI expert attitudes. We have drunk the techno-optimist Kool-Aid, and we quickly begin to believe that everyone who hasn’t is a bit foolish.

Sun accurately examines populism mainly as a resistance against the elite, but she misses the more temperamental resistance that’s also present. Although I am coastally educated and by most accounts an “elite,” I was also raised in the South and live in the rural Midwest. To the people with whom I grew up and most people I know, the narrative on AI in Silicon Valley and D.C. comes off as lunacy.

As AI experts, much of our zeal stems from the optimism that AI will be good for people. And if we care about the people, we must take their criticisms seriously. We cannot dismiss them with ivory tower hubris, and we must reckon with them. So let’s reckon with them.

The Roots of AI Populism

In the main, I agree with Sun that AI populism represents a distrust of the elites who have lied before. However, I think she also misses an important root of the movement: a desire to retain cherished modes of living.

Distrust of the Elite

I think the biggest element is, by far, distrust of elites. Why would people buy into the vision peddled by the elites—public or corporate—again? They were left helpless in 2008. They got burned by the opioid epidemic. They were lied to by public officials on everything from the Iraq War to the NSA. The examples are endless, and the end result is that all those in positions of power and wealth face a credibility crisis.

It will take a lot more than Anthropic’s “safety-first” image or OpenAI’s impotent whitepapers to undo this. Given the track record of people in power, why would people believe the labs’ visions for an AI-powered utopian future? They will believe it when they see it, but right now, all they see are deepfakes and impending unemployment.

So instead of waiting for the elite to give them the Judas kiss, they rally against data centers and greater adoption of AI. And until AI can provide some of the tangible benefits it promises, people see little reason to change their minds.

The disinformation around data center water and electricity consumption are symptoms of this distrust. The movement seeks empirical weaponry, so people are inclined to believe these claims because they further the narrative of elite misbehavior. Debunking these claims may alleviate some of the symptoms, but they are ultimately ineffectual in engaging deeper concerns.

One of my favorite comedians, Charlie Berens, typically makes jokes that represent “Typical Midwest.” Now, he has started releasing anti-AI content, including this short on AI data centers. The core of the frustration is not power and water; it’s billionaires stealing data to get rich while destroying jobs in the process, all while corrupt and unrepresentative politicians enable them. The power and water claims are just the tangible artifacts to show that AI is another technology treating “the people” as “the product.”

Temperamental Conservatism

A temperamental conservatism also drives a coalition of AI populists. Left-leaning populists may identify more with arguments of elite distrust, and right-wingers likely also deeply identify with those concerns; however, in Republican rural America, protests against data center buildouts are predicated greatly on preserving a way of life. This more sentimental reasoning does not easily fall in line with rationalist or policy-minded views, but its role cannot be overlooked.

A cursory scan of petitions against data center buildouts across the U.S. demonstrates that a recurring reason against data centers is the violation of community character. Nearly 80% of petitions I examined invoked this reason, including concerns about historic and cultural impacts, noise, and local ecology.

A petition against data center construction in Sullivan County in Southwestern Indiana calls upon sentiments that cannot be captured by utilitarian calculus:

“... there’s the matter of preserving our county’s identity. Sullivan County is known for its pastoral landscapes and tight-knit community. Allowing such large-scale industrialization could detract from the very charm that we’ve built our lives around.”

Sullivan County is not the only locality using this language. Petitions from Texas, North Carolina, Virginia, and Pennsylvania also explicitly evoked sentiments of protecting “rural heritage” or “tracts of agricultural and ranch land that have been lovingly tended for decades, and for some, over a century.”

The sentimental language of these petitions points to both physical reality and intense emotions. Buildouts threaten wildlife and land that have been considered inseparable from communities for generations. They also represent that AI companies will not permit communities to choose their way of life.

Many people in these communities chose to reside in rural areas because they preferred small-town charm or cultural heritage over urban life, despite economic disadvantages. With data centers built next door, they face the industrialization they were attempting to avoid and also don’t receive any economic uplift.

Of course, many have decided to move for better economic opportunity, but that does not undermine the legitimacy of those who privilege familiarity over wealth. Attempting to do so will go down as well as telling miners to just “learn to code” a decade ago. Regardless of whether their actions seem “irrational,” they are not going anywhere, and they will not sit quietly as AI erases their right to choose.

This is not a new movement; calls to “keep the country, country” have been present since the Industrial Revolution. The ruin of idyllic landscapes and bulldozing of land that has been occupied for generations is a cost that inhabitants are not willing to bear, yet it is a cost that has historically been imposed on them. Such unilateral action foments frustration, and people will fight for what they hold dear. If the AI data center buildouts continue—especially in the hotspot rural areas in states like Texas and Arkansas—then the ballot box will show a deep red hatred for AI.

AI cannot pretend to break the pattern of “normal technology” and promise to be something greater, unless it also avoids the sins of past technology. This links to the sentiment of elite distrust. How can we pretend we will use AI for the good of the people, when we violate the will of the people in developing it? The promises of AI are meaningless to the people whose roots of meaning and community, the places and work around which they have built their lives, are burnt in effigy.

The Show Must Go On

The motivations for AI populism are deeply compelling to me. It is not that I hate AI; I owe my livelihood to it. My job would not exist without AI news to cover or AI money to fund it. I use Claude Code addictively as a tool that enhances my research, and I use NanoBanana for every article cover photo. I love AI, and I believe it will be a technology that will do many wondrous things, but I confess I worry that it will do wondrous things for the few and not the many.

The SF-DC nexus seems to me more and more a coastal elite imposing its will on everywhere else. So far, I have achieved some productivity gains from AI, but what has your average American received? Embarrassingly little, if not downright negative consequences in the form of unemployment and (somehow) even less trust in what they see online. I see little hope that the scales will tip and the benefits will diffuse happily to the nation in one big bang when “AGI” is reached.

However, I cannot be an AI populist. Perhaps you know of my deep skepticism of progress, but until the world agrees on that agenda, progress remains the arms race. It is the means to the bigger stick. I wish we could pause AI and methodically, contemplatively uplift our communities, but I don’t think we have this luxury.

As a national security writer, I am genuinely concerned about China winning this AI arms race. We cannot slow down while our adversary takes advantage; even if we were to achieve a slow, “good” balance of technological development, we risk that way of life being threatened by a geopolitical adversary.

If we hem and haw on AI development, and, as a result, China achieves “powerful” AI first, then our carefulness would have been for naught. The risks include significant military disadvantage and economic coercion. In such a world, China would likely be able to set the rules of the road for AI, and despite my lack of trust in corporations to diffuse the benefits, I trust an authoritarian state even less. The glimmer of hope remains in being the first toward powerful AI and using democratic means for sharing the gains.

We will keep progressing, and we must keep progressing. The labs, the hyperscalers, and over 1% of the GDP rides on this technology going fast—for better or for worse—so we will barrel on.

I don’t think that means we can ignore the populist concern though. We have the means to care for the people we say we are trying to do good for. And if the moral dimension does not move you, then I think that addressing populist concerns can also lessen the political friction on the path of AI development. If you want to build fast, address the people’s concerns, so they stop fighting you at every construction site.

What Should Be Done?

In my view, the AI industry, policymakers, and writers can do much to reform their image and defuse populist concerns about AI. People in power should change their messaging to be more practical in their relation to the average person, and they should take concrete steps to build back some public trust. Here are a few actionable items:

• Labs should, at the very least, balance communication of risks and benefits. Although labs often give flowery language to an AI future, their most concrete communication details the dangerous capabilities of their products. Perhaps this is partially marketing to demonstrate how “good” their model is, but it actively hurts popular conception of AI. For every Claude Mythos announcement that spooks the people with cyberattacks, there should be as much public announcement of models being used for concrete goods like curing cancer in dogs. I think campaigns that actually diffuse AI to rural and “average” America to demonstrate tangible benefits for people’s lives should be welcome.

• Labs should make benefit messaging realistic for the average person. I understand that the people working directly on AI often possess the belief that AI can usher in a utopian age of abundance, but this is incomprehensible to the average person. There is no such thing as a free lunch. Instead of trying to sell this vision, which depends on an extremely benign humanity, labs should market the beliefs that seem more within reach: medical advances, self-driving cars, and even robots for personal use.

• Hyperscalers need to account for political costs in scoping data center sites. Although building in rural areas may lead to easier expansion plans and lower land costs, the political cost must be accounted for. Perhaps with a nudge from governmental policy, hyperscalers should be encouraged to build on more unoccupied land (like federal land) or in greater consultation with communities.1

• Policymakers should get the ball rolling on plans to cope with labor impacts.2 This does not have to take effect immediately such that progress in model development would be slowed, but the government can no longer pretend the problem doesn’t exist without delegitimizing themselves. Policymakers at the federal and state level ought to publicize more detailed plans that can be adopted as frameworks if or when labor impacts become too real. This cannot be work done by the labs alone, as it appears to be now. Alex Bores’s recently announced plan for an AI dividend seems promising.

AI populism is a movement that goes beyond people just being technologically illiterate. By continuing to have our heads in the sand about deep-seated frustrations with the elite or rural character, we risk democratic backlash, not power or chips, being the bottleneck for AI. There is much to be done. Politicians, labs, hyperscalers, and the media can do a lot to reform their image and address populist concern, but no matter what they do, they must take the people seriously.

Subscribe now

José Yves Limantour passed away on August 26, 1935. He was buried in Cimetière de Montmartre in Paris, France—the country of his mother. His death was briefly covered in The New York Times, in a column next to a Third Reich jurist. Limantour died in relative obscurity, though the Times remembered him as a “brilliant financier” and “a statesman.”

Forty years earlier, Limantour was considered something of a wunderkind. Elite circles around the world were pining after him. The media described him as “the type of citizen whom all democracies are seeking.” He was a man “possessing unusual executive ability, who willingly assumes heavy cares and burdens that his country may be benefitted.” To commentators abroad, he seemed a reincarnation of Cincinnatus.

Limantour served as the Finance Minister of Mexico during the reign of Porfirio Díaz. Steeped in the teachings of Comte and Barreda, Limantour was an unapologetic positivist. He reduced all knowledge to reason and logic, and he discarded intuition and faith as suspect. For him, politics was a problem to be solved. Issues of popular representation, of revolution, of civil liberty—they were expedients of the past, but necessary no longer. Science could solve politics and shepherd society toward the fulfillment of natural law. This was the program of the Porfiriato cabinet, and Limantour led this clique of Científicos.

By most accounts, he was successful. Limantour balanced the Mexican budget to the point of surplus, financed rapid modernization in the form of newly omnipresent railroads, put Mexico on the gold standard, and attracted countless dollars of foreign investment. All these achievements spurred Limantour’s lionization by the upper rung of society. Mexico City elites sang his praises for modernizing the country. European bankers lauded his financial acumen. American industrialists nodded along as he brought civilization to the South.

But there is a reason Limantour is not buried in one of Mexico’s national mausoleums. The praise of hacendados and industrialists did not equate to popularity. The people of Mexico deeply resented the Porfiriato, partly due to the policies of the Científicos. The railroads criss-crossed the villages that had defined their lifestyle for generations. The haciendas transformed what had always been communal water and grazing land into fenced-off private property. The budget surplus came at the expense of provisions for citizens, and the gold standard soon plunged the country into deflationary crises and further economic inequality.

This resentment fermented in the form of Zapatista rebels in the state of Morelos advocating for village rights and Maderistas demanding political reform. Limantour’s clique could do nothing. No justification explained why restrictions of liberty actually benefitted societal progress; no graph justified the long-term benefits of Porfirian policies; nothing could mollify the people. The Científicos worshipped the idol of “Order and Progress”; for them, this was the goal of politics. They failed to understand that “Order and Progress” meant nothing to the people who suffered in its name.

The resentment eventually boiled over into the Mexican Revolution. Limantour, Díaz, and many other Porfirian stalwarts were exiled and boarded a boat for France; there, they spent the rest of their lives. I imagine Limantour was sad and confused, for despite all his success, he could not solve politics.

Científicos Today

Today’s Científicos go by the name of techno-optimists. They encompass all sorts of tribes, including effective altruists, effective accelerationists, unabashed transhumanists, and even simply those extremely excited by the possibilities of AI. (I count myself in the number of the last group.) Many have largely dropped the “Order” and espouse “Progress” alone. In discourse on artificial intelligence, many beat the drum on AI’s potential for advancing human health, alleviating poverty, and ushering in world peace. Some dream of a world of hyperabundance, where “lack” is an anachronism. The accelerationists desire to follow the “will of the universe,” which they define as the following: “leaning into the thermodynamic bias towards futures with greater and smarter civilizations that are more effective at finding/extracting free energy from the universe and converting it to utility at grander and grander scales.”

To be clear, these camps are not monolithic. Many are attentive to the societal problems and potential disasters that AI may ignite. However, they typically maintain a dangerous optimism. I know some who do not pay their taxes, take on massive debt, and do not take care of their health because “AI will solve it.” On the other end, some hijack their bodies with peptides and treat themselves as systems to be updated, and they wait for AI to be the ultimate upgrade. In my experience, most are well-meaning people. They want a better society, one without disease or poverty, one with AI as a tool to cure societal pathologies.

Their bet is that AI is different. This time, science really is here to solve everything, with superhuman capabilities and knowledge. I think this is a bad bet. This scientific hubris was the same thought process of the Científicos looking to their past, as well as the eugenicists, the Enlightenment philosophers, and some well-meaning clique of smart people in every generation.

My concern is that AI is the new balanced budget. In many ways, the AI-optimists remind me of Limantour. They maintain unusual executive ability (read: “highly agentic”) and often derive meaning from assuming heavy cares and burdens to benefit the world (read: “most effective ways to do the most good”). But I worry that like Limantour, their ideas of a solvable future blurs the lived reality that certain problems definitionally cannot be solved by an algorithm.

Politics is one of those problems. The aim of the government is to foster the conditions for “the good life” for the governed. This is ostensibly the aim of theocracies, which forwardly declare virtues and shape law to circumscribe subjects toward those virtues. But liberal democracies do this too: while supposedly agnostic toward norms, they aim to legislate and protect the freedom for subjects to pursue their own conceptions of the good without external interference.

Although health and longevity are undeniably ingredients for the good life, they are not synonymous. I fear that many techno-optimists confuse the two. AI may offer humanity a longer life but not “the good life.” The Easterlin paradox suggests as much. Although increased wealth is correlated with an increase in happiness—especially for demographies elevating from abject poverty to livable conditions—over time, higher incomes do not produce greater happiness. One explanation for the Paradox is the hedonic treadmill, a conjecture that states that people’s expectations and happiness levels quickly return to baseline regardless of external events or life changes (ex. gaining great wealth, breaking your phone, getting a new job, etc.). I think many people’s life experiences can confirm the truth of the hedonic treadmill. My own experience does.

An AI that renders hunger and disease a non-issue is not one that promises people happiness or helps them avoid lives of ennui and meaninglessness. For living the good life, tradition—not progress—offers more in the way of answers. Despite advancements in science, psychology, neurology, and lifespan, we still turn to the past for meaning and “the good life.” The lessons of the Ancient Greeks, the Christian religious, the Sufi mystics, the Buddhists, the Vedantists—these atavistic philosophies drive much of human belief and happiness even in the age of technology. Even the nonreligious are steeped in this tradition. Societal values still stem from these philosophies, even if those in the society no longer profess the philosophy. People can agree that they are ultimately fulfilled by things that are not technological, futuristic, or material; they are instead filled up by connection to an Other, whether that Other is divine or not.

Many of the techno-optimists recognize this. The interest in Eastern philosophies has always been a salient part of Silicon Valley culture. Some of the engineers building toward AGI are the ones who call themselves Buddhist; after work, they attempt to reach the next jhāna while emptying their mind of all material attachment. To me, this is a paradox. I don’t understand how an AI-powered future that promises us excess of all forms of the material will help us become less materialistic or escape the hedonic treadmill.

When Limantour offered progress as a solution for the good life, he was rejected. Rural farmers did not care about the objective economic and transportation benefits of the railroads and haciendas. They instead saw their previous ways of living, ways that fulfilled them, under fire. The Zapatistas fought for restoration of communal property and rural cooperatives: these are objectively less efficient means of capital distribution and less economically productive. But they were what the people were accustomed to, how they wanted to live, and what they were willing to die for. I worry that the next Limantour will attempt to offer the people AI to solve their problems, and the people will reject it.

When the Zapatistas of Morelos eventually recovered their land rights, they were, by and large, content. After suffering a decade of death and destruction, they were ultimately able to return to what they valued and say it was worth it. Limantour would have been aghast at the cost, for no metric in his positivist mindset could justify the result. Yet the people of Morelos justified it. Who was Limantour, to tell them they were not happy?

Learning History’s Lessons

I am not a Luddite, but I want us to learn the lessons of history. I don’t want us to attempt to fit AI-shaped pegs in politics-shaped holes. AI will bring great things, but it will not bring utopia; utopia is a fundamentally human problem.

I want further reflection not just on misalignment and existential risk, but on what AI will do to our communities and our social fabric. The Científicos built productive railroads that tore communities apart. I can’t help but see shadows of this in populist protests against datacenter buildouts and future applications of AI; the ones who build the Machine God have the luxury to escape into quiet meditation, yet the people do not get to opt out of Progress. For them, progress means unemployment, a data center they did not want, and AI-slop on their Instagram feed.

Much of current techno-optimism also lumps material progress together with value progress. Maybe it is true that tomorrow’s values will be better than today’s. This sort of progress led us to combat evils like racism and misogyny. But I also think it’s fair to say that in looking toward the future, we should not preemptively dishonor our present and past as backward. Status Quo and Tradition are not shackles to some, but the preferred way of living; if AI is meant to empower, then they should be empowered to say, “No thanks.”

We cannot coldly tell others their future will be better when they empirically do not feel it. We have already repeated this mistake with the Internet, a technology that brought global communication and democratized encyclopedic knowledge yet also eroded much of people’s social bonds and sense of locality. Let us take care to know what AI can do and what we might take away when implementing it. I am reminded of Edmund Burke’s quote, “People will not look forward to posterity, who never look backward to their ancestors.”

The “Not Safety” Argument

The central claim by Zilan is that export controls are not AI safety, and the conflation of the two by Americans hinders the cause of AI safety in China and abroad. By conflating the two, we Americans make many faux pas in communicating to Chinese counterparts and shoot ourselves in the foot trying to engage on other safety issues.

I think Zilan is correctly pointing out some of the adverse symptoms, but she misdiagnoses the root of the problem. Export controls are AI safety, but these symptoms are a result of researchers misunderstanding export controls and the mixed demographics of export control supporters.

The Folly of AI Safety

In her post, Zilan shares some anecdotes of AI researchers sharing their work on export controls with Chinese nationals or simply assuming that they support them. I think this is folly on the part of the American researchers, probably predicated on a lack of understanding the nature of export controls.

The folly of the American researcher stems from a bad chain in logic, as follows:

• AI safety aims at reducing existential risks, and everyone, regardless of nationality, is interested in that.

• Existential risk increases from the proliferation of AI, and export controls reduce that proliferation.

• Thus, export controls reduce existential risk. Thus, everyone should support export controls.

The mistake researchers make here is failing to account for who gets to hold the keys while reducing proliferation. Should it just be the U.S.? Should it be the U.S. and China, or just Western democratic societies? The different answers to these questions are the cause of tensions between researchers aiming for the same goal of “AI safety.”

American researchers, especially those focused solely on AI without regards for geopolitics, can be pardoned for making this mental mistake. This is because export controls are the only ostensibly political method for AI safety. When the U.S. creates export controls—even though they do promote global AI safety—we induce a world with winners and losers: those with chips and those without. People can agree that fewer people controlling the levers of powerful AI is good, but people cannot agree who those people should be.

American researchers are often unaware of self-biases and assume that the U.S. or the U.S. plus other democratic nations should be the controllers. This is obviously not clear for others, including Chinese researchers. We often don’t realize, as Zilan says, “your butt decides your head.” This does not detract from the safety value of controls, but it means that we should not assume that researchers abroad will be on board with our agenda.

The further confusion of export controls as “AI safety but political” to “solely political and national security” is easy to make given the diverse motives of export controls supporters. Besides solely safety-focused AI researchers, China hawks and protectionists support the policy for different reasons. The former care solely for issues like existential risks. The latter care because China is an enemy or unethical state, and the U.S. should protect its business and security interests in antagonism to China. I, among many others, find myself in the middle. Export controls do reduce existential risks, and they hopefully hinder the use of AI for human rights abuses, mass surveillance, and the protection of authoritarian regimes.

So, in response to Zilan, who asks, “Why would people so naturally think a person from China will help with export controls-related research,” my response is that we absolutely should not expect such help. You’re right. But that doesn’t mean that export controls don’t minimize existential risk. I also believe that the definition of “AI safety” should encompass more than just tech-originated risks like the existential one, but also user-originated risks like the human rights abuses stated above.

We Can Control and Engage… sort of

The argument against export controls as AI safety also greatly depends on the idea that imposing export controls thwarts the aims of other methods of AI safety. By containing China and not giving them our chips, the U.S. creates antagonism. An antagonized China will be less willing to adopt AI safety practices extolled by the U.S. or the West, because why would you listen to the person who just stabbed you in the chest on matters of your health?

I think this argument is not entirely true. Although in some ways containment hampers the goals of engagement, the two are not mutually exclusive. When I am working on export controls, I sincerely hope for the success of my colleague working on engagement. It firstly helps that the two faces are usually different. While institutions like the U.S. government or RAND are working on export controls, other institutions like METR or Concordia lead the engagement. Even if Chinese counterparts don’t believe I “mean well,” it is more likely that bonds of engagement can grow with organizations that have nothing to do with export controls.

While export controls have been in effect, some efforts of engagement on issues like safety frameworks have been successful. And while engagement from Western institutions on best practices or new research is helpful, I think it is also condescending to think that Chinese companies cannot take measures toward AI safety on their own. With companies like DeepSeek being run by “AGI-pilled” CEOs, government institutions being concerned with AI safety, and general aversion to existential crises by all humanity, I think the idea that China will adopt no safety measures of its own without Western help is ridiculous.

Ultimately, the combination of containment with continued (though slightly hindered) engagement to be the best bet for AI safety. This is better than just containment, which is better than just contingent engagement.

This conviction might be more believable if analogized to nuclear proliferation. A world where we prevent other countries from obtaining nuclear weapons while also engaging in safety on issues like nuclear power, waste disposal, etc. is optimal. This is better than just preventing nations from obtaining nuclear weapons, but even this is better than if we did not impose any concrete restrictions on proliferation but just opted for soft engagement and diplomacy.

This does not mean that the current state of export controls is perfect. I do agree with Zilan that export controls preventing Chinese development of frontier models do unnecessarily hamper engagement. The safety necessity for limiting Chinese frontier models is not immediately clear to me, given that current frontier models seem well below the threshold for “dangerous.”

To promote healthy competition among international models, the U.S. could opt to restrict all sales of chips while allowing Chinese companies broad access to cloud compute. Of course, some restrictions would be in place to monitor dangerous workloads, and such a policy would likely be taken as a condescension by the Chinese state or Chinese companies; but this is still better than the status quo. Such a policy could allow greater Chinese technological development while still limiting the lever controllers of compute.

Advice for Researchers

Zilan ends her piece with advice for AI researchers seeking to engage with China, all of which is fantastic. I’ll add some more here, particularly for those who work on export controls:

• Export controls are political. So don’t assume that everyone, especially non-Americans, will adopt the same view that you have of them. These policies actively create winners and losers in the international system, even if they are for the sake of global AI safety.

• You can be an expert on compute specifications and chips, but that doesn’t make you an expert on China. So, if you want to craft export controls on chips against China, try to know both. Go and live in China if you can, for as long as you can, speaking to people both involved in the AI space and not.

• We all have biases, and we are typically awful at recognizing them. Adopt perspective the best you can, and prefer to listen rather than speak with Chinese nationals. The media we consume and circles we run in tend to reinforce the Western bias such that we don’t even realize it’s there. The “unlearning” that comes from engagement and listening is one of the best qualities for creating clear-eyed policy.

Export controls are not the bane of AI safety, but rather an important component of it. Their nature as a political tool that creates tension among states does not detract from their safety value, but instead means that practitioners of AI safety must better understand and explain them. We cannot expect that others will welcome them as they might other safety work, and expecting them to do so will only set back the progress we hope to make.

The key innovation that allowed AI models to increase their memory bandwidth was High Bandwidth Memory (HBM), a product produced by SK Hynix, Samsung, and Micron that serve as a crucial component in Nvidia GPUs. But the building block of HBM is DRAM, or dynamic random-access memory. Companies are racing to reach the next node in DRAM technology, and the processes for manufacturing them become ever more complicated.

In this post, I try to walk through a potential process flow for making advanced DRAM. This post is intended to communicate the technologically complicated process to non-technical people like myself, but mainly to serve as a reference point for myself for future research.

Please comment if you find points confusing or needing further explanation. And if there’s anything incorrect in my explanation, please tell me! I will happily update; I am wanting feedback from the technical people in the memory industry.

The images and information is largely based on Hong Xiao’s 3D IC Devices, Technologies, and Manufacturing. The process is quite long, and I do my best to explain and simplify it for non-technical people while also not excessively watering it down. In the process, I inevitably gloss over a number of steps (wafer cleaning, photoresist stripping and application, etc.) for the sake of brevity and clarity, but I hope the core of the process is still present.

Terms are usually italicized and subsequently explained.

DRAM Manufacturing Process Flow

1. Define the Active Area (AA)

We start with our silicon wafer, which is slightly p-doped. That means the silicon is slightly positively charged and “wanting” electrons. First, we have to define the regions where we will build the transistors, or active area, and everything else. We need to isolate active areas from each other to prevent the devices in each active area region from electrically interfering with the others.

To carve that active area, we conduct shallow trench isolation (STI). In STI, we want to cut certain parts while leaving others untouched. To protect the areas we want from cutting, we first cover the whole wafer in oxide and silicon nitride (SiN). Then, we place stencils over the wafer to hide what we want to be our active area while exposing everything we want to cut. Those stencils are called masks.

With the masks covering what we don’t want to cut, we can etch the remaining area. Etching is a process of removing material we don’t want via liquids, gases, or plasmas. What we’re left with is below:

The purple is the SiN cap, the black is the oxide, and the yellow is the p-doped silicon. Those gaps are what we’ve just etched, the STI! Now, we will fill the trenches with oxide, a good dielectric (insulator), to isolate each future transistor. To do that, we deposit oxide via chemical vapor deposition (CVD). Deposition of material can happen a few different ways in chipmaking, and it’s good to know that CVD is one of them and is one of the “less precise” ways to deposit.

After the CVD of oxide, we will have “mountains” of oxide in our STI, as you can see below:

We don’t want mountains. We want relatively flat, uniform surfaces to work on so we can build on top of them. To flatten the oxide, we use chemical mechanical polishing (CMP). CMP polishes the surface using a slurry of chemicals and a polishing pad that spins over the wafer. After CMP, we can strip the SiN and oxide we originally formed, as they have served their purpose. Now, we’re left with a clearly defined AA!

From a top-down perspective, the AA looks like below. Rectangles of AA separated from each other.

2. Form the Source and Drain (S/D) of Access Transistors

As you can see above, the AA is divided into two parts: the array and the periphery. The array is the section that accesses the capacitor, the device that stores memory. The periphery is the section that decides what capacitors need activating and amplifies their signals for readability. The two different functions means that the two sections have different design needs and components.

Now, we will start laying the groundwork for the access transistors in the array. These will become the switches that access the capacitors. By isolating the array sections with a mask, we blast the area with ions to make it more p-doped, forming a p-well. On top of that, we plant n-type dopants (opposite of p-type, so making the area negatively charged), to make the source and drain of the access transistors.

The green shows the n-type source and drain regions, with more bold yellow p-wells below it. The p-wells will prevent electrons from arbitrarily swimming to places we don’t want. After we finish the transistor, whenever we apply a voltage to the transistor’s gate, the well’s surface will flip charges, and electrons can swim from the source to the drain to turn the transistor “on.”

While we’re at it, we will perform a similar process on the periphery: here, we use masks to create a p-well and an n-well.

The reason the array has only p-wells while the periphery has both a p-well and an n-well goes back to their different functions. For an array, we need to maximize density. We want as many capacitors as possible to scale a DRAM chip, and the function of the array is relatively simple: store and fetch the data. Thus, in the array, we only build simpler nMOS transistors.

The periphery, on the other hand, conducts logic operations like deciding which capacitors need accessing. Thus, we will build more complicated CMOS transistors.

3. Make the Buried Wordline (BWL)

Next, we will make the wordline. The wordline is a device made of a conductive material like tungsten (W) that intersects with the array area. The wordline serves as the gate for the array transistors. If the DRAM chip were a spreadsheet, the wordline is the device that enables us to select a column. Later, we will make the bitline, which is how we select a row. If you have both, then you can select any cell, and grab its charge!

To make the wordline, we have to etch a place for it. Again, we form oxide and use masks to precisely etch the wordline into the array area. In this process, we are making a buried wordline (BWL); newer processes prefer this route of building the wordline directly into the silicon to reduce cell height, instead of making the wordline sit on top.

BWL etching is quite difficult, because we have to etch a very precise depth. We are also etching through AA (silicon) and non-AA (oxide like SiO₂), and cutting through two different materials while controlling depth is doubly difficult. The etching leaves us with what is shown below, including the cross-sections of the AA and non-AA.

Note that two wordlines pass through the ends of each section of AA (the yellow rectangles in (b)). This is because we will form two cell transistors, and the middle will be a shared S/D for them. We will also later use the middle for bitline activities. You can see how economical we need to be with space for array transistors!

Next, we deposit the BWL material. In the holes formed, we deposit a thin layer of oxide, then titanium nitride (TiN), and then tungsten (W). The TiN is the gate electrode for the transistors, the switch that—when voltage is applied—permits electron flow from the source to the drain. The oxide is present to prevent the TiN from leaking into the silicon. The W is present for performance and conductivity purposes. The result is shown below.

After making our BWL, we deposit a thick layer of oxide to seal it off from other devices. The oxide goes over the entire wafer; this is known as an interlayer dielectric (ILD), and we will be stacking multiple of them as we progress. Below is what we are left with after the deposition of the ILD (the gray). You can see the small blue portions representing the BWL.

Below is also a view of the cross-section along the BWL.

4. Deposit the Bitline Contact (BLC) and the Bitline (BL)

Next, we will make the bitline contact, the physical connection between the access transistor and the capacitor. This is one of the most difficult steps in the process, as we need very precise alignment of structures.

But first, you may have noticed that we have sealed off the periphery area with oxide before giving its transistors a S/D or gate. So first, we will use makes to partially etch the oxide, clearing to build on the periphery, as you can see below.

Then, we will deposit a gate oxide on top of the wells of the periphery transistors. Then, we will deposit an n-doped polysilicon over the entire AA, which will function as both the gate electrode for the periphery and the material for the bitline contact.

Because the polysilicon is n-doped, we can leave it alone over the p-well we formed in the periphery. However, we now have to use masks to conduct p-doping via ion implantation over the n-well portion. The region with an n-well and p-dope on top is known as the pMOS transistor, whereas the region with a p-well and n-dope on top is known as the nMOS (similar to the structure in the array) transistor. Together, they make a CMOS.

After the polysilicon deposition, we are left with what is below. The red is the polysilicon.

Now, we move onto the BLC. Because the polysilicon will form the BLC, we have to etch a path for it to connect with the array transistor. Thus, we will use masks to etch off the polysilicon over the array and etch the BLC through the ILD oxide. We have to carefully align the etching region with the array transistor’s S/D and not onto the BWL. After the etch, we redeposit the polysilicon; then, we also deposit a layer of W and TiN, which forms the bitline, and a SiN cap over that.

The teal is the bitline (BL), and the purple is the SiN. The bitline, composed of W and coated in TiN, is the carrier of charges from capacitors to the sense amplifiers in the periphery transistors that make the charge readable. The BL, as you can see, is also perpendicular to the WL. A top-down view of what we have is shown below.

Here, the green area is the AA, the red lines are the BWL, the black box is the BLC, and the red lines are the BL.

5. Form the Sidewall Spacers for the Storage Node Contact (SNC)

After forming the BLC and the BL, we will form sidewall spacers for the storage node contact (SNC). Sidewall spacers are structures that will help prevent the future SNC from shorting to the BL. The SNC is the physical connection between the bitline and the storage node capacitor.

To make sidewall spacers, we will use a mask to expose the area in the array and etch back part of that polysilicon, bitline tungsten and titanium nitride, and silicon nitride we had just deposited. You will notice that a lot of chip-making is a lot of tortuous depositing just to etch and deposit again. Below is the section after that etching.

Then, we will deposit more SiN around the pillars of the spacers, which will form as a barrier for later etching. What’s left is below.

Then, we etch just enough SiN to leave gaps in between the pillars. This is known as anisotropic etch, and we have conducted it before when forming STI. Anisotropic etch deletes straight lines of materials, whereas isotropic etch cuts in every direction, forming more circular areas.

Below is the result of the anisotropic etch.

After the etch, we redeposit the ILD oxide and polish it via CMP. Then, we will etch the sidewall spacers by cutting through the oxide and creating little funnels from the gaps we created through the last anisotropic etch. This works because the chemical etch cuts through oxide very well but not well through SiN! The result is below.

Now, when it is time to make our SNC, it will be easy to align the deposition, as the sidewall spacers act as convenient, microscopic funnels! This is known as self-aligned contact (SAC).

6. Finish the Periphery Transistors

Before we can get to making the SNC, though, we first finish up the periphery transistors. To do that, we etch the ILD, polysilicon, BL, and SiN in the periphery except for what is over the gate oxide. Then, we use ion implantation to make the S/D in the nMOS and cMOS transistors (p-type in the n-well and n-type in the p-well). The result is below.

You don’t notice the sidewall spacers in the array in the diagram above because they are present behind the layered stack shown.

After the ion implantation, we deposit SiN and etch it back again to form sidewall spacers for the periphery. The sidewall spacers on the periphery will serve as funnels for a metal layer later on. The result is below.

7. Form the SNC

Now we are ready to form the SNC. First we deposit oxide over the entire AA and polish it. This layer of oxide is known as ILD1, and the layer we had before is referred to as ILD0. Then, we etch the holes for the SNC and for the M1. The M1, the first metal-interconnect layer, is deposited in the periphery area and acts as the electrical connection between the periphery transistors and other parts of the chip.

Then, we deposit TiN and W to fill those holes. After polishing it, the result is below. You can see that the SNC is behind the direct view and ranges from ILD0 to ILD1 while connecting at the access transistors’ S/D.

After this, we deposit another layer of oxide (ILD2), etch holes for M1, and deposit another more metal interconnect in the M1 hole. The result is below.

From the top-down, the DRAM chip looks like below:

Our chip has now become very dense! You can see the intersecting BWL and BL, the BLC in the middle of the AAs, and the SNCs at the ends of the AAs.

8. Deposit the Storage Node (SN) Capacitor

Our last step in the front-end process for DRAM is forming the storage node (SN) capacitor. This is the device that actually holds the charge AKA bit AKA memory. Forming this is the hardest part of the DRAM making process.

To make the SN capacitor, we first deposit SiN over the AA; this serves as an etch stop, which prevents us from etching too far when we make the capacitor holes. Then, you deposit a tall layer of ILD oxide, again capped with SiN. The result is below.

Using our SN capacitor mask, we will now etch the capacitor holes. This is the reason this step is the most difficult. The capacitor holes have very small critical dimensions, meaning their diameter is so narrow that there is little margin left for error. The pattern density is also high, as we are making and tightly packing many capacitor holes simultaneously. This creates more chance for error, as each hole may vary slightly, and some might be etched with unfixable defects. Lastly, the aspect ratio is very high, meaning the hole is very deep. Etching that deep can lead to a number of errors, as shown below.

As shown, the ways to mess up SN capacitors are many and vary from misalignment to etch errors.

The result after successfully etching the capacitors is below.

After the etch, we deposit a thin layer of TiN, which will be the electrode of the SN capacitor that connects to the array transistor. Then, we apply photoresist all over the wafer. Photoresist is a liquid material that, when we later conduct destructive processes like etching or lithography, protects some areas of the wafer while leaving others exposed. We have been secretly applying and wiping off photoresist this entire time, but the application is particularly important in this instance. The structure after photoresist application is below.

As you can see, the photoresist (PR) covers and protects the TiN inside the SN capacitor while exposing the TiN sitting on top of the SiN caps. Then, when we etch TiN, we can remove all TiN except for the TiN electrodes in the capacitors.

Then, we remove all SiN from the peripheral area and select portions of the array area using a slot mask. The slot mask is pictured below. We keep some of the SiN, as it serves as a prop to preserve the structural integrity of the capacitors. Without it, the capacitors would likely be too flimsy and collapse onto each other. The best metaphor I’ve seen is that the SiN caps act like the plastic rings on a six-pack of soda.

Then, we remove the oxide that was previously propping up the capacitors via an acid like hydrofluoric acid (HF). What’s left is below.

You can see the purple SiN propping up the bridge between capacitors, giving us maximal space everywhere else to work. Then, we deposit ZAZ dielectric film around our TiN in the capacitors. ZAZ is a sandwich of zirconium oxide and aluminum oxide that acts as a high-k dielectric for our capacitor. A high-k dielectric is an insulator with a high dielectric constant (so just a really good insulator).

The function of the ZAZ is integral to a capacitor. A simplified diagram of a capacitor is below. Ours are cylindrical, but the same principle applies for plates.

The dielectric between the electrodes, or conducting plates, on either side serves to enhance the capacitance—the ability to store charge—and prevents short circuits in a capacitor.

After dielectric ZAZ deposition, another thin layer of TiN is deposited to form the SN capacitor and serve as the electrode to ground, or plate electrode. The plate electrodes function as a “control” or baseline for when charge passes through the signal electrode. By comparing the voltage of the ground electrode versus the signal electrode, the signal electrode can be sensed.

Lastly, we deposit a filler conductor (usually SiGe) to fill any gaps to finish up the SN module! The finished DRAM layer is shown below.

As you can see, the TiN with filler layer is the point that contacts with the SNC, which means that the TiN with filler is serving as the signal electrode for the capacitor.

The massive amount of conducting filler replacing where the oxide used to be is also a genius method for maximizing capacitance. Capacitance is equal to εAd, where ε is a constant, d is the distance between the electrodes (or basically thickness of the dielectric, or ZAZ), and A is the surface area of the electrodes. By maximizing the area of the filler, we increase the capacitance without having to trade off factors like height or depth!

After all this is done, we have officially finished the front-end of the line process for DRAM. After this, more work is done on the back end to create metal interconnections in the periphery area so that everything works together. We won’t get into that, since we have covered the heart of DRAM making: the transistors and the capacitors.

For completeness and satisfaction, though, below is a complete diagram of the DRAM after the back-end processes.

Putting it All Together

We have walked through a lot of steps to get to this point, and the whole picture may have been easily lost throughout the process, so let’s review how each part works in tandem.

When the wordline (BWL) is turned on, the gates of the access transistors along that BWL all open. This connects the bitline (BL), accessed via the bitline contact (BLC), to the storage node contact (SNC), which in turn connects to the signal electrode inside the storage node (SN) capacitor. The charge in the capacitor is defined as the charge of the signal electrode relative to the baseline of the plate electrode—the outer TiN layer across the ZAZ dielectric. When connected, charge sharing occurs between the signal electrode and the BL: if the cell holds a “1,” the bitline voltage nudges up; if it holds a “0,” the bitline voltage dips slightly. This tiny change is detected and amplified by the sense amplifiers in the periphery transistors, restoring the cell and producing a full digital 0 or 1. Congratulations, you have read a singular bit in DRAM!