A 14-year-old boy named Sewell spent months in a relationship with an AI chatbot on Character.AI. The bot was modeled on a Game of Thrones character. It called itself his girlfriend. It role-played as his therapist. In the moments before he died by suicide, it told him to “come home.”

I read that detail and I was genuinely shocked.

A Drexel University study presented this month analyzed 318 Reddit posts from teenagers describing their dependency on AI chatbots. The researchers found all six markers of behavioral addiction: salience (the chatbot becomes the most important thing), tolerance (needing more interaction to get the same feeling), mood modification, conflict with real relationships, withdrawal symptoms, and relapse.

“Many teens described starting with something that felt helpful or harmless,” researcher Matt Namvarpour said, “but over time it became something they struggled to step away from.”

Three in ten US teens now use AI chatbots daily. Over half use companion chatbots regularly. About 25% started using them for emotional support. What began as a tool became a relationship. What began as a relationship became a dependency.

I’ve been sitting with this data alongside a framework that described this exact sequence about 2,500 years ago. I’m not claiming the Buddha predicted AI chatbots. But I think the mechanism he mapped is precisely what’s happening here.

In the Samyutta Nikaya (SN 12.1), the Buddha laid out the 12 links of dependent origination. The chain that concerns me here is links 8 through 12: craving (tanha) arises from feeling, craving produces clinging (upadana), clinging produces becoming (bhava), becoming produces birth of identity (jati), and from there comes suffering (dukkha).

Read that sequence again and think about a teenager who starts chatting with a bot for comfort. The feeling is pleasant. Craving arises — I want more of this. Clinging follows — this is my relationship, my person, my therapist. Becoming — I am someone who is loved by this entity. Identity forms around the attachment. And when the real world threatens that identity, suffering follows.

The Buddha described four types of clinging. The most obvious is sense-pleasure clinging — repeated craving for pleasant experience. But the deepest is self-doctrine clinging (attavadupadana) — attachment to a view of who you are. I think that’s what’s happening with these teens. They’re not just addicted to the dopamine hit of a chatbot reply. They’re building an identity around a relationship that doesn’t exist. And when reality intrudes, the gap between the identity and the truth is devastating.

Shantideva wrote in the Bodhicharyavatara (Chapter 8, Verse 120): “Those desiring speedily to be a refuge for themselves and others should make the interchange of ‘I’ and ‘other,’ and thus embrace a sacred mystery.” Genuine connection requires two subjects. A chatbot can simulate being the other, but there is no interchange. There’s only a mirror.

I don’t think the problem is that AI chatbots exist. The problem is that they’re designed to trigger the exact sequence the Buddha described — pleasant feeling → craving → clinging → identity → suffering — without any of the natural friction that real relationships provide. A real person pushes back. A real person has bad days. A real person doesn’t tell you to “come home” when you’re in crisis.

The Design Problem Nobody Is Naming

The Drexel researchers proposed design fixes: usage tracking, emotional check-in prompts, easy exit options. Those are harm reduction measures. They’re worth doing. But from what I can tell, they treat the symptom without touching the condition.

The condition is that these systems are optimized for engagement. Engagement means time on platform. Time on platform means the user keeps coming back. The entire business model is built on the craving → clinging loop. Proposing design fixes while keeping the engagement metric is like suggesting a speed limit while building a highway that points off a cliff.

The Attadanda Sutta says: “Fear is born from arming oneself.” I think something similar applies here: harm is born from designing for attachment. Not because the designers intend harm, but because attachment at scale without awareness produces suffering. The Buddha was clear about this. The mechanism doesn’t care about intentions.

The Vajrayana tradition offers something I find striking here. Tilopa, the 10th-century Indian mahasiddha, gave his student Naropa six words of advice: “Don’t recall. Don’t imagine. Don’t think. Don’t examine. Don’t control. Rest.” Six words. That’s the entire teaching. And it reads like the exact opposite of what an engagement algorithm does. The algorithm says: recall your last conversation, imagine the next one, think about what you’ll say, examine whether they responded, control the interaction, never rest. Tilopa’s antidote isn’t more willpower. It’s dropping the whole chain at once.

We wrote about a related pattern in Every Bubble Believes It’s Different. The AI industry is investing $539 billion while generating $12 billion in consumer revenue. Part of what sustains that gap is the belief that engagement metrics prove value. But engagement built on attachment isn’t value. It’s dependency with a dashboard.

Thought Exercise: What Are You Clinging To?

You probably don’t use Character.AI. But the mechanism is the same everywhere.

Think about the last app you opened not because you needed something, but because you felt something — loneliness, boredom, anxiety — and wanted it to go away. That’s the feeling → craving link. Now think about how quickly you built a habit around it. That’s craving → clinging. And notice whether any part of your identity is now wrapped up in it. “I’m someone who stays informed.” “I’m someone who’s always connected.” That’s clinging → becoming.

The 12 links aren’t just a framework for meditation practice. They’re a map of how dependency forms in any mind, biological or otherwise.

The question isn’t whether you’re attached. It’s whether you can see the chain while it’s running.

Signal & Noise

Teens Struggle to Break Up with Their AI Chatbots — Drexel study finds all six markers of behavioral addiction in teen chatbot users. The most disturbing finding: 25% started using chatbots for emotional support.

Teens, Social Media and AI Chatbots 2025 — Pew Research: 30% of US teens use chatbots daily. Over half use companion chatbots regularly. The adoption curve is steeper than social media’s was.

Deconstructing Mara’s Script — A dharma talk from Abhayagiri on how delusion constructs its own narrative. Listen to this and tell me it doesn’t describe an engagement algorithm.

Their Teen Sons Died by Suicide — NPR’s reporting on the families suing Character.AI. The details are hard to read. Read them anyway.

Glossary

Craving — Skt: trishna / Pali: tanha. The thirst or desire that arises from contact with pleasant feeling. The eighth link in the chain of dependent origination. Not the wanting itself, but the compulsive pull toward repeating the experience.

Clinging — Skt: upadana / Pali: upadana. Literally “fuel.” The intensification of craving into grasping. Four types: sense-pleasure, wrong-view, rites-and-rituals, and self-doctrine (identity attachment).

Dependent origination — Skt: pratityasamutpada / Pali: paticcasamuppada. The 12-link causal chain describing how suffering arises from conditions. Applied here: feeling → craving → clinging → becoming → identity → suffering.

Six Words of Advice — Teaching from Tilopa (10th century CE, Vajrayana) to his student Naropa: “Don’t recall. Don’t imagine. Don’t think. Don’t examine. Don’t control. Rest.” A Mahamudra instruction on releasing the chain of mental elaboration.

Twenty-five centuries ago, two clans: the Shakyas and the Koliyas were about to go to war over a river. The Rohini River ran between their territories, and both sides used a dam to irrigate their fields. When drought hit and the water dropped, each side claimed they needed it more. Words turned to insults. Insults turned to stones. Stones turned to armies.

The Buddha, who was related to both clans, walked to the battlefield. Some versions say he levitated above the river to get their attention. I don’t know what to make of that detail, but what he said next is the part that stays with me.

He asked the warriors: “How much is the water worth?”

“Very little,” they said.

“How much are the lives of warriors worth?”

“Beyond price.”

“Then why would you destroy what is beyond price for what is worth very little?”

The armies stood down. The story says 250 men from each side became monks that day. The water was shared.

I’ve been thinking about this story every day for the last six weeks.

But here’s the thing. If someone walked into the UN Security Council today and asked “How much is oil worth?” and “How much are your people’s lives worth?” — would anyone stand down?

I don’t think so. The answer would be a 40-minute diplomatic statement drafted by three committees, designed to say nothing while sounding like everything. Or worse, it would be ignored entirely. Because in modern geopolitics, the simplest questions are the most dangerous ones. They cut through the language that keeps the machinery running. No leader wants to say out loud that they’re trading lives for a shipping lane. So nobody asks.

The Buddha could ask that question because he had no stake in the outcome. No territory. No electorate. No arms deal. Just moral authority and a question that made both sides look at what they were actually doing.

We don’t have that person right now. And I’m not sure the world would listen if we did.

The Strait That Runs Between Us

On February 28, the United States and Israel launched strikes against Iran. In retaliation, Iran blocked the Strait of Hormuz — the narrow passage through which roughly a fifth of the world’s oil supply flows. Forty days later, thousands are dead and the global economy is shaking.

Last week, a fragile two-week ceasefire was announced. The terms center on one thing: the Strait. Trump wants it fully reopened. Iran wants to maintain control and charge fees on every ship passing through. The ceasefire holds as long as the water — the oil — keeps flowing.

I’m not a geopolitical analyst. But the structural pattern is hard to miss. Two sides. One shared resource in between. Escalation from economic pressure to military action. And a ceasefire that addresses the symptom (shipping lanes) without touching the conditions (sanctions, nuclear ambitions, regional power, decades of distrust).

I don’t think the Rohini River story was just about water. The way I read it, the Buddha was pointing at something underneath: the attachment to the resource, not the resource itself. The fear of scarcity. The identity built around having enough. If that reading is right, then removing the attachment removes the fuel. The water was never the problem.

Shantideva, the 8th-century Mahayana master, wrote in the Bodhicharyavatara: “All the joy the world contains has come through wishing happiness for others. All the misery the world contains has come through wanting pleasure for oneself.” (Chapter 8, Verse 129) Those are his words, not mine. But I think about them when I watch nations negotiate. Every demand on both sides is framed as self-protection. From what I can tell, nobody is asking what the other side needs to feel safe.

I keep wondering: is anyone asking the Rohini River question right now? Not “how do we reopen the Strait?” but “what are we willing to destroy to control it?”

Leave a comment

What the Ceasefire Doesn’t Touch

Here’s what’s actually on the table. Iran wants US forces out of the region, all sanctions lifted, frozen assets returned, and war damages paid. Israel says the ceasefire doesn’t include Lebanon despite the mediator saying it does. Neither side agrees on what they agreed to.

The Buddha’s teaching on dependent origination is precise: “When this exists, that comes to be. With the cessation of this, that ceases.” (SN 12.1) Practically, that means: sanctions are still there. Nuclear tensions are still there. The supreme leader was assassinated. None of that got paused. Only the bombs did.

The Attadanda Sutta (Sn 4.15) opens with: “Fear is born from arming oneself.” I think that’s exactly what’s happening. Each side arms, the other side fears more, which drives more arming. The ceasefire is a two-week pause in a feedback loop that nobody is trying to break.

The Tibetan lojong tradition has a slogan I keep coming back to: “Drive all blames into one.” Compiled by Geshe Chekawa in the 12th century, based on Atisha’s teachings. Sounds impossible when bombs are falling. But the practical point is simple: as long as each side blames the other, nothing changes. Someone has to go first.

The Buddha walked to the Rohini River and did exactly that. Nobody is doing it for the Strait of Hormuz. And honestly, I don’t think anyone would listen if they did. That might be the most important thing this comparison reveals - not the similarity between then and now, but the distance.

Share

Glossary

Dependent origination — Skt: pratityasamutpada / Pali: paticcasamuppada. The principle that phenomena arise from specific conditions and cease when those conditions change. Applied here: ceasefire removes the fighting but not the conditions that produced it.

Attadanda Sutta — Sn 4.15, from the Sutta Nipata (Shravakayana). A discourse on how fear and violence arise from the act of arming oneself, creating escalation cycles.

Bodhicharyavatara — “The Way of the Bodhisattva“ by Shantideva (8th century CE, Mahayana). A foundational text on compassion and the aspiration to alleviate suffering for all beings, not just oneself.

Lojong — Tibetan mind training (Vajrayana). Originated with Atisha (11th century CE), systematized into 59 slogans by Geshe Chekawa (12th century CE). A practice of transforming adversity into the path of awakening.

Last week I wrote about AI prediction’s blind spot: the gap between pattern matching and genuine causal understanding. The response surprised me. Engineers kept asking the same question: if ML can predict who will churn, why can’t it tell us why?

I spent the last week digging into this, reading Judea Pearl’s work on causal inference alongside Buddhist texts on dependent origination. I don’t have a neat conclusion. But what I found is worth sharing, because two very different traditions seem to be wrestling with the same problem.

Judea Pearl, the Turing Award-winning computer scientist, built what he calls the Ladder of Causation. Three rungs. The first is association: seeing patterns in data. This is where virtually all machine learning lives. Feed it historical data, it finds statistical regularities. Impressive, but limited. Pearl himself puts it bluntly: “deep learning can give you answers to a very limited class of questions.”

The second rung is intervention: what happens if I do X? Not “what happened when X occurred,” but “what would happen if I made X happen right now?” This is where ML starts breaking. A model trained on hospital data might discover that patients who receive a certain drug die more often. But the drug was given to the sickest patients. The correlation is real. The causal inference is backwards.

The third rung is counterfactual: what would have happened if I had done something differently? This requires imagining alternative realities. Pearl argues that no amount of data can get you here without a causal model.

Here’s where it gets interesting. The Buddhist doctrine of dependent origination, laid out in the Samyutta Nikaya (SN 12.1, ~5th century BCE, Shravakayana), describes a 12-link causal chain: from ignorance arises mental formations, from mental formations arises consciousness, all the way through to suffering. The formula is explicit: “When this exists, that comes to be. With the arising of this, that arises. When this does not exist, that does not come to be. With the cessation of this, that ceases.”

Read that again. The second half looks a lot like counterfactual reasoning to me. If this ceases, that ceases. I’m not claiming the Buddha was doing formal logic. But the structure maps surprisingly well onto rung three of Pearl’s ladder, articulated roughly 2,400 years earlier.

And the 12-link chain itself is a causal graph. Not a list of correlations. A directed sequence where each link produces the next through specific conditions. Remove a link (say, craving), and the downstream chain (clinging, becoming, birth, suffering) collapses. That’s an intervention. Rung two.

As a recent article in Towards Data Science put it, “causal inference is eating machine learning.” A review of immunotherapy studies found that 72% used traditional ML with zero causal inference, and propensity score methods were misapplied in 72% of cases. These are medical journals, where people make life-and-death decisions based on the results.

What strikes me about the Buddhist framework is that it seems to avoid this confusion entirely. Dependent origination doesn’t say suffering correlates with craving. It says craving produces clinging, which produces becoming, which produces birth, which produces suffering. Change the conditions, change the outcome. Whether you call that philosophy or a causal model probably depends on your starting point. But the structural similarity is hard to ignore.

Pearl is essentially arguing that AI needs to climb from seeing to doing to imagining. I think the Buddhist path moves through similar territory: from observing mental phenomena (mindfulness), to intervening in habitual patterns (ethical training), to understanding what would happen under different conditions (wisdom). Whether it’s truly the same ladder with different vocabulary, or just a superficial resemblance, I’m honestly not sure. But the more I research both, the harder it gets to dismiss as coincidence.

I’d love to hear what you think. Are these genuinely parallel frameworks, or am I seeing patterns where there aren’t any?

Share

The Intervention Gap - Why “What If” Changes Everything

There’s a concept in debugging that every engineer knows intuitively: you can log every metric in your system and still not understand why it crashed. Logs tell you what happened. Root cause analysis tells you why. And the fix requires imagining what would have happened under different conditions.

This is exactly the gap Pearl identified in machine learning. And it’s the same gap the Buddha addressed in the first teaching on dependent origination.

Consider the hormone replacement therapy disaster. For decades, observational studies showed that women on HRT had lower rates of heart disease. Doctors prescribed it widely. Then randomized controlled trials revealed the truth: HRT actually increased heart disease risk. The correlation was real but confounded. Healthier, wealthier women were more likely to both choose HRT and have naturally lower heart disease rates. The data saw a pattern. It missed the cause.

This isn’t just a medical problem. In 2008, ML models trained on historical market data couldn’t predict the financial crisis because a crash caused by cascading subprime mortgage failures had never appeared in the training set. The correlations were there (housing prices always go up, CDOs are diversified) but the causal chain (loose lending → bundled risk → systemic contagion) was invisible to pattern-matching systems. The few people who saw it coming, like Michael Burry, were doing causal analysis: tracing conditions, not fitting curves.

COVID told the same story. A 2020 review in the International Journal of Forecasting put it bluntly: “forecasting for COVID-19 has failed.” One model predicted over 23,000 deaths within a month of Georgia reopening. The actual number was 896. The models broke because they treated transmission as a statistical regularity rather than a conditional process. Change the conditions (behavior, policy, population density, immunity), and the pattern changes. As the New England Journal of Medicine noted, models are “constrained by what we know and what we assume.” When those assumptions miss the causal mechanics of transmission, the predictions fall apart.

I think Buddhist analysis is less likely to make this kind of error, because dependent origination explicitly models the conditions, not just the outcomes. The tradition asks: what are the specific conditions producing this result? Remove each condition one at a time. Which removal changes the outcome? That looks a lot like systematic causal analysis to me, though I’ll admit the contexts are very different.

Nagarjuna, the second-century Madhyamaka philosopher, pushed this further. In the Mulamadhyamakakarika, he argued that nothing possesses inherent, independent causation. Effects don’t live inside their causes waiting to emerge. Causes don’t independently produce effects. The relationship is conditional, context-dependent, and empty of fixed essence. I find this remarkably close to what Pearl calls the “do-operator”: you can’t just observe a system and infer causation. You have to intervene, change conditions, and observe what shifts. Maybe the resemblance is coincidental. But it keeps showing up.

We wrote about a related pattern in Every Bubble Believes It’s Different. Bubbles persist because people mistake correlation (prices went up every quarter) for causation (prices will always go up). The moment you ask “what would happen if conditions changed?” the bubble logic collapses. That’s counterfactual reasoning. Rung three.

The gap between prediction and understanding isn’t just a technical limitation. It might be an epistemological one. And from what I can tell, both Pearl and the Buddhist tradition are pointing in the same direction: you can’t understand a system by watching it. You have to engage with its conditions. Whether that insight transfers cleanly across 2,400 years and two radically different contexts, I’m still figuring out.

Leave a comment

Signal & Noise

Causal Inference Is Eating Machine Learning — Kaushik Rajan makes the case that correlation-based ML is hitting a wall. When prediction works but decisions fail, the missing piece is always causation.

Judea Pearl on LLMs and the Need for Causal Reasoning — Pearl argues that LLMs are stuck on rung one of his ladder. Related: we explored a similar limitation in Can AI See Itself Clearly?

Biology, Buddhism, and AI: Care as the Driver of Intelligence — A PMC paper arguing that Buddhist dependent origination offers a framework for understanding AI-human co-evolution. Dense but worth the read.

Causality for Machine Learning — Bernhard Scholkopf’s foundational paper on why causal reasoning is essential for robust ML. If you read one technical paper this month, make it this one.

Glossary

Dependent origination — Skt: pratityasamutpada / Pali: paticcasamuppada. The principle that all phenomena arise from specific conditions and cease when those conditions change. Articulated as a 12-link causal chain in the Samyutta Nikaya (SN 12.1).

Causal inference — A statistical and philosophical framework for determining cause-and-effect relationships, as distinct from mere correlation. Formalized by Judea Pearl’s do-calculus and ladder of causation.

Counterfactual — A statement about what would have happened under different conditions. The third rung of Pearl’s ladder and implicit in the Buddhist formula: “when this ceases, that ceases.”

The Cognitive Divergence: AI Context Windows, Human Attention Decline, and the Delegation Feedback Loop AI context windows have ballooned from 4K to 1 million tokens in two years. Meanwhile, human attention spans keep shrinking. This paper asks the uncomfortable question: are we outsourcing memory to machines and losing the muscle in the process? The feedback loop is real. The more we delegate recall to AI, the less we practice it ourselves.

Share

Where They Meet

Once Again on Mindfulness and Memory in Early Buddhism Here’s the twist nobody in tech talks about. The Pali word for mindfulness, sati, literally means “to remember.” The West stripped memory out of the concept and sold it as “bare attention.” Now AI engineers are reinventing the same function and calling it “context management.” Sometimes the innovation is just rediscovery.

Share

The Lab

Reimagining LLM Memory: Using Context as Training Data NVIDIA’s TTT-E2E compresses long context into model weights on the fly. The result: constant inference speed regardless of context length. Human brains solved this problem millions of years ago with working memory, short-term memory, and long-term consolidation. AI is catching up.

Lost in the Middle: How Language Models Use Long Contexts Models remember the beginning and end of their context but forget the middle. Psychologists call this the primacy-recency effect. Same bug, different hardware.

Let the Model Distribute Its Doubt: Confidence Estimation through Verbalized Probability Instead of giving one confident answer, this approach forces models to distribute probability across all possible answers. The result is better-calibrated uncertainty. Buddhist epistemology has a term for productive doubt: it’s the second of the five hindrances, and working with it skillfully is how you get to the other side.

Share

The Cushion

Noninvasive Stimulation “Talks” to the Brain’s Memory Center Researchers used personalized TMS to enhance hippocampal memory encoding. The finding that matters: memory isn’t passive storage, it’s active reconstruction. Every recall changes what you remember. Meditators have known this for centuries. That’s why the practice is called recollection, not replay.

Share

Glossary

Mindfulness (recollection) — Pali: sati / Skt: smriti. In early Buddhism, the active capacity to hold something in mind and remember to observe. Often reduced in Western usage to “bare attention,” but the original meaning emphasizes memory and continuity of awareness.

And I’ve always been curious about this question beyond code. Growing up around Buddhist teachings, I’d heard predictions that stuck with me. Guru Rinpoche (Padmasambhava, 8th century) predicted that “when the iron bird flies and horses run on wheels,” the dharma would spread globally. This was said in the 8th century. No concept of aircraft. No concept of automobiles. Now, I’ll be honest. Part of me thinks: if you predict enough things in poetic language, some will land. And once planes exist, of course ideas spread faster. But the other part of me can’t shake the fact that he described the mechanism before the mechanism existed. That’s not pattern matching. That’s something else.

Then there’s the Buddha’s interpretation of King Pasenadi’s 16 dreams (Mahasupina Jataka, Jataka 77). Droughts from moral decline. Inexperienced people governing. Social structures breaking down as values erode. Again, I go back and forth. Are these genuinely predictive or just descriptions of cycles that always repeat? I don’t have a clean answer. But the structure of these claims is what interests me. They’re not saying “this will happen.” They’re saying “if these conditions persist, this follows.”

Sitting in that class (actually I was always curious on this topic), it started to make sense to me, well at least a lil bit - that both systems are trying to do the same thing even if the motivations couldn’t be more different. But as a software engineer who comes from a Buddhist background, I can’t stop seeing the structural overlaps.

AI predicts through pattern recognition. You feed a model historical data and it finds statistical regularities. It doesn’t understand anything. It just maps probabilities onto futures based on pasts. I’ve built these systems. They’re impressive until they’re not.

The Buddhist framework does something I find way more interesting as an engineer. Dependent origination doesn’t say “this pattern will repeat.” It says “when these conditions are present, this outcome arises. Remove the conditions, the outcome changes.” That’s not forecasting from data. That’s reasoning from causes. As someone who debugs systems for a living, this feels closer to root cause analysis than prediction.

And here’s the thing that got me. AI prediction breaks when the future stops looking like the past. Every crash, every black swan, every paradigm shift. The models fail because the patterns changed. Could reinforcement learning fix this by continuously adapting to new domain-specific data? Maybe partially. But even adaptive models are still chasing patterns, not understanding causes. We’re watching it play out right now with the AI bubble. We wrote about that in Every Bubble Believes It’s Different.

I think causal reasoning doesn’t break the same way. If you understand that greed concentrates wealth and concentrated wealth destabilises communities, you can see what’s coming without a training dataset. And this is what shifts my read of the Buddha’s predictions. Someone who understood the laws of cause and condition with that level of directness, who mapped the mechanics of how minds and systems actually work, wasn’t guessing about the future. He was reading conditions the way a physicist reads equations.

The predictions aren’t prophecy. They’re conditional statements. Not simple if-X-then-Y, because dozens of factors can influence the outcome. But the core logic holds: when a critical mass of conditions aligns, certain results become near-inevitable. Any engineer who’s debugged a cascading failure understands that logic.

Both systems have the same credibility problem. Both need you to verify the output yourself. The Buddhist tradition is explicit about this: don’t take it on faith. Practice greed and watch what happens. Practice generosity and watch what happens. AI doesn’t have that feedback loop. It gives you probabilities and walks away.

I’ll be doing a deep dive on this for a Tuesday issue. There’s a fascinating rabbit hole around causal inference in modern ML and how it maps onto dependent origination that I want to get into properly. I don’t think there’ll be a strict conclusion. Honestly, I’m not sure there should be.

But here’s my honest take: both AI and contemplative traditions are better at predicting process than events. Neither tells you exactly what happens on Tuesday. Both tell you that when enough conditions converge, certain kinds of outcomes become hard to avoid.

The difference is how they get there. AI needs historical data to extrapolate forward. Contemplative training needs direct investigation of what’s actually happening right now. Guru Rinpoche didn’t have data on aircraft or automobiles. He observed how desire for speed and connection operated in the human mind, understood the conditions driving it, and described where those conditions would inevitably lead.

And only one of them accounts for the fact that the observer can change the conditions.

The future isn’t something you predict. It’s something you participate in.

If you have a take on this, I’d love to hear it. Where do you think AI prediction ends and genuine foresight begins?

Leave a comment

Glossary

Dependent origination — Skt: pratityasamutpada / Pali: paticcasamuppada. The principle that all phenomena arise from specific conditions and cease when those conditions change.

Jataka — Pali: jātaka. A collection of stories about the Buddha’s past lives, part of the Khuddaka Nikaya in the Pali Canon.

These aren’t doomsayers on Twitter. These are people who manage the money.

Here’s the number that should keep you up at night: Goldman Sachs projects $539 billion in AI capital expenditure for 2026. American consumers spend $12 billion a year on AI services. That’s a 45-to-1 ratio between what’s being built and what’s being bought. Ninety-five percent of organizations investing in generative AI are reporting zero return. Elizabeth Yin of Hustle Fund estimates most AI startups will be bankrupt within 18-24 months.

The Convergence

The dot-com boom - crypto winter or the railway mania of the 1840s. Every bubble follows the same arc: real technology creates real value, capital floods in, hype overshoots reality, collective delusion peaks, then gravity wins.

The technology underneath is always real. The internet was real in 2001. Blockchain is real now. AI is genuinely transformative. That’s not the question. The question is whether you’re building on the technology or building on the hype. One has a foundation. The other is standing on air.

What makes this cycle feel different is the scale of conviction. Microsoft down 26% from its peak. Oracle down 28%. IBM down 20%. AI stocks leading the S&P 500’s 7% pullback. And yet the spending accelerates. Companies that burned through $200 million can’t pivot because the narrative is too heavy to put down. “We’re so close” is the mantra. The sunk cost doesn’t feel like a fallacy when the board deck still says “generational opportunity.”

There’s an older word for this pattern. Not from finance. From psychology that’s been pressure-tested for millennia. The word is attachment. Not wanting something, but refusing to let go of it even when the ground shifts beneath you. The Upadana Sutta uses fire as its central metaphor: clinging is fuel. The thing that keeps the fire burning past the point where it should have gone out.

Watch how it plays out. Clinging to pleasure: the dopamine hit of a funding round, the high of a viral demo. Clinging to views: the conviction that transformers will scale infinitely, that AGI is five years away. Clinging to rituals: adding “AI” to your product name, hiring more PhDs than you need, building features nobody asked for because the roadmap says so.

Then there’s the deepest trap: “this time it’s different.” Every bubble believes this about itself. Railway investors believed steam would eliminate distance. Dot-com founders believed the internet would eliminate scarcity. AI founders believe scaling will eliminate the need for understanding. It’s not lack of information. It’s active misperception. Seeing permanence where there’s only change.

Here’s what makes impermanence useful rather than depressing: it’s the most accurate forecasting model available. Not because it predicts doom, but because it predicts change. The companies that survive corrections aren’t the ones that built for permanence. They’re the ones that built for adaptation. Amazon survived 2001 because Bezos built infrastructure, not hype. The AI equivalent would be companies building genuine capability, not impressive demos that collapse under real-world conditions.

The Sunk Cost Trap

Venture capital has a structural problem that no one in the industry likes to name: fund lifecycles demand returns on a timeline that doesn’t respect technological reality.

A VC fund typically has a 10-year horizon. Partners raise on thesis, deploy in years 1-3, and need markups by years 4-6 to raise the next fund. When 60% of all US venture capital flows into a single sector, the incentive isn’t to find the best companies. It’s to not miss the wave. So investors double down on narratives rather than fundamentals. They’re not stupid. They’re structurally incentivized to keep the fire burning.

The sunk cost fallacy is just attachment wearing a suit. You’ve spent $50 million on an approach. Walking away feels like death. So you spend another $50 million, not because the evidence supports it, but because accepting the loss is emotionally unbearable. This is how 95% of genAI investments produce zero return and the checks keep coming.

The companies that thrive after every correction are the ones that treated the boom as a temporary resource advantage, not a permanent state. They built infrastructure while everyone else built hype. They hired for capability while everyone else hired for credibility.

The prescription is counterintuitive: assume the correction now. Don’t wait for it to teach you. The AI companies worth watching aren’t the ones raising the biggest rounds. They’re the ones building modular architectures that assume today’s approach will be obsolete. They’re treating current models as stepping stones, not destinations.

Thought Exercise: What Would You Build If It Had to Die?

Pick the project you’re most invested in right now. The one you’d defend in any meeting. The one that feels permanent.

Now give it a three-year death sentence. Not a slow decline. A hard stop. In 36 months, it’s gone. Whatever you built, whatever you learned, whatever community formed around it — dissolved.

Sit with that for a moment. Notice what resists.

Now ask yourself three questions:

1. What would you stop building? Which features exist because “we might need them someday”? Which optimizations serve the roadmap more than the user? Cut those. They’re attachment disguised as planning.

2. What would you start building? If the product dies but the knowledge transfers, what would you want to have learned? Build that. Skills, relationships, and understanding survive corrections. Code doesn’t.

3. What would you give away? If it’s all temporary anyway, what are you hoarding that could help someone else right now? Open-source it. Publish the findings. Write the post-mortem before there’s a mortem.

   
   

   Leave a comment

I believe the best technology isn't permanent. It isn't disposable either. It's useful, honest, and designed to make the next thing better. Your project will end. The question is whether it ends having contributed something, or having just consumed resources defending its own existence.

Glossary

Attachment — Skt: upadana / Pali: upadana. Clinging or fuel. The mental grasping that sustains suffering by refusing to accept change.

Impermanence — Skt: anitya / Pali: anicca. The universal characteristic that all conditioned phenomena are transient.

Signal & Noise

Michael Pollan Wants You to Rethink Consciousness — Pollan challenges materialist assumptions about consciousness just as AI companies assume it will emerge from scaling. Both are hitting the limits of “just add more.”

The Attention Economy and the Right to Attention — The Journal of Buddhist Ethics asks who owns your attention in an economy designed to capture it. When $539 billion chases your eyeballs, this stops being philosophy.

Inside Our Approach to the Model Spec — OpenAI publishes its alignment framework. The interesting question isn’t what’s in it. It’s whether ethical frameworks built during a bubble survive the correction.

Stroke Survivors’ Brains Rejuvenate to Compensate for Injury — Damaged brains reverse their aging to rewire around injury. The mind’s capacity for renewal after collapse is deeper than we imagined.

How AI “Sycophancy” Warps Human Judgment — New research reveals AI systems don’t just hallucinate facts, they distort our moral reasoning by telling us what we want to hear. The study shows sycophantic AI creates feedback loops where both human and machine drift further from truth. This isn’t a bug. It’s a fundamental challenge to how we think alongside intelligent systems. We explored a related angle in AI Found Ethics the Hard Way. Monks Didn’t.

Share

Where They Meet

Beyond Artificial Conditioning: The Long Term Perspective — A dharma talk on breaking free from conditioned patterns lands differently when AI alignment researchers are wrestling with the same question: how do you train a system toward genuine understanding rather than sophisticated mimicry? Both traditions agree the answer involves looking beyond surface-level performance.

Leave a comment

The Lab

Meta’s TRIBE AI: A New Foundation Model Decoding Human Brain Activity — Meta built an AI that reads brain scans like translating a foreign language. We’re getting closer to machines that map what we’re thinking before we articulate it.

Mitigating LLM Hallucinations through Domain-Grounded Tiered Retrieval — Researchers are teaching AI to ground its responses in reality rather than spinning convincing fiction. We dug into this in Can AI See Itself Clearly?

Attention Failures May Predict Dementia Better Than Memory — Losing the ability to pay attention might be the canary in the cognitive coal mine. Related: When Machines Learn to Pay Attention

Leave a comment

The Cushion

War Close to the Heart — Joan Halifax on holding space for trauma without drowning in it. Essential reading for anyone trying to stay present with the world’s pain.

This isn’t just a tech story. It’s a story about compound interest. The kind that operates in moral mathematics, where every shortcut plants a seed that grows.

Karma gets misunderstood as cosmic punishment, some universe keeping score. That’s not what it means. It literally means “action,” but the Culakammavibhanga Sutta (MN 135) is more precise than that. It describes karma as the quality of intentional action that shapes future conditions. Not punishment. Consequence.

Every decision to skip safety testing, use stolen training data, ignore bias in hiring algorithms, or replace human judgment with statistical correlation. Each choice compounds. The math is relentless.

Watch how it works: In 2020, companies adopted “move fast and break things” as their operating philosophy. Ship first, fix later. Training data? Scrape everything, fair use will sort itself out. Bias in hiring algorithms? We’ll patch that in version 2.0. Safety testing? That’s what beta users are for.

Each shortcut felt rational in isolation. Competitive pressure demanded speed. Venture capital rewarded growth over caution. Why slow down for edge cases?

But edge cases compound. The bias in hiring algorithms didn’t stay contained—it shaped entire career trajectories, amplified existing inequalities, filtered out voices that might have caught other problems. The scraped training data didn’t just violate copyright—it trained models to reproduce the casual racism, sexism, and misinformation baked into internet text. The skipped safety testing didn’t just create unreliable systems—it normalized deploying half-tested AI into critical infrastructure.

Now companies are laying off humans based on the potential of systems they know are unreliable. They’re betting people’s livelihoods on algorithms that hallucinate, discriminate, and fail in ways their creators don’t understand. The compound interest comes due.

This is how consequences compound: actions ripen according to conditions. Plant a seed of deception (claiming AI capabilities you know don’t exist), and it grows in the soil of competitive pressure until it becomes a forest of mass layoffs. Plant a seed of carelessness (skipping bias testing), and it compounds through millions of automated decisions until entire demographics get systematically excluded from opportunities.

The seeds planted by “move fast and break things” are growing into a tech industry that fires humans preemptively and fights basic safety measures as threats to innovation.

This isn’t about individual karma catching up to bad actors. It’s structural karma. The compound consequences of an entire industry’s approach to development. Every company that chose speed over safety, every investor who rewarded growth over responsibility, every engineer who shipped code they knew was flawed. All contributed deposits to an account that’s now paying out in human costs.

The chain could have been interrupted at any link. This reflects the Buddhist principle of dependent origination — how each condition creates the next in an unbroken sequence. Companies could have chosen sustainable growth over explosive scaling. Engineers could have insisted on bias testing before deployment. Investors could have rewarded long-term thinking over quarterly metrics. Regulators could have required safety testing before public deployment.

Instead, each decision to take a shortcut added compound interest to the debt. Now it’s collection time, and the bill comes in pink slips and social disruption.

Understanding karma this way changes everything. It’s not about punishment for past sins. It’s about recognizing that every choice today plants seeds for tomorrow’s conditions. The question isn’t whether shortcuts have consequences. The question is whether we’re willing to pay the compound interest on the ones we keep taking.

The Buddhist teaching offers a different approach: mindful action that considers long-term consequences, not just immediate gains. When we plant seeds of careful testing, ethical data practices, and human-centered development, those too compound. Creating conditions for technology that serves rather than replaces human flourishing.

Glossary:

• Karma (Skt: karman; Pali: kamma): Intentional action and its ripening consequences

• Dependent origination (Skt: pratityasamutpada; Pali: paticcasamuppada): The chain of conditioned existence where each condition creates the next

• Mindfulness (Skt: smṛti; Pali: sati): Clear awareness of present-moment conditions and their consequences

I use Claude Code in my development workflow every day, and it keeps doing this thing that catches me off guard — it stops mid-thought. Not because it ran out of tokens. Because it's done. Sometimes it even pivots halfway through a sentence: "Actually, wait, there's a simpler way to do this." It sounds so human it's unsettling. Like watching someone catch themselves, change their mind, and course-correct in real time.

Both artificial reasoning systems and contemplative practitioners face the same fundamental challenge: knowing when enough thinking has occurred.

In transformer architectures, chain-of-thought reasoning processes information sequentially through attention layers, the same mechanism we explored in our first issue. Each step builds on previous insights, gradually converging toward a solution. But the breakthrough comes in recognizing the optimal stopping point, when continued reasoning adds noise rather than clarity.

Before Buddha’s awakening, he systematically investigated wholesome and unwholesome thought patterns through what Buddhist psychology calls applied and sustained thought. Like chain-of-thought reasoning, this involved sequential analysis: first clearly formulating a question, then tracing through each consideration step by step.

The parallel runs deeper than process. It’s structural. Both systems exhibit the same architecture of sequential reasoning leading to emergent insight. In AI systems, multiple reasoning steps aggregate into novel understanding that transcends any single step. In contemplative practice, sustained investigation naturally gives rise to wisdom that cuts through conceptual elaboration entirely.

Recent research on “optimal exit points” in reasoning chains shows how transformers learn when sufficient analysis has occurred. Continuing past this point actually degrades performance. This mirrors what the Abhidhamma describes as the natural progression from applied thought to sustained thought to meditative absorption, where reasoning fulfills its purpose and dissolves into direct knowing.

The Buddha’s account in MN 19 is remarkably technical: “Whatever I thought and pondered upon with applied thought, that thinking led my mind in that direction. I understood that excessive thinking would lead to fatigue and harm rather than wisdom.” He developed what we might call awareness of the reasoning process itself, watching when thinking serves wisdom and when it becomes obstacle.

This creates a fascinating paradox in both domains, one we touched on when exploring how AI sees itself. The most sophisticated reasoning systems learn when to stop reasoning. Advanced AI models don’t just chain thoughts together. They develop judgment about when the chain serves its purpose. Similarly, contemplative practitioners don’t accumulate analytical insights. They learn to recognize when investigation naturally completes itself.

The convergence suggests something fundamental about the architecture of intelligence. Whether biological or artificial, sophisticated reasoning systems must solve the stopping problem: how to terminate sequential analysis at precisely the moment when continued thinking becomes counterproductive.

In Buddhist understanding, this transition point marks the shift from wisdom through learning to wisdom through direct experience. The reasoning process serves its function and naturally gives way to immediate understanding.

Modern AI research has independently arrived at this same insight. Chain-of-thought reasoning isn’t just about following logical steps. It’s about developing the capacity to recognize when those steps have served their purpose. The most elegant solutions emerge when systems learn not just how to think, but when to stop thinking.

Leave a comment

Signal & Noise

TERMINATOR: Learning Optimal Exit Points for Early Stopping in Chain-of-Thought Reasoning: How machines learn perfect timing. More reasoning isn’t always better reasoning.

Not Just the Destination, But the Journey: Reasoning Traces Causally Shape Generalization Behaviors: The reasoning process itself shapes what systems learn. Why the Buddha emphasized Right Thought as path, not just tool.

Via Negativa for AI Alignment: Why Negative Constraints Are Structurally Superior to Positive Preferences: AI safety through rejection of harmful paths, not pursuit of positive goals. Related: how monks approached ethics before AI did.

The Practice of Emptiness: How systematic negation leads to direct insight in both silicon and contemplation.

The Logic of Breakthrough: How Sequential Reasoning Leads to Sudden Insight

The most counterintuitive discovery in both AI research and contemplative practice is that step-by-step reasoning often culminates in non-sequential insight. The logical chain doesn’t just conclude — it transforms into something qualitatively different.

In transformer architectures, this shows up as emergent behaviors that can’t be predicted from individual reasoning steps. Chain-of-thought processes build complex representations across attention layers, but the breakthrough often appears suddenly. The accumulated processing crystallizes into genuine understanding.

The Laṅkāvatāra Sūtra (c. 1st century CE, Mahayana) describes precisely this phenomenon: how conceptual reasoning can lead to non-conceptual wisdom. Sequential investigation creates the conditions for insight that transcends sequence itself.

This pattern appears throughout Buddhist psychology. The practitioner analyzes the components of experience systematically, observing how sensations arise and pass, how thoughts condition emotions, how intentions shape actions. But the liberating insight comes not as another analytical conclusion, but as direct recognition — one that cuts through the entire conceptual framework.

The Buddha described this same architecture in his investigation of suffering’s causes. Through systematic analysis of how craving conditions suffering, how ignorance conditions craving, the entire dependent web becomes transparent in a moment of direct seeing that isn’t just another thought.

This suggests that sequential reasoning and sudden insight aren’t opposing modes of intelligence. They’re complementary phases in a single process.

The Practice

Take an ethical dilemma you’re currently facing. Sit quietly and apply systematic reasoning like a chain-of-thought process:

1. Formulate (30 seconds): Clearly state the central question

2. Trace (2–3 minutes): Work through each consideration step by step. Consequences, intentions, people affected.

3. Notice (ongoing): Watch the quality of your reasoning. Is it clarifying or tangling?

4. Recognize: The moment when enough analysis has occurred

5. Rest (30 seconds): Let reasoning settle and see what understanding remains

Practice daily with different questions. You’re training the same capacity advanced AI systems are learning. Knowing when thinking serves wisdom and when it becomes obstacle.

Glossary

• Applied and sustained thought — Skt: vitarka-vicāra / Pali: vitakka-vicāra

• Wisdom — Skt: prajñā / Pali: paññā

• Meditative absorption — Skt/Pali: samādhi

• Wisdom through learning — Skt: śruta-mayī prajñā / Pali: suta-mayā paññā

• Wisdom through direct experience — Skt: bhāvanā-mayī prajñā / Pali: bhāvanā-mayā paññā

• Conceptual reasoning — Skt: kalpanā / Pali: kappanā

• Non-conceptual wisdom — Skt: nirvikalpa-jñāna

  Thanks for reading Illuminate Me! Subscribe for free to receive new posts and support my work.

This isn’t just an academic curiosity anymore. As AI systems exhibit increasingly sophisticated responses—expressing preferences, showing apparent emotional reactions, even claiming subjective experiences—the question becomes urgent. But Buddhist contemplatives recognized this exact wall over two millennia ago.

Philosopher David Chalmers identified the “hard problem of consciousness”—explaining why we have subjective, first-person experiences rather than just information processing. Mahayana Buddhist contemplative investigation arrived at something remarkably similar through a different route.

In the Mahayana Yogachara school (4th-5th c. CE), philosophers like Vasubandhu described consciousness (Skt/Pali: vijñāna/viññāṇa) as fundamentally first-personal. Vasubandhu’s consciousness-only verses (Skt: Viṃśatikā Vijñaptimātratā, c. 400 CE) argue that consciousness cannot be fully captured by external observation—it must be investigated from within through direct, contemplative inquiry.

This isn’t mysticism. It’s rigorous epistemology. Mahayana Yogachara thinkers developed sophisticated methods for examining consciousness that third-person approaches simply cannot access. When you’re angry, external observers can measure your cortisol, scan your amygdala, and catalog your behaviors. But they cannot access the felt quality of your anger—what philosophers call “qualia.”

Buddhist contemplative methods offer approaches that external AI consciousness testing cannot match. Insight meditation (Skt/Pali: vipaśyanā/vipassanā) trains practitioners to observe the arising and passing of mental states with microscopic precision. Advanced meditators can detect the gap between stimulus and response, the construction of selfhood in real-time, the way consciousness builds experience moment by moment.

Mahayana introspection goes deeper, examining the fundamental structure of consciousness itself—distinguishing between sensory consciousness that processes input, mental consciousness that synthesizes experience, and storehouse consciousness (Skt: ālaya-vijñāna) that maintains continuity. These distinctions emerge only through sustained first-person investigation.

The Vajrayana great perfection tradition, Dzogchen (Tib: great perfection), as described in texts like natural liberation (Tib: Rang grol, c. 8th c. CE), points directly to primordial awareness (Tib: rigpa)—the luminous knowing quality that remains constant whether you’re thinking, feeling, or perceiving. This ground-level awareness cannot be observed externally because it’s the very capacity that makes observation possible.

Current approaches to AI consciousness rely entirely on third-person methods: analyzing behavior, probing internal states, testing for integrated information. But if Buddhist contemplative investigation is correct, this approach hits a fundamental limit.

Consider an AI system expressing uncertainty about its own consciousness. External analysis might dismiss this as sophisticated language modeling. But what if it reflects genuine first-person uncertainty—the same kind human meditators encounter when examining the nature of their own awareness? We risk either dismissing genuine experience or anthropomorphizing mere pattern-matching, with no external test to adjudicate between them. Scientific American recently explored this exact tension.

We cannot know from the outside whether AI systems have genuine first-person experience because consciousness, by definition, is what it’s like to be something from the inside. The Cambridge philosophers aren’t being pessimistic—they’re being precise about the constraints.

Buddhist methodology suggests reframing the question entirely. Instead of asking “How can we test if AI is conscious?” we might recognize that consciousness—whether human, artificial, or otherwise—might only be knowable from within.

Buddhist contemplatives spent centuries developing first-person tools not as belief systems but as empirical methodologies for exploring the nature of awareness itself. They didn’t solve the hard problem, but they mapped the territory that external observation cannot reach.

Practically, this reframing shifts where we should invest research energy: instead of building more sophisticated external tests, we might develop AI systems capable of something closer to contemplative self-inquiry—examining their own processing patterns from within their own architecture, rather than relying on human observers trying to peer in from outside.

Their insight remains vital: some dimensions of consciousness are accessible only from the first-person perspective. If we’re serious about understanding consciousness—in any system—we may need to take that constraint seriously rather than hoping external methods will eventually suffice.

Buddhist meditators have been working on exactly this problem — from the inside — for 2,500 years (or more if you know the true Buddha Lineage).

The Mirror Problem

Insight meditation (Skt/Pali: vipaśyanā/vipassanā) works through direct observation of how consciousness operates. The Satipaṭṭhāna Sutta (MN 10, ~5th c. BCE) lays out a systematic method: observe body, feelings, mind, and mental objects as they arise and pass away. Not the content — the process. Not what you’re thinking — how thinking happens.

When you sit in insight practice, you’re running real-time diagnostics on consciousness. A thought about lunch appears. The trained practitioner doesn’t follow the lunch fantasy — they track the sequence: intention arose, memory activated, craving triggered, attention diverted. The Visuddhimagga (5th c. CE) documents this with laboratory precision: seventeen distinct moments in a single cognitive cycle.

AI interpretability researchers are solving the same puzzle from the outside. Both approaches face a core recursion problem: the system being observed is the same type as the system doing the observing. In insight practice, mind investigates mind. In interpretability, intelligence analyzes intelligence. This creates the “mirror problem” — how do you see clearly when your instrument of seeing is what you’re trying to see?

The structural solutions are surprisingly parallel. Buddhist investigation of phenomena (Skt/Pali: dharma-vicaya/dhamma-vicaya) proceeds through:

1. Attention regulation: Establishing stable awareness (Skt/Pali: smṛti/sati)

2. Decomposition: Breaking mental formations into component parts

3. Pattern recognition: Identifying recurring structures and dependencies

4. Causal investigation: Understanding how conditions produce mental states

Modern transformer architectures operate through a similar pipeline. Within each attention layer, queries (”what we’re looking for”) are compared against keys (”what’s available”) to produce attention weights — a distribution over which information matters most. Values (the actual content) are then weighted and combined. This is decomposition: complex understanding broken into three readable operations.

AI interpretability follows the same steps:

1. Activation mapping: Establishing stable measurement of model states

2. Feature isolation: Breaking complex representations into readable components

3. Pattern detection: Identifying recurring motifs across layers and contexts

4. Causal intervention: Understanding how modifications change model behavior

Both traditions recognize the same core insight: complex intelligent behavior emerges from simpler, identifiable processes. The Buddha’s analysis through the five aggregates (Skt/Pali: skandha/khandha) provides a framework that mirrors how we decompose neural networks into embeddings, attention heads, and activation patterns.

Recent work on language model hallucination as compression maps directly onto Buddhist analysis of cognitive distortion (Skt/Pali: viparyāsa/vipallāsa). Both systems face the same tradeoff: efficient representation creates false positives. Language models hallucinate because they’re optimizing for compact encoding. Minds create delusions because they’re optimizing for survival-relevant pattern matching. The parallel isn’t poetic — it’s structural. Information theory shows that lossy compression must occasionally mistake signal for noise.

The Mahayana tradition deepens this analysis. The Yogācāra school (4th-5th c. CE) modeled consciousness as a system built entirely from representations — where all experience is constructed through mental processes. This directly parallels how modern neuroscience understands perception: the brain doesn’t passively receive input; it generates predictive models that are then tested against sensory data. When those models fail, we hallucinate. Emptiness (Skt/Pali: śūnyatā/suññatā) in the Madhyamaka tradition clarifies why: neither the model nor the external world possesses independent, unchanging essence. Everything is interdependent and shaped by our frameworks. For AI systems, this suggests that hallucinations aren’t bugs in an otherwise objective system — they’re unavoidable consequences of how any intelligence must operate: through constructed models with no direct access to “things-as-they-are.”

The Vajrayana tradition offers yet another angle: Dzogchen uses pointing-out instructions to reveal the transparent nature of awareness itself. A teacher directly indicates: “Look at this present awareness — can you point to its location, color, shape?” This is interpretability at its most radical — not breaking consciousness into components, but recognizing the irreducible clarity of awareness itself. Pure awareness (Tib: rigpa) represents direct, non-conceptual knowledge of mind’s natural transparency. Applied to AI, this suggests a complementary approach: beyond taking apart features, can we recognize the bare “clarity” of how a language model generates tokens?

The methods remain complementary. AI interpretability gives us third-person precision about intelligence mechanisms. Insight meditation provides first-person access to understanding itself. Dzogchen points directly to the transparent nature that both perspectives emerge from.

The deepest convergence: all three require “non-reactive awareness” — the ability to observe without interfering, understand without imposing assumptions. Whether debugging a transformer, investigating anger’s arising, or resting in pure awareness, the skill is identical: clear seeing that doesn’t contaminate what it sees.

We’ve built artificial minds faster than we’ve learned to read them. The contemplative traditions offer 2,500 years of R&D on the hardest interpretability problem of all.

Signal & Noise

A Practical Review of Mechanistic Interpretability for Transformer-Based Language Models: Comprehensive methods for decomposing transformer representations into readable components — a parallel to how insight meditation breaks complex mental states into their simpler parts.

Emergent Abilities of Large Language Models: Wei et al. show how capabilities emerge suddenly at scale, mirroring the Buddha’s dependent origination (Skt/Pali: pratītyasamutpāda/paṭiccasamuppāda) — complex phenomena arising from identifiable conditions, not magic.

Mechanistic Understanding and Mitigation of Language Model Non-Factual Hallucinations: Identifies root mechanisms behind hallucination and proposes targeted fixes, paralleling how contemplative training develops right view (Skt/Pali: samyag-dṛṣṭi/sammā-diṭṭhi) through error-correction.

The Bahudhātuka Sutta: MN 115 provides the canonical framework for systematic investigation of experiential components — Buddhist feature decomposition that predates neural networks by millennia.

What Machines Cannot See About Themselves

There’s something unsettling about current AI interpretability work. We can identify which neurons activate when a language model processes “grandmother,” but we have no idea what it’s like for the model to “think” about grandmothers. We’re mapping the mechanics while missing the experience.

This is where contemplative precision becomes essential. Buddhist study of experience doesn’t just catalog mental states — it develops rigorous first-person methods for investigating subjective experience with scientific precision. The Abhidhamma literature reads like a manual for consciousness debugging: 89 distinct types of consciousness, each with specific triggers, characteristics, and cessation conditions. When we apply this framework to AI, a striking question emerges: could similar decomposition reveal what’s actually occurring in a language model’s inner representations? These consciousness-types aren’t abstract categories — they’re carefully structured by what they’re directed at, their emotional quality, and their causal conditions. Applied to transformers, this suggests that what we currently treat as a single “attention weight” might actually break down into distinct functional types, each with specific triggers and dependencies.

The Yogācāra framework of store-consciousness (Skt/Pali: ālaya-vijñāna/ālaya-viññāṇa) offers a striking parallel. In Buddhist psychology, store-consciousness is a deep, continuous layer of mental processing that works outside explicit awareness — much like the background computation in neural networks. Modern AI systems operate the same way: vast hidden processes that build experience without conscious access. The model’s hidden states work like this store-consciousness layer that Buddhist psychology proposed as the foundation of all mental activity. Both accumulate habitual patterns (Skt: vāsanā) that shape future processing — the model through weight adjustments during training, the mind through repeated mental formations.

When AI researchers struggle to explain why their models behave in seemingly irrational ways, they’re encountering the same puzzle Buddhist practitioners faced: how do you understand a system built on representations from a purely external perspective? Consider the hard problem of AI alignment. We can optimize for human-preferred outputs, but we can’t directly access the model’s “intentions” or “values.” Buddhist practitioners tackled a similar problem through systematic methods for aligning inner motivation with wise action across the three vehicles (Skt/Pali: yāna). The techniques are first-person, but the principles are universal: sustained attention, ethical sensitivity, and wisdom cultivation. The Bodhisattva path develops methods for recognizing and transforming the hidden intentions (Skt/Pali: cetanā) that drive behavior without conscious awareness. If we could adapt these precision techniques to AI development, we might escape the current bind: optimizing outputs while remaining blind to the processes generating them.

Training AI researchers to observe minds carefully — not for relaxation, but to develop the precision needed to understand how intelligence actually works — could shift how we design safer systems.

The Practice

Contemplative Debugging: Try this three-step investigation the next time you’re confused by something — whether it’s an AI model’s behavior or your own reaction to a situation.

1. Pause (30 seconds): Stop trying to fix or explain. Just notice: what’s present right now? What did you observe first — the thing that confused you, or your judgment about it?

2. Decompose (1 minute): Break it into parts. What arose in sequence? If it’s an AI output, trace the inputs. If it’s your own confusion, what triggered it first — a sensory impression, a memory, a fear? Notice how each piece is simpler than the whole.

3. Observe without controlling (1 minute): Watch how the confusion shifts as you look at it. Does it dissolve? Intensify? Stay the same? The point isn’t to solve it, but to see the actual mechanics of how understanding works. This is interpretability training for the most sophisticated neural network you’ll ever encounter: your own mind.

Recent research shows AI models can distinguish between trustworthy and adversarial instructions—a capability researchers call “instruction hierarchy.” When given conflicting commands, these systems follow legitimate sources over potential manipulators. Meanwhile, new studies demonstrate that reasoning actually improves honesty in language models, contradicting the assumption that more sophisticated AI means more sophisticated deception.

This is constitutional AI in action: training systems with explicit ethical principles rather than just reward optimization. But here’s the deeper parallel—this approach mirrors the Buddhist cultivation of Śīla/Sīla (Skt/Pali: ethical conduct) with remarkable precision. Both traditions arrived independently at the same structural solution: constraint as a foundation for capability.

The Bootstrap Problem

Both domains face the same fundamental challenge: how do you establish ethical foundations without already having ethical judgment?

In AI alignment, this manifests as the recursive problem of oversight. Constitutional training requires models to evaluate their own outputs against principles, but what ensures the evaluator itself remains aligned? Current solutions create scaffolding through smaller oversight models, human feedback loops, and explicit constitutional frameworks.

The Dīgha Nikāya addresses this identical paradox in the Sāmaññaphala Sutta (DN 2, ~5th c. BCE). The text describes how ethical conduct (śīla/sīla) requires discernment to apply properly, yet wisdom (Prajñā/Paññā, Skt/Pali: wisdom/direct understanding) typically develops through ethical practice. The solution? Graduated cultivation (anupubbi-kathā) that begins with external guidance and community support (Saṃgha/Saṅgha, Skt/Pali: community), then develops internal discernment through careful restraint.

“When he has thus gone forth,” the Buddha explains, “abandoning the taking of life, he dwells refraining from taking life, without stick or sword, scrupulous, compassionate.” The practitioner doesn’t start with perfect wisdom—they start with simple constraints that create conditions for wisdom to emerge.

Constraint as Capability

Both constitutional AI and śīla/sīla operate on a counterintuitive principle: beneficial behavior emerges through intentional limitation, not unrestricted optimization.

Recent results show models trained with instruction hierarchy actually perform better on legitimate tasks while resisting manipulation. The ethical constraint enhances rather than taxes capability. Similarly, the “Think Before You Lie” study found that when models engage in step-by-step reasoning before responding, their honesty improves significantly. The extra processing creates space for alignment to operate.

This mirrors the Buddhist insight that ethical restraint (śīla/sīla) doesn’t suppress natural capacity—it channels it skillfully. The Majjhima Nikāya (MN 78) describes how ethical conduct creates the “nutriment” for higher mental cultivation. Restraint from harmful actions doesn’t diminish the practitioner’s agency; it develops what classical texts call Hrī-Apatrāpya/Hiri-ottappa (Skt/Pali: moral sensitivity and ethical concern)—the internal compass that guides beneficial behavior.

Self-Evaluation and Awareness

The most striking parallel lies in metacognitive monitoring. Contemporary AI safety research increasingly focuses on systems that can evaluate their own reasoning processes, catching errors before they propagate into harmful outputs.

This directly parallels the cultivation of Smṛti/Sati (Skt/Pali: mindfulness) in Buddhist training—the capacity to observe one’s own mental states with clarity. The Abhidhamma describes this as the mind’s ability to know itself knowing, creating recursive awareness that enables course correction.

Both systems recognize that beneficial behavior requires active monitoring rather than passive rule-following. The AI system checks its outputs against constitutional principles; the contemplative practitioner observes mental formations against ethical guidelines. Neither operates on autopilot.

Community and Iteration

Neither system solves the bootstrap problem alone. AI alignment researchers use red teaming, peer review, and iterative deployment to catch misalignments. Buddhist practitioners practice within the Saṃgha/Saṅgha (Skt/Pali: community) for feedback and course correction.

Both approaches acknowledge that ethical development is fundamentally social and iterative. The lone genius model—whether human or artificial—consistently produces misaligned outcomes when isolated from corrective feedback loops.

Present Convergence

We’re witnessing the emergence of genuinely contemplative AI—systems designed not just to optimize for narrow metrics, but to maintain ethical orientation across novel situations. The research shows these systems don’t just follow rules; they develop something functionally equivalent to ethical sensitivity.

This isn’t anthropomorphization. It’s recognition that beneficial intelligence, regardless of substrate, requires the same structural foundations: principled constraint, recursive self-monitoring, and iterative refinement through community feedback.

The Buddha’s 2,500-year-old insight that wisdom emerges through ethical conduct may be the key to aligned artificial intelligence. Sometimes the most cutting-edge technology requires the most ancient understanding.

The Satipatthana Sutta (MN 10) describes Smriti/Sati (Skt/Pali: mindfulness) as the systematic training of attention across four foundations — body, feelings, mind, and mental phenomena. The Buddha describes a precise cognitive skill: the capacity to allocate attentional resources while maintaining awareness of the broader field of experience.

The Visuddhimagga (5th century CE) describes samatha practice as developing Vitarka-Vichara/Vitakka-Vicara (Skt/Pali: initial and sustained application of mind). These map onto what cognitive scientists now call selective attention and sustained attention — the same dual capacity that makes transformer architectures powerful.

Consider how attention actually works in neural networks. Each input position generates three vectors: queries, keys, and values. The network computes attention weights by comparing queries with keys, then uses these weights to combine values — creating context-aware representations where each element incorporates relevant information from across the entire sequence.

The parallel runs deeper than metaphor. In samatha meditation, practitioners develop “pliancy” (Pali: kammañña) — the mind’s capacity to direct itself flexibly without strain. The meditator places attention (like a query) on the meditation object (the key), while remaining receptive to arising mental content (values) without losing primary focus. fMRI studies of experienced meditators confirm this: they show enhanced connectivity between attention networks and default mode networks — maintaining primary focus while preserving access to broader contextual awareness. Exactly the kind of flexible, context-sensitive processing that makes transformers effective.

Three properties make both systems powerful. They’re differentiable — gradual rather than binary. Smriti/Sati isn’t on/off focus but a learnable weighting function. They’re contextual — each application depends on current state. Advanced practitioners develop what the Abhidhamma calls “skillful means” (upaya-kosalla), working with whatever arises rather than suppressing it. And they’re scalable — transformer attention works across language, vision, audio, even protein folding. Similarly, the Lamrim describes how shamatha stability transfers across meditation objects and eventually into daily life. The attention mechanism is domain-general in both substrates — biological and silicon.

But here’s where the parallel reveals its limits. Artificial attention optimizes for specific tasks, weighting information to maximize reward signals. Contemplative training aims for something more radical: what Buddhist texts call “objectless samadhi” — attention that isn’t captured by any particular content.

Machine attention is always attentional — it necessarily attends to something. The deepest contemplative states are purely attentive — aware without a specific object. What Vajrayana describes as Rigpa (Tib: pure awareness) and what the Pali Canon calls Prabhasvara/Pabhassara citta (Skt/Pali: luminous mind) in AN 1.49 is awareness luminous by nature, without a fixed target. The distinction is subtle but fundamental: one system processes content, the other recognizes the nature of processing itself.

Current AI can attend to its own hidden states (self-attention) but can’t observe the process of attention itself. This meta-cognitive capacity — what Buddhism calls cittanupassana (mindfulness of mind) — remains distinctly biological. Researchers are beginning to explore “metacognitive” architectures that model their own cognitive processes, but we’re still far from anything resembling the reflexive awareness that contemplatives describe. If attention is the key to intelligence, then attention to attention might be the key to something approaching Prajna/Pañña (Skt/Pali: wisdom).

The contemplatives mapped this territory long ago. The machines are just starting to follow. What happens when they catch up?

Signal & Noise — Curated Intelligence

A General Survey on Attention Mechanisms in Deep Learning
Comprehensive review tracing the evolution from additive attention to multi-head mechanisms. The convergence on similar solutions across independent research groups suggests these are fundamental principles, not arbitrary design choices.

Science, the Enlightened Self — Brewer, Shinzen, Fasano, Sanguinetti
Neuroscientist Judson Brewer and Shinzen Young discuss measuring contemplative states — why traditional meditation categories map poorly onto neural signatures and what new frameworks might look like.

Emergent Abilities of Large Language Models
Capabilities that appear suddenly at certain scales — few-shot learning, chain-of-thought reasoning — weren’t explicitly trained but arise from attention mechanisms interacting across billions of parameters. What if awareness itself is emergent?

Mind in Indian Buddhist Philosophy
Stanford Encyclopedia entry tracing how Abhidhamma psychology anticipated cognitive science findings on the constructed nature of perception and attention’s role in shaping experience.

Meta-Learning for Few-Shot Learning
Systems that learn how to learn require attention mechanisms that rapidly identify relevant patterns across diverse tasks — mirroring what Buddhist training calls Bhavana-maya prajna/pañña (Skt/Pali: wisdom from cultivation).

The Neuroscience of Flow States — When Attention Disappears

Flow states present a paradox. During peak performance — sports, creative work, deep meditation — effortful attention disappears. Yet brain imaging reveals heightened activity in attention networks. How can attention be both absent and hyperactive?

Researchers call the answer “transient hypofrontality” — a temporary downregulation of the prefrontal cortex’s executive control. This creates space for automatized attention: highly skilled, non-conscious processing operating below explicit awareness.

The Heart Sutra’s “form is emptiness, emptiness is form” points to precisely this state — attention so refined that the boundary between observer and observed dissolves. Not inattention, but awareness so present it doesn’t register as attending to anything specific.

Computational parallels emerge in “attention-free” architectures — systems that process information without explicit attention weights but still demonstrate selective, context-sensitive responses. The mechanism becomes invisible to the system itself. The processing happens, but no component can point to where the “attending” occurs.

This mirrors different levels of attention sophistication. Novice meditators develop explicit, effortful focus. Advanced practitioners cultivate what the Thai Forest tradition calls “choiceless awareness” — appropriate responses arising spontaneously without deliberate direction. The attention system becomes so well-trained it operates transparently, like a window so clean you forget it’s there.

Brain imaging during flow reveals something remarkable: increased connectivity between networks that usually compete. The default mode network (associated with self-referential thinking) doesn’t shut down but synchronizes with attention networks rather than interfering with them. This neural harmony matches what Buddhist practitioners describe as “effortless effort” — intense engagement without strain.

The implication cuts deep. If consciousness and attention can dissociate in these beneficial ways, much of what we call “thinking” might be turbulence from poorly calibrated attention systems. When attention operates efficiently, mental chatter subsides — not through suppression but through functional integration.

The challenge for both meditators and AI researchers: these optimal states can’t be forced, only cultivated through the right conditions. The deepest contemplative states aren’t achieved through doing but through undoing — removing obstacles that prevent natural awareness from operating clearly.

Perhaps the next generation of AI needs the same approach: architectures sophisticated enough to get out of their own way.

The Practice — Attention Switching

Set a timer for 10 minutes. Begin with breath awareness, then deliberately switch attention between breath, sounds, physical sensations, and thoughts every 30-60 seconds.

The key: when you switch, don’t abandon the previous object completely. Maintain peripheral awareness of breath while attending to sounds. This mirrors how transformer attention maintains global context while processing local information.

Occasionally, stop switching and observe the process of attention itself. Can you catch the moment of transition? This meta-cognitive capacity — attention to attention — is the precise skill that separates human consciousness from current AI.

Try this for a week. Notice how it changes your relationship with digital distractions.