How reliable are our models today? At 50% reliability, frontier models complete 5-hour METR tasks. At 90%, they collapse to just 10 minutes.

If you do some back-of-the-envelope math, 50% → 99% reliability unlocks 3 orders of economic magnitude: O($10Bs) → O($10Ts). Different research agendas are the bottleneck. Full breakdown below.

Imagine you have a coworker named Alex. Impressive resume. Holds his own in technical discussions. Occasionally drops insights that make you think, “Huh, never considered that.”

But here’s the thing: Alex makes up random stuff 30% of the time.

Not lying, exactly. Alex genuinely believes what he’s saying, he’s just confidently, eloquently wrong. He’ll swear Q3 revenue is up 40% (it’s down 15%). He’ll insist the Johnson account closed last month (still in negotiation). He’ll draft customer emails with completely fabricated product features.

What would you do with Alex?

You’d fire him, obviously! It doesn’t matter how smart Alex seems, nobody can work with someone who makes up stuff 30% of the time.

Here’s the uncomfortable question: If your AI were a person, would you hire them? Not just “are they impressive in interviews.” Would you trust them with your customers? Your team? Your reputation?

If your AI were Alex — seemingly smart, occasionally brilliant but completely bullshitting you 30% of the time — you’d fire them next week.1

Today's AI valuations only make sense if AI replaces vast swaths of everyday work. So why are we deploying models that fail 30-70% of the time and then wondering why there was no ROI?

Welcome to the AI reliability crisis — and the reason R-AGI (Reliable Artificial General Intelligence) will matter more than AGI.

For years, we’ve been asking “When will AI be generally intelligent?” Wrong question. Here’s the question that unlocks trillions: “When will AI be generally reliable?”

Reliability Matters

Folks in AI are continually surprised that study after study shows minimal economic impact. “But it’s magic!” they protest, showing you another demo that nails some seemingly impossible task.

Folks outside AI are continually surprised by the numbers. “It’s a bubble!” they declare, watching billions pour into technology that doesn’t move revenue.

Both sides are right. And both sides are missing the point. As Dwarkesh Patel put it a couple weeks ago:

Sometimes people will say that the reason that AIs aren’t more widely deployed across firms […] is that technology takes a long time to diffuse. I think this is cope.

The gap is explainable by one simple concept: reliability. There are lots of fancy technical definitions of reliability, but here’s my favorite one:

Roughly, accuracy is hitting the bullseye on average but reliability is hitting the same spot repeatably. You can be unreliably accurate (centered on average but inconsistent), reliably inaccurate (consistent but off-center) or any of the other combinations. The AI industry has been optimizing for accuracy. Economics demands reliability.

At Opendoor, we ran one of the largest-scale real-world AI systems on the planet: spatiotemporal world models pricing hundreds of billions of dollars of homes every year. Everyone there learns a brutal lesson fast: a model can be 97% accurate but only 70% reliable — and the gap can bankrupt you.

Consider two AI pricing models, both with 3% MAPE (mean absolute percentage error):

• Model I: Usually amazing, occasionally very wrong (prices a house for ±25% of market)

• Model II: Consistently “fine” (off by 2-3%)

Most people pick model I. It feels better. The median performance is amazing! The demos are incredible! Then you deploy it… and adverse selection eats you alive. You overpay for bad houses and underpay for good ones, the good ones are insulted (“WTF, how dare you”) and the bad ones are happy to offload their lemons (“um, ok, sure I’ll take an extra $100K”). Congratulations, you just picked yourself off into bankruptcy.

(Fun fact: This is exactly what bankrupted Zillow’s iBuyer program. They bet on mean accuracy. Their tails swung every which way — over time, across neighborhoods, between home conditions — and killed them.)

On the other hand: Model II doesn’t feel as good, but it lets you build a business. You can plan against consistent 2-3% error. You know exactly what guardrails to add. You know what bias terms to adjust. You know which edge cases need human review. The errors are small enough that they don’t trigger adverse selection: sellers of all kinds stay in the pool. Put another way, engineerable consistency beats unpredictable brilliance because you can design systems around it. It’s the difference between a solid collaborator you can trust and a schizophrenic genius who might ruin you.

This isn't just about AI home pricing. Consider Project Vend, Anthropic’s newest eval for agentic AI. Phase 1's AI literally tried calling the FBI over a $2 fee. Phase 2 upgraded to newer models and improved at basics — yet still agreed to illegal onion futures contracts, offered below-minimum-wage security jobs, and fell for imposter CEO schemes. The AI and its AI-CEO spent entire nights rhapsodizing about “eternal transcendence.” As Anthropic concluded: “The gap between ‘capable’ and ‘completely robust’ remains wide.” A-player capability, intern-level dependability.

Here’s the thing about modern AI: We’ve obsessed over mean accuracies when tail reliabilities are what actually matter in the real world. 96% of lm_eval metrics measure average performance.2 At the other end of the lifecycle, virtually all pretraining uses cross-entropy loss — what that masks is that it’s actually mean cross-entropy loss. A model can become SOTA by being slightly better everywhere while failing on the tail cases that represent real-world reliability.

And, just like Project Vend, it gets worse when you add agentic loops.

We talk breathlessly about AI’s ability to handle exponentially longer tasks. “Look, it can do 5-hour METR tasks!” (Opus 4.5 METR results came out over the weekend.) But, if you read the fine print, METR shows median success by default. What happens when you need >50% reliability?

At 80% reliability, 5 hours of agentic task completion collapses to 29 minutes. And at 90%, it’s 10 minutes. At 99%, it’s less than 1 minute.3

Conversely, here’s some middle-school math: If your model is 90% reliable per step, after 100 steps you’re at 0% total reliability. That’s not a rounding error. That’s “demo magic, production tragic.”

Check out the tweets I’ve added throughout. They use different words, but they’re all making the same underlying observation about reliability. (There are lots of serious stories, but we all need a moment of levity so I’ll just stick to a funny few.)

Here’s the brutal truth: Being smart on average isn’t good enough. If you’re smart but 30% of the time you just make random shit up, you get fired. Period.

You have to be reliably smart. Alas, today’s models are unreliably smart.

9s Unlock Economic Outcomes

If you buy my basic thesis that AI reliability matters, here’s the follow-on: different 9s of reliability unlock different trillions of economic value.

At Amazon in the early 2000s, my team (personalization) ran AI systems handling 100,000+ queries/second. Alongside Google and Meta, we learned what Internet-scale reliability really meant — more importantly, we learned different tasks demanded different reliability SLAs. Order systems needed 5 9s (99.999% uptime). High-traffic product pages and features needed 3-4 9s (99.99%). Editorial content? 2 9s (99%) was fine. The more valuable your task, the more reliable your systems needed to be.

Depending on task and context, today’s AI models are about 30-70% reliable. And though you’d fire poor Alex at 70% reliability, tomorrow’s trillion-dollar IPOs (and trillion-dollar data center buildouts and, hell, even today’s basic economic engine) depend on them replacing vast swaths of everyday work.

Borrowing from the SREs who taught us about uptime reliability, let’s work backwards from the humans doing that work today. What 9s of reliability do we actually expect from our co-workers?4

Think about it. You don’t hire the same way for every role. A fresh intern gets different responsibilities than your fellow A-player teammate than your battle-tested VP. It’s not really about intelligence — an intern might be brilliant — but about proven reliability. You’ve watched your VP make good decisions for three years straight. The intern? Still TBD.

The same logic applies to AI deployment. Each additional 9 unlocks orders of magnitude more economic value because it unlocks entirely different classes of work we’re willing to trust it with.

Right now, most AI is stuck between 30-70% reliability. This is why it feels simultaneously magical (“wow, it drafted my email!”) and worthless (“...but then insulted my boss”). We’re trying to build a $10T AI economy with technology that’s reliable enough for maybe $10B of use cases.

The question isn’t if AI can do something; the question is if you trust it to.

Research Unlocks 9s

Now, a final leap: if you buy that different 9s of reliability unlock different economic outcomes, I posit that different research agendas unlock different 9s of reliability.

To get from 70% reliability to 90% (or 99% or 99.9%) reliability, what actually needs to happen? I claim we need three things: (1) better systems, (2) better inference-time adaptability, and (3) better training-time fundamentals. Let’s work it backwards, here’s my quick-and-dirty Fermi decomposition of how reliability manifests for LLMs:

Now, I know what you’re thinking: “Great, another researcher with a made-up formula.” Fair. But, humor me for a second because this decomposition tells us where to invest R&D dollars.

Let’s unpack where the reliability gradients are largest and what research agendas move which needle:

Compositional Reliability: 99% Systems from 90% Models

Remember that middle-school math? 0.9^100 = 0% reliability. Episode length kills exponentially.

What can you do? Instead of trying to make one monolithic model hit 99.9% per-step reliability (good luck), break tasks into verifiable chunks: sequential decomposition via Tree of Thoughts, hierarchical decomposition via Least-to-Most prompting. Ground facts in RAG instead of fuzzy parameters. Use tools for math and logic (DSPy makes this programmable).

Plus, look at that max() term. You can win with either a strong verifier or strong generator, but verification is often easier than generation — think P vs. NP vibes. You can build reliable systems from unreliable components if you verify intelligently.

The key insight: systems amplify reliability. You don’t need a perfect model — you need a reliable system.

Adaptive Reliability: Inference-Time Tail Management

Models don’t know when they’re going astray. This is the Alex problem: he’s confidently wrong because he can’t detect when he’s in over his head.

So, how do you solve AI Dunning-Krueger? OpenAI’s process supervision catches errors mid-reasoning before they compound. Google’s self-consistency flags when models disagree with themselves — disagreement signals tail risk. Conformal prediction provides calibrated confidence. Constrained decoding limits yourself to valid subspaces. Selective abstention says “I don’t know” on OOD inputs.

This matters because if you can’t detect when you’re unreliable, you can’t build guardrails. Confident wrongness is worse than admitted uncertainty.

The key insight: detection enables protection. Self-aware AI beats confident AI.

Intrinsic Reliability: Training-Time Tail Awareness

See that auto-regressive noise term? It’s about the signal to noise ratio. Managing the second moment (noise) has a quadratic impact on downstream reliability.

Historically, we’ve almost exclusively focused on average scores (first moment). What percentage of evals do you pass? What’s your MMLU score? Your Arena Elo? This is what killed Zillow: optimizing mean accuracy while the tails swung wildly.

A few forward-thinking folks have been looking ahead. Percy Liang’s WILDS benchmark was a wake-up call for evals, testing models on in-the-wild distribution shifts, revealing fragility that MMLU can’t see. FAIR’s Focal Loss showed you can reshape training losses to down-weight easy examples and hammer the hard ones. Ethan Perez’s Red Teaming LMs with LMs demonstrated you can use models to automatically probe each other’s failure modes at scale, while Redwood’s adversarial training for high-stakes reliability showed you can systematically harden systems against catastrophic failures.

We took a different angle. You can use a 70-year-old (Nobel Prize-winning) theory to reduce hallucinations by ≥25% at just +0.004% additional pretraining, if you’re simply willing to re-think the typical framing of hallucinations. And the math isn’t just elegant, it’s empirically tight: our theoretical framework explains 94% of observed reliability variation across configurations. All in, it suggests a third scaling axis beyond parameters and data for reliability: not “how big?” (parameters) or “how much?” (data) but “how different?” (neural diversity).

The key insight: prevention beats detection. Bake in reliability, don’t bolt it on.

To date, the overwhelming majority of AI research has chased means: bigger models, more data, better benchmarks. Reliability has been an afterthought. And the few who have thought about it have mostly focused on systems (compositional architectures, RAG, tool use) and inference-time adaptations (process supervision, constrained decoding, selective abstention).

Training-time reliability is the under-explored frontier. Look back at that formula: Compositional Reliability attacks the exponent (episode length), Adaptive Reliability improves the max() term (verifier capability), and Intrinsic Reliability reduces the denominator (auto-regressive noise). All three matter. But only one compounds before you ship and that’s the one getting the least attention.

The companies that figure this out — those that shift from purely chasing capability to engineering reliability — will be the ones that actually capture the economic value everyone’s been promising. Here’s a diagnostic: if your AI roadmap is 90% capability R&D and 10% reliability R&D, you’ve got it wrong. The $10B → $10T jump isn’t about smarter models — it’s about trustworthy ones.

The rest will keep wondering why their magic demos don’t turn into revenue.

Conclusion

Different use cases — entire industries — don’t gradually open up. They unlock at specific 9s of reliability.

The first wave of generative AI was about discovering what’s possible. It was — and is — magical. But while it’s been technically fascinating, it hasn’t yet proven commercially valuable at scale yet.

The next wave will be about making it trustworthy. We’ve spent years racing toward Artificial General Intelligence. The real prize is Reliable Artificial General Intelligence (R-AGI). This’ll be both technically fascinating and commercially valuable.

The question for leaders is no longer “What can AI do?” but “What can I trust it to do?” The answer — and the trillion-dollar AI economies we’re all depending on — will be written in 9s.

Translation: Double your GPUs? You need 40% more data or you're just lighting cash on fire. But there's no 2nd Internet (we’ve already eaten the first one). The path forward: data alchemists (high-variance, 300% lottery ticket) or model architects (20-30% steady gains), not chip buyers. Full analysis below.

The Gospel of Scale

For almost a decade, the most important essay in AI has been Rich Sutton’s Bitter Lesson. And for years, I think we’ve profoundly misunderstood it. Talking to dozens of researchers at top labs over the past year, I think this might be the single-most common (and single-most dangerous) fallacy in AI today.

The lesson, as we all learned it, was a beautifully reductionist gospel: stop trying to be clever. Here’s Sutton’s opening sentence:

The biggest lesson that can be read from 70 years of AI research is that general methods that leverage computation are ultimately the most effective, and by a large margin.

— Rich Sutton

It taught us that our elegant, hand-crafted features and clever inductive biases were ultimately just temporary crutches. It told us our ingenuity was a bug, not a feature. That the true path to AGI wasn’t more sophisticated algorithms, but "general methods that leverage computation."

And the gospel worked. For years, the only slide that mattered was the one with the loss curve going down and to the right on a log-log plot. We became scaling maximalists, and our faith was rewarded.

I was there. I lived this transition — twice. Before the Bitter Lesson even had a name, I cut my teeth designing and training large-scale multi-sequence HMMs, surfing the paradigm shift from hand-curated models built on thousands of tokens to generalized training on billions of them. And in the last several years, from training statistical models on billions of tokens to training world models1 and early agentic systems on trillions of tokens. We were right to listen.

But we were wrong about what we heard.

The Heresy We All Missed

The Bitter Lesson lays out the spirit of the laws, but the Scaling Laws specify their letter. And the letter, in the form of Chinchilla, revealed that the god of Compute we worship is actually yoked to an even greater one — the god of Data.

Here’s the quarter-sentence from Sutton’s original text we all latched onto:

general methods that leverage computation are ultimately the most effective

For years, we read that and concluded the primary bottleneck was access to computation. We begged, borrowed, and stole our way into multi-billion dollar deals for more TPUs and GPUs. But the scaling laws from DeepMind and OpenAI gave us the fine print on those contracts. They showed that, for a given train-time compute budget (C), there was an optimal size of model (N) and training dataset (D).

But that frames the question we face backwards.

If you invert the logic, something interesting happens. Simplifying for clarity, the usual approximation of the Scaling Laws is that compute scales as the product of model parameters and training tokens (C ~ 6 N⋅D). Chinchilla showed that, for a given compute budget, model size N should scale roughly with data size D (N ~ D).

If you do some basic algebra, you’re left with a starker, more profound relationship:

Simply put, your train-time compute budget scales quadratically with the size of your dataset.2 Conversely, if you double your GPUs, you also have to increase your data by ~40% or you’re just lighting cash on fire.

Let that sink in. This is the real lesson, the one we all missed.

It was never about compute for its own sake.

The heroic feats of computation — whether to radically increase memory, parallelize across thousands of GPUs or invent entirely new silicon just to speed up matmul — were trailing indicators.

Instead, it was always about having a dataset so massive that it enabled a model so massive that together they demanded that level of compute that massive.

The Great Data Bottleneck

And herein lies the problem — we’ve basically ingested the entire Internet, and there is no second Internet. For language models, there’s no obvious way to scale the data.

A study from Epoch AI unpacks things a bit more. In short, while the total sum of human-generated data is vast (~100T to ~1000T tokens), the stock of high-quality text and code suitable for training is 2-3 orders of magnitude smaller. After filtering for quality and duplicates, you’re staring at a pool of ~10T useful pre-training tokens.3

We can use specialized datasets for mid-/post-training, but we’re even reaching the end of that. GitHub is big, but it’s not infinite: we’re strip-mining these resources and the veins of high-quality data for bug-fixes (let alone longer tasks) are running dry. For GPT-6, when you think about METR curves going to 4h and 8h, that maybe you need another O(100Bs) parameters for that, each requiring ~20 tokens per parameter, where do you think that data is coming from?

This is the bottleneck. Without more data, you can’t increase model parameter sizes and without either increasing computation doesn’t make sense. In the wake of GPT5, there’s been a lot of chatter about the Bitter Lesson plateauing. There’s nothing wrong with the Bitter Lesson. At this particular point in time, at this point in data, it’s the central assumption that most people make about The Bitter Lesson is holding us back.

A Longitudinal Prophecy, a Cross-Sectional Fallacy

The Bitter Lesson is a prophecy about the long arc of history, but many misread it as a battle plan for today.

Talking to dozens of AI researchers and leaders over the past year, this misinterpretation is maybe the single-most common (and single-most dangerous) fallacy in AI today.

Fundamentally, Sutton’s lesson is a longitudinal lesson. It correctly observes that, when you zoom out and view progress across decades, simple, general methods that utilize increasing data and compute win out over clever, specialized methods. Over the past decades, it’s been proven true again and again.

But the Bitter Lesson is not a cross-sectional statement. At any fixed point in time, for any fixed amount of data, the Bitter Lesson makes no claim. While incisive, Sutter’s observation is an asymptotic one. And if you’re in charge of an AI team today, asymptopia isn’t useful — you need to survive long enough to capture it, and the markets (and VCs) are fickle, impatient masters.

Think of it like this:

Being grossly reductionist, progress in AI is driven by two fundamental forces:

1. More Data: As you move (longitudinally) to the right, more (good) data allows you to scale up models, capture novel emergent behaviors and create the magic we’ve all seen over the past few years. In this graph, it looks like a jaunt to the right and up. In a nutshell, the Bitter Lesson.

2. Better Architecture: But, at a fixed (cross-sectional) point in data, you can also go up by improving the model architecture. Better architectures can enable core capabilities, like Differential Transformer enabling sparser/de-noised attention; enable novel reasoning paradigms, like HRM’s multiple reasoning time-scales; and even multiply usable data, like CLIP basically doubling data by learning from both images and text simultaneously by exploiting spatial proximity.

Translation: If you can’t go to the right (because there’s no more data), there’s still a way up (via architecture). That’s where the Architect comes in.

A Crossroads: The Architect vs. the Alchemist

Combined, the Bitter Lesson and Scaling Laws teach us that more data trumps all else. But with training data for language models now functionally fixed, today’s battle landscape changes. We must either wring more intelligence from every bit (the Architect) or conjure new bits from the void (the Alchemist).

Path #1: An Architectural Renaissance

The Architect's mandate is to discover and embed superior inductive biases into the model’s functional form. The results of this search are redefining the state-of-the-art:

• Breaking the O(L²) Bottleneck: Standard Transformers require all-to-all attention, a quadratic bottleneck. The shift to structured State-Space Models (like Mamba) uses linear-time recurrence and selective state compression to handle vastly longer contexts with greater efficiency

• Unlocking Deeper Reasoning. Instead of homogenous layers, architectures like Hierarchical Reasoning Model (HRM) from Sapient use conditional computation to route inputs through cheap, shallow paths for simple tasks or expensive, deep paths for complex logic, mimicking System 1/System 2 reasoning.

• Discovering New Scaling Laws. And we see it in the most recent ParScale attack on the memory-vs-intelligence bottleneck from Qwen. By running P parallel forward passes of a single model (with shifted learned POVs), it re-uses weights to scale computation instead of memory. It’s the equivalent of scaling the model up by O(N log P) parameters at O(P) memory cost.

This isn’t nostalgia for the hand-tuned features of a bygone era. It’s about structural advantages — symmetries, invariances or causal relationships inherent to the data — that bend the scaling curve in your favor.

Path #2: An Alchemical Breakthrough

The Alchemist’s path is more radical: if the universe of high-quality data is finite, then why don’t we create new universes? Instead of passively consuming data, build engines that generate it, creating feedback loops where models learn from their own experiences and shape data distributions with maximal information gain relative to their current knowledge.

Some of the biggest breakthroughs in recent years have happened here:

• Self-Play in Simulated Worlds: This is the canonical example. AlphaGo used Monte Carlo Tree Search (MCTS) to explore the game space of Go, generating novel, superhuman strategies that existed in no human dataset.

• Synthesizing Preference and Behavior: The engine of modern alignment. Techniques like RLHF and DPO are fundamentally data-generation processes, creating ranked preference datasets to instill desired behaviors into models.

• Agentic Feedback Loop: The frontier. We can achieve credible long reasoning traces by interacting with APIs, databases and code interpreters, but when you combine them with revealed (human) preferences in vertically-integrated tools like Cursor, these long-horizon traces become the new, high-quality dataset to bootstrap next-gen long-reasoning models with real reward signals.

However, this path is a razor's edge. Between model autophagy, specification gaming and the sim-to-real gap, it’s easy to get stuck in dead-ends or get lost in degenerative feedback loops if you don’t have enough ground truth.

Grand Slams, Doubles & Strike-Outs

I pulled ~10 high-profile papers over the last 2 years from each category and then calculated a (quick-and-dirty) Fermi estimate of the % gain in intelligence per unit cost (±% IpC):4

When you look at it, a few things jump out immediately:

1. First, the Alchemist's Path is a high-variance lottery ticket—when it hits (see DeepSeek R1 and DreamerV3), it completely reshapes the landscape, but it's a game of grand slams or strike-outs.

2. The Architect's Path is the king of compounding returns; with the exception of the HRM outlier, it’s mostly a steady drumbeat of 20-30% doubles that continually push forward the state-of-the-art.

3. But the most important insight is this: these two paths aren’t in opposition, Alchemists and Architects are locked in an intertwined dance. Architectural breakthroughs like Mamba are precisely what make running Alchemical bets like DreamerV3 computationally tractable in the first place.

A Leader’s Gambit

So, which do you choose?

This is the wrong question. A research leader’s job isn’t to predict a single future, but to design a research portfolio that can win in different possible futures. The real question is your tolerance for risk.

The Alchemist’s Path offers a high-risk, high-return path. Although a breakthrough in general-purpose RL hasn’t happened yet, it probably will — if you survive long enough to see through the dead ends first. In contrast, the Architect’s Path seems like the “safe” bet. It’s a steady 30% gain, every few quarters. The only problem? It hides the extreme opportunity risk of being made obsolete overnight by a competitor’s 3x breakthrough. As DeepSeek and Qwen showed not so long ago, leapfrogs are real.

Faced with this, I think folks have two primary gambits:

1. The Incumbent's Gambit (Risk-Off): Allocate most resources (say, 70%) to the Architect's Path. This ensures a steady stream of predictable wins, efficiency gains and product improvements to keep the business running. The Alchemist’s Path is the "lottery ticket"— a smaller, speculative R&D budget (the other 30%) to hedge against being blindsided and start the alchemical dance. Make sure you can fast-follow the rest.

2. The Challenger's Gambit (Risk-On): Overweight the Alchemist's Path — 70% again, say. This is an aggressive, high-stakes bet to leapfrog incumbents and redefine the market. Here, the 30% allocation to the Architect's Path serves as runway extension — it ensures you’re still shipping something of value regularly and creating enough traction to buy yourself time to land the (eventual) moonshot.

The real risk is not taking a risk. The scaling maximalism of the last decade allowed us to avoid many hard choices — now, we have to think strategically.

Modernizing the Bitter Lesson

If I could update Sutton’s gospel with what we’ve learned from the Scaling Laws, it’d read something like this.

General methods that maximally leverage today’s finite data — subject to tomorrow’s compute limits — are ultimately the most effective, and by a large margin.

— Kushal Chakrabarti

The tragedy is that most teams are still fighting the wrong battle. They’re running the “more GPUs” playbook in a world where the real bottleneck is the data supply chain. If your team is asking for more compute but can’t explain their data roadmap, send them back to the drawing board.

The leaders who win the next era of AI won’t just be scaling models — they’ll be scaling data itself, whether by architecting better use of limited data or alchemizing entirely new unlimited data universes.

The Bitter Lesson wasn’t an excuse to avoid hard choices. It was a warning. With everything we’ve learned in the last 6 years, we should heed an updated warning.

My bottom-line hypothesis: The AI companies that survive 2026 will be the ones who figured out data scarcity in 2025.

Both as an advisor and executive, I've seen this movie before. You don't like how it ends. A few weeks ago, the CEO of a Fortune 500 tech company and I were chatting in his office overlooking the Bay and he was complaining about his GenAI team. "We've invested $50 million, it's been two years and they still haven't moved the needle." The pattern is always the same: massive technical investment, brilliant technical breakthroughs that look great in demos and papers and then… crickets on business impact. Boards get impatient. Talent hemorrhages. Teams get "reorganized" (read: disbanded).

The Problem: Winning at the Wrong Math

Here's the thing: You're not failing at AI. You're winning at the wrong math. Most AI initiatives fail not because research or engineering or business is hard, but because teams optimize for different loss functions that don't exist in the real world.

Everyone knows what a P&L is. You take revenue, subtract costs, get profit (or loss). There are two ways to improve it: make revenue bigger or costs smaller. But doing that takes effort. Project effort gets decided at the team level. Business teams want big results yesterday. Engineers want to solve today's practical problems. Researchers want to solve tomorrow's grand challenges. These aren't the same. The gap between them is where AI goes to die.

dP&L/dt: A Common-Sensical Loss Function

At Amazon, we taught everyone to work backwards. So, let's apply some common sense and work the problem backwards.

You want to improve the P&L. To do that, you need to make revenue bigger or costs smaller. Doing that takes effort and different things take amounts of different effort. Given that you have finite effort (i.e. your employees), what do you focus on?

Whatever has the highest dP&L/dt!

What is dP&L/dt, you ask? Mechanically, it's just the derivative of the P&L relative to effort, i.e. change in your P&L divided by effort. If you do some math, you can show that if you keep picking projects with the highest dP&L/dt until you run out of effort [*], you'll get the highest P&L impact for your effort. This isn't rocket science but, done right, it's profound.

[* Yes, effort isn't fungible, projects have different risk profiles, balancing dependencies is tricky and 17 other things — I'm simplifying out a bunch of things for now. See below.]

Spiritually, it's organizational DNA — a search and destroy mission. By identifying places with high dP&L/dt and systematically working backwards to break bottlenecks, you not only unleash P&L potential but also build a common vocabulary. By continually doing it over time, you build muscle memory and relationship habits.

Factions: The Root Cause of Failure

Why do common vocabulary and relationship habits matter? Let's break it down another level.

At its root, a common failure mode in many AI organizations today is the divide between research, engineering and business. We see the symptoms everywhere today: skyrocketing AI investments but backtracking on AI promises, revenue churn at AI SaaS startups, lack of AI revenue impact at big companies.

As any seasoned manager (or parent) knows, incentives don't just produce outcomes — they have hundreds of knock-on behavioral effects. When researchers, engineers and business leaders are incentivized for different outcomes, they not only produce different outcomes (duh) but they also make value judgments about other outcomes and break up into factions.

Organizations break down when factions stop talking. Failing communication causes frustrations to fester. Festering frustrations lead to loss of trust, which means people stop even trying to communicate. Thus begins the downward spiral. The research team thinks business is short-sighted. The business team thinks research is building toys. Engineering thinks they're the only ones doing anything "real." When factions emerge, everyone's usually (partially) right. But, ultimately, it doesn't matter who's right because everyone gets fired (read: disbanded).

Beyond a metric, what dP&L/dt DNA does is incentivize different factions to think through the same end-to-end system and bottlenecks to drive ultimate impact. Researchers start thinking about customer pain, engineers start incorporating longer-term research strategy, and business leaders start considering technical challenges.

And when you deeply think about someone else's challenges, a funny thing happens — empathy emerges. In one swoop, dP&L/dt DNA solves the revenue problem and breaks down factions, helping teams move from adversarial value judgments to empathetic problem-solving.

What Does This Look Like In Practice?

If you're seeing a discrepancy between technical innovation and business impact, this is highly solvable. I've implemented it many times, from 10-person pizza teams to 50-person startups to large Fortune 500 orgs, both in executive and advisory roles.

What does it look like in practice? As with most management, it's highly contextual but the spirit is the same: integrating the P&L work-back into your organization's day-to-day rhythms and team-level metrics. Your researcher isn't just showing BLEU scores — they're explaining the work-back from customer churn to model explainability to the tradeoffs for different mechanistic interpretability techniques. Your product manager isn't asking for "better accuracy" — they're quantifying exactly how much customer lifetime value each percentage point unlocks. Your executive team isn't just saying "AI is strategic" — they're debating which directions unlock the richest areas of dP&L/dt under different forward scenarios.

To be sure, there are sharp edges and judgment matters. Short-term dP&L/dt vs. long-term dP&L/dt portfolio mix. Non-fungible effort — your star researcher can't just be reassigned to close deals. Correlated risks — putting all your dP&L/dt eggs in one basket. Platform effects where today's investment creates tomorrow's dP&L/dt multipliers (d²P&L/dt² for the nerdy).

Like many things management, dP&L/dt is 40% science, 60% art — and 100% doable.

Hot Take: The Ultimate Winners

Here's my hot take: The companies that win this AI cycle won't have the smartest researchers or biggest budgets — they'll be the ones where everyone executes with dP&L/dt DNA.

Historically, some of technology's breakthroughs have came from vertically-integrated teams with unified objectives. SkunkWorks' SR-71, AWS' two-pizza teams, and DeepSeek's efficiency play — they all succeeded because they were small, multi-functional teams where everyone ruthlessly optimized against a single mission. The converse proves the point: Google invented transformers and had web-scale data, but lost to OpenAI out of the gate because their research, engineering and product teams were fragmented; only as it's reorganized into vertically-integrated units has Google been begun catching up. During rapid change, horizontal coordination breaks down.

The next AI winners won't be silo'd research divisions or traditional enterprises bolting on AI — they'll be companies that (re-)build themselves as vertically-integrated dP&L/dt machines, where AI teams sit next to customer success, where engineering works directly with research, where business strategy and technical architecture are co-designed together.

In the 1970s, it was assembly → C. In the 1990s, it was C → Java. In the 2000s, it was Java → Python, Ruby & JavaScript. Now, it’s prompting.

What do I mean? Well, you can write 10,000 lines of C that can be compiled to 1,000,000 lines of assembly. You can also write 100 lines of Python, Ruby and JavaScript that can be interpreted to 10,000 lines of C (and 1,000,000 lines of assembly). And now, imagine your favorite LLM as a compiler: you can write 1 line of prompt that can be converted to all of the above too. This is what programmers call abstraction.

When revolutionary abstractions emerge — from Java's cross-platform usability to AWS' cloud infrastructure to Objective-C's mobile SDKs — a predictable cascade of transformations follow. Accessibility explodes as barriers fall. New failure modes replace old ones: buffer overflows give way to dependency hell, multi-threaded UI race conditions and network partitions. Engineering culture and organizations shift from individual proofs to collaborative, iterative problem-solving. Economics transform as middleware commoditizes as infrastructure and application layers gain value. This pattern of abstraction shifts and cascading changes has repeated throughout computing history: assembly to C (1970s), C to Java (1990s), interpreted languages (2000s), and cloud/mobile (2010s). Now, we're about to witness the next stage in real-time.

Taking the long view, prompt engineering is the next stage in a long, inevitable series of abstraction shifts and cascading transformations. It's neither a passing fad nor a once-in-a-lifetime revolution. If you dismiss it, you’ll be left behind, just like the “scripting languages aren’t real languages” programmers of the 2000s. If you work yourself into a frenzy, you’ll miss the real opportunities because you’re too tired or distracted.

As a kid, I learned to surf on the beaches of Los Angeles. Anyone who’s surfed can tell you how humbling it is to try to paddle out the first few times and come up snorting saltwater. I learned an important lesson that summer: You can’t fight the ocean but, if you figure out how it behaves, you can learn to duck under the wave and surf the next one to glory.

Three Causal Mechanisms

So how do abstraction shifts behave? As one of my favorite mentors says, “just stare out the window for a few minutes.” The world isn’t random; there are some basic physics to how it all works.

I claim three basic mechanisms explain a bunch of things about abstraction shifts:

1. The Accessibility Flywheel: With lower technical barriers, more people can build more things; with new economies of scale, others build better tools that lower barriers even further.

2. The Complexity Paradox: Each new abstraction simplifies development by hiding historical complexity, but creates newer and harder sources of complexity (read: bugs) at higher levels of abstraction.

3. The Value Barbell: As the meat-and-potatoes work of yesteryear gets commoditized away, economic value gets pushed to the extremes: low-level system builders and high-level domain applicators.

If you work out the math (and make some assumptions), it turns out the Value Barbell explains 32% of the value created from M&A and IPOs in the past 45 years, the Accessibility Flywheel explains 27% and the Complexity Paradox explains 15%. More on that in a minute.

Does This Actually Work?

Let’s be honest, this is just a hypothesis. Everyone and their mother has a hypothesis for AI right now. Good hypotheses should at least be accurate — can it pass a backtest? Let’s consider a few examples:

Remember when two founders in a garage became the Silicon Valley archetype? Y Combinator launched in 2005, but Ruby on Rails launched in 2004. It wasn't magic — it was the Accessibility Flywheel in action. Suddenly, developers could launch MVPs in weeks not months, with teams of 3 not 30.

But for every action, there's an equal and opposite reaction. Alongside this democratization came a predictable explosion of XSS vulnerabilities, CSRF attacks, and SQL injections, AKA the Complexity Paradox, as we traded yesterday's bugs for tomorrow's. I downloaded all reported NIST exploits over the past 35+ years, cross-referenced assigned root causes related to dynamic web frameworks and calculated the percentage — just look at the step change:

The mobile revolution tells the same story. Before iOS, developing mobile applications required arcane knowledge of embedded systems, cryptic SDKs, and carrier relationships that kept innovation locked behind practical and economic barriers. When location services became accessible through simple SDKs, we didn't just get better maps — we got entirely new categories of businesses. Uber, Pokémon GO, and countless others emerged from that abstraction shift, creating a $20+ billion market in a decade.

But it's not just about moving up. While WhatsApp built an app serving 900 million people with only 50 engineers, NVIDIA created enormous value at the lowest abstraction levels. Their stock has risen 17x in the last five years by understanding that as middle-layer tasks get commoditized, exceptional value emerges at both extremes. They saw that as AI abstractions simplified development, whoever built the specialized silicon foundations would become indispensable infrastructure — capturing massive value across an entire technological revolution. In a nutshell: The Value Barbell.

Downstream, as value, difficulty and people evolve, so do culture and organizations. When I started at Amazon in the early 2000s, the entire company ran on a single C++ monolith called Obidos. Back then, bi-weekly release cycles on 50+ million lines of code caused ~1000 engineers to operate in lockstep — God help you if you broke the build. A decade later, with microservices and cloud infrastructure, small, fast-moving, autonomous teams could build and launch (and break and fix) functionality without talking to anyone else.

There are dozens of other examples: enterprise apps going 70% low-code; US engineers tripling; microservices reaching 85% of big companies; left-pad breaking the Internet; cloud spending exploding to $595 billion; Apple paying app developers $300+ billion; Tesla vertically integrating from chips to self-driving algorithms. Each showcases the same pattern — as we abstract away yesterday's complexity, we unlock more builders, create new kinds of bugs and push economic value to the extremes.

This isn't random — it's physics. When you abstract away today’s challenges, you necessarily create new failure modes, new audiences and new economic winners and losers. The same forces that transformed programming from assembly to JavaScript are now transforming it from JavaScript to prompts.

What Comes Next?

“OK, let’s say I believe you. What does this mean?” A useful hypothesis makes useful predictions. If you play out the above to their logical conclusions, a few things emerge. Let’s consider the Complexity Paradox and the Value Barbell first:

• In an AI-augmented world, what do tomorrow's bugs look like? Modern LLMs are designed to mimic (and please) humans, suggesting future systems may inherit human cognitive errors. Imagine an AI zero-day that exploits reciprocity bias — tricking your AI customer service agent into issuing a $1000 voucher because a scammer was "nice" to it first. You can vet standard code for SQL injection vulnerabilities, but how do you vet AI code for psychological vulnerabilities?

• Will we accept distributional failure in AI-driven businesses? Mostly, traditional software produces binary outcomes — it either works or doesn’t. AI systems produce distributional outcomes and have irreducible uncertainty in any single case. Will businesses and users accept this? My bet: human psychology says no. We punish negative outcomes disproportionately, even when average performance improves. This mismatch between technical reality and human expectations will drive a wave of innovation in AI safety, testing and containment. Companies that solve the “reliability gap” will win.

When you work out the math, it turns out the Value Barbell is the #1 driver of technology value via M&A and IPOs in the past 45 years (32% of all value), but the Complexity Paradox is the fastest-growing driver, growing from 5% to 20% of tech valuations in the past 25 years. Check out the below — I just downloaded all US tech IPOs and acquisitions >$1B from Crunchbase, asked Claude to categorize everything and then plotted it.

While the Accessibility Flywheel is a solid, steady driver of shareholder value over the decades, my guess is its impact is felt more organizationally and socially.

• Who’ll be the 10x engineers of 2030? Probably not today's JavaScript wizards. The skill ceiling will rise dramatically while the floor drops to near zero. We're already seeing signs of the new elite: people who can architect complex request-response chains, prompt ensembles, and system-level AI interactions — people who are part psychologist, part philosopher, part systems architect. My bet is that it’ll be the folks who can personally embody the Value Barbell, simultaneously thinking about low-level system constraints and high-level LLM psychology.

• How will organizations resolve the tension between democratized development and centralized control? If everyone writes code, is everyone an engineer? (And does everyone report to the CTO?) No. Like data teams today, AI will likely end up following a hybrid model: centralized platforms (under CTOs), centralized risk management (under CISOs), and federated AI-enabled operations throughout the company. As more code gets written by more AI-enabled teams, my guess is that technology organizations will look very different than today.

Who Cares?

We're all just going to get eaten by the singularity anyways. Why does it matter?

I joke that 80% of my job reduces to "I've seen this movie before — I don't think you'll like how it ends, you should do something different." In the past year, I've probably talked to hundreds of technology executives, startup founders and investors. Many fall into two camps:

• "Move along, nothing to see here." Dismissing it as just another passing fad, they’re trying to focus on next quarter’s P&L. Let’s name it: they’re the IBMs of today.

• "OMG, the train is leaving, EVERYONE LET'S GO!" Frenzied by FOMO, these leaders are throwing money at anything with "AI" in the pitch deck. They're the WeWorks of today.

We've invested and committed $1 trillion dollars into modern AI, e.g. Stargate, GenAI startups. Power laws being power laws, most of that won’t generate a profit (but some of it will probably generate a lot). We’re not in Kansas anymore, Toto.

I claim there’s a third path: strategic patience + leveraged action. Identify where complexity will migrate next; get there first. Red-team how the Value Barbell might upend your industry; disrupt yourself. Find your organizational points of leverage and kickstart the Accessibility Flywheel. Some teams might need to guard against psychological vulnerabilities of AI systems, whereas others might need to look for disruptive

The question today isn't if modern AI or prompting will transform the technology ecosystem — it will; the tide is coming in and the ocean always wins. Just as interpreted languages, mobile SDKs and cloud infrastructure transformed the technology ecosystem, so will prompting. The fundamental physics of abstraction shifts and cascading transformations guarantee it.

Instead, the question is how will you approach the ocean? Ignore it, you'll get hammered by a wave and end up with your face in the sand snorting saltwater. Jump in at the wrong place (or wrong time), the waves will drag you against the corals and you’ll end up snorting even more saltwater. Or... just hang back for a beat, learn the ocean’s rhythms, strategically dive in (and under) a wave and then ride the next one to glory.

I don't know if there'll be a singularity or not in the next N years or what p(doom) is. What I do know is if there is a singularity, ignoring one of the most disruptive forces we’ve seen in a generation (or not taking a few minutes to properly understand it) isn’t a recipe for success for you, your family or your team.

I don't have a crystal ball — anyone who says they do is lying — but I do think reality follows some basic physics. Data, people, organizations behave in (mostly) predictable ways. While the past is no guarantee of future performance, it is the best predictor of future performance — ignoring it is a good way to end up with a mouthful of sand and a noseful of saltwater.

Hang five, my friend.

• 2020-21 (AKA "Wait, we're doing WHAT?"): Started as an advisor making color-coded trade-off tables from the sidelines. Joined full-time because, well, the tables weren't enough. Made 17 contingency plans. Threw them all out. Built new ones. Convinced ourselves we weren't crazy. IPO'd. Quintupled revenue ($2.6B ⇒ $8.0B ⇒ $15.6B!). 🚀

• 2022 (AKA “So, about those interest rates…”): Enter the fastest series of rate hikes in 40 years. Installed more than a few "days since last incident" boards. Worked backwards from market fundamentals and forwards from algorithmic frontiers. Learned there's only so much you can do when "once in 40 years" happens twice in 2 years. 🤦🏾‍♀️

• 2023-2024 (AKA "Hold my whiteboard marker"): Merged engineering, product & AI/ML research into one beautifully chaotic family. Shipped breakthrough AI/ML innovations at a pace that'd make even the most cynical Twitter/X ML threads proud. Proved that anti-fragile isn't just a fancy word in library books (≤1%/yr regretted attrition!). Turns out great teams really do get stronger under pressure. Helped re-start growth. 🚀

Working backwards from what makes great R&D organizations tick: you need technical excellence that delivers real impact, teams that grow stronger under pressure, and the ability to unite different perspectives toward an ambitious goal. And sure, there's some secret sauce as well: maximizing throughput beats minimizing latency, designing uncorrelated risks in research portfolios is how you beat the house, dancing the three-step between commercial viability ⇔ technical feasibility ⇔ org health is art not science, etc. But mostly, you need extraordinary people who somehow make the impossible look routine. I've been incredibly fortunate to work with some of the brightest minds in the AI/ML industry who showed up every day ready to tackle the seemingly impossible challenges I threw their way.

To the incredible team who turned my grossly reductionist frameworks into reality (and politely nodded at all my "Kushal"-isms) — thank you. You taught me more than any color-coded table could capture. The resilience, creativity, and sheer determination you've shown through every market cycle is inspiring. I'm excited to see what the team will achieve next — the foundation we've built together is just the beginning.

What's next? Well, me being me, let’s work it backwards and forwards:

• Step 3: Tackle another "impossible" problem

• Step 2: After I figure out which one

• Step 1: But first, a (doctor-mandated) nap...

So, funny story: I broke my ankle last week. Between surgery and the cast, I'm basically under doctor's orders to relax for the next 3 months.

Some people think as long as you work hard and follow the rules, you’ll make it. You’ll hear them say that this generation is lazy or that others are taking advantage of the system. They say you need to pull yourself up by your bootstraps.

Others reject the American Dream as a nepotistic nightmare. You’ll hear them say that it’s not what you know, it’s who you know. They say the system is stacked against you.

Yet others say it’s all just luck.

I disagree. Over the last 20 years, here’s what I’ve learned:

American Dream = Hard Work * Big Bets

But here’s the twist: You don’t get to make the big bets, your teachers, bosses and friends do. I’ve worked hard and taken some risks, sure, but if it weren’t for people I admired being willing to bet on me, none of it would’ve mattered.

When I was struggling with depression in high school, sleeping through classes and failing exams, Mr. Z and Mr. Murray took me aside and told me about “potential.” (They gave me the inspiring version. Later, I learned that “has potential” is how teachers describe kids who are wasting their potential.) At a time I was being put into remedial math classes, they staged an intervention and taught me that second chances exist.

At Berkeley, I discovered statistical models — it was love at first sight. I worked hard, working two jobs and taking every class I could find on AI/ML, computer science and statistics. But ultimately, it was Lior Pachter who bet on 19-year-old me and gave me my first big chance: working on the Human Genome Project, helping write some of the foundational systems and models. Getting to see Lior in action, alongside Richard Karp, Eric Xing and Michael Jordan, was mesmerizing. They taught me how to think systematically about the world and that math can be beautiful.

Around Amazon, Kim Rachmeler, Jeff Holden, Joel Spiegel, Karen Van Dusen and Rick Dalzell took me under their wing. Leading a $1B+ data team — the original Amazon personalization engine — at the age of 22 was no small feat and I was more than a little cocky. As a kid who’d walked to school on streets where you didn’t wear red, they showed me a totally different world — the corporate boardroom — with its own unique characters, customs and culture. They taught me good leadership habits and about the speed of innovation.

In startups, I found a brotherhood of channeled intensity. Hadi Partovi, Rich Barton, Mitch Kapor, Mike Murray, Dave Peery and Vinod Khosla bet on an intense kid with a big idea. Glenn Kelman, Geoff Davis, Jim Fruchterman and Sebastien de Halleux showed me I wasn’t alone in my intensity, and by example showed me who I could be. Tim O’Reilly, Saar Gur and Auren Hoffman unveiled to me a world of thought leadership that I didn’t know existed and then invited me into it. And for the first time, I learned how people, strategy and emotions can work together: Olivia Fox Cabane taught me that the whole is bigger than the sum of parts, and that true leadership means inspiring hearts as well as minds. I wasn’t ready for all of it but, when I needed it most, I finally understood. They all taught me I wasn’t alone.

At Opendoor, Daniel Morillo and Raji Subramanian saw something — ”potential,” perhaps ;) — and not only created opportunities for me but also pushed me to continually focus on the fundamentals. I will forever have Daniel’s “just stare out a window for 10 minutes” and Raji’s “why, what and how” imprinted in my brain. Only with time will I be able to fully understand what I’ve been learning from them (and fully express my gratitude). Daniel and Raji taught me that simplicity wins, even (and especially) at the highest levels.

Looking back, I’m not sure anyone ever bet on my father. He arrived in this country with an engineering degree but worked as a night watchman for years because no one would hire a brown man. He doesn’t talk about his past much, but from what I’ve pieced together, he and my mother had to do it all themselves.

In recent years, I’ve tried to repay the favor. Although I still do some research and engineering, these days I’m more coach than player. Both technically and managerially, my role today is to see how things might play out several moves ahead and coach my leaders through the right choices today. I joke that a lot of my job reduces to “I’ve seen this movie before, you don’t like how it ends — do something different.” In that capacity, I’ve gotten to work with dozens of emerging leaders, from startup founders to corporate executives and technical leaders. I’m incredibly excited to see what the future holds for them.

Among other things, if we play our cards right, this generation of artificial intelligence will revolutionize whole industries, both for Opendoor and beyond. Having lived through the last three AI winters, I’ve seen this kind of movie before. There are many similarities: there is a lot of snake oil and many of us are seeing worrying signals in the leading indicators. However, for the first time, for select players in strategic use cases, I believe the emergent capabilities outweigh the hype.

At least for my family, the American Dream was possible. In a single generation, we went from being refugees to leaders. Sure, I worked hard and took some risks, but so did others; the difference was that a few people were willing to bet on me. In any other country, our story wouldn’t have been possible.

This Thanksgiving, what am I thankful for? The American Dream and the many people over the years who’ve bet on me. Thank you.

And most of all, to my Baba — thank you for making that original bet and boarding that plane to the United States of America, and for the countless times you’ve doubled down on that bet with me since then.

As I was leaving, I asked lots of people I admired for advice. “What should I do next?” As every entrepreneur knows though, getting advice is a pain in the ass. (Let’s be honest.)

One (smart) lady says one thing—

“Go travel for a couple years, don’t worry about it— Vittana is amazing, you’ve earned it.”

Another (equally smart) dude says another thing—

“Don’t send yourself into exile, start your next thing now— I’ll write you a blank check today.”

What do you do? (Both real quotes, by the way.) They both sound great, don’t they?

But then, once upon a cloudy Seattle day, you talk to someone who gives you some of the best advice you’ve gotten in your life. Here’s what he said—

Kushal—

You talk about starting a family, starting an even bigger company. Here’s the thing— you’re always on. That white-hot intensity of yours never turns off.

What’ll your team do if you’re always on? If you’re always on, you’ll burn yourself out, you’ll burn your team out.

But, worse: How’ll your family be if you’re always on? Those are going to be some seriously fucked-up kids.

Here’s my advice—

Go relax until you learn how to relax.

Damn.

He was dead right. I’d been clocking 100-hour weeks since I was 15. First, at Berkeley: an EECS and biology double major, plus two research jobs. Then Amazon, helping lead the recommendations team. And finally Vittana. What I lacked in experience or intelligence, I made up in sheer intensity and grit.

But, it came at a cost. I delivered results, but I drove people crazy doing it. I pushed people away— some of my closest mentors, friends and colleagues.

In recent months, I’ve been asking those folks for feedback. The three most common replies? “Inspiring.” “Infuriating.” “Exhausting.”

Sometimes, there isn’t much to say except, “Yup. You’re totally right. I’m going to go do exactly what you said now.” And that’s exactly what I did.

I broke my lease in March, put all my stuff in storage— my only link to Seattle being that my billing address was my assistant’s home address. The plan? A one-month sublet in every place I’d ever wanted to live.

I’ve seen the midnight sun set over Priekestolen, hiked the sea cliffs of Molokai, gone skinny-dipping in the Arctic, thrown back vodka shots on a Cold War-era Russian base with old Cold War-era Russian guides, been proposed to on a midnight bus on the way to a Turkish monastery, been straight-up vetoed by the American Embassy in Algeria, lived in a UNESCO world heritage site in the 1,000-year-old Marrakech medina, discovered true awe amongst God’s fingers in St. Peter’s Basilica, and much, much more.

Here’s what I’ve learned:

• People are the same everywhere: moms worry, boys try to impress girls, and everyone just wants to be seen.

• ”We cannot provide consular support” is code for “The SEALs aren’t coming for you. Don’t fuck up.”

• Crayons are really good for drawing in the shower.

• Always jump off the cliff, always dance with the girl, always go for the kiss, always say yes — and always chase bubbles.

• Corollary: Skinnydipping in the Arctic is always a good idea.

• Learn “thank you” in every language you can. Smiling and pointing is good enough for everything else.

• Discover awesome— the classic definition, not the surfer one.

• Never try to bribe a Bangladeshi border guard about a motorcycle.

• Smile, even (especially) if you don’t feel like it.

But, most of all? Breathe, be present and see the moment.

Back in Berkeley, I fell in love— that you could build something from nothing was intoxicating, electrifying. As entrepreneurs, we’re always trying to mold reality to our dreams. And that’s beautiful: it lets us do beautiful things, create things others can’t even imagine, solve problems the world thinks are unsolvable.

But it has a dark side too. Sometimes, it means we don’t celebrate the little things. Sometimes, it means we’re blind to others’ emotional state. Sometimes, it means we miss the beauty of the moment. Sometimes, it means we don’t listen when we really should. Sometimes, it means we forget we already live in the future.

I was lucky in an infinite number of ways— not everyone can leave everything behind and travel for a year. Go for a meditation retreat instead. Take an interpretive dance class. Draw in the shower. Try acroyoga. Chase bubbles. Cliffdive. Dance. Live.

After a year of living, do I have any advice? Sure. Here’s my advice—

Go relax until you learn why.

Ignore the hippie bullshit— yes, you’ll be happier. But, you’ll also be a better friend, better partner, better leader and all-around better human being.

And me? Did I finally learn how to relax? Yes. It took eight months and more dollars than I care to admit, but I finally learned how in the Sahara this September. No laptop, no phone, no camera— just me and a sketchpad with a lot of sand and sky.

There is one quote that was on our first business plan, strategic plan & every other document since founding:

Ridiculing idealism is shortsighted, but idealism without the rigors of pessimism is misleading.

We need very hard-headed idealists who can look into the worst and best of humanity and can create and implement strategies of success.

— State of the Future 2007, Millennium Project

At its heart is the idea of hard-headed idealism. It’s an idealism grounded in the brutal facts of reality, one willing to make the tough decisions while relentlessly fighting to achieve its original aim — a better world.

You can’t separate the two.

You can’t just be a mercenary corporation, arbitraging market gaps to make a buck. What’s the point? You can’t take all that money with you in the end.

You can’t just be an innocent non-profit, believing good things will come because you’re a good person. To create real change at scale, you have to make the tough calls.

In the end, how do we create the world we want to live in? By having an unshakable faith that our world is possible and putting in the hard work & tough decisions to make that world a reality.

Reality is wrong. Dreams are for real.

— Tupac

Non-profit folks sometimes tease me about tech entrepreneurs building useless stuff. (I’m the token techie at a lot of conferences.) But really, non-profits aren’t much better. So often, we just put in bandaids.

How long did it take us to find a malaria vaccine? We’re so close. Mosquito nets are great, but what if we’d really taken a chance and thrown our weight behind a vaccine earlier?

A few days ago, I attended Seattle’s 2nd-annual Social Innovation Fast Pitch. Nathan Myhrvold, former CTO of Microsoft and the keynote speaker, came out guns blazing:

It’s great you can kill aliens with Xbox at a faster rate than ever before, but we didn’t really need our lives changed. The fact is, there are people who do need their lives changed.

You know something? Most of your ideas will fail and I’m completely OK with that.

Whatever you think of sometimes-controversial Nathan (he’s also the CEO of Intellectual Ventures, which has an aggressive patent stance), this is something near and dear to my heart. And he’s dead right.

When I was starting Vittana in 2008, I remember getting laughed out of several meetings. “This’ll never work,” people said. “If I give you money and you fail, that means I wasted my $1,000 — I could’ve given to Red Cross instead and been sure of impact.” (It turns out education does work and it’s hard to be sure.)

Nathan and I got a chance to talk for a few minutes before his keynote. He’s a razor-smart guy (obvious, really — not anyone ends up Microsoft’s CTO) and I’m excited he’s challenging folks to think differently. A bit of our discussion:

Yes, there’s a direct cost of failure. But, there’s also the opportunity cost of not trying. Looking for a malaria vaccine might be riskier than simply buying a mosquito net, but once you find the cure, you’re done — once and for all. But, if you never try, you’ll be buying mosquito nets forever.

If you never put down a big bet, you’ll never change the world. It’ll just stay the exact same place. What’s the point of that?

And let’s not even get started about trying to build (another) link shortener. For-profit, non-profit, B-corps — it’s tax status, not religion. Anyone can build cool stuff.

Make meaning, not money

— Guy Kawasaki

One of the SIFP finalists, Microryza, has an awesome project up right now to find a cure for parasitic worms that affect malnourished children in Africa. (Disclosure: I’m an advisor.)

These are exactly the sorts of innovative projects that’ll never get funded through usual means because they shake things up. Check it out and see if you can contribute — I just made a $100 donation.

UPDATE 10/22 10am: Woah! Hello Hacker News! Sorry for the slow site — I wasn’t expecting 4,000+ hits/hour.

UPDATE 10/22 12pm: Caching enabled and available memory quadrupled — site should be stable now. Let me know if you run into trouble.