I do not think the most important question is which technology will win.
The more useful question is: how do I get meaningful exposure to a possible future without doing something stupid if I am wrong?
That is how I think about asymmetric bets.
A good asymmetric bet has a downside I can survive and an upside that can materially change my life, my work, or the options available to my family. It does not require certainty. It requires a sensible price for being wrong.
This is not only about investing. Choosing what to learn is a bet. Building a product is a bet. Writing publicly is a bet. Running my own infrastructure is a bet. Spending nights understanding AI agents, Bitcoin, payment systems, or a new programming language is a bet.
Most of them will not produce a dramatic result. That is fine. If the downside is limited, the learning remains useful, and one successful bet can pay for many failed ones, the portfolio still makes sense.
As a father, this distinction matters to me. I am interested in ambitious outcomes, but I am not interested in risking the stability of my family for a dramatic story. The goal is not to go all in. The goal is to stay in the game long enough to recognize a real shift and have enough position when it arrives.
The biggest changes usually look smaller at the beginning
We tend to remember technological revolutions as obvious turning points. In reality, they usually arrive as incomplete, expensive, and slightly awkward tools.
Steam power did not appear and instantly produce the modern factory. One of its early advantages was much more practical: factories no longer had to sit beside the right source of water. Research on the Corliss steam engine shows how better power technology helped manufacturers move toward cities and markets. The engine changed more than energy production. It changed where industry could exist.
Electricity followed a similar pattern. Replacing a steam engine with an electric motor was useful, but it did not capture the full opportunity. Factories had to be redesigned around smaller motors, new workflows, different layouts, and different organizations. Historical work on electrification and productivity finds that the gains arrived together with capital investment and organizational change.
The technology was only one part of the revolution. The complementary systems were the rest.
This pattern is common enough that economists describe steam, electricity, and information technology as general-purpose technologies. They do not improve one product. They become inputs into thousands of other products, processes, and business models.
That also explains why their impact can feel slow at first. Companies need new infrastructure, skills, interfaces, regulations, and habits before the new capability becomes normal. The initial tool can look overhyped while the deeper transformation is still underestimated. Research on these adoption cycles describes it as a period of sowing before reaping.
The internet was similar. A modem and a static website did not look like the end of newspapers, software distribution, retail, entertainment, and large parts of the advertising industry. The first version was not the final value. The first version gave builders a new primitive.
The interesting question is rarely, "Is this useful today?"
It is, "What becomes possible when this gets cheaper, easier, and connected to everything else?"
2018 was already telling us
People often talk about large language models as if they suddenly appeared with ChatGPT in 2022.
The public interface arrived then. The technical direction was visible years earlier.
The Transformer paper was published in 2017. In 2018, OpenAI published the first GPT work on combining transformer models with unsupervised pre-training. The same year, Google's BERT showed how far large-scale pre-training could push language understanding across many tasks.
None of those releases looked like the complete product we use today. The models were smaller, the interfaces were technical, and the business use cases were uncertain. But the important ingredients were becoming visible: a general architecture, large-scale pre-training, transfer across tasks, and evidence that more compute and data could continue improving results.
In hindsight, 2018 looks like an obvious moment to pay attention.
At the time, it was much easier to dismiss the work as impressive research that still made basic mistakes.
Both statements were true.
That is the uncomfortable part of spotting an asymmetric opportunity. The technology can be genuinely limited and historically important at the same time. Waiting until every limitation disappears usually means waiting until everybody understands the opportunity and the asymmetry is gone.
Why most people miss the larger change
We naturally judge a new technology by comparing its first product with the mature system it wants to replace.
The early car is compared with a reliable horse. The first website is compared with a polished newspaper. A 2018 language model is tested as if it should already be a perfect employee. A Bitcoin payment is compared with decades of banking infrastructure, customer support, fraud handling, and accounting software.
That comparison is useful for deciding what to deploy today. It is weak for predicting what will matter in ten years.
The bigger effects arrive when a new primitive changes the surrounding system. Steam changed factory location. Electricity changed factory design. The web changed distribution. Smartphones combined cameras, sensors, internet access, identity, and payments into a platform for businesses that would have sounded strange a decade earlier.
A technological big bang is often not one invention. It is the moment several maturing pieces finally connect.
My current portfolio of bets
I have spent more than thirteen years building backend systems. That remains my base. Reliable software, payments, logistics, operational tooling, and clear system boundaries are not going away because a new model can write code.
But the way I allocate my attention has changed.
I use modern AI tooling every day, not because I believe every claim around AI, but because the cost of learning it deeply is low compared with the possible upside. If agents become a standard layer of software, I want to understand their failure modes, context, permissions, memory, and economics from practical work rather than from headlines.
I build products such as PLAYGRND because shipping a real system teaches me things that consuming other people's products cannot. Product decisions, distribution, user behavior, infrastructure, and maintenance all become concrete.
I run HILLS Lab and some of my own infrastructure because ownership creates options. It gives me a place to experiment, deploy quickly, understand the whole stack, and keep useful capabilities independent of a single platform.
I write because clear thinking compounds, and because distribution is an asset. An idea left in my notes has limited reach. A useful article can create conversations and opportunities years later.
I keep studying Bitcoin because a scarce, permissionless, digitally native asset and settlement network is a serious candidate for a world where software increasingly moves value. I do not need every coffee to be priced in satoshis tomorrow for that bet to remain interesting.
None of these requires me to abandon my career, ignore my family, or bet the house on one prediction. They overlap. The backend knowledge helps with agents. The agent work helps with products. Products improve my judgment. Infrastructure gives the products somewhere to run. Writing makes the learning legible. Bitcoin forces me to think carefully about ownership, security, and money.
The individual bet may fail. The capability stack remains.
The convergence I am watching
I do not think the next big bang will be another chatbot with a higher benchmark score.
My current guess is that it will come from the convergence of several systems:
- Cheap intelligence: models capable enough to understand messy goals and environments.
- Autonomous action: agents that can use tools, coordinate work, and complete multi-step tasks.
- Persistent context: memory, identity, reputation, and permissions that survive beyond one chat window.
- Programmable money: payment rails that software can use globally, with budgets and audit trails.
- Embodied execution: robots and machines that can turn digital decisions into physical action.
- Abundant infrastructure: enough compute, connectivity, and energy to make all of this economically ordinary.
We already have imperfect versions of every piece.
Models can reason but still hallucinate. Agents can act but remain brittle. Memory systems exist but identity and authorization are messy. Bitcoin and stablecoins can move value globally, but custody, tax, accounting, and regulation are still difficult. Robots can perform impressive demonstrations, but reliability and cost limit general deployment.
That does not weaken the thesis. It looks exactly like the early phase of a general-purpose technology: the central capability exists, while the complementary systems are still being built.
The next major transition may happen when software can reliably perceive, decide, pay, and act.
At that point, an AI agent is no longer only an interface. It becomes an economic actor. It can buy data, rent compute, pay another agent for a result, order a physical component, schedule a machine, and reconcile the outcome. Connect that to robotics and the boundary between software and labor changes materially.
This is a prediction, not a certainty. The exact breakthrough may come from energy, biotechnology, robotics, cryptography, or something that currently looks like a research toy. But the agent-money-robotics convergence is where I currently see unusually strong asymmetry.
Payment rails are one missing complementary system
My interest in this became more concrete because I work around systems where money is never just a number in a database.
At Tilt, a payment is connected to an auction, a seller, shipping, fulfillment, refunds, fees, timing, ownership, and an audit trail. A successful API response is only the beginning. The business still has to know what happened when one part fails.
Human payment products assume a person is available to log in, pass a challenge, approve a transfer, and explain it later. An agent buying twenty cents of data or three seconds of compute cannot wait for a human approval flow on every transaction.
Giving it unrestricted access to a company bank account or wallet would be equally absurd.
The agent needs a programmable spending envelope:
- hard limits per transaction, task, vendor, and period;
- allowlisted destinations and assets;
- purpose binding between a payment and the work being purchased;
- idempotency so a retry cannot pay twice;
- a double-entry ledger and exchange-rate evidence;
- human escalation for unusual behavior;
- an emergency stop that does not require rebuilding the treasury;
- reconciliation across the invoice, settlement, fee, refund, and delivered service.
This is mostly familiar backend engineering. The new part is that the caller is autonomous and the counterparty can be anywhere.
Stripe is an excellent product, but its own availability map demonstrates that global internet access does not create globally available banking relationships. Bitcoin begins with a different primitive: permissionless settlement. Stablecoins add a familiar unit of account. Lightning and projects such as Taproot Assets point toward networks where stable value and open routing may coexist.
None of this makes regulation or accounting disappear. In the EU, MiCA makes custody, exchange, transfer, and issuance boundaries increasingly explicit. A payment system still needs records, tax treatment, recovery procedures, and a clear answer to who controls the keys.
That is why the opportunity is larger than adding a wallet to an agent. The real product is a reliable authority, policy, settlement, and accounting layer for machine commerce.
Asymmetry cuts both ways
Not every asymmetric outcome is positive.
A tiny implementation mistake in wallet entropy can put a life-changing amount of money at risk. A leaked signing key can compromise years of work. One agent with excessive permissions can spend more in an hour than it creates in a month. One missing idempotency key can turn a harmless retry into repeated settlement.
The downside can also be nonlinear.
This is why I do not separate optimism from engineering discipline. The more leverage a technology creates, the more carefully its authority has to be bounded. Good architecture is partly the work of converting dangerous asymmetric risks into limited, recoverable failures.
For self-custody, that means verified devices, tested recovery, current firmware, strong operational practices, and risk distributed across locations or signers where appropriate. For agents, it means least privilege, budgets, audit logs, deterministic controls, and human escalation. For my own decisions, it means never placing one bet that can remove my ability to place the next one.
How I decide whether a bet is worth making
I do not have a formula, but I keep returning to a few questions:
- Is the downside survivable? Money, time, reputation, and family attention all count.
- Is the upside nonlinear? Can a modest input create a much larger increase in capability or options?
- Does the learning transfer? If the specific product fails, do I keep valuable skills, relationships, or infrastructure?
- Are the ingredients improving independently? Falling costs and better adjacent tools can compound faster than one roadmap suggests.
- Can I build before consensus? Practical work creates better information than passive prediction.
- Am I betting on a primitive or a fashion? Interfaces change quickly. New capabilities and lower costs tend to persist.
The last question is especially useful.
I do not need to know which AI company wins. I want to understand what happens when intelligence becomes cheaper. I do not need to know which wallet becomes dominant. I want to understand digitally native ownership and settlement. I do not need every product I build to become a company. I want to become better at turning an idea into a reliable system people can use.
That is a more durable position than predicting a ticker, a model name, or a launch date.
I do not need certainty. I need position.
In 2018, the important signal was not that a language model could already replace a person. It was that a general architecture was beginning to improve across tasks as data and compute scaled.
The comparable signal today may not be one spectacular demo. It may be the quiet connection of intelligence, persistent context, autonomous action, programmable money, and physical machines.
If that connection becomes reliable, it will create products and business models that are difficult to see from the current interface. Just as electricity was bigger than the first electric motor and the web was bigger than the first website, autonomous intelligence will probably be bigger than the chatbot.
I may be wrong about the timing. I may be wrong about which rail, model, or device becomes standard.
That is acceptable.
My responsibility is not to predict the future perfectly. It is to build a life and a capability stack where being wrong is affordable, learning keeps compounding, and being right once can matter.
That is how I think about asymmetric bets.