Live in the Future, Then Work Backwards
Buy tomorrow's cost curve today, watch what becomes possible, then build backwards.
- Difficulty
- Moderate
- Time to result
- ~ongoing to results
- Steps
- 6
- Confidence
- 70%
The method treats the future as something you can rent rather than forecast. Pick the input whose price is falling fastest, then deliberately pay today's inflated price to consume it at tomorrow's volume. In this episode that means spending roughly a hundred thousand dollars a year on tokens so that your daily working life resembles a normal citizen's in 2028. Living at that consumption level surfaces the capabilities and workflows that only become obvious when scarcity is removed. You then work backwards: the things that felt effortless at future prices become the product you build for everyone else once the cost curve catches up. The output is a roadmap grounded in lived experience rather than speculation, and a head start measured in years rather than features.
Origin
Naval and Gary Tan credit the underlying maxim to Paul Buchheit — build the future by living in it and working backwards. In this episode it is applied to AI: pay 2026 prices to consume 2028 volumes of inference, then build what you discover.
Core principles
- 01The cheapest way to predict the future is to overpay for it early.
- 02Cost curves fall predictably; capability curves surprise you.
- 03Every order of magnitude of compute unlocks capabilities, not just volume.
- 04You cannot reason about a technology you refuse to use daily.
- 05Complaining about a trend is not a substitute for standing inside it.
How to run it
- 1
Identify the collapsing input
Find the single input in your domain whose cost is falling by orders of magnitude rather than percentage points. In this conversation it is inference, projected to expand by roughly five orders of magnitude within two to three years.
Pro tip Plot the curve rather than the point — the slope matters more than today's price.
- 2
Buy the future price on purpose
Deliberately overspend so that the input is effectively unlimited for you now. The spend is a research budget, not an operating cost.
Pro tip Compare the spend to a headcount, not to a software line item — a hundred thousand a year is roughly one hire.
Watch out If the spend genuinely threatens runway, scope it to one team or one workflow rather than skipping the exercise.
- 3
Live at that consumption level daily
Run your real work through the abundant input for a sustained period. Spin up agents per task, per person, per question, until scarcity stops shaping your choices.
Pro tip Use it for mundane work too — the workflow discoveries usually come from volume, not from the showcase use case.
- 4
Log what only works at future prices
Keep a running list of things that became viable purely because the input was free at the margin. These are the candidate products.
Pro tip Note which items would still work at ten percent of the capability — those ship soonest.
- 5
Work backwards into a shippable product
Take the most valuable discovery and engineer the cost down until an ordinary customer can afford it. This is the actual company.
Pro tip Cost engineering is a product surface, not an infrastructure chore — treat the price drop as the feature.
Watch out Do not ship the version that only works at your subsidised spend; it will not survive contact with real unit economics.
- 6
Re-run on the next order of magnitude
Each further drop in price unlocks a new tier of capability. Repeat the loop rather than treating the first discovery as the destination.
Pro tip Set a calendar trigger tied to model or price releases, not to your planning cycle.
In the wild
Gary Tan describes going from not coding at all to shipping one of the most-used open source packages for vibe coding, after months of extremely heavy AI use. His conclusion is that anyone willing to spend around a hundred thousand dollars a year on tokens can already work the way a normal person will in 2028, because token cost falls while available compute rises by orders of magnitude.
→ A working preview of the 2028 default workflow, years before it is affordable to the median user.
One founder on the panel gives every single user in their app a dedicated agent. At launch, running that on a frontier model cost roughly a hundred dollars per person per month. Over three to four months the team built a fleet manager and an eval harness that spins agents up and down elastically, driving the cost to two dollars and eighty-four cents per person.
→ A workflow that started as an unaffordable experiment became a shippable per-user feature within one quarter.
Common mistakes
Debating the trend instead of inhabiting it
Scoring points about whether AI is overhyped produces no information. The panel's view is that the only way to develop a sense of where things are going is to use the technology at saturation.
Treating the overspend as an operating cost
Budgeted as normal opex, the spend gets cut in the first review. Framed as the price of a year of foresight, it survives — and it is cheaper than a single hire.
Stopping at the first discovery
Each order of magnitude of cost decline unlocks a different tier of capability, so a single pass captures only one wave and leaves the next to a competitor.
Is it for you?
Best for
Founders and operators building on a fast-deflating input like inference, where today's luxury spend is tomorrow's default.
Not ideal for
Businesses in mature, price-stable markets where no core input is dropping by orders of magnitude.
From the transcript
“if you're just willing to spend like a hundred thousand dollars a year on tokens, you can basically live like you uh are a normal…”
“the way to, you know, build the future is just like live in the future and then work backwards”
“to predict the future, you just live in the future”
From the episode
Live in the Future