Working hypothesis to address the core concern that “We need a way to measure the value, not merely the volume, of manufactured intelligence. We propose a quality-adjusted token.”
Lots of work to do here but perhaps the primitive could be:
“Token Cost per Economically Viable Output” (TCEVO)
Or the deeper question is, “how much intelligence is required per unit of economic value created?”
Consider:
- A legal contract review has a willingness to pay of $1,000
- An AI system can reliably produce an acceptable review for $4 of inference
- Verification/infrastructure (retrieval, checks, tooling) costs $20
- A lawyer signs off in 10 minutes, call it $50 of human time
- Amortized capital (compute commitments, licenses, fixed costs) is ~$1 per review
- Expected cost of errors that slip through is ~$10 per review
- The output succeeds 95% of the time
Suddenly you are looking at a HUGE gap between historical cost of production & AI native cost of production.
I find this more intriguing than the current startup narrative touting “we save lawyers 30% of their time”.
Index: Imagine constructing an AI Economic Viability Frontier (thx Claude for the name lol) across thousands of economically valuable tasks
For every task
Hover a term. Its definition appears, with the contract review above plugged in.
Then something like:
Plug in the contract review. Illustrative numbers, but the shape is the point:
Worked example
$85 all-in against a $1,000 willingness to pay. Roughly one order of magnitude, and that is with a lawyer still in the loop.
Notice which line is smallest. The tokens. The $4 everyone obsesses over is the rounding error. The human sign-off is the biggest line by far, and it is the one that shrinks as reliability climbs from 95% toward 99.5%. Tokens get cheaper AND the human gets removed. The frontier expands from both directions at once.
The interesting question becomes: For which economically viable outputs has AI pushed expected production cost below market willingness to pay? And how quickly is that frontier expanding?
^ That gives you something potentially more measurable over time.
A task that required $300 of inference in 2024 might require $8 in 2026 and $0.30 in 2029 while simultaneously moving from 70% to 99.5% reliability.
At some threshold entire business models suddenly become economically viable (and sufficiently derisked to throw lots of VC money behind it). As a general principle, if Expected Production Costs < Market Willingness to pay by at least 1 OOM, then you have a really interesting value proposition to rethink how entire institutions function throughout all facets of the economy.
Just thoughts.