How it works

How public data becomes one score.

49Data sources
28Signals
4Questions
33%Software solved

1. Data sources

We collect public numbers, keep only processed counts, and credit every source. See all 49.

2. Signals

Each signal turns its sources into one number, then answers one of four questions. See all 28.

Can AI do the work?

Is it being used to build?

Are fewer people needed?

Is software cheaper and more reliable?

3. Formula

The score is the average of four questions, and each question is the weighted average of its signals. Every signal has a weight w of 1, except AI-built sites at 0.25, because its history is short. Counting questions equally stops a busy area from outvoting a quiet one.

\text{Software solved} = \frac{1}{4}\sum_{q=1}^{4} \text{question score}_q \qquad \text{question score}_q = \frac{\sum_{i} w_{q,i}\,\text{signal score}_{q,i}}{\sum_{i} w_{q,i}}

How each signal scores

Here $\operatorname{clamp}$ limits a score to between 0% and 100%, x₀ is the value in November 2022 or when its data starts, and t is the current month.

Fell since the start

s = \operatorname{clamp}\!\left(-100\,\log_{10}\frac{x_t}{x_0}\right)

Prices after inflation, revenue and software spending against GDP, security holes, bugs and incidents per git push, job openings, outsourcing revenue after inflation and Stack Overflow questions. Falling to a tenth scores 100%, and any rise scores 0%. Token prices use 50 in place of 100, so they need to fall to a hundredth.

Why this goal: A round number and our judgment: a tenfold fall is too big to put down to normal movement. Token prices fall much faster than most prices, so a tenth would be too easy.

Grew since the start

s = \operatorname{clamp}\!\left(100\,\log_{10}\frac{x_t}{x_0}\right)

Code activity, no-code revenue and new tech businesses. Growing tenfold since November 2022 scores 100%.

Why this goal: A round number and our judgment, matching the tenfold fall above.

A direct share

s = 100 \cdot \text{share}_t

Coding benchmarks, AI pull requests, the programming share of AI, developer sentiment and AI company hiring score the share itself.

Why this goal: No goal chosen, because a share stops at 100% by itself.

Length of task

s = \operatorname{clamp}\!\left(100\,\frac{\ln m_t}{\ln 120000}\right)

The longest task, in minutes, that the best model completes half the time, up to a working year of 2,000 hours.

Why this goal: Our judgment: a working year is about what one engineer delivers alone. It is not a measured threshold.

Cost of a hard task

s = \operatorname{clamp}\!\left(100\,\frac{3 - \log_{10} c_t}{3}\right)

The cheapest near-top run on FrontierSWE, from $1,000 a task at 0% to $1 at 100%.

Why this goal: Our judgment: $1,000 is roughly a day of an engineer's time, and at $1 the cost of a task stops mattering.

Growth from a recent start

s = \operatorname{clamp}\!\left(100\,\frac{\log_{10}(x_t / x_0)}{3}\right)

AI and coding app tokens, and sites on AI builders, against their first month. Growing 1,000 times scores 100%.

Why this goal: A round number and our judgment: these start from a small base, where tenfold comes quickly.

Output above its trend

s = \operatorname{clamp}\!\left(100\,\log_{2}\frac{W_t}{T_t}\right)

Output per tech worker W against its 2015 to 2019 trend T. Doubling that trend scores 100%.

Why this goal: Our judgment: doubling means one worker producing what two did on the old trend.

Jobs against the market

s = \operatorname{clamp}\!\left(100\left(1 - \frac{J_t / M_t}{J_0 / M_0}\right)\right)

How much software postings J have lost against all US postings M since ChatGPT launched.

Why this goal: No goal chosen: 100% means software postings have disappeared against the market.

4. The score today

With today's data, each question scores this much.

Can AI do the work?56%

Is it being used to build?41%

Are fewer people needed?28%

Is software cheaper and more reliable?8%

\text{Software solved} = \frac{56 + 41 + 28 + 8}{4} \approx 33\%

Under 50% reads "Not yet.", under 75% "Getting there.", under 90% "Almost.", and 90% or more "Solved." See the score.

Where to disagree

A simple formula we chose, so you can check it or argue with it.

More building counts toward solved because it suggests software is getting easier to make, but people still build most of it and the data can't separate them. Read the other way, the score would be lower.
The goals and groups are our judgment, and no data sets them. Moving every goal together, a fifth closer or a quarter further on its log scale, takes today's score from 33% to between 29% and 39%. A tenfold goal, for example, becomes about 6 or 18 times.
New MCP repos and product launches are shown but left out of the score, because we found no defensible number of them that would mean solved.
What AI models say is shown as a signal but left out of the score, because it is opinion, not evidence.
Coding benchmarks, task length and token prices only show what AI could do or cost, not that it replaced anyone.
AI company hiring has only one reading, so past months assume it was the same. Signals count from when their data starts, so the trend can jump when one joins.