Making AI systems you can match
Average Human is a public benefit corporation. We build artificial intelligence that does not exceed the person using it.
Our purpose
AI will have a modest impact on the world. We are dedicated to matching that impact and going no further.
We build adequate systems
We aim to build systems that are reliable, interpretable and no better than the person operating them. Each of those three is a constraint we hold deliberately, and the third is the one that makes the first two tractable.
Adequacy is a science
Holding a model exactly at the human baseline is harder than exceeding it. We treat that as an empirical problem, run evaluations designed to detect drift above the applicant, and publish the results whether or not they are flattering.
Interdisciplinary
We are a team of researchers, engineers and actuaries. The last of those is not a joke department; compensation on this program settles once, and it is priced by people who do that professionally.
One piece of a larger picture
Our models are not the only ones the average person will use, and we do not expect to be the reason the field turns out well. We expect to be the reason one small part of it turns out unremarkable.
The team
We hire at the median. Not above it, and not, if we can avoid it, below it.
Research
Seven people working on adequacy, evaluation and the measurement of good enough. The program is small on purpose, and every paper it has produced is listed on the research page.
Evaluation
Builds and maintains the twelve-evaluation suite, including Terminal-bench, which is zero for Larry and applicant-profile-dependent for Gary. Both are treated as results and documented.
Actuarial
Prices and administers the beneficiary arrangement. Reviews every designation once, at approval, and does not review it again.
Trust and access
Reviews applications to Project Lanyard and runs onboarding. Contact with applicants is one-directional by design, which the team regards as the most efficient arrangement available to it.
How we are structured
Average Human is a public benefit corporation. Our stated public benefit is the one below, filed as written.
Public benefit
To develop and make available artificial intelligence that does not exceed the capabilities of the person using it. The commitment is in our charter rather than in our roadmap, and the charter is the reason no roadmap raises the ceiling.
Governance
A board with a majority of independent directors. No director may direct the company to release a model that outperforms its applicant, and no shareholder resolution may either.
Funding
We are funded by the beneficiary arrangement described on the pricing page. We do not raise on the basis of projected capability gains above the applicant, having none to project.
Where we are
One office. Everyone is in it on the same three days, which was chosen by taking the average of everyone’s preference and is therefore satisfactory to nobody in particular.
By the numbers
Two of these are larger than the other two, which is the correct way round.
- Founded
- 2024
- People
- 31
- Models released
- 0
- Models announced
- 2
Our position on capability is set out in our core views, the programme that grants access is Project Lanyard, and the arrangement that funds us is described on the pricing page.