What we believe at ORO

ORO’s mission is to give AI more and more autonomy by running businesses while making sure AI is aligned with strong moral and ethical values.

That is why we’re building these benchmarks. That is why we’re evaling.

What is it that we believe about the world that does not exist yet?

Deployment of AI will take far too long and will be far too laborious from existing individuals who do not possess the plasticity/trust to give AI all the autonomy it can be given to run their business well.

So we cannot spend our time trying to sell agents that help the bottom line of existing companies. We must take the time to build these businesses, in this new paradigm, from bottom up. How will AI run this business? What are the key decisions that experts make? Can we use those key decisions to calibrate verifiers/judges in order to incentivize the correct behaviour within this domain?

Why do I believe so strongly in this thesis?

1. The bitter lesson teaches us that any and all intermediate harness improvements are not good long term plays because the baseline model capability is going to continue gobbling up any harness improvements we make. The cheapest and lightest harness, focussed purely on model learning and search, will be the one that prevails in the long run. What follows here is that if there’s a capability gap, the longer term play is to sell a task that improves that capability gap back to the model provider rather than build a harness.
2. People selling AI agents to existing players are playing a short-term game. I believe there is very little point to playing short term games and durable businesses are built while balancing both short term and long term needs.
3. AI needs to be given autonomy in a business to pick up the necessary tacit knowledge that is tucked away in each and every knowledge worker’s brain. The only true way for this deployment to happen is to let AI learn this tacit knowledge by letting AI be in the arena. Sims are good abstractions for arenas but we will also need to take this a step further. We will have to monitor and run existing businesses using AI, mining and deeply understanding traces in order to continue moving towards our goal.
4. Large companies lack the agility required to be truly AI-native, from the ground up. I’m unsure if this company exists at all today, an AI-native company but ofc, I am not very well researched in this domain. Likely the big labs, spin-offs, neo-labs are AI-native. This is easy to do in software, but what does this look like in the physical world? Likely the thesis that Travis is testing at atoms.

Why are we the ones to enable this?

Because we thought of it.