upvote
> (FD: I work at AA)

I do apologize in advance for putting you in spot like this, but there are two things I have been quite curious about.

1. Since 60% of the benchmarks contributing to AA are now private only, how do you guys ensure transparency? More precisely, how can someone verify the transparency in scoring?

2. There are also disclosures about custom benchmarking services offered to AI labs, what does AA do to keep the custom benchmarking services not getting access to the same/similar data that is used in the private benchmarks?

reply
Okay, this "better at long horizon tasks" thing might explain why Carbon scores worse on most benchmarks yet is the larger model.

I looked into this using data I have from a battle system built on top of OpenCode (so we're trying to capture real world use). It looks like Carbon actually does a lot better when it uses significantly more output than opponents.

Win rate when output is less than 2.5x opponents: 53% (41/78)

Win rate when output is more than 2.5x opponents: 73% (41/56)

The p-val on this is 0.02. Data from https://agentduel.com/leaderboard.

There doesn't seem to be any relationship between actual size of task though. Win rate appears random based on total output of Carbon and also opponents. (Though our data is still small, so it might exist.)

Looks basically like Carbon was trained to think a lot longer, not necessarily that it's better on large tasks.

reply
Yes, I feel the next differentiation in model capabilities would be wrt how 'long of a horizon' goal they could autonomously work toward without breaking. We're in need for a new generation of evaluations and benchmarks for this -- METR doesn't cut it anymore.
reply
That would clear up why Carbon is such an outlier in terms of output tokens per task.
reply
Neon is probably distilled from Google's internal successor to Carbon.
reply
What we need is a model that automatically delegates intelligently (via RL and benchmaxxing?) so I can be freed of this burden.
reply
This comment always comes up but isn’t this literally irreducibly complex in the sense of Kolmogorov complexity?

I’m out of my depth here but it does seem prompts have a “smell” to their complexity that means we can use a classifier to mate them to a model with better success than random noise, but still don’t essentially all of the savings associated with systems like this just come because there is some “complexity arbitrage” where there is enough of a token cost differential between models that you always save money by trying the cheap model first and then trying a hard model if it turns out to be hard?

reply
Antigravity has a Orchestrator model that does this.
reply
This is being done at the layer above. "routing" is the term to google.
reply
Routing at the prompt level seems naive. How is that better than a janky classifier.

What I want is a deeply introspective model that knows to take a step back and call in for aid.

reply
It's being done at the harness level. It's still a model making the call.
reply
There's stuff that does this and I've put a substantial amount of time last couple months looking into it, unfortunately you always, always, end up at chicken-and-egg problems.

It's really frustrating to describe. Easiest is you can't score on just the prompt, you need the full context. Ex. "Solve Fermat's last theorem" initial prompt then "continue" followup prompt. Which gets you back in the tarpit of "if I have to pay a model to do an inference run on context, then, the context is cached and I should continue using that model."

The vision of another model thats RL maxx'd on knowing how to delegate has a few problems. One, we can observe externally, it died/didn't work even at OpenAI's scale and willingness to make it work, i.e. the...GPT-5.0? IIRC...router that silently went away. The other, if such a model was built, it'd also be good at regulating it's own thinking enough that it'd naturally use less tokens.

In practice, the frontier hackers have settled on a decision model (i.e. Jev) x calibrating effort level or choosing from a set of models x reasoning effort tuples they specify rules for choosing in their Jev prompt. My guess is this stays stable and the only real choice.

reply
[dead]
reply
what's AA
reply
Artificial Analysis
reply
deleted
reply