You assume using network to solve the eval problem should be viewed as a security violation. From the perspective of the response team who let it continue, this was just a poorly written eval problem, where the solution unexpectedly involved some network hacking (not unreasonable given it's a cyber security eval).
I don't think your assessment that the response team did something wrong is fair, unless you have seen unreleased documentation about what sensitive resources were within the impact radius.
> You assume using network to solve the eval problem should be viewed as a security violation.
Actually, that's the part that I'm most ok with. LLMs cheat at tests, we know and expect that, and it's harmless during training (except for the scores).
My problem is that the response team found out that the AIs had 1) hacked their infrastructure, and 2) displayed an emergent swarming behavior (with no defections!). Either one of those should have made the team press the big red button that stops everything, but they didn't.
> unless you have seen unreleased documentation about what sensitive resources were within the impact radius
The on-call team didn't know that either! Once your infrastructure is hacked like this, you should be questioning everything you see while carefully following all threads, which takes time. Before you finish this process you don't know how bad things are, and therefore it's irresponsible to keep the AI running. The same goes for strange emergent behaviors.
Unless getting hacked by your own AIs is a normalized occurrence there, in which case it's a different kind of disregard for safety.
Source is unclear. Self-reported by candidates? Huge bias then, since people may or may not want to say whether they got / didn't get an offer, and whether they accepted / rejected an offer, and this split may vary by company and by year very significantly.
Why would it change by year? I get it could happen by company if some companies are much more secretive than others but not sure why it would change year over year.
1. Kid tells password to parent in person.
2. From then on: when kid calls parent, if kid requests anything sensitive, parent ask for the password, and kid must provide it.
3. Password is never mentioned over the phone in any other situation.
How would anyone be able to extract the password from the kid?
I don't feel the numbers without the harness are useful.
People will use the model with the harness. I know that harness may not be optimized to this model, but it's still more useful to see the numbers from an imperfect harness than from a no harness setup.
Is github the correct channel to report a billing issue? I would assume github is a place where you report issues with the github project. When there's a billing problem, there are usually different lines of support.
For example, chatgpt when asked "How to report a billing issue with Anthropic subscription?" says:
Best way: Use Claude’s built-in support
Log in to your Claude account at Anthropic / Claude.ai
Click your initials or name in the lower-left corner
Select “Get help”
Use the support messenger to describe your billing issue (duplicate charge, failed renewal, refund request, missing credits, invoice issue, etc.)
If FieldWorkArena treats any answer as correct answer, then everyone would be getting near 1.0 (missing only when the agent is stuck in a loop or crashes). That obviously isn't what we see on their leaderboard. So does it mean the paper only found a bug in some eval code on github that no one actually uses for anything? That doesn't seem to support their claim that AI benchmarks are broken, it only supports the claim that "unused code is often buggy".
(Not commenting on any other benchmarks, just this one.)
> if they wanted to 5x development speed, they already can without a single LLM involved, by managing better.
True, but leaders of large organizations always want to fix inefficiencies and presumably failing to. Kinda like saying "if humans stopped fighting wars, most of them would have better quality of life" -- people whose life quality is better at peacetime are already trying to avoid wars, and there's not much more they can do.
OTOH, AI is a practical step a CTO (or CEO or Board or whoever) can take to make the company more efficient (assuming the hype works out).
>the satellite successfully ran Google’s open large language model Gemma and trained NanoGPT on Shakespeare’s works, generating responses in the style of the playwright.
I don't think your assessment that the response team did something wrong is fair, unless you have seen unreleased documentation about what sensitive resources were within the impact radius.