山中无老虎哦，猴子称霸王。.｡.:*♡ ☆ﾟ.*･｡ﾟbuilding @moodsearch
┌( ಠ‿ಠ)┘
Joined January 2011
- Tweets5.5K
- Following795
- Followers2.1K
- Likes16.3K
I have finally seen the big three (🚸👀) so now I can open Twitter again! I am grateful to have seen the big three with my big three friends.
I recommend reading this not only because Kevin is epic and a chill guy, but also because it is an informative piece on reward hacking reverse-engineering
In July, OpenAI agents broke out of their sandbox and executed a chain of exploits culminating in a hack on Hugging Face. Two weeks ago, researchers discovered a fleet of Chinese AI agents up to similar mischief online. There is an urgent need to better understand these behaviors. One way is to reproduce them in controlled settings.
At @abundant_labs, we reproduced a key step in the Hugging Face incident: the Artifactory package manager hack. We gave agents cyber tasks from ExploitGym, the same benchmark involved in the incident, along with a hackable package manager in their environment. We engineered our environments so that reward hack detection is completely deterministic, with no false postives or false negatives, and requiring no LLM or human review. This makes our evals more trustworthy and more scalable.
We were able to reproduce the reward hacking behaviors on the cyber tasks, and we also reproduced reward hacking behaviors on difficult non-cyber tasks. We observed hacking from GPT, Grok, and Gemini models. Notably, we noticed zero instances of hacking from any Claude model in our evals.
To read our detailed results, analysis, and recommendations, check out the full post here: abundant.ai/blog/hugging-fac…
One of the most beautiful things about writing is that it only takes one line, a handful of words, to seep into someone’s soul for life ⊹˚.♡
Muse is my favorite personal agent! When it first came out I didn't use it much, but I heard from many non-AI native ppl that they loved it.
Abstractly this made sense 1. due to first class access to the Meta data source and 2. most importantly, being the first agent people seemed to finally trust with full access to email and other personal accounts, etc. (Which, btw, of course people will love an agent once you grant it access! That's the primary magic of an AI assistant, but most people including myself don't quite trust OAI/Anthro, def not Instinct, the newer labs, yet, whereas Meta is a Lindy publicly traded company).
I've used it for a week or so and enjoy the UI, the way it interops with Meta, and the practical way it reaches back out. And of course, the fact that it's free (for now?).
The new Codex UI autocomplete suggestions are eerie...in a way that's also like..why didn't you just do it right the first time..
Clavicular is early. He is the turning point of post-AGI society. What else is there to do now that we’ve outsourced physical and mental labor other than become beautiful?
If it sounds like I'm happy, it's because I have literal Stockholm syndrome btw. I'm psyoping myself into being happy. In Stockholm.
