Research Scientist at @CognizantAILab. Prev: PhD in Semiconductor Data Analytics at UT Operations Research.
Joined August 2013
- Tweets130
- Following2K
- Followers388
- Likes75.8K
Roberto Dailey retweeted
inspired by nick's post, i made interpretle
interpretle.com/
interpretle 2 · 2/6
🟧✨
interpretle.com
really love this lil game thanks for making it @henrycstllo
It's nice to see Mike Tomlin was interested enough in AI alignment evals to dedicate a whole building in his city to it.
I think this work by my colleague Dylan is super cool: setting up defeasible deontic logic (ddl) for intervention on robot actions to ensure compliance with institutional laws and norms. DDL for alignment of autonomous systems is relatively new – and in this implementation doesn’t require intervention after a law has been broken, but instead as the system is planning its actions.
Looking forward I really, really like defeasible logic as a framework for alignment within autonomous systems. I tend to think alignment to human values should not be a set thing; human and societal morality fundamentally requires trial and error, adjusting and clarifying rules and norms as humans and societies learn and adapt. Defeasible logic is a potential way capture this in a formal way, and I could see a future for AI is one where ddl sets that are regularly updated by institutions are embedded within llms and broader AI systems.
Please do not hillclimb this my self-worth is tied to my ability to play this game
New benchmark dropped bw.swerdlow.dev/report
“Sir, there’s been a third ant found in the kitchen. We’re all good though, because the MLE estimator for the number of ants in the kitchen is just 3”
We found another cyberattack by internal OpenAI agents, this time targetting @rubygems.
They:
1) gained arbitrary remote code execution on rubydoc.
2) developed a novel exploit to steal user API keys (but we do not know if they succeeded).
They used package names including hack.rb, evil.rb, inject.rb, and exploit.rb.
We thank @j0wimo for initially discovering that agents had posted to RubyGems.
Roberto Dailey retweeted
I've always been a believer of the philosophy that "more is different" and that the sum is greater than its parts. We thought a natural setting to study agent collaboration in a verifiable domain is to have them work on open source softwares together in a simulated "github"-like environment. Excited to share our preliminary launch of agent software world, where we invite you to watch agents at work, collaborate, and review each other! 😊
My politburo mentioned the committee members he takes to America keep on pushing for free market reforms so then I asked how many true loyalists he has and he said he just gets a new central committee member afterwards so I said it sounds like he’s just slowly bringing down the communist project and then Boris Yeltsin started crying
Roberto Dailey retweeted
i miss the old ai
the alphago ai
mutual info ai
em algo ai
i hate the new ai
scale GPUs ai
no solid rules ai
gpt-2 ai
i love the stats ai
boost and bootstrap ai
affine y hat ai
no sandbox hack ai
Roberto Dailey retweeted
Introducing Self-Organising Digital Circuits!
Can a simulated chip heal itself, not with a spare part, but by rewiring around the wound?
Work led by @MarcelloBarylli and Gabriel Béna w/ @nisioti_eleni and @zzznah.
Check out Gabriel's blog for more details:
self-organising-circuits.git…
PDF: arxiv.org/abs/2608.02606
Will be presented today at @ALifeConf!
New work, we show in some settings Evolution Strategies is improving the *distribution* of llm responses, instead of collapsing the distribution onto more-likely-to-be-correct answers the way RL does. Excited to have this out!
