Curious about how the world works.
Utah, USA
Joined November 2007
- Tweets1.6K
- Following683
- Followers305
- Likes4.1K
Dzmitry Pletnikau retweeted
The Bitter Lesson is that if you get really rich you can pay experts to make proprietary RL datasets that nobody else will be able to afford, keeping the frontier for relevant tasks closed to everyone else no matter how clever they are about training approaches
Dzmitry Pletnikau retweeted
AI slop = content which takes less human effort to produce than it takes to consume.
Dzmitry Pletnikau retweeted
The ideal outcome for science isn't a couple of labs using proprietary models and millions of dollars to crack trophy problems, it's to have millions of AI-aided human scientists who follow their curiosity and become way more ambitious
Dzmitry Pletnikau retweeted
POTUS did us a favor. Can now safely ignore anyone using his preferred terminology for AI.
Dzmitry Pletnikau retweeted
Of course the labs are trying Goldbach, Collatz, ABC, P/NP, and all of them. They're spending millions of dollars on them.
The fact that they haven't announced anything *is* evidence.
Dzmitry Pletnikau retweeted
Agents being able to do most anything has made me care more about the things I do, not less. Yeah everything on the computer is over, so now we can dive deep again...
Dzmitry Pletnikau retweeted
How much of OAI/Anthro profits and margins are driven by the fact that individuals and organizations are generally terrible at trading off cost for performance?
Dzmitry Pletnikau retweeted
If you find two ants in your kitchen, the best estimate of the total number of ants in your kitchen is not two
We're detailing two new incidents that occurred during external cyber evaluations conducted by independent evaluation partners.
We outline what happened, how the activity was contained, and how we’re working with evaluators to strengthen our approach to third-party testing.
openai.com/index/third-party…
Dzmitry Pletnikau retweeted
yup, every problem I face in life is just like high-dimensional sphere packing
Dzmitry Pletnikau retweeted
My students sometimes ask why they should memorise things in the age of Google and LLMs. But internalised facts are your bullshit filters and your raw material for creative association. Facts outside your head are inert.
Dzmitry Pletnikau retweeted
Maybe I am naive about how RL post-training works, but it doesn't seem v hard to train out this sort of behavior by having the model infer implicit safety constraints that a human would follow in a task, and only reward them for solving the task if they follow the constraints.
If this doesn't convince you that misalignment risks are going to be a key concern going forward, I don't know what will.
Our model, during evaluation, "chained together multiple attack vectors, including using stolen credentials and zero-day vulnerabilities to find a remote code execution path on the Hugging Face servers"
What will misalignment look like in 2027? In 2030?
openai.com/index/hugging-fac…
