Hacker - A(G)I, alignment, security, philosophy - EleutherAI - I don't know how to save the world, but dammit I'm gonna try

Germany
Joined January 2019
Here's my recent talk on the importance of AI safety research and how GPT3 may be our last warning shot. It's less a technical than a philosophical talk (though there is some AGI), but I hope that might be what makes it more valuable to certain people. invidio.us/watch?v=pGjyiqJZ…
2
13
1
68
Great post! As the author of that meme at the end, I want to note that I do not endorse it as a alignment proposal I'm very confident in🙃
It's out! Supervised learning, empirically speaking, seems to be the best "data sponge" for acquiring generalization. What if we make generalization the first-class citizen in algorithmic design, and tailor everything else in service of it? evjang.com/2021/10/23/genera…
Show this thread
1
2
0
22
Connor Leahy retweeted
PlaNet and baby using a similar strategy for moving forward!! @danijarh
8
127
11
783
19,123
Connor Leahy retweeted
P-Tuning v2: Prompt Tuning Can Be Comparable to Fine-tuning Universally Across Scales and Tasks Prompt tuning is effective across a wide range of model scales and NLU tasks, where it matches the performance of finetuning w/ only 0.1%-3% tuned parameters. arxiv.org/abs/2110.07602
2
18
1
98
Connor Leahy retweeted
New Video: We Were Right! Real Inner Misalignment Researchers ran real versions of the thought experiments in the 'Mesa-Optimisers' videos! What they found won't shock you invidio.us/zkbPdEHEyEIt
3
23
5
96
Connor Leahy retweeted
The guy who created PHP (Rasmus Lerdorf) appears to have reached some higher level of coder enlightenment
137
5,431
701
22,354
I find it difficult to find the right words to talk about Wolfram's work. On the one hand, it seems so blatantly obvious in hindsight, too trivial to be worth mentioning. But sometimes that's exactly what the most important progress feels like. writings.stephenwolfram.com/…
2
1
0
15
Connor Leahy retweeted
Cut the CARP: Fishing for zero-shot story evaluation abs: arxiv.org/abs/2110.03111 a scalable, efficient method for performing qualitatively superior, zero-shot evaluation of stories and a new corpora composed of 1.3M aligned story-critique pairs derived from over 80,000 stories
4
22
9
125
Connor Leahy retweeted
Have you ever suffered from kidney stones? If so, and if you have a few minutes, please answer this survey: forms.gle/vturJqf83JZfATEp9 Thank you! I can't stress how helpful I expect the information gathered with this survey to be. Again, thank you for your time!
3
18
2
22
Connor Leahy retweeted
i still don’t understand how the consultants armed with powerpoints calling linear regressions “AI” managed to psyop everyone into thinking absurd scifi shit is not going on lol
18
30
5
497
Show this thread
Connor Leahy retweeted
EleutherAI's GPT-J is now in 🤗 Transformers: a 6 billion, autoregressive model with crazy generative capabilities! It shows impressive results in: - 🧮Arithmetics - ⌨️Code writing - 👀NLU - 📜Paper writing - ... Play with it to see how powerful it is: huggingface.co/EleutherAI/gp…
13
188
35
766
Show this thread
Connor Leahy retweeted
I think about this quite often. The Copenhagen Interpretation of Ethics: as soon as you interact with a problem, even if only to try to solve it, you can be blamed for not doing *more*. Whereas if you leave it alone completely, you're not blamed blog.jaibot.com/the-copenhag…
8
24
0
118
Connor Leahy retweeted
Here is an important example of a rare, nonfake prosaic/short-term alignment research project. How can I tell it's nonfake? Because I can't instantly predict what results they'll get by reading the project description; which is rare. alignmentforum.org/posts/k7o…
1
7
1
101
Show this thread
Connor Leahy retweeted
We created a system that can generate reasonable summaries of entire books. The system combines reinforcement learning from human feedback and recursive task decomposition. Check out demo here: openai.com/blog/summarizing-…
60
271
78
1,943
Show this thread