🔬 AI safety & transformations of capitalism @AIObjectives. Prev. led tech @EFF & research @PartnershipAI. Cofounder @letsencrypt #certbot & @privacybadger
San Francisco & Australia
Joined June 2015
- Tweets269
- Following656
- Followers1.2K
- Likes138
The probability distribution of all art nouveau images!
ALT The probability distribution of all art nouveau images...
Peter Eckersley retweeted
You can see Earendel as it's lensed by the cluster, WHL0137-08, in the center of the image! Here is a zoom-in on the star itself! Stay tuned for an update from the science team!
We are so lucky to have such a flukishly, magnificently lensed view of such a distant star as Earendel! Eagerly waiting for spectroscopy...
There some strange travel crises in the world these days. I've seen friends and colleagues affected by a couple of them, but are there others? 1/n
There's an airline luggage handling crisis that sounds close to a positive feedback loop tipping point theguardian.com/travel/2022/… 3/n
Peter Eckersley retweeted
We need a logo for the AI objectives institute... a thread 🧵to pick one! Retweets and likes are votes :)
So generative models dream up objects like this at the drop of a whimsical hat. Perhaps one day there will be a way to make them, so that our world is filled with such things?
Excited to have worked on a little corner of this grand, ambitious project! @machinaut @realSharonZhou & I added some simple calibration measures to BigBench. ML models are usually overconfident, it's past time to start measuring & mitigating that!
After 2 years of work by 442 contributors across 132 institutions, I am thrilled to announce that the github.com/google/BIG-bench paper is now live: arxiv.org/abs/2206.04615. BIG-bench consists of 204 diverse tasks to measure and extrapolate the capabilities of large language models.
Speaking of which, there's also a great new paper from Stephanie Lin, Jacob Hilton and @OwainEvans_UK on calibration in words rather than logits, an approach that should really be in future versions of BigBench... arxiv.org/pdf/2205.14334.pdf
I used to think that laypeople would firmly believe that software can't be conscious, until long after conscious AI was present to argue with them. 1/2
Great write up of the situation 🙏✨Language models are almost certainly slightly conscious, and to a degree we don't yet understand or know how to measure properly.
Now promoting a consciousness piece I wrote before:
If language models can generate a "stream of consciousness" indistinguishable from humans, why aren't they conscious?
jasonwei20.github.io/files/a…
Peter Eckersley retweeted
Reminder: you should almost never deploy a high-stakes AI or decision-making system that has a single, definite objective. 1/3
Peter Eckersley retweeted
OpenAI's recent models are well-trained to disclaim or avoid answering political questions 1/2
Peter Eckersley retweeted
🧵Conclusion🧵
If democracy is to survive and thrive, we must be deeply intentional about the algorithms that mediate our communication.
Engagement-based ranking isn't working for us.😠
Bridging-based ranking can help.🌉
Read the full report here ⤵️
belfercenter.org/publication…
If you haven't seen this inadvertently truth-filled confession by President Bush, it is the mother of all Freudian slips youtube.com/watch?v=ZEg6Ht2p… 1/2