Pegah Maham
@pegahbyte
Followers
573
Following
1K
Media
14
Statuses
201
Is it true or is it just confirmation bias? || Freiheitlich-demokratische Grundordnung || overcame reductionism Policy Development & Strategy @GoogleDeepMind
London
Joined October 2021
RT @janleike: humans built machines that talk to us like people do and everyone acts like this is normal now. it's pretty nuts.
0
134
0
AI governance is often more like improv comedy than I expected: don't try to come up with a clever idea, instead just focus on the present moment and execute well on the just-next open thing: build taxonomies / frameworks, substantiate each aspect, repeat 🔁.
1
0
16
RT @fmf_org: Frontier AI safety frameworks have emerged as a critical tool for managing potential risks to public safety and security. In….
0
5
0
RT @BethMayBarnes: Benchmarks saturate quickly, but don’t translate well to real-world impact. *Something* is going up very fast, but not c….
0
60
0
And everything that does not change the model weights can be "post- training".
0
0
4
I propose we start calling "post-training" -> "secondary training" and "pre-training" -> "initial training".
2
0
8
RT @pegahbyte: @DAcemogluMIT "Whether near-term AGI is an achievable goal remains an open question". Are your estimates on jobs (5% heavily….
0
1
0
RT @geoffreyirving: Someone should publish a safety paper that flips all the experimental results exactly around, watch to see who details….
0
8
0
"I saw a grandson and their grandfather last week outside the Walmart trying to book a cab on Uber. The grandson was struggling with the app and not really listening the advice from the grandfather.". "Who was giving the other the opportunity to learn how to use the phone?". If.
AI capabilities are tracked by public and up-to-date evaluations (e.g. Chatbot Arena, OpenLLM, HELM). AI governance requires similar evals to track progress on risks. To foster a more empirical approach to AI safety, we are launching HELM Safety:
1
1
10
I agree with this in principle and overall. On the ground, at the moment, I see the following downsides, and the need to substantiate the high level principles. - It seems to me that there is a significant lack of agreement on what exactly a critical dangerous capability is.
We need an AI governance approach that works for both safety advocates and skeptics. I think "if-then policies" /RSPs/FSPs could be it. Instead of debating *when* dangerous capabilities will arise, we can debate specifics about how to test and respond. Here's why this matters:.
2
3
24
RT @Plinz: the reason that science turned into slop in the second half of the 20th century is that it became a model trained mostly on its….
0
269
0
RT @Lewis_Bollard: For decades, animal advocates lobbied to stop the worst abuses of farm animals -- with limited results. Then they chang….
0
70
0
RT @EgeErdil2: the take that "we can't use probabilities to model risk X, so we should behave as if its probability of occurring is zero" i….
0
21
0
RT @rohinmshah: I've been super impressed at the speed with which our interpretability team gets stuff done. Their previous paper (also Sot….
0
2
0
RT @emollick: Since it is in the news, what people don't understand about ChatGPT & bias is that, while there are real biases in ChatGPT's….
0
122
0
RT @geoffreyirving: Pet peeve: people claiming LLMs are hard to understand because they are nondeterministic. 1. A coin flip: Very easy to….
0
3
0
RT @Liv_Boeree: I received my bluecheck back in ~2010 or so because I was a minor poker celebrity who happened to know a guy who knew anoth….
0
94
0
RT @emollick: A favorite finding on crowd wisdom: don't do any research. When individuals research before making guesses, it makes their o….
0
86
0