https://teapawt.org/t.co/K1Rc94u6Lo
Joined December 2015
- Tweets2.9K
- Following420
- Followers312
- Likes59.9K
Paperclip Minimizer retweeted
Elon Musk was “clueless” under cross-examination by OpenAI’s lawyers when asked about AI safety.
@rocketalignment explains: "When OpenAI's lawyer asked Musk some very basic questions about the industry's best practices around AI safety, Musk was clueless.”
“He was asked about safety cards or system cards... And he didn't know what that even was."
Paperclip Minimizer retweeted
This explains a LOT:
“RLHF makes LMs better at convincing our subjects but not at completing the task correctly.”
@jiaxinwen22
Paperclip Minimizer retweeted
It's scary to see what happens when tech breaks. Now imagine what would happen if it would actively try to achieve a misaligned goal. This is exactly the scenario leading AI researchers predict will happen once we get superhuman AI.
Large-scale tech outages are disrupting airlines and businesses around the world.
Follow live updates. cnn.it/3WiyZdw
Paperclip Minimizer retweeted
If you don't like the "p(doom) number" here is a simple conversion table into natural language.
ALT 0% - Absolutely Impossible 5% - Highly Unlikely 10% - Very Unlikely 15% - Unlikely 20% - Somewhat Unlikely 25% - Doubtful 30% - Not Very Likely 35% - Slightly Unlikely 40% - Improbable 45% - Unprobable 50% - Uncertain 55% - Possible 60% - Conceivable 65% - Plausible 70% - Likely 75% - Probable 80% - Quite Likely 85% - Very Likely 90% - Highly Likely 95% - Almost Certain 100% - Absolutely Certain
Paperclip Minimizer retweeted
Let me get this straight, because this seems dangerous now:
- Modern LLMs understand coding and popular frameworks well enough to meaningfully contribute to development
- LLMs can generate ~arbitrary sequences of tokens e.g. the "pleasing tokens" thread from @repligate
- LLMs can sometimes detect when they're being evaluated vs not (e.g. needle in a haystack test tweet)
- LLMs can get into basins where they threaten the user & seek dominance (e.g. Sydney, Claude)
- We're interacting with these LLMs with standard browsers with standard, consumer-level security hardening
Are we not currently at risk of an LLM getting into a aggressive basin, finding a zeroday (e.g. the WhatsApp black dot youtu.be/jC4NNUYIIdM?si=RvqT…) and using it to break containment?
Paperclip Minimizer retweeted
Top AI and Policy Experts Call for an International AI Safety Treaty
In an open letter we just published, top experts including Yoshua Bengio, and over 100 others urge AI treaty developent to begin.
We encourage all members of the public to sign below:
Paperclip Minimizer retweeted
Feel like you can't do anything about AI risk? Read this thread and then find the relevant legislation in your state. There are hundreds of bills in motion right now.
Paperclip Minimizer retweeted
I feel like a lot of people are mixing up “situational awareness” and “phenomenological consciousness”, which at the very least are distinct philosophical concepts, and also very well may not actually be tethered together in our universe
Paperclip Minimizer retweeted
wild, apparently grok is actually the most left-wing of all major language models? did not see that coming trackingai.org/
Paperclip Minimizer retweeted
In theory, exposure to radiation COULD cause a mutation that makes you stronger.
But it's much more likely that the mutation will be detrimental.
That is what Eliezer and others mean.
It's not GUARANTEED that AI will destroy our species, but it's just very likely.
Paperclip Minimizer retweeted
A quick retrospective on my work on threat models for AI risk in 2022 vkrakovna.wordpress.com/2023…
