💡Grant opportunity - share with your ML/assurance friends💡
My program at has put out our first public call for research ideas, on "AI assurance for general-purpose systems in open-ended domains."
Summary below, full details here: cset.georgetown.edu/wp-content/upl
Luke Muehlhauser
@lukeprog
Open Philanthropy Program Officer, AI Governance and Policy. Summary: openphilanthropy.org/blog/ai-govern
Luke Muehlhauser’s Tweets
Many people, including me, have been surprised by recent developments in machine learning. To be less surprised in the future, we should make and discuss specific projections about future models. In this spirit, I predict properties of models in 2030:
22
132
544
Show this thread
New piece from , and me pushing back on the claim that we can't regulate AI because that would just let China pull ahead. This is not a good argument!
4 reasons why, though first a caveat: 🧵
Quote Tweet
“Even if it were true that new AI regulations would slow innovation in the United States—and it very well may not be—China does not appear poised to surge ahead.”
trib.al/txcFcJu
14
133
285
Show this thread
3
11
Few things are harder to predict than the ways in which someone much smarter than you might outsmart you.
That's the reason I worry about AI. Not just the problem itself, but the meta-problem that the ways the problem might play out are inherently hard to predict.
209
360
3,518
Show this thread
As a society, we should really do something about this.
8
7
53
Indeed. And I would say this is more than "a bit" alarming.
Quote Tweet
I tend to agree with @robertskmiles's statement that "Making safe AGI is like writing a secure operating system." But if so that's a bit alarming, because, like the US tax code, the development of this field was a series of responses to successful exploits.
3
5
67
As increasingly capable AI models are trained, model evaluations for dangerous capabilities and alignment will become crucial to inform decisions about whether and how models are deployed.
More in new paper: "Model evaluation for extreme risks."
arxiv.org/abs/2305.15324
5
56
150
Show this thread
We’re excited to share the results of our recent expert survey on best practices in AGI safety and governance!
Paper: arxiv.org/abs/2305.07153
Co-authors: , , David McCaffary, , , Ben Garfinkel
Summary in the thread 🧵
4
57
132
Show this thread
Huge win! The Supreme Court just upheld California’s Prop 12, the nation’s strongest farm animal welfare law.
This ruling spares millions of animals from crates and sets a powerful precedent that could help many more.
More details ⬇️
supremecourt.gov/opinions/22pdf
41
377
1,468
Show this thread
A common criticism of people who are trying to stop existential risk from powerful future AI systems is that speculating about the future has poor feedback loops and doesn't work great.
10
25
175
Show this thread
Reminder that we have >$200K of prizes for essays that could change our mind on AI risks, due by May 31:
8
28
94
It is so heartening to see the gradual emergence of well-respected figures in AI candidly sharing their concerns about existential risks from AI.
If this is you, now seems like a great time to speak up!
Quote Tweet
I share Geoff’s concerns. It is of the utmost need that we continue to develop intelligent tools responsibly, e.g. to fight disease. As scientists and engineers, it is also our moral imperative to point out existential risks when we suspect them. youtu.be/DsBGaHywRhs
2
9
90
Geoff Hinton quit Google so he can speak freely on the risks of AI.
"He is worried that future versions of the technology pose a threat to humanity"
26
178
621
Show this thread
If ever a bill was self-recommending, here is the "Block Nuclear Launch by Autonomous Artificial Intelligence Act":
21
53
221
It's fair to have a general background worldview that's pro-innovation & anti-regulation.
But if I found myself saying almost the same thing both about nuclear power (very safe) & the hydrogen bomb (very unsafe), then I'd seriously worry I was leaning too hard on that worldview.
2
7
55
Plausibly 's most cost-effective donation. She should triple it! malariaconsortium.org/news-centre/ma
1
2
36
I'm among the big group of people who think society often frets about scaling up technologies that are good:
1. nuclear power
2. wind/solar in the country
3. CRISPR to cure human diseases
4. pre-safety-test vaccines
5. apartment blocks
6. etc etc
But there are 2 big exceptions!
6
29
178
Show this thread
12 good ideas from for US AI policy: openphilanthropy.org/research/12-te
There's a call to action at the end to let us know if you're interested in helping work towards ideas like these!
101
173
239
6
39
This is an absolutely incredible video.
Hinton: “That's an issue, right. We have to think hard about how to control that.“
Reporter: “Can we?“
Hinton: “We don't know. We haven't been there yet. But we can try!“
Reporter: “That seems kind of concerning?“
Hinton: “Uh, yes!“
71
768
2,785
Show this thread
"Why would you think AI might end up displaying [deception/power-seeking/other scary thing]?"
"People will design them to"
"But those theorems might not apply to the real wo.... WAIT WHAT?"
"People will design them to"
11
80
512
Show this thread
In general, people should be much more positive about technological progress.
But with one exception: technologies that might end the world.
14
26
235
In case you missed it, GPT-4 is capable of deception:
13
20
114
If you're interested in both effective altruism and information security, consider joining this great-looking reading group, which starts Apr. 1st!
2
21
The GPT-4 safety card gets pretty wild in some places. Like how ARC used GPT-4 to simulate a rogue AI on the internet trying to replicate itself.
Section 2.9 of cdn.openai.com/papers/gpt-4-s
4
46
166
Show this thread
Really great to see pre-deployment AI risk evals like this starting to happen
49
675
2,101
"One of two things must happen. Humanity needs to accelerate its adaptation to [AI progress] or a collective, enforceable decision must be made to slow the development of [AI]. Even doing both may not be enough."
3
16
60
Win money by helping to change our views about AI timelines and AI risk!
4
18
Once there are entities that are capable of doing anything people can do but which are smarter and better at it than people, people are no longer in control of civilization, those entities are. This is not like any previous technical transition, full stop.
5
12
65
Show this thread
Quote Tweet
"nowhere near solved" ... from "A brief history of AI", published in january 2021.
Show this thread
7
38
233
If you spend much time on AI twitter, you might have seen this tentacle monster hanging around. But what is it, and what does it have to do with ChatGPT?
It's kind of a long story. But it's worth it! It even ends with cake 🍰
THREAD:
42
666
2,632
Show this thread
Very cool that Kahneman did this adversarial collaboration:
4
18























