David Duvenaud @DavidDuvenaud

Machine learning prof @UofT. Working on generative models, inference, & latent structure. cs.toronto.edu/~duvenaud/ Joined July 2013

Tweets

736
Followers

28K
Following

3K
Likes

12K

Richard Ngo @RichardMCNgo

6 days ago

“Hate speech” is so subjective that penalizing it will always be an exercise in power politics. It should never be the target of legislation. The recent spate of authoritarian policies proposed in its name across the west are a lesson in the importance of free speech absolutism.

Muriel Blaive, PhD @MurielBlaivePhD

7 days ago

118 896 3K 334K 626

Download Image

11 4 97 10K 13

Sam Bowman @sleepinyourhat

4 weeks ago

Interesting and concerning new results from @cem__anil et al.: Many-shot prompting for harmful behavior gets predictably more effective at overcoming safety training with more examples, following a power law.

Anthropic @AnthropicAI

4 weeks ago

11 37 298 100K 81

Download Image

1 9 76 11K 17

Richard Ngo @RichardMCNgo

2 months ago

You might think that prediction markets are about prediction. But that would be insufficiently cynical. Instead I expect their biggest medium-term effect will be to create common knowledge about what most people already know but can’t express publicly.

20 21 257 24K 47

Holly ⏸️ Elmore @ilex_ulmus

2 months ago

Job loss from AI is bad, imo, bc jobs are a major way that ppl have ownership and security in society. It would be great if ppl didn’t have to work to live, but obsoleting their jobs doesn’t just take away their burdens— it strips them of their security and bargaining power.

13 5 73 11K 14

Daniel Johnson @_ddjohnson

2 months ago

New paper: How can you tell when a model is hallucinating? Let it cheat! An expert doesn't need to cheat, so if your model learns to cheat, there must be something it doesn't know. Our general new approach for measuring uncertainty: arxiv.org/abs/2402.08733

5 60 320 48K 233

Download Image

Ronny Fernandez 🔍⏸️ @RatOrthodox

5 months ago

I have a lot of sympathy for a certain kind of objection to my worldview. To summarize: "You literally want to use the threat of government force to stop me from doing certain kinds of computations because of your half baked pet theory built on an edifice of independent abstract…

61 30 388 422K 66

David Duvenaud @DavidDuvenaud

4 months ago

People who say "There is a 10% to 20% chance", what are you trying to communicate by giving a range? Why not just marginalize and give a single probability? I've heard super-forecasters say that probability ranges are useful, but I don't understand what they mean.

94 2 146 41K 23

Xander Davies @alxndrdavies

4 months ago

The UK's AI Safety Institute is hiring. In my (biased) view, this one of the best places to do AI research/engineering for the public good, with top talent & the resources / backing / access of gov. Super excited for even more technical ppl to join gov :) 🧵 on open roles! 1/9

9 83 231 115K 122

David Duvenaud @DavidDuvenaud

4 months ago

Love this new paper. They define deception, intent, and belief precisely enough to code them up. The definitions are mostly behavioral, so they can even be somewhat meaningfully applied even to messy unstructured agents like LLMs.

AI Safety Papers @safe_paper

5 months ago

1 5 45 26K 57

Download Image

2 6 95 18K 98

Andrew Critch (h/acc) @AndrewCritchPhD

6 months ago

In 2021, I publicly released these AI disaster scenarios that I found especially plausible: "Production Webs", "Flash Wars", and "Flash Economies". Now in 2023, these scenarios have stood the test of time — they're plausible to many more people now that GPT-4 is out, and…

7 10 53 10K 24

Dwarkesh Patel @dwarkesh_sp

6 months ago

Talked with Paul Christiano (world’s leading AI safety researcher) about: - Does he regret inventing RLHF? - What do we want post-AGI world to look like (do we want to keep gods enslaved forever)? - Why he has relatively modest timelines (40% by 2040, 15% by 2030), - Why he’s…

6 33 186 31K 66

Download Video

Richard Ngo @RichardMCNgo

6 months ago

I find the techno-optimist philosophy defended by @pmarca inspiring in many ways. I think it *used* to be the best approach to improving the world. But as I argue in the post linked below, the biggest problem we face today is not lack of growth but rather lack of understanding.

11 20 123 24K 46

Download Image

Daniel @growing_daniel

7 months ago

Every person who works in AI is like “I’m super worried about a god AI that destroys humanity so my solution is build it”

89 155 2K 304K 74

Yo Shavit @yonashav

7 months ago

To quote @jeremyphoward, there are no perfect metaphors for AI. AI isn’t nukes (civilian apps are central and risks more complex). AI also isn’t the printing press, bc it’s not a pure reflection of human operators’ will (& will grow less so). My suggestion: AI is more like LLCs.

12 19 106 33K 30

Samuel Albanie @SamuelAlbanie

10 months ago

Can you prove which data was used to train an AI? New techniques from @damichoi95, @yonashav and @DavidDuvenaud suggest the answer may be "yes" An overview of the work 👇 youtu.be/89I1XXdMMvw