Video wird geladen...

Video konnte nicht geladen werden

Zur Startseite

Safety-oriented interpretability researchers should be focused on AI systems, not individual model artifacts. A snippet from the NeurIPS CogInterp workshop panel on Sunday:

ChrisGPotts's profile picture

Christopher Potts

16,356 subscribers

16,504 Aufrufe • vor 9 Monaten •via X (Twitter)

LIVE

Anya Rossi• Live Now

Private livecam show

6 Kommentare

Profilbild von ‏عبد الهادي‏
‏عبد الهادي‏vor 9 Monaten

Hello Prof. Potts, you often emphasize compound systems—how challenging are they to interpret mechanistically?

Profilbild von Christopher Potts
Christopher Pottsvor 9 Monaten

@AlgerianAb26687 I think they will be challenging to interpret because they have all the complexity of traditional software and all the complexity of foundation models.

Profilbild von arya
aryavor 9 Monaten

I'm confused why would these systems be fundamentally that different from individual model artifacts? And also aren't systems being studied via creating tool calling envs?

Profilbild von Christopher Potts
Christopher Pottsvor 9 Monaten

We have an overview of the differences between models and compound AI systems here: My concern is specifically about what interp researchers are doing; I know security folks and others are focused on the nature and integrity of compound AI systems.

Profilbild von arya
aryavor 9 Monaten

read the article - do you have examples of interp work that could be done on a compound system?

Profilbild von Himanshu Kumar
Himanshu Kumarvor 9 Monaten

Christopher is right, focusing on the AI system is crucial for safety, not just individual components.

Ähnliche Videos

Snippet from The Forever Model workshop this morning… Should I drop the full video this week on Youtube?
1:37

Snippet from The Forever Model workshop this morning… Should I drop the full video this week on Youtube?

$𝘯𝘪𝘱𝘦𝘳

304,622 Aufrufe • vor 2 Jahren

We discussed AI safety with Yoshua Bengio at NeurIPS 2024 - he proposes that we should design agency out of AI systems. Releasing on MLST soon.
2:22

We discussed AI safety with Yoshua Bengio at NeurIPS 2024 - he proposes that we should design agency out of AI systems. Releasing on MLST soon.

Machine Learning Street Talk

21,429 Aufrufe • vor 1 Jahr

“The hope is that ... just optimizing something to be sparse—without optimizing it to be interpretable—will stumble across that interpretable decomposition.” — Neel Nanda on sparse autoencoders for mechanistic interpretability and AI safety at the Vienna Alignment Workshop.
1:36

“The hope is that ... just optimizing something to be sparse—without optimizing it to be interpretable—will stumble across that interpretable decomposition.” — Neel Nanda on sparse autoencoders for mechanistic interpretability and AI safety at the Vienna Alignment Workshop.

FAR.AI

1,431,043 Aufrufe • vor 2 Jahren

The o3 model is in 'preview' and only open to safety and security researchers who apply through the link on their site. Recently, Sam Altman said there should be a federal testing framework to ensure safety before release, so the cautiousness makes sense
0:23

The o3 model is in 'preview' and only open to safety and security researchers who apply through the link on their site. Recently, Sam Altman said there should be a federal testing framework to ensure safety before release, so the cautiousness makes sense

Rowan Cheung

69,710 Aufrufe • vor 1 Jahr

No child should be a guinea pig for unregulated AI. That's why I am calling for an AI Child Safety Pledge built on three rules: safety testing & oversight, zero tolerance for abuse & crisis support systems. When a child is harmed, the answer must never be “the algorithm did it.”
1:38

No child should be a guinea pig for unregulated AI. That's why I am calling for an AI Child Safety Pledge built on three rules: safety testing & oversight, zero tolerance for abuse & crisis support systems. When a child is harmed, the answer must never be “the algorithm did it.”

António Guterres

70,936 Aufrufe • vor 2 Monaten

Should we be worried about the 'mind' of an AI model? What does an LLM think but not say? Join Inner Cosmos this week with Anthropic researcher Jack Lindsey as we discuss the mind-bending world of AI interpretability.
1:20

Should we be worried about the 'mind' of an AI model? What does an LLM think but not say? Join Inner Cosmos this week with Anthropic researcher Jack Lindsey as we discuss the mind-bending world of AI interpretability.

David Eagleman

13,587 Aufrufe • vor 17 Tagen

Meta president Dina Powell McCormick on AI safety and guardrails: "Safety is not something you think about after you build the model. It is embedded in the model."
3:47

Meta president Dina Powell McCormick on AI safety and guardrails: "Safety is not something you think about after you build the model. It is embedded in the model."

CSPAN

12,628 Aufrufe • vor 1 Tag

Yesterday, a federal judge barred Elon Musk's lawyers from arguing that AI could threaten humanity in his lawsuit against OpenAI. OpenAI was founded as a nonprofit focused on developing AI safely. But our recent The New Yorker investigation documented how some researchers at the company have raised concerns about safety being sidelined.
2:52

Yesterday, a federal judge barred Elon Musk's lawyers from arguing that AI could threaten humanity in his lawsuit against OpenAI. OpenAI was founded as a nonprofit focused on developing AI safely. But our recent The New Yorker investigation documented how some researchers at the company have raised concerns about safety being sidelined.

Ronan Farrow

255,661 Aufrufe • vor 4 Monaten

Rep. Smith says Trump is too focused on renovation projects: This is not what a democratically elected president in a representative democracy should be focused on. This is much more what an authoritarian dictator would be focused on.
1:42

Rep. Smith says Trump is too focused on renovation projects: This is not what a democratically elected president in a representative democracy should be focused on. This is much more what an authoritarian dictator would be focused on.

Acyn

48,953 Aufrufe • vor 2 Monaten

Trump should be focused on lowering costs for Americans not on a vanity remodel to make the White House look like Mar-a-Lago.
0:36

Trump should be focused on lowering costs for Americans not on a vanity remodel to make the White House look like Mar-a-Lago.

Meghan Hays

28,960 Aufrufe • vor 11 Monaten

Trotskyists: “The left should not be focused on winning back Pauline Hanson voters” Great idea geniuses
0:51

Trotskyists: “The left should not be focused on winning back Pauline Hanson voters” Great idea geniuses

Drew Pavlou 🇦🇺🇺🇸🇺🇦🇹🇼

16,807 Aufrufe • vor 3 Monaten

This might be the beginning of a new field.. Researchers just reconstructed sounds from human brain activity using an fMRI and a generative AI model.
0:41

This might be the beginning of a new field.. Researchers just reconstructed sounds from human brain activity using an fMRI and a generative AI model.

Lior Alexander

208,384 Aufrufe • vor 3 Jahren

OpenAI is focused on “AI safety” while falling very short on AI security. This incident is a *security* problem that requires *security* solutions. And if there’s anything to learn from it, we need the audit log.
3:10

OpenAI is focused on “AI safety” while falling very short on AI security. This incident is a *security* problem that requires *security* solutions. And if there’s anything to learn from it, we need the audit log.

Zack Korman

17,294 Aufrufe • vor 1 Monat

As AI moves into the physical world, safety must be the top priority. Leaders from Zoox, Aurora, Boston Dynamics, ANSI and NVIDIA discuss the innovations and safety practices shaping the next generation of autonomous vehicles and robotics. 🎥 Watch the full panel on-demand:
0:58

As AI moves into the physical world, safety must be the top priority. Leaders from Zoox, Aurora, Boston Dynamics, ANSI and NVIDIA discuss the innovations and safety practices shaping the next generation of autonomous vehicles and robotics. 🎥 Watch the full panel on-demand:

NVIDIA DRIVE

12,446 Aufrufe • vor 9 Monaten

Leader Jeffries: Republicans should not be focused on massive tax cuts for the wealthy and the well-off and the well-connected. We should all be focused on making life better for working-class Americans, including protecting and strengthening social security.
0:38

Leader Jeffries: Republicans should not be focused on massive tax cuts for the wealthy and the well-off and the well-connected. We should all be focused on making life better for working-class Americans, including protecting and strengthening social security.

House Judiciary Dems

135,829 Aufrufe • vor 1 Jahr

The Vice President knows my name. But that’s not the point. He should be focused on removing the thousands of unnecessary troops from the streets of Los Angeles, not petty slights.
1:33

The Vice President knows my name. But that’s not the point. He should be focused on removing the thousands of unnecessary troops from the streets of Los Angeles, not petty slights.

Senator Alex Padilla

814,257 Aufrufe • vor 1 Jahr

Everyone’s focused on what the latest AI model can do. Flick is focused on what creators can actually make with it. 17 filmmakers created Real stories, real characters, real creative vision. Made entirely on Flick. Fully open-sourced. Not just AI demos—finished films.Flick
0:46

Everyone’s focused on what the latest AI model can do. Flick is focused on what creators can actually make with it. 17 filmmakers created Real stories, real characters, real creative vision. Made entirely on Flick. Fully open-sourced. Not just AI demos—finished films.Flick

William AI

105,023 Aufrufe • vor 17 Tagen