Accenture Will Help Anthropic Test AI Model Safety
Read the Portuguese translation →
Author: Bloomberg Television · Watch on YouTube
Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.
Original language: English
Original video: Accenture Will Help Anthropic Test AI Model Safety · Channel: Bloomberg Television · 02:38 · Watch on YouTube
Topics: AI evaluation and benchmarks, AI governance and regulation, AI safety
Summary
Section titled “Summary”The discussion focuses on how companies are trying to defend against risks from increasingly powerful AI models, as cyber threats evolve and firms patch their systems. Bloomberg’s Ed Ludlow says the debate has moved away from the previous week’s panic about existential risk, with some experts expressing skepticism about that threat. Anthropic has agreed to let Accenture personnel access and red-team its models before deployment, in a non-exclusive arrangement. The broader challenge is establishing independent oversight without allowing frontier AI labs to set rules that serve their own interests; the speakers say this work will require multiple parties.
Key points
Section titled “Key points”- Citigroup CEO Jane Fraser says AI-related threats are evolving quickly and businesses need to strengthen their defenses.
- Companies are undertaking extensive patching to secure their perimeters against potential negative effects of AI models.
- The discussion of existential risks from AI has become less alarmed, with some people saying such a threat is near zero.
- Anthropic and Accenture have an agreement for Accenture personnel to access and red-team Anthropic models before deployment.
- Red teaming means deliberately trying to break models and identify failures.
- The Anthropic–Accenture arrangement is non-exclusive, and Accenture is not Anthropic’s sole auditor.
- The industry is considering how to secure independent evaluations without creating a cartel or regulatory capture.
- The labs want outside parties to check their work, and the discussion suggests that oversight will require multiple parties.
Chapters
Section titled “Chapters”| Start | Chapter | Summary |
|---|---|---|
| 00:00 | AI risks and corporate cyber defenses | Jane Fraser describes the rapidly evolving threat and the effort to patch and strengthen companies’ defenses. |
| 00:30 | Debate over existential AI risk | Ed Ludlow says the discussion has moved away from last week’s panic, with some experts questioning the scale of existential risk. |
| 01:03 | Anthropic’s red-team agreement with Accenture | The companies’ non-exclusive agreement gives Accenture personnel access to test Anthropic models and identify failures before deployment. |
| 02:07 | Independent oversight and industry rules | The speakers discuss avoiding regulatory capture and the need for multiple outside parties to review AI labs’ work. |
Quotes
Section titled “Quotes”there is a race to defend and shore up all of the firm’s perimeters — 00:00
they try and break them — 01:03
Accenture doesn’t become the sole auditor of anthropic. — 01:35
how do they get to a place where there is not a cartel formed? — 02:07
It will take multiple parties. — 02:38
Concepts and entities
Section titled “Concepts and entities”Anthropic, Accenture, red teaming, OpenAI, Sam Altman, Jane Fraser, Jensen Huang, Dario Amodei, David Sacks, captura regulatória
Full transcript (original language)
Section titled “Full transcript (original language)”00:00 Also on the docket for this week, talks around AI in his AI models grow more powerful. Companies are racing to make sure their cyber defenses can keep up. Citigroup CEO Jane Frazier says the threat is evolving quickly and businesses need to be ready. There is a race to defend and shore up all of the firm’s perimeters, and there is a tsunami of patching going on in the world at the moment in all companies, as we make sure that we are trying to secure the defense around the perimeter
00:30 of all our firms against the negative side of the models. Let’s bring in Bloomberg Tech host Ed Ludlow. Ed, it’s a discussion that will be on everyone’s minds as the UN and China in the U.S. meet. Sam Altman of OpenAI also set to address the U.N. Security Council. How is this conversation taking shape? We are at least far away from the panic that we started last week with. Yeah, we ended the week with people saying, actually, the idea that there’s sort of an existential threat to human life from AI is not universally believed, and lots of people came out and said that we believe it’s near to
01:03 zero. From leading academics to Jensen Wang himself speaking over the weekend. And what’s interesting is like, you know, how is this being acted on? So the big news that broke Friday and is carried over into this Monday equity session is anthropic doing an agreement with Accenture where Accenture, people go into anthropic get employee level access and red team models before they’re deployed. Red teaming in that context means they try and break them. They try and identify failures deliberately. But that is an interesting development and an interesting story, because if you
01:35 remember Dario Almodovar’s essay from, uh, Saturday prior to last, he said, we want third party independent evaluation of the models. And the deal between anthropic and Accenture is not exclusive meaning. And very significantly, Accenture doesn’t become the sole auditor of anthropic. Um, but these kind of arrangements, um, you know, expect to see them and it is, uh, substantive. so we knew a century already had somewhat of a relationship. There is there up to for the other sort of consultancy firms are we expect this
02:07 more to be siloed. And basically this is going to be a winner take all situation. Yeah, it’s why I ask I flagged the exclusivity thing. There seems to be, uh, talks between anthropic and other third parties for people that could come in. The main point, uh, the, the industry’s grappling with is how do they get to a place where there is not a cartel formed? In other words, um, there is no regulatory capture by which the frontier labs are setting their own rules to their own benefit. And, you know, you have people like David Sachs saying pretty robustly,
02:38 like, you guys, sort this out yourselves. This is an engineering problem. According to Jensen Long, if you’re so good at I, you should be able to make sure the models are safe prior to release. The other side of the argument is that some oversight is needed, right? And the labs themselves want somebody to come in and just check their work and like, that’s a big body of work. It will take multiple parties.