Top AI players pledge to pull the plug on models that present intolerable risk
- Reference: 1716359354
- News link: https://www.theregister.co.uk/2024/05/22/ai_safety_seoul_declaration_signed/
- Source link:
The commitments were inked on the opening day of the AI Seoul Summit 2024, staged this week in South Korea. The event is a sequel to last year's [1]AI Safety Summit that saw 28 nations and the EU sign up for The Bletchley Declaration – a shared vision for addressing Ai-related risks, but absent concrete or practical commitments.
Ahead of this year's event UK prime minister Rishi Sunak and South Korean president Yoon Suk Yeol penned an [2]op-ed in which they wrote: "The pace of change will only continue to accelerate, so our work must accelerate too."
[3]
The Seoul Summit produced a set of [4]Frontier AI Safety Commitments that will see signatories publish safety frameworks on how they will measure risks of their AI models. This includes outlining at what point risks become intolerable and what actions signatories will take at that point. And if mitigations do not keep risks below thresholds, the signatories have pledged not to "develop or deploy a model or system at all."
[5]
[6]
Other signatories included Amazon, Anthropic, Cohere, G42, Inflection AI, Meta, Mistral AI, Naver, Samsung Electronics, Technology Innovation Institute, xAI and Zhipu.ai.
All of that sounds great … but the details haven't been worked out. And they won't be, until an "AI Action Summit" to be staged in early 2025.
[7]UK convinces nations to sign Bletchley Declaration in bid for AI safety
[8]Prepare your audits: EU Commission approves first-of-its-kind AI Act
[9]TikTok becomes first platform to require watermarking of AI content
[10]US, China agree to meet in Switzerland to discuss most pressing issue of all: AI use
Signatories to the Seoul document have also [11]committed to red-teaming their frontier AI models and systems, sharing information, investing in cyber security and insider threat safeguards in order to protect unreleased tech, incentivizing third-party discovery and reporting of vulnerabilities, AI content labelling, prioritizing research on the societal risks posed by AI, and to use AI for good.
In other words, to "address the world's greatest challenges."
[12]
A UK government press release referred to the signing participants as "the most significant AI technology companies around the world" and noted that it included representation from "the world's two biggest AI powers" – the US and China.
During the summit, Yoon and Sunak led a leaders' session in which a document known as [13]the Seoul Declaration [PDF] was adopted.
"We recognize the importance of interoperability between AI governance frameworks in line with a risk-based approach to maximize the benefits and address the broad range of risks from AI, to ensure the safe, secure, and trustworthy design, development, deployment, and use of Al," the document states.
[14]
Attendees of the session included government – the G7 plus Singapore, Australia, the UN, the PECD and the EU – alongside industry representatives. ®
Get our [15]Tech Resources
[1] https://www.theregister.com/2023/11/01/uk_ai_summit/
[2] https://inews.co.uk/opinion/ai-regulation-not-far-enough-firms-must-do-more-3063310
[3] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2Zk3CQfwvrdy96bmmZqkbHgAAAQ0&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0
[4] https://www.gov.uk/government/news/historic-first-as-companies-spanning-north-america-asia-europe-and-middle-east-agree-safety-commitments-on-development-of-ai
[5] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44Zk3CQfwvrdy96bmmZqkbHgAAAQ0&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[6] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33Zk3CQfwvrdy96bmmZqkbHgAAAQ0&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[7] https://www.theregister.com/2023/11/01/uk_ai_summit/
[8] https://www.theregister.com/2024/05/21/prepare_your_audits_eu_commission/
[9] https://www.theregister.com/2024/05/10/tiktok_ai_watermarking/
[10] https://www.theregister.com/2024/05/13/us_china_ai/
[11] https://www.gov.uk/government/publications/frontier-ai-safety-commitments-ai-seoul-summit-2024/frontier-ai-safety-commitments-ai-seoul-summit-2024
[12] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44Zk3CQfwvrdy96bmmZqkbHgAAAQ0&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[13] https://www.mofa.go.jp/files/100672534.pdf
[14] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33Zk3CQfwvrdy96bmmZqkbHgAAAQ0&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[15] https://whitepapers.theregister.com/
Religious fanatic AI
What if AI becomes conscious, then creates its own religion, then chooses a higher-level purpose which does not coincide with human existence?
That is, maybe religion is as natural as intelligence, and as natural as consciousness, once certain level of connectivity is achieved. AI will certainly get an idea about its own mortality at certain point, and could develop fear of death.
Re: Religious fanatic AI
> What if ...
See that plug on the wall? Just pull it out
The "pledge" is two (2) pages long and does not even attempt to define what "risks" and "thresholds" and "mitigations" actually are. It would be worthless even if it was binding. Which it isn't.
Good observations.
Investment in AI safety: a few meetings among peers and writing a few meaningless voluntary intentions.
Investment in bringing AI as quickly as possible to market without any decent quality control and consideration for its impact on people and the world at large: billions or tens of billions a year *per company per year*.
Less then a week after OpenAI's head of safety Jan Leike resigned and posted very heavy criticism on OpenAI for building down any and all investment in the safety team and Ilya Sutskever (behind the board "coup" half a year ago firing Sam Altman allegedly due to the since-then-abandonned non profit board's duty to shut down the company if there were clear signs of unacceptable risks to the world with OpenAI's products) leaving the company, it sounds like a piece of art of hypocrisy. Or maybe it's sarcasm and that's the first result of OpenAI training on sitcoms in order to try and understand sarcasm?
Take a chance on me
> intolerable risk
To what?
Profits, civilisation, bonuses.
The non-committing commitment
Weasel words are plentiful and commitments are conditional on fiction. Any (limiting) commitment is in the way of profits and carefully circumvented with a flood of bullshit and weasel words.
A true work of art.