Will OpenAI sharing future AI models early with the government improve AI safety, or just let it write the rules?

Date:
Sat, 03 Aug 2024 00:00:30 +0000

Description:
OpenAI to share AI models early with US AI Safety Institute

FULL STORY ======================================================================

OpenAI CEO Sam Altman shared in a post on X this week that the company is teaming up with the U.S. AI Safety Institute and will give the government agency early access to its next major AI model for safety testing.

Altman described the plan as part of a larger new push for AI safety
measures, which could significantly impact ChatGPT and other OpenAI products in the years ahead. It also might be part of a PR and policy push against critics who say OpenAI no longer prioritizes AI safety.

One big reason for the criticism is that OpenAI created and then largely dissolved an internal AI safety team. The disbanding led to the resignation
of top executives at the company, like Jan Leike and Ilya Sutskever, both of whom were instrumental in the companys early safety research efforts. Many other current and former employees publicly wondered if OpenAI had decided to ignore safety in favor of product development.

Another reason is highlighted in a note appended directly to Altmans post. He mentions pledging to allocate 20% of its computing resources to safety research last year, but that allocation at the time was specifically for the same committee whose leaders left the company. Still, the deal with the government isnt the only genuine change. Altman said OpenAI has now voided restrictive non-disparagement clauses that discouraged whistleblowing.

If the deal with the U.S. AI Safety Institute is a strategic move by OpenAI
to regain trust, it is a significant one. The Institute operates under the National Institute of Standards and Technology (NIST) within the Commerce Department. It has been working on developing standards and guidelines for AI safety and security and is already working with many major AI and tech firms like Microsoft, Google, Meta, Apple, and Nvidia.

OpenAI sharing early access to its forthcoming AI model will let the
Institute conduct safety testing and evaluations before the public release. This move is intended to ensure that any potential risks are identified and mitigated early in the development process, aligning with the broader goals outlined in President Joe Bidens AI executive order last year.

Its not the only regulatory encouragement by OpenAI in recent days. The company endorsed the Senates new Future of Innovation Act, which would give the AI Safety Institute responsibility for creating federal regulations for
AI safety. The obvious connection between pushing for the bill and making its models central to the Institutes work could be seen as an attempt to unduly influence regulation or a more innocent broadstroke effort to up AI safety.

Both seem plausible, and OpenAI does have a similar partnership with the
U.K.s AI safety body. Then again, OpenAI has spent more than triple on lobbying in the first half of this year compared to all of last year. Safe, Profitable, Both?

OpenAIs partnership with the U.S. AI Safety Institute will likely be judged
by just how safe the eventual models actually are and whether there are any real restrictions to make them safer. Thats where the real question lies for consumers. As AI becomes increasingly integrated into everyday life, the tension between safety and profitability might grow even as OpenAI and its rivals hope to get the public to trust them and their AI models.

A government agency stamp of approval would certainly help on the trust
front, and OpenAI wouldnt want other companies to get it before them. The involvement of an independent safety body in evaluating AI models before they are released could provide greater assurance that the tools they use are secure and reliable. Growing concerns around AI issues, like data privacy, bias, and deliberate misuse of AI, might be mitigated by the proactive approach. OpenAIs lobbying and other efforts to make those rules favorable to itself could undermine that and the entire point of the AI Safety Institute
if theyre not careful. You might also like... If you think GPT-4o is something, wait until you see GPT-5 a 'significant leap forward' ChatGPT
just (accidentally) shared all of its secret rules here's what we learned OpenAI co-founder's new company promises 'Safe Superintelligence' a
laughably impossible dream



======================================================================
Link to news story: https://www.techradar.com/computing/artificial-intelligence/will-openai-sharin g-future-ai-models-early-with-the-government-improve-ai-safety-or-just-let-it- write-the-rules


--- Mystic BBS v1.12 A47 (Linux/64)
* Origin: tqwNet Technology News (1337:1/100)