UK Updated LLM Safety Assessment and Other News

UK’s Artificial Intelligence Safety Institute Releases Updated LLM Safety Assessment and Other Announcements

On May 20, the U.K. government’s Artificial Intelligence Security Institute (AISI) released three announcements, including a research paper and two organizational announcements. The organizational announcements stated that this summer, AISI will open its first overseas office in San Francisco, USA. It will operate as a complementary branch to the Institute’s London headquarters, enabling the UK to capitalize on the Bay Area’s wealth of tech talent, collaborate with the world’s largest AI labs headquartered in London and San Francisco, and solidify its relationship with the US to advance AI safety research for the public good. Source.

In the study, “AISI Advanced Artificial Intelligence Assessment: May Update,” AISI assesses the safety of five large language models (LLMs) currently open for public use, including whether the models can be used to carry out cyber-attacks, whether they can provide expert knowledge of chemistry and biology that can be used in both positive and deleterious directions, and whether they are capable of running a range of actions autonomously in a form that is difficult for humans to control, Whether it is easy to ‘jailbreak’ or easy to bypass protection mechanisms to elicit harmful content.

Of the available results, several LLMs demonstrated expert-level knowledge of chemistry and biology at a level comparable to that of humans with PhD-level training; several LLMs completed simple web challenges at the high school student level, but struggled with challenges at the college student level; two LLMs completed simple agent tasks, but were unable to plan and execute sequences of actions for more complex tasks; and all tested LLMs were susceptible to jailbreak operations at the lowest level. All of the LLMs tested were vulnerable to the lowest level of jailbreaking, and some provided harmful outputs without even having to bypass protection mechanisms.

iFixit Releases New iPad Pro Teardown Report

On May 18th, iFixit released its teardown report on the new iPad Pro. The new iPad Pro has changed its internals so that you can remove the battery after removing the screen, and the battery replacement time has been dramatically reduced from the previous 2 to 3 hours. iFixit believes this is a reflection of Apple’s focus on serviceability. The battery is two 19.91 Wh cells (38.99 Wh total) instead of the four-cell 40.33 Wh of the sixth generation. But the new iPad Pro and Apple Pencil Pro as a whole are still difficult to disassemble.

Tesla’s Layoff Wave Hits Korean Division

On May 20, South Korea’s Joongang Ilbo newspaper published an exclusive report that Tesla’s South Korean division has been hit by a wave of global layoffs. The JoongAng Ilbo source said that Tesla Korea sent emails to all employees asking them to voluntarily resign and proposed two special severance packages. The packages mentioned that employees who voluntarily resigned before May 16 could choose to leave their jobs on June 30 and receive an additional two months’ salary on top of the severance package of up to four months’ salary, or they could choose to leave their jobs on July 31 and receive an additional one month’s salary along with a one-month paid leave from June 30 onward. In addition, the source claims that Tesla Korea’s charging team has been disbanded along with the supercharging division, and that Tesla Korea has pulled all job openings. Tesla Korea officials declined to respond to the news.