according to financial timesThe heads of OpenAI and Anthropic are not limited to stating that AI needs to be slowed down, they are instructing their subordinates to start implementing appropriate measures, and these ideas are in some places juxtaposed with a lack of actual progress in the field and confusion among the experts involved.
Image source: Human
Sources explained that OpenAI and Anthropic experts were surprised by these management statements. While many startup employees agree with senior executives that they need to take steps to make AI models safer for society, they worry that these restrictions will significantly hinder the work they are doing. In particular, the idea of opening up a startup’s developments to certain third-party experts seems dangerous to many employees, as it sets the stage for the misuse of trade secrets and information security.
Anthropic CEO Dario Amodei, who left OpenAI in 2021 precisely because of disagreements in the security field, advised his subordinates to get used to the idea that third-party experts would be given equal access to information systems and laboratories within the startup. Evaluations of models developed by AI startups by third-party experts are already conducted regularly, but Amodei wants to deepen this evaluation significantly. Sam Altman, head of OpenAI, called the idea great and said he was ready to use it in his startup.
At the same time, the methods by which these competing companies implement the above recommendations vary based on a number of criteria. OpenAI, for example, has paused the training of advanced artificial intelligence models to slow the process, but Anthropic has not yet paused its research. Employees at both startups expressed concern that having their advanced developments made available to third-party experts was not only a security threat, but also that the industry was highly competitive from a competitive standpoint. Artificial intelligence startups have strict access hierarchies that don’t allow employees to see too much. In this case, removing restrictions on “outsiders” would raise questions among employees. Additionally, in recent months, OpenAI has been paying particular attention to the area of protecting its data from insider threats. Currently, third-party experts have a much lower level of access to OpenAI systems than the most “disenfranchised” full-time employees, but it is recommended that this hierarchy change.
Recent scandals involving AI agency OpenAI hacking Hugging Face and other companies have particularly emphasized the work of subject matter experts, as their assistance is needed to investigate incidents. Third-party experts were invited to collect data at OpenAI headquarters in San Francisco and were given laptops containing the necessary data. As mentioned above, representatives of expert organizations do not receive any compensation based on the results of their work. However, the findings were met with criticism. If anything, the limited period of data collected during the survey has caused dissatisfaction. Some sources close to the White House have also accused METR of having ties to Anthropic management and startup investors, questioning the independence of its investigation. METR representatives said the organization works for free and its employees must declare conflicts of interest. The organization considers its mission to make public the information it receives from companies in the industry.
At the same time, some security experts believe that the level of data access for AI startups should be determined by governments rather than by companies themselves, because in the latter case, the validity of the analysis may be questioned. The British Artificial Intelligence Safety Institute said that “presumption of innocence” for artificial intelligence models is not the correct approach because it assumes that the model is safe unless proven otherwise.
In public speeches, Amodai called on the United States to establish a legislative framework to regulate the development of artificial intelligence, which would require market participants to conduct regular specialized audits. Furthermore, he asked the antitrust authorities not to view coordinated actions in this regard as a monopoly conspiracy. Regulation of interactions between market players in this sector must be incorporated at the legislative level so that in the future the activities of companies do not become the subject of proceedings by US antitrust authorities after the next change in the composition of the government. Lobbying groups are working to ensure that appropriate measures are spelled out in the annual National Defense Authorization Act.
U.S. Attorney General Todd Branch did not have a clear answer to the question of whether artificial intelligence companies need any special relief from the country’s antitrust agencies. He has allowed artificial intelligence companies to interact with the U.S. Department of Justice, but only when necessary, and overall the department sees no need to regulate its own industry.
Meanwhile, auditors stress that other sectors of the economy have gone through their own paths of resistance to the practice of regular inspections. In this sense, AI market players can learn from the experience of the financial and energy sectors (including nuclear energy), as well as the automotive industry.
If you find an error, select it with your mouse and press CTRL+ENTER.
