Biden's AI Actions: Will They Survive a Trump Presidency?

Recent Developments in US AI Policy

This newsletter provides an update on recent artificial intelligence policy decisions. To receive these updates weekly, you can subscribe here.

Biden Administration's Final Actions

The past week saw significant actions from the Biden administration regarding AI regulation. These actions may represent a culmination of their policy efforts.

On Monday, the White House unveiled substantial new limitations on the export of AI chips. These restrictions drew criticism from major technology companies, notably Nvidia, as their operations could be significantly impacted if implemented as proposed.

Subsequently, on Tuesday, an executive order was issued authorizing the use of federal land for the development of AI data centers.

Impact and Potential Reversal Under a New Administration

A key question arises: will these measures have a lasting effect? The incoming Trump administration could potentially reverse these policies when taking office on January 20th.

Currently, Trump has not publicly indicated his intentions concerning these AI policies. However, he possesses the authority to nullify Biden’s recent actions.

The export controls are slated to become effective following a 120-day period for public comment. The Trump administration will have considerable discretion in implementing these measures and making any necessary modifications.

Executive Order on Federal Land Use

Regarding the executive order concerning federal land allocation, Trump could choose to revoke it. David Sacks, formerly of PayPal and now Trump’s advisor on AI and cryptocurrency, has already pledged to repeal a separate Biden executive order focused on AI safety and security standards.

Potential for Continuity

Despite the possibility of reversals, there are indications that the incoming administration might not drastically alter all existing policies.

Similar to Biden’s initiative to facilitate access to federal resources for data centers, Trump has previously committed to accelerating permit approvals for companies investing at least $1 billion within the U.S.

Furthermore, the appointment of Lee Zeldin to lead the EPA, with a stated commitment to reducing regulatory burdens on businesses, suggests a potential continuation of a pro-business approach.

Focus on China and Israel

Certain aspects of Biden’s export rules, particularly those targeting China, may remain in effect. Trump has consistently identified China as the primary competitor to the U.S. in the field of AI.

However, the inclusion of Israel on the list of countries subject to AI hardware trade restrictions is a point of contention. Trump has repeatedly expressed strong support for Israel and is likely to adopt a more lenient stance towards its military operations.

Looking Ahead

Greater clarity regarding the future of these policies is expected within the coming week.

AI News Update

A new capability has been added to ChatGPT. Subscribers can now utilize the AI assistant to set reminders and schedule recurring tasks.

This feature, known as Tasks, is currently in beta and will be progressively released to ChatGPT Plus, Team, and Pro users globally this week.

Meta's Competitive Drive

Internal communications revealed a strong focus at Meta on surpassing OpenAI’s GPT-4 model. Executives and researchers were intensely dedicated to developing Meta’s Llama 3 models to achieve this goal.

These messages were recently made public following a court order.

Changes at OpenAI's Leadership

OpenAI has expanded its board of directors with the appointment of Adebayo “Bayo” Ogunlesi, a key executive from BlackRock.

The composition of OpenAI’s board has significantly changed since the late 2023 events surrounding the temporary dismissal and subsequent reinstatement of CEO Sam Altman.

Blaize to Enter Public Market

Blaize is preparing for an initial public offering (IPO) in 2025, marking a significant milestone as the first AI chip startup to do so.

Established in 2011 by former Intel engineers, the company has secured $335 million in funding from investors, including Samsung, to support its development of chips for edge devices like cameras and drones.

Mysterious Multilingual Reasoning

OpenAI’s o1 AI model exhibits an unusual behavior. It occasionally processes information and “thinks” in languages other than English, such as Chinese, French, Hindi, and Thai, even when prompted in English.

The underlying reasons for this phenomenon remain unclear and are currently under investigation.

Weekly Research Highlight

A newly published study, with contributions from Dan Hendrycks – who serves as an advisor to xAI, the AI firm founded by Elon Musk – indicates a correlation between AI safety assessments and the inherent abilities of AI systems.

Essentially, the research demonstrates that as an AI model’s overall performance increases, its results on safety benchmarks also tend to improve, potentially creating a misleading impression of enhanced safety.

The study’s authors state that approximately 50% of existing AI safety benchmarks unintentionally measure underlying factors directly linked to general capabilities and the amount of computational resources used during training.

It is challenging to isolate the measurement of safety without also assessing the fundamental capabilities of the AI model, according to the findings.

Developing Improved Safety Metrics

The researchers put forward a framework grounded in empirical data for the creation of more robust and significant safety metrics.

This initiative aims to foster progress in the scientific understanding and evaluation of AI safety protocols.

Key takeaway: Current safety benchmarks may not accurately reflect true safety improvements, but rather correlate with increased AI capabilities.

Featured AI Model: Transformer²

Sakana AI, a Japanese artificial intelligence firm, recently released a technical document on Tuesday outlining their innovative AI system, Transformer² (also known as “Transformer-squared”). This system exhibits a unique capability to adapt itself to varying tasks.

The process begins with Transformer² analyzing a given task, such as code generation, to fully comprehend its specific demands.

Following analysis, the system implements “task-specific adaptations” and optimizations, effectively calibrating itself for optimal performance on that particular task.

According to Sakana AI, the underlying techniques utilized in Transformer² are applicable to open-source models like Meta’s Llama.

They suggest this represents a potential future direction for AI, moving away from static models towards systems capable of dynamic adjustment.

Key Capabilities of Transformer²

• Dynamic Adaptation: The system isn't fixed; it changes based on the task.
• Task Analysis: It first understands the requirements of each new challenge.
• Optimization: It fine-tunes itself for peak performance.

This approach offers a significant departure from traditional AI models, which are typically trained for a specific purpose and remain relatively inflexible afterward.

The implications of this technology could be far-reaching, potentially leading to more versatile and efficient AI solutions across a wide range of applications.

An Open-Source Alternative to AI Search Engines

A dedicated group of developers has recently launched an open-source project designed to function as an alternative to AI-driven search platforms such as Perplexity and OpenAI's SearchGPT.

This new initiative, named PrAIvateSearch, is hosted on GitHub and released under a permissive MIT license. This licensing allows for broad usage and modification with minimal restrictions.

Core Technologies

The project leverages freely accessible AI models and services to deliver its functionality. Specifically, it utilizes the Qwen model family developed by Alibaba, alongside the privacy-focused search engine, DuckDuckGo.

Project Goals and Implementation

According to the PrAIvateSearch team, their primary objective is to replicate the features found in SearchGPT. However, they aim to achieve this in a manner that prioritizes openness, local operation, and user privacy.

Detailed instructions and guidance for setting up and running PrAIvateSearch can be found in the team’s most recent blog post. This resource provides a comprehensive overview for potential users.

PrAIvateSearch offers a compelling option for those seeking an AI-powered search experience that emphasizes control and data protection.