Hello and welcome to TechCrunch's regular artificial intelligence newsletter. If you'd like to receive this information in your inbox every Wednesday, sign up here .
This week is something of a swan song for the Biden administration.
On Monday, the White House announced sweeping new restrictions on the export of artificial intelligence chips — restrictions that have drawn strong criticism from tech giants including Nvidia. (Nvidia’s business would be severely impacted if these restrictions go into effect as proposed.) Then on Tuesday, the administration issued an executive order opening up federal lands to artificial intelligence data centers.
But the obvious question is, will these initiatives have a lasting impact? Will Trump, who takes office on January 20, simply overturn Biden's bill? So far, Trump has not made his intentions known in any way. But he certainly has the ability to undo Biden’s final AI actions.
Biden's export rules are scheduled to take effect after a 120-day comment period. The Trump administration will have broad leeway on how to implement these measures and whether to change them in any way.
As for executive orders related to federal land use, Trump could repeal them. Former PayPal COO and Trump’s AI and cryptocurrency “czar” David Sacks recently pledged to rescind another AI-related Biden executive order, the one on AI safety and security Standards were set.
However, there is reason to believe that the incoming administration may not cause too much trouble.
Following Biden's move to free up federal resources for data centers, Trump recently pledged to provide fast-track licensing to companies investing at least $1 billion in the United States. He also chose Lee Zeldin, who has vowed to cut regulations he sees as burdensome to businesses, to lead the EPA.
Some aspects of Biden's export rules may also hold true. Some of these regulations target China, which Trump has made no secret of viewing as the United States’ biggest competitor in artificial intelligence.
One of the issues is Israel’s inclusion on a list of countries subject to trade restrictions on artificial intelligence hardware. As recently as October, Trump described himself as Israel's "protector" and suggested he might be more tolerant of Israeli military operations in the region.
Regardless, we'll get a clearer picture this week.
information
ChatGPT, remind me…: Paid users of OpenAI's ChatGPT can now ask the AI assistant to schedule reminders or repeat requests. A new beta feature called "Tasks" will begin rolling out to ChatGPT Plus, Team and Pro users around the world this week.
Meta and OpenAI: Executives and researchers leading Meta's artificial intelligence efforts were bent on beating OpenAI's GPT-4 model while developing Meta's own Llama 3 series of models, according to court documents released on Tuesday.
OpenAI board of directors expands: OpenAI has appointed investment firm BlackRock executive Adebayo “Bayo” Ogunlesi to its board of directors. The company's current board bears little resemblance to OpenAI's board in late 2023, when board members fired CEO Sam Altman and reinstated him days later.
Blaise disclosed: Blaize will become the first artificial intelligence chip startup to go public in 2025. The company, founded in 2011 by former Intel engineers, has raised $335 million from investors, including Samsung, for chips used in cameras, drones and other edge devices.
A “reasoning” model for thinking in Chinese: OpenAI's o1 artificial intelligence inference model sometimes "thinks" in languages such as Chinese, French, Hindi and Thai, even when questions are asked in English, but no one really knows why.
Research Paper of the Week
A recent study co-authored by Dan Hendrycks, an advisor to billionaire Elon Musk's artificial intelligence company xAI, shows that many of the security benchmarks for AI are inconsistent with AI systems. function related. That is, as the system's overall performance improves, it "scores better" on the benchmark -- making the model appear "safer."
"Our analysis shows that many AI security benchmarks (roughly half) often inadvertently capture underlying factors that are closely related to general capabilities and raw training computations," the researchers behind the study wrote. "Overall, in Measuring the capabilities of upstream models is difficult to avoid in AI security benchmarks.”
In the study, the researchers present what they say is an empirical basis for developing "more meaningful" safety metrics, which they hope will "(advance) the science of AI safety assessment."
Model of the week
In a technical paper published on Tuesday, Japanese artificial intelligence company Sakana AI detailed Transformer² ("Transformer-squared"), an artificial intelligence system that can dynamically adjust to new tasks.
Transformer² first analyzes a task (such as writing code) to understand its requirements. It then applies "task-specific adaptation" and optimization to tune that task.
Sakana said the methods behind Transformer² can be applied to open models such as Meta's Llama, and they "give us a glimpse of a future where AI models are no longer static."
Capture packets
A small group of developers have released open alternatives to AI-powered search engines, such as Perplexity and OpenAI's SearchGPT.
The project, called PrAIvateSearch, is available on GitHub under an MIT license, which means it can be widely used without restrictions. It is powered by publicly available artificial intelligence models and services, including Alibaba's Qwen series of models and search engine DuckDuckGo.
The PrAIvateSearch team says its goal is to "achieve similar functionality to SearchGPT" but in an "open source, native and private way." For tips on getting it up and running, check out the team’s latest blog post.