Hello everyone, welcome to TechCrunch's regular AI newsletter. If you'd like to have this sent to your inbox every Wednesday, sign up here.
The AI war between the US and China is escalating, or at least the rhetoric surrounding it is escalating.
On Tuesday, a U.S. Congressional committee proposed a “Manhattan Project-style” effort to fund the development of AI systems with human-level or superhuman intelligence.
In its annual report, the U.S.-China Economic and Security Review Commission (USCC) found that policymakers approved funding for “major AI, cloud, and data center companies” and that AI development is a “national priority.” It recommended that the Secretary of Defense instruct the US Secretary of Defense to ”
Reuters quoted USCC Commissioner Jacob Helberg as saying, “Throughout history, we have seen that countries that are the first to take advantage of periods of rapid technological change can often lead to shifts in the global balance of power.'' I'm here,'' he said. “China is [AI superintelligence]. …It’s important to take them very seriously. ”
The USCC was established by Congress to make recommendations on U.S.-China relations, and its recommendations tend to be hawkish. But the commission is not alone in calling for more aggressive action to slow China's technology ambitions.
For example, Commerce Secretary Gina Raimondo has proposed that the United States share AI technology with foreign allies to counter the rise of China. Meanwhile, Pentagon officials are pushing for safeguards to prevent technology from leaking to China through overseas data centers and chip suppliers.
The United States has already adopted a number of policies aimed at curbing China's AI advances, including a ban on exports of hardware infrastructure in the region and investments in AI technology. China has avoided some of these. But the impact is clear and far-reaching. For example, China's access to the most sophisticated chips needed to train AI, including next-generation GPUs, has been completely cut off.
And considering that, the USCC's statement seems like a bit of a stretch.
It's not even clear what superintelligent AI will look like. But assuming this includes so-called inferential models, as some have suggested, Chinese laboratories appear to be lagging behind, rather than leading the way. One analysis shows that the models of top Chinese companies are about six to nine months behind their American counterparts.
It is necessary to consider the possibility that the USCC's recommendations are self-serving. Mr. Helberg is a senior advisor to the CEO of Palantir, a company with many AI defense contracts. And of course, government funding for AI will benefit US AI companies.
So the calls for a Manhattan Project-type program for superintelligent AI seem more alarming than anything else.
news
Image credit: Mark Kauzlarich / Bloomberg / Getty Images
AI at Ignite: Microsoft announced a number of AI products at Microsoft Ignite 2024 on Tuesday, including voice cloning and an AI development platform called Azure AI Foundry.
Advanced voice mode on the web: OpenAI has extended ChatGPT's advanced voice mode capabilities to the web, allowing users to converse with AI chatbots directly from their desktop browser.
Indian news agency sues OpenAI: As for OpenAI, Asian News International, one of India's largest news agencies, is suing the startup in a potentially precedent-setting case over the use of copyrighted news content. I did.
Gemini now remembers: Google's Gemini chatbot can now remember information about your life, work, personal preferences, and more during conversations.
UK greenlights investment in Anthropic: The UK Competition and Markets Authority has approved Alphabet's partnership and investment in AI rival Anthropic, concluding that it is not subject to investigation under current merger rules.
Perplexity launches shopping: AI-powered search engine Perplexity debuts the ability to provide e-commerce recommendations and the ability to order without going to a retailer's website. However, Stripe seems to be doing the heavy lifting here.
Altman joins Team SF: San Francisco Mayor-elect Daniel Lurie has appointed OpenAI CEO Sam Altman to help run his transition team. Altman, along with nine other San Francisco leaders, will coach Lurie's team on ways the city can innovate.
New Mistral Model: French AI startup Mistral released major new products and tools this week. This includes the chatbot platform's “Canvas” feature, which allows users to transform and edit content such as web mockups.
This week's research paper
The UK AI Safety Institute, a UK government agency that studies risks in AI systems, has published its first academic paper proposing ways for AI developers to prove that their models do not pose “unacceptable cyber risks”. .
Co-authors from the AI Safety Institute note in their paper that “safety cases” — structured, substantiated arguments for why the risks associated with a model are acceptable — are gaining traction. However, there is no “out-of-the-box” safety case methodology for frontier AI.
The co-authors propose a safety case template that focuses on cyber capabilities, arguing that short-term risks are well-established. The template is designed to inform deployment decisions, such as whether to start or continue training a model.
“This template serves as a proof of concept,” the co-authors wrote. “There is no guarantee of safety. Some of the claims in the template may not actually hold true, invalidating the conclusions. Still, even these imperfect safety cases can make the development We expect this to increase the level of rigor in the reasoning surrounding adoption decisions.”
this week's model
Image credit: Suno v4
Controversial music generation startup Suno today released its latest music generation model, Suno v4.
Suno claims that v4, which is only available to paid users of the platform, offers clearer audio, better lyrics, and “more dynamic” song structures than the previous version, v3. Suno's v4 enhances the company's Cover feature, which “reimagines” uploaded audio, and its Persona feature, which captures the vocals, style, and “vibe” of a track to incorporate into other works.
Suno's forward thinking is remarkable in many ways, considering he's been sued by three major record labels for copyright infringement. Sony Music Entertainment, Universal Music Group, and Warner Music Group filed suit this summer against Suno and rival company Udio, alleging that they trained models on music without permission.
In their response to the lawsuit, Mr. Suno and Mr. Udio more or less acknowledged that the models may have ingested copyrighted music during their training, but the fair use doctrine under U.S. copyright law. claimed to be protecting them.
grab bag
HarperCollins has signed a three-year data licensing agreement with Microsoft, allowing the tech giant to train AI based on the company's nonfiction works.
HarperCollins' parent company News Corp. has a similar agreement with OpenAI, but says authors must opt in and that the agreement only covers “select nonfiction backlist titles.” Ta.
The authors are dissatisfied, and it doesn't help that HarperCollins offers meager compensation. One of the authors, Daniel Kibblesmith, said he was offered a flat rate of $2,500 per book.
“I would probably do $1 billion,” Kibblesmith wrote in a post Tuesday. “I'm going to do it for enough money that I don't have to work anymore, because that's the ultimate goal of this technology.”