Keeping up with an industry that changes as quickly as AI can be a challenge. Until AI can do it for you, here's a roundup of recent buzz in the world of machine learning, along with notable research and experiments that we wouldn't have covered on our own.
By the way, TechCrunch will be launching an AI newsletter soon. Stay tuned. In the meantime, we're increasing the semi-regular cadence of our AI columns from twice a month to weekly, so keep an eye out for upcoming issues.
In AI news this week, OpenAI announced it had inked a deal with emerging publishing giant News Corp to train generative AI models developed by OpenAI on articles from News Corp brands, including The Wall Street Journal, Financial Times, and MarketWatch. The deal, which both companies described as “multi-year” and “historic,” also gives OpenAI the right to display News Corp's name within apps like ChatGPT in answers to certain questions, presumably when the answers quote in whole or in part from News Corp publications.
Doesn't this sound like a win for both sides? News Corp will receive a $250 million+ cash infusion for its content at a time when the outlook for the media industry is even bleaker than usual (Generative AI threatens to make things worse, drastically reducing referral traffic for publications). Meanwhile, OpenAI, which is battling copyright holders on multiple fronts over fair use, now has one less expensive legal battle to worry about.
But the devil is in the details. Keep in mind that the deal with News Corp has an end date, as does all of OpenAI's content license agreements.
That's not, in itself, bad faith on OpenAI's part — perpetual licenses are unusual in the media, given the incentives of all parties involved to leave room for renegotiating the contract — but it's a bit suspicious, given OpenAI CEO Sam Altman's recent comments about the declining importance of training data for AI models.
“Definitely,” Altman said on the “All In” podcast. [doesn’t] I think there will be an arms race. [training] “Once the models get smart enough, at some point, they don't need more data, at least in training,” he said, telling MIT Technology Review's James O'Donnell that he's “optimistic” that OpenAI, and the AI industry as a whole, “will find a way out of this.” [needing] The training data continues to grow.”
The models aren't very “smart” yet, so OpenAI is reportedly experimenting with synthetic training data and scouring the far reaches of the web and even YouTube for organic sources. But let's assume there will come a day when barely any additional data is needed to make leaps of improvement. What will happen to publishers once OpenAI has scraped their entire archives?
My point is that publishers and other content owners with whom OpenAI has worked are merely short-term partners of convenience. Through licensing agreements, OpenAI effectively neutralizes its legal threats (at least until the courts decide how fair use applies in the context of AI training), gets to celebrate a PR victory; publishers get much-needed funding; and work on AI that could seriously harm those publishers continues.
Other notable AI-related news in the past few days include:
Spotify's AI DJ: Spotify's addition of an AI DJ feature was a step toward the company's AI future, showcasing personalized song selections to users. Now, Spotify is developing a Spanish-speaking DJ alternative, writes Sarah. Meta's AI Council: Meta announced the formation of an AI advisory council on Wednesday. There's a big problem, though: the members are exclusively white and male. It feels a bit outdated, given that marginalized groups are most likely to be affected by flaws in AI technology. FCC Proposes AI Disclosure: The Federal Communications Commission (FCC) has proposed requirements to disclose AI-generated content in political ads, but it doesn't ban it. Devin reports the details. Answer Calls with Your Voice: Truecaller, the widely known caller ID service, will soon allow customers to use its AI-powered assistant to answer calls with their own voice, thanks to a newly inked partnership with Microsoft. Humane Considers Sale: Humane, the company behind the much-hyped Ai Pin, which launched last month but was poorly received, is looking for a buyer. The company is reportedly priced at between $750 million and $1 billion, and the sale process is in the early stages. TikTok turns to generative AI: TikTok is the latest tech company to bring generative AI to its advertising business, announcing the release of its new TikTok Symphony AI suite for brands on Tuesday. The tools will help marketers write scripts, produce videos, and enhance current ad assets, Aisha reports. Seoul AI Summit: At an AI Safety Summit in Seoul, South Korea, government officials and AI industry executives agreed to apply basic safety measures to the rapidly changing field and establish an international safety research network. Microsoft's AI PCs: During a keynote address at its annual developer conference Build this week, Microsoft unveiled a new lineup of Windows machines (and Surface laptops) it's calling Copilot+ PCs, along with generative AI-powered features like Recall, which helps users find apps, files, and other content they've viewed in the past. OpenAI voice issue: OpenAI is removing one of the voices for ChatGPT's text-to-speech feature. Users found the voice, called Sky, eerily similar to Scarlett Johansson (who has played AI characters). Johansson herself also released a statement saying she had hired a lawyer to inquire about the Sky voice and get precise details about its development. UK self-driving car law: The UK's regulations for self-driving cars have been officially enacted after receiving Royal Assent, the final approval that all legislation must pass before becoming law.
More Machine Learning
Here are some interesting AI-related research findings this week: Renowned researcher Shyan Gollakota from the University of Washington is making waves again with his noise-canceling headphones that block out everything except the voice of the person you want to hear. When you put the headphones on and look at the person you want to hear, you press a button and the headphones sample the sound coming from that particular direction and use it to power a hearing rejection engine that filters out background noise and other sounds.
The researchers, led by Gollakota and some graduate students, called the system “Targeted Speech Hearing” and presented it at a conference in Honolulu last week. It's both a useful accessibility tool and an everyday option, and it's sure to be a feature big tech companies will include in the next generation of premium headphones.
Chemists at EPFL are apparently tired of performing 18 tasks in particular, so they're instead training a model, called ChemCrow, to perform them—not the actual tasks like titrating or pipetting, but the planning tasks like reviewing the literature or planning reaction chains. Of course, ChemCrow doesn't just do all that for the researchers, it also acts as a natural language interface to the whole set, with search and calculation options when needed.
Image credit: EPFL
The lead author of the paper introducing ChemCrow says that ChemCrow is “analogous to human experts with access to computers and databases,” i.e. like a graduate student. So they can work on more important things and skip the boring parts. It reminds me a bit of Coscientist. As for the name, “because crows are known to be good with tools.” Enough!
Robotics experts at Disney Research are hard at work making their creations move more realistically without having to animate every movement by hand. A new paper being presented at SIGGRAPH in July shows a combination of procedurally generated animation and an artist interface for fine-tuning it, all working on an actual bipedal robot called Groot.
The idea is to let artists create a type of bouncy, stiff, erratic movement, and engineers don't have to implement every single detail, just make sure it fits within certain parameters. The proposed system can then essentially improvise precise movements and execute them instantly. You'll probably see this at Disney World in a few years…