Hey guys, and welcome to TechCrunch’s regular AI newsletter.
This week in AI, genetic AI is starting to spam academic publications—a chilling new development on the disinformation front.
In a post on Retraction Watcha blog that tracks recent academic retractions, assistant professors of philosophy Tomasz Żuradzk and Leszek Wroński wrote about three journals published by Addleton Academic Publishers that appear to consist entirely of AI-generated articles.
The journals contain papers that follow the same pattern, filled with buzzwords like ‘blockchain’, ‘metaverse’, ‘internet of things’ and ‘deep learning’. They list the same editorial board – 10 of which have died – and a nondescript address in Queens, New York, which appears to be a home.
So what’s the big deal? you might ask. Isn’t browsing AI-generated spam just the cost of doing business on the Internet these days?
Well yes. But the bogus journals show how easy it is to game the systems used to evaluate researchers for promotions and hires — and that could be a wake-up call for knowledge workers in other industries.
In at least one widely used rating system, CiteScore, journals are ranked in the top 10 for philosophical research. How is this possible? They interbreed extensively. (CiteScore takes citations into account in its calculations.) Żuradzk and Wroński find that, of the 541 citations to one of Addleton’s journals, 208 are from other bogus publications by the publisher.
“[These rankings] they often serve universities and funding agencies as indicators of research quality,” wrote Żuradzk and Wroński. “They play a critical role in decisions about academic awards, recruitment and promotion, and thus can influence researchers’ publication strategies.”
One could argue that CiteScore is the problem – it’s clearly a flawed metric. And this is not a wrong argument. But it’s also not wrong to say that genetic AI and its abuse are disrupting the systems on which people’s livelihoods depend in unexpected—and potentially quite damaging—ways.
There is a future in which genetic AI forces us to rethink and re-engineer systems like CiteScore to be more fair, holistic and inclusive. The gloomier alternative—and the one playing out now—is a future in which generative AI continues to rage, wreaking havoc and destroying professional lives.
I sure hope we can correct course soon.
News
DeepMind Soundtrack Generator: DeepMind, Google’s artificial intelligence research lab, says it’s developing AI technology to create soundtracks for videos. DeepMind’s AI takes the description of a soundtrack (eg “jellyfish pulsating underwater, sea life, ocean”) combined with a video to generate music, sound effects and even dialogue that match the characters and the tone of the video.
A robot chauffeur: Researchers at the University of Tokyo developed and trained a “musculoskeletal humanoid” called Musashi to drive a small electric car on a test track. Equipped with two human-eye cameras, Musashi can “see” the road ahead as well as views reflected in the car’s side mirrors.
A new AI search engine: Genspark, a new AI-powered search platform, uses artificial intelligence generation to write custom summaries in response to search queries. So far it has raised $60 million from investors including Lanchi Ventures. The company’s latest funding round valued it at $260 million post-money, a respectable number as Genspark goes up against competitors like Perplexity.
How much does ChatGPT cost?: How much is ChatGPT, OpenAI’s ever-expanding AI chat platform? It’s a harder question to answer than you might think. To keep track of the various ChatGPT subscription options available, we’ve created an up-to-date guide to ChatGPT pricing.
Research paper of the week
Autonomous vehicles face an endless variety of edge cases, depending on location and situation. If you’re on a two-lane road and someone turns their left blinker, does that mean they’re going to change lanes? Or that you have to pass them? The answer may depend on whether you’re on I-5 or the Autobahn.
A team of researchers from Nvidia, USC, UW, and Stanford show in a paper just published in CVPR that many ambiguous or unusual circumstances can be resolved, if you can believe it, by reading the local driver’s manual from an AI.
Theirs Large Language Driving Assistant, or LLaDa, gives LLM access — not even optimization — to the driver’s manual for a state, country, or region. Local rules, customs or signage are found in the literature and when an unexpected situation such as honking, a large staircase or a herd of sheep occurs, an appropriate action is generated (pull up, stop turning, honk back).
It’s by no means a complete end-to-end driving system, but it shows an alternative route to a “universal” driving system that still has surprises. Plus, maybe a way for the rest of us to know why we get honked at when visiting unfamiliar places.
Model of the week
On Monday, Runway, a company that makes productive artificial intelligence tools aimed at film and image content creators, introduced the Gen-3 Alpha. Trained on a huge number of images and videos from both public and internal sources, Gen-3 can create video clips from text descriptions and still images.
Runway says the Gen-3 Alpha offers a “significant” improvement in production speed and fidelity over Runway’s previous video model, the Gen-2, as well as detailed control over the structure, style and movement of videos that creates. Gen-3 can also be tweaked to allow for more “stylistically controlled” and consistent characters, Runway says, targeting “specific artistic and narrative requirements.”
The Gen-3 Alpha has its limitations — including the fact that its shots are as long as 10 seconds. However, Runway co-founder Anastasis Germanidis promises she’s just the first of many video production models to come in a family of next-generation models trained on Runway’s upgraded infrastructure.
The Gen-3 Alpha is just the latest video production system of many to hit the scene in recent months. Others include OpenAI’s Sora, Luma’s Dream Machine, and Google’s Veo. Together, they threaten to upend the film and television industry as we know it – assuming they can win copyright challenges.
Grab bag
Artificial intelligence will not take your next McDonald’s order.
McDonald’s this week was announced that it would remove automated order-taking technology, which the fast-food chain had been testing for the better part of three years, from more than 100 of its restaurant locations. The technology — co-developed with IBM and installed in drive-thrus restaurants — went viral last year because of its tendency to misunderstand customers and make mistakes.
Recent piece at Takeout suggests that AI is generally losing its grip on fast food operators, who not long ago expressed excitement about the technology and its potential to boost efficiency (and lower labor costs). Presto, a major player in the AI-assisted drive-thru space, recently lost a major customer, Del Taco, and is facing mounting losses.
The point is the inaccuracy.
McDonald’s CEO Chris Kempczinski he said CNBC in June 2021 reported that voice recognition technology was accurate about 85% of the time, but that human staff had to help with about one in five orders. The best version of Presto’s system, meanwhile, completes only about 30% of orders without human assistance, according to Takeout.
So while the AI is decimating Some parts of the gig economy, it seems, are some jobs — particularly those that require understanding a variety of accents and dialects — that can’t be automated. For now at least.
