Hey guys, and welcome to TechCrunch’s regular AI newsletter.
This week in AI, the US Supreme Court struck down Chevron deference, a 40-year-old ruling on federal agency power that required courts to uphold the agencies’ interpretations of congressional laws.
Chevron’s deference allowed agencies to shape their own rules when Congress left aspects of its statute unclear. Courts will now be expected to exercise their own legal judgment — and the results could be far-reaching. Scott Rosenberg of Axios writes that Congress — almost the most operating body today — must now effectively try to anticipate the future with its legislation, as agencies can no longer apply basic rules to new enforcement circumstances.
And that could kill efforts to nationally regulate AI for good.
Already, Congress has struggled to pass a basic AI policy framework — to the point that state regulators on both sides of the aisle have felt compelled to step in. Now any regulation he writes will have to be very specific if it is to survive legal challenges — a seemingly insurmountable task given the speed and unpredictability with which the AI industry moves.
Justice Elena Kagan addressed artificial intelligence specifically during oral arguments:
Let’s imagine that Congress passes an AI bill and has all kinds of delegations. Just by the nature of things and especially the nature of the matter, there will be all sorts of places where, although there is no express agency, Congress has actually left a loophole. … [D]o Do we want the courts to fill this gap or do we want an agency to fill this gap?
The courts will fill that gap now. Otherwise, federal lawmakers will deem the exercise futile and kill their AI bills. Whatever the outcome, regulating artificial intelligence in the United States just got orders of magnitude harder.
News
Google’s AI Environmental Cost: Google has released its 2024 Environmental Report, an 80-plus page document outlining the company’s efforts to apply technology to environmental issues and mitigate its negative contributions. But it avoids the question of how much energy Google’s AI uses, Devin writes. (AI is notoriously power hungry.)
Figma disables the drawing feature: Figma CEO Dylan Field says Figma will temporarily disable its AI ‘Make Design’ feature, which is said to be removing designs from Apple’s Weather app.
Meta is changing its AI tag: After Meta began tagging photos with the “Made with AI” label in May, photographers complained that the company had mistakenly tagged real photos. Meta is now changing the label to “AI info” on all its apps in an attempt to appease critics, Ivan reports.
Robot cats, dogs and birds: Brian writes about how New York state is handing out thousands of robot animals to seniors amid a “loneliness epidemic.”
Apple brings AI to Vision Pro: Apple’s plans go beyond the previously announced versions of Apple Intelligence on iPhone, iPad and Mac. According to Bloomberg’s Mark Gurman, the company is also working to bring these features to its Vision Pro mixed reality headset.
Research paper of the week
Text-generating models like OpenAI’s GPT-4o have become table stakes in technology. Applications that no use them these days, for tasks ranging from filling out emails to writing code.
But despite the popularity of the models, how these models “understand” and create text that is audible to humans is not clear in science. In an effort to peel back the layers, researchers at Northeastern University looked in tokenization, or the process of breaking up text into named units brands where models can work more easily.
Today’s text generation models process text as a series of tokens drawn from a set of “token vocabularies”, where a token may correspond to a single word (“fish”) or part of a larger word (“sal” and ” mon’ to ‘salmon’). The vocabulary of tokens available in a model is typically specified before training, based on the characteristics of the data used to train it. But the researchers found evidence that the models also develop one spoken vocabulary which maps groups of tokens—for example, multipoint words like “northeast” and the phrase “break a leg”—into semantically meaningful “units.”
On the back of this evidence, the researchers developed a technique to “explore” the implicit vocabulary of any open model. From Meta’s Llama 2, they pulled phrases like “Lancaster”, “World Cup players” and “Royal Navy”, as well as more obscure terms like “Bundesliga players”.
The work has not been peer-reviewed, but the researchers believe it could be a first step towards understanding how lexical representations are formed in models – and serve as a useful tool for uncovering what a given ‘knows’ model.
Model of the week
A research team at Meta has trained several models to generate 3D elements (ie, 3D shapes with textures) from textual descriptions, suitable for use in projects such as apps and video games. While there are many shaper models out there, Meta claims they are “high-end” and support physical rendering, which allows developers to “re-light” objects to give the appearance of one or more light sources.
The researchers combined two models, AssetGen and TextureGen, inspired by Meta’s Emu image generator into a single line called 3DGen to generate shapes. AssetGen converts text messages (eg “a t-rex wearing a green woolen sweater”) into a 3D mesh, while TextureGen improves the “quality” of the mesh and adds a texture to give the final shape.
3DGen, which can also be used to reconstruct existing shapes, takes about 50 seconds from start to finish to generate a new shape.
“With the combination [these models’] advantages, 3DGen achieves very high quality 3D object synthesis from written prompts in less than a minute,” the researchers wrote in a technical paper. “When evaluated by professional 3D artists, 3DGen output is preferred most of the time compared to industry alternatives, particularly for complex prompts.”
Meta seems poised to incorporate tools like 3DGen into their metaverse game development efforts. According to a task listthe company seeks to research and create prototype VR, AR, and mixed reality games built with the help of genetic AI technology — including, presumably, custom shape generators.
Grab bag
Apple could get an observer seat on OpenAI’s board as a result of the two companies’ partnership announced last month.
Bloomberg References that Phil Schiller, an Apple executive responsible for leading the App Store and Apple events, will join OpenAI’s board as its second observer after Microsoft’s Dee Templeton.
If the move goes ahead, it would be a notable show of strength on the part of Apple, which plans to integrate OpenAI’s ChatGPT AI chat platform with many of its devices this year as part of a broader suite of AI features.
Apple won’t be beneficiary OpenAI for its ChatGPT integration reportedly made the argument that PR exposure is as valuable as – or more valuable than – cash. In fact, OpenAI may end up paying off apple; Apple is said to be mulling a deal in which it would receive a cut of revenue from any premium ChatGPT features offered by OpenAI on Apple’s platforms.
So, as my colleague Devin Coldewey pointed out, this puts close OpenAI partner and major investor Microsoft in the difficult position of effectively subsidizing Apple’s ChatGPT integration — with little to show for it. What Apple wants, it gets, apparently — even if that means quibbles that its partners have to iron out.
