March 15, 2024

Ina here, getting excited for Bay FC's inaugural NWSL game this weekend. Today's AI+ is 1,020 words, a 4-minute read.

1 big thing: AI's mind-body problem

Illustration: Natalie Peeples/Axios

Scientists using AI and other tools to simulate fruit flies, rodents and human toddlers aim to understand a key aspect of natural intelligence β€” and that research could move today's generative chatbots up the AI ladder, as Axios' Alison Snyder reports.

Why it matters: ChatGPT doesn't have a body β€” but some AI researchers think "embodied cognition" is a necessary ingredient to achieve the field's holy grail of artificial general intelligence, or AGI.

  • Others, including ChatGPT creator OpenAI, are betting all they need to reach AGI is to keep scaling up today's large language models with more data and more computational power.

The big picture: Language, reasoning and other abstract skills tend to get the most credit for human intelligence. But gaining knowledge of how the world works by walking, crawling, swimming or flying through it is an important building block of all animal intelligence.

  • A group of prominent AI researchers last year advocated for an "embodied Turing test" to shift the focus away from AI mastering games and language, which are "well-developed or uniquely human, to those capabilities β€” inherited from over 500 million years of evolution β€” that are shared with all animals."

How it works: Teams of neuroscientists, anatomists and machine learning researchers around the world are building detailed virtual models of rodents, flies and human infants.

  • Researchers from Google DeepMind and HHMI's Janelia Research Campus have built a virtual fruit fly by combining an anatomical model of the fruit fly skeleton, simulations of the physics a fly experiences (such as fluid dynamics, adhesion and gravity) and an artificial neural network trained on fly behaviors.
  • The behavior of the virtual fly is compared to the behavior of a real fly to update the virtual model until it matches the real bug's actions β€” walking, flying and crawling upside down.
  • Members of the team previously built a virtual rodent.
  • Researchers at EPFL in Lausanne, Switzerland, also published a virtual fly model late last year.

The goal is to understand "how the body mediates between the brain and the world," says Srinivas Turaga, a neuroscientist at Janelia and co-author of the preprint paper about the virtual fly posted yesterday.

  • Eventually, these models might be combined with diagrams of how neurons in the brain are connected with one another β€” "connectomes "β€” to try to understand how a network of neurons gives rise to a particular behavior.
  • "The body and the nervous system evolved together," Turaga says. "And so intelligence, in some sense, isn't just in the brain. There's also mechanical intelligence" that helps animals move.

The intrigue: Embodied cognition also helps animals understand how the world works β€” by experiencing it.

  • "There's an argument to be made that biological systems learn from interacting with the world," says Jochen Triesch of the Frankfurt Institute for Advanced Studies.
  • "Most machine learning systems today learn by basically passively absorbing large data sets, whether it is video or images or captioned images," Triesch says.
  • Learning through interaction with the world is something "really essential that most of the machine learning community is right now completely missing."

Triesch and his colleagues are interested in human cognitive development and have developed MIMo, a virtual human model with the body of an 18-month-old child with five-fingered hands. Its virtual body senses its surroundings with binocular vision, proprioception and a full-body virtual skin.

  • MIMo isn't as detailed as the fruit fly model, but that makes it much faster to simulate, Triesch says. There is a tradeoff between the level of realism and the computation required, and the MIMo researchers believe the critical part of their model is the touch-sensitive skin rather than the exact body shape.

Reality check: It's an open question β€” and debate β€” whether information about the brain-body relationship gleaned from neuroscience studies can be used to teach machines to work in the physical world.

  • "In AI, that's the hardest problem yet," says Aran Nayebi, a postdoctoral researcher at MIT who works at the intersection of AI and neuroscience to try to reverse-engineer neural circuits, including the visual system in mice.

2. Apple buys Canadian AI startup

Illustration: AΓ―da Amer/Axios

Apple has acquired DarwinAI, a Canadian startup focused on using AI to improve inspections in manufacturing operations.

Why it matters: Apple has said it is investing heavily in generative AI, but has yet to reveal any major uses of the technology.

Driving the news: As Bloomberg first reported, Apple bought DarwinAI earlier this year.

  • Apple essentially confirmed the acquisition, offering the statement it gives when it buys small companies.
  • "Apple buys smaller technology companies from time to time and we generally do not discuss our purpose or plans," it said in a statement to Axios.

The big picture: CEO Tim Cook said on Apple's most recent earnings call that the company would have more news on its generative AI plans later this year.

  • "We have a lot of work internally as I've alluded to before," Cook said, adding later that he sees a "huge opportunity" for Apple in AI, broadly, and generative AI, specifically.

Meanwhile, in a sign Apple is continuing to explore generative AI, yesterday researchers at the company published an academic paper on multimodal large language models.

3. Training data

  • Adobe's Firefly is having trouble with the balance between diversity and historical accuracy when generating images with AI. (Semafor)
  • The U.S. is pushing a resolution at the UN that would call for AI development that is "safe, secure and trustworthy" and gives people in all parts of the world equal access. (AP)
  • Experts explain why the government is afraid of TikTok. (Axios)
  • X removed a post shared by its owner Elon Musk that contained an unverified video that purported to show cannibalism in Haiti. (Axios)
  • Snapchat is testing a feature that lets users save their direct messages, executives told Axios' Sara Fischer.
  • Taiwan is building a satellite internet service that it can control instead of relying on Elon Musk's Starlink service. (New York Times)
  • Trading places: Morgan Stanley has promoted Jeff McMillan to be its firmware head of artificial intelligence.

4. + This

I've heard of rain delays, but a pro tennis match in Southern California had to be postponed yesterday because of a swarm of bees.

Thanks to Scott Rosenberg and Megan Morrone for editing this newsletter and to Carolyn DiPaolo and Caitlin Wolper for copy editing it.