Must Read
Osaka University scientists used AI models to turn human brain scans into images that resemble what the person viewed.
The researchers used the Stable Diffusion image generator from startup Stability AI to help reconstruct the images. The AI method could re-create what test subjects saw by reading their brain scans.
More:
The researchers used functional Magnetic Resonance Imaging (fMRI) scans that were processed while participants from a prior study were shown distinct images, such as a teddy bear and an airplane.
They initially trained a model to link the fMRI data shown in the participants' early visual cortex to the images they had viewed.
They then trained a second model to connect image descriptions and fMRI data to the ventral visual cortex.
Once those models were created, they fed the brain-imaging data into Stable Diffusion, which was able to reconstruct images viewed by the participants at around 80% accuracy.
Zoom out:
Notably, the study only involved brain scans from four participants, and the AI models had to be customized to each person, which required "lengthy brain-scanning sessions and huge fMRI machines," according to Sikun Lin at the University of California, Santa Barbara.
The preprint paper, "High-resolution image reconstruction with latent diffusion models from human brain activity," was co-authored by Shinji Nishimoto and Yu Takagi of the Graduate School of Frontier Biosciences (FBS) at Osaka University.
SCIENCE
AI company D-ID has introduced a web app for users to chat face-to-face with photorealistic AI humans.
Dubbed chat.D-ID, the app combines OpenAI's ChatGPT with D-ID's text-to-video capabilities.
More:
The web app can now allow users who can't write or read to communicate with ChatGPT through synthetic voice and face technology.
Chat.D-ID, which is now in beta, lets users ask questions via either voice or text with an avatar named Alice.
"Video is more effective than text, so the app increases the power of large language models by adding a face," D-ID CEO Gil Perry told TechCrunch.
Eventually, D-ID plans to add more digital characters, as well as enable people to upload photos of individuals to be used as the face on the app.
Zoom out:
D-ID's AI software is also used by online genealogy service MyHeritage for its "Deep Nostalgia" tool, which makes still photos appear as if they're moving. People use it to animate old photos of their ancestors.
D-ID also recently launched Creative Reality Studio, a platform for creating videos of a person from text and a single image of that person. Users simply upload a photo and input text or an audio clip. They then choose a voice, language, and style, such as cheerful or excited.
Meta's latest language model, called LLaMA, was leaked online despite the company trying to limit its availability to academic researchers.
A 4Chan user shared details for obtaining the model through peer-to-peer file sharing. Download instructions later appeared on the developer platform GitHub.
More:
Reports have since confirmed the legitimacy of the LLaMA copies on GitHub.
In its response, Meta stated that it will still release AI models to researchers. It did not deny the leak.
"It's Meta's goal to share state-of-the-art AI models with members of the research community to help us evaluate and improve those models," Meta said.
The company first announced LLaMA (Large Language Model Meta AI) in February, saying the collection of models can outperform OpenAI's GPT-3 even though they are smaller.
It said it would only release the models to researchers "affiliated with organizations in government, civil society, and academia; and industry research laboratories."
DuckDuckGo is the latest search engine to trial a generative AI feature.
The "DuckAssist" AI tool, now in beta, is powered by technology from ChatGPT creator OpenAI and AI startup Anthropic.
More:
Using AI, the privacy-focused search engine can generate answers to queries using natural language tech.
Users click on a blue "Ask" button, and the AI service will deliver summarized information mostly from Wikipedia but occasionally similar sources such as Encyclopedia Britannica.
The answers are more conversational and often contain more information without showing multiple links.
According to founder Gabe Weinberg, DuckDuckGo is using OpenAI's Davinci model and Anthropic Claude models to generate the natural language summaries.
DuckAssist beta is currently free to use in the search engine's mobile apps and browser extension, with no login required.
What's next:
If the trial goes well, AI could be added to its web-based search engine at a later date.
The capability is "the first in a series of generative AI-assisted features we hope to roll out in the coming months," Weinberg said.
CNET
Humane Inc., a startup working on a secretive consumer device that uses AI, unveiled an investment round of $100M.
The company, founded by former Apple executives, also announced "strategic collaborations" with Microsoft and OpenAI, Qualcomm, and SK Networks.
More:
Humane was formed in 2018 by Bethany Bongiorno, a former Apple software engineering director, and Imran Chaudhri, an Apple design director from 1995 to 2016.
The wife and husband co-founders have yet to disclose what the startup's first product will be, only saying that it involves a "software platform and consumer device built from the ground up for artificial intelligence." The product is due out this spring.
A patent filing and videos indicate that it could be working on a wearable that would project a display onto a user's surroundings and let them interact with and manipulate that information using "air and touch gestures."
The company is now partnering with OpenAI to incorporate its AI technology into its device. It will also use Microsoft's cloud for its software services platform.
"Our first device will enable people to bring AI with them everywhere," Chaudhri said.
Funding:
Kindred Ventures led the recent Series C round, which included participation from Microsoft and early Human investor Sam Altman, the CEO and founder of OpenAI, among others.
Humane has now raised at least $230M to date.
