by A small new open-source AI model performs as well as powerful large models
Melissa Heikkiläarchive page | MIT Technology Review
“[The Allen Institute for Artificial Intelligence (Ai2)] claims that its largest Molmo model, which has 72 billion parameters, outperforms OpenAI’s GPT-4o, which is estimated to have more than a trillion parameters, in tests that measure things like understanding images, graphs and documents. Meanwhile, Ai2 says a smaller Molmo model, with 7 billion parameters, comes close in performance to OpenAI’s state-of-the-art model, a feat it attributes to much more efficient data collection and training methods.
Hands-on with Orion, Meta’s first AR glasses
Alex Heide | The edge
“They’re watching almost like normal glasses. That’s the first thing I notice when I walk into a conference room at Meta’s headquarters in Menlo Park, California. The black Clark Kent-style frames on the table in front of me look modest, but they represent CEO Mark Zuckerberg’s multi-billion dollar bet on the computers that will come after smartphones. They’re called Orion and they’re Meta’s first augmented reality glasses.’
Startup says it can make a 100x faster CPU
Dina Genkina | IEEE spectrum
“Rather than trying to speed up computation by putting 16 identical CPU cores in, say, a laptop, a manufacturer could put 4 standard CPU cores and 64 so-called parallel processing unit (PPU) cores from Flow Computing in the same footprint , and achieve up to 100 times better performance.”
OpenAI will become a profitable company
Deepa Seetharaman, Berber Jin, Tom Dotan | The Wall Street Journal
“OpenAI plans to transform from a nonprofit to a for-profit company while simultaneously undergoing major personnel shifts, including the abrupt resignation Wednesday of its chief technology officer, Mira Murati. Becoming a for-profit company would mark a seismic shift for OpenAI, which was founded in 2015 to develop AI technology “for the benefit of humanity as a whole, unconstrained by the need to generate financial returns,” according to a statement which it published at launch. ”
Detachable robot hand crawls around on finger legs
Evan Akkerman | IEEE spectrum
“One of the great things about robots is that they don’t have to be limited by our limitations, and this week at ICRA@40 in Rotterdam we saw another new thing: a robot hand that can detach itself from its arm and then crawl around to to grab objects that would otherwise be out of reach, designed by roboticists from EPFL in Switzerland.”
SECURITY
Remember that DNA you gave to 23andMe?
Kristen V. Brown | The Atlantic Ocean
“Things are not going well for 23andMe. The share is about to be delisted from the stock exchange. It shuttered its own drug development unit last month, only the latest in a series of layoffs. Last week, the entire board of directors resigned, except for Anne Wojcicki, the company’s co-founder and CEO. Amid this downward spiral, Wojcicki has said she will consider selling 23andMe — meaning the DNA of 23andMe’s 15 million customers would also be for sale.”
An ultra-thin graphene brain implant has just been tested in a person
Emily Mullin | Wired
“Twenty years [after its discovery]Graphene is finally finding its way into batteries, sensors, semiconductors, air conditioners and even headphones. And now it’s being tested on people’s brains. This [week]surgeons at the University of Manchester temporarily placed a thin, tape-like implant made of graphene on the patient’s cortex – the outermost layer of the brain. The technology, created by Spanish company InBrain Neuroelectronics, is a type of brain-computer interface, a device that collects and decodes brain signals.
The first ever 3D printed hotel is underway in Texas
Evan Garcia | Reuters
“It looks like any other 3D printer, except it’s the size of a crane and it’s building a hotel layer by layer in the Texas desert. El Cosmico, an existing hotel and campsite on the outskirts of the city of Marfa, is expanding. 43 new hotel units and 18 residential homes will be built on an area of 24 hectares, all with a 3D printer.”
AI bots now beat 100% of those CAPTCHAs in terms of traffic
Kyle Orland | Ars Technica
“While there have been previous academic studies that attempted to use image recognition models to solve reCAPTCHAs, they were only successful 68 to 71 percent of the time. The increase to a 100 percent success rate “shows that we are now officially in the era beyond captchas,” say the authors of the new paper.”
Image credits: Victor / Unsplash
