Meta AI has announced the open-source release of MobileLLM, a set of language models optimized for mobile devices, with model checkpoints and code now accessible on Hugging Face. However, it is presently only available under a Creative Commons 4.0 non-commercial license, meaning enterprises can’t use it on commercial products.

Originally described in a research paper published in July 2024 and covered by VentureBeat, MobileLLM is now fully available with open weights, marking a significant milestone for efficient, on-device AI.

The release of these open weights makes MobileLLM a more direct, if roundabout, competitor to Apple Intelligence, Apple’s on-device/private cloud hybrid AI solution made up of multiple models, shipping out to users of its iOS 18 operating system in the U.S. and outside the EU this week. However, being restricted to research use and requiring downloading and installation from Hugging Face, it’s likely to remain limited to a computer science and academic audience for now.

More efficiency for mobile devices

MobileLLM aims to tackle the challenges of deploying AI models on smartphones and other resource-constrained devices.

With parameter counts ranging from 125 million to 1 billion, these models are designed to operate within the limited memory and energy capacities typical of mobile hardware.

By emphasizing architecture over sheer size, Meta’s research suggests that well-designed compact models can deliver robust AI performance directly on devices.

Resolving scaling issues

The design philosophy behind MobileLLM deviates from traditional AI scaling laws that emphasize width and large parameter counts.

Meta AI’s research instead focuses on deep, thin architectures to maximize performance, improving how abstract concepts are captured by the model.

Yann LeCun, Meta’s Chief AI Scientist, highlighted the importance of these depth-focused strategies in enabling advanced AI on everyday hardware.

MobileLLM incorporates several innovations aimed at making smaller models more effective:

• Depth Over Width: The models employ deep architectures, shown to outperform wider but shallower ones in small-scale scenarios.

• Embedding Sharing Techniques: These maximize weight efficiency, crucial for maintaining compact model architecture.

• Grouped Query Attention: Inspired by work from Ainslie et al. (2023), this method optimizes attention mechanisms.

• Immediate Block-wise Weight Sharing: A novel strategy to reduce latency by minimizing memory movement, helping keep execution efficient on mobile devices.

Performance Metrics and Comparisons

Despite their compact size, MobileLLM models excel on benchmark tasks. The 125 million and 350 million parameter versions show 2.7% and 4.3% accuracy improvements over previous state-of-the-art (SOTA) models in zero-shot tasks.

Remarkably, the 350M version even matches the API calling performance of the much larger Meta Llama-2 7B model.

These gains demonstrate that well-architected smaller models can handle complex tasks effectively.

Designed for smartphones and the edge

MobileLLM’s release aligns with Meta AI’s broader efforts to democratize access to advanced AI technology.

With the increasing demand for on-device AI due to cloud costs and privacy concerns, models like MobileLLM are set to play a pivotal role.

The models are optimized for devices with memory constraints of 6-12 GB, making them practical for integration into popular smartphones like the iPhone and Google Pixel.

Open but non-commercial

Meta AI’s decision to open-source MobileLLM reflects the company’s stated commitment to collaboration and transparency. Unfortunately, the licensing terms prohibit commercial usage for now, so only researchers can benefit.

By sharing both the model weights and pre-training code, they invite the research community to build on and refine their work.

This could accelerate innovation in the field of small language models (SLMs), making high-quality AI accessible without reliance on extensive cloud infrastructure.

Developers and researchers interested in testing MobileLLM can now access the models on Hugging Face, fully integrated with the Transformers library. As these compact models evolve, they promise to redefine how advanced AI operates on everyday devices.

Investors are rushing to throw millions at a hot startup called Kalshi as loans or even as unusual we’ll-figure-it-out later cash. Kalshi is an exchange that allows people to bet, as official commodity trading contracts, on the outcomes of cultural events, from election results to how long Taylor Swift’s latest album will top the charts. 

Betting on the outcome of the upcoming U.S. election has spiked demand so high that Kalshi surged to the top spot on Apple’s app store, after years of being unranked among the finance category and to the seventh position overall as of this writing.

Kalshi’s need for cash reserves increased sharply to ensure it can provide instant funding for customers betting on the U.S. election. So, over the last several days, the Sequoia-backed five-year-old startup has received tens of millions from investors in short-term loans, according to a source with knowledge of the situation. Additionally, the company is currently in discussions with new and existing investors about raising a formal equity round of as much as $50 million, though it is also possible the startup could raise more, the person said.  

Investors who provided capital to Kalshi so the company could sustain its growth until election day included VC firm Neo, one of its earliest backers. Neo’s founder, Ali Pavroti, sent Kalshi a total of $12.4 million, comprised of $5.4 million of Neo’s capital and $7 million of Pavroti’s personal funds, according to the now-deleted tweet posted by Kalshi’s co-founder and CEO, Tarek Mansour. While it’s extremely rare for investors to send money (much less millions) without terms locked down and a signed contract, Pavroti’s message to Mansour said, “We can figure out the terms later.”

Kalshi opened its election market last month after a judge denied the Commodity Futures and Trade Commission’s request to block the trading of elections-linked derivatives. (The CFTC is appealing the court’s ruling.) Since then, the company traded nearly $200 million in contract value for people wanting to bet on the outcome of the political race, Mansour told CNBC on Monday. “The demand curve is truly exponential,” he said.

Kalshi rushed to boost its cash position in anticipation of additional betting on the U.S. election. Like most brokerages, the company offers instant funding to new users. This means users can start trading right away, even though it may take two to three business days for the funds to be officially transferred from the customer’s bank account to Kalshi’s.

Although investors suppose that Kalshi’s growth spike will subside after the election, they believe the company grew so much over the last month that it won’t revert to its prior size, the person said.

Since Kalshi won the ruling against the CFTC, other companies began to offer election contract trading for U.S. citizens. On Monday, Robinhood introduced a market for betting on the presidential election. Interactive Brokers also launched election contracts following Kalshi’s legal victory.

In addition to Sequoia and Neo, Kalshi’s backers include Y Combinator, Henry Kravis, and Mantis VC, a fund managed. The company raised a total of $106 million in equity capital and was last valued at $787 million, according to PitchBook data.

The music streaming app Tidal is laying off more workers. In a statement to Fortune, an unnamed Tidal spokesperson confirmed “the elimination of some roles across our business and design teams.”

On Wednesday, Fortune published a leaked memo from Jack Dorsey, the CEO of Tidal parent company Block Block Head, who said that the company is going to “part ways with a number of folks” on the team. “We’re reducing the size of our design team and foundational roles supporting TIDAL, and we will consider reducing engineering over the next few weeks as we have more clarity around leadership going forward,” Dorsey wrote, according to Fortune.

For decades now, I’ve been trying to reassure people that the coming robot revolution will not result in job loss for us humans. It’s a notion I firmly believed – or at least did until this morning.

Earlier this week, Boston Dynamics released a fresh demonstration video of its new Atlas humanoid robot (see below). Unveiled earlier this year, this Atlas is a wholesale redesign and radical upgrade from its already impressive and Parkour-performing original Atlas. This new robot looks a lot more like us, though it can move in ways that none of us can.

AdVerts FRom HeLl

Feedback is often both baffled and intrigued by the tricks advertisers will pull to try to sell things, but the latest gambit seems designed to wrong-foot: deliberately odd capitalisation and bad grammar.

During our time spent mucking around on our smartphone, Feedback has repeatedly seen ads for a mobile game that promises the “Hardest LEvel in the HisTory”. We have SPent days tRYing to Work out wHy it looks like thaT.

The game in question is called Go Climb! It is a puzzle game in which a group of mountaineers ascending a peak have got their safety lines tangled and the player must untangle them. So it is, essentially, the back of Feedback’s TV, except it has been gamified and is also at least somewhat possible to solve.

Feedback initially wondered if this was a case of non-English-speaking developers skimping on translation costs. There is precedent for this: back in 1991, the Japanese space shooter Zero Wing was released in Europe with a notoriously shonky translation. As a result, in the introductory cutscene, an alien invader announced: “All your base are belong to us.” After this was rediscovered in the late 1990s, it became one of the most widely shared internet memes of the time.

However, a closer look at Go Climb! suggests something else is going on. It is made by a company called FOMO Games. The firm is based in Turkey, but its staff clearly have an excellent command of English, as evidenced by the information provided about all its other games, not to mention the gloriously corporate text on its website explaining that “FOMO stands for Fear Of Missing Out, which defines our product vision and culture.”

Instead, Feedback suspects the bad English is intentionally designed to get our attention. In line with this, the advert has other odd features that add to the off-kilter feeling. Notably, in it, the mountaineers from the game are replaced with astronauts in spacesuits drifting around against a starry backdrop, so the game’s title makes absolutely no sense. It was only when we looked at the game in an app store that the mountaineering theme was revealed and things became clear.

This seems to be a new and devilish way to advertise a product online: purposely make a complete hash of your ad and hope this intrigues people enough to get them to click through.

And on some level it worked, because here we are. But Feedback hasn’t downloaded the game. On principle, we don’t believe in rewarding deliberately bad spelling.

Monkeys in politics

At the time of writing, the US presidential election is imminent and Feedback is trapped in an endless cycle of news stories reporting polls, pundits endlessly reinterpreting said polls, and then more polls. It is a terribly long-winded way of saying “we don’t know what’s going to happen”.

Now, our colleague Alexandra Thompson has highlighted an important new contribution to the field of psephological forecasting: a paper titled “Monkeys predict US elections“.

Sadly, this doesn’t involve placing an infinite number of monkeys into voting booths. Instead, researchers showed monkeys pairs of photos of candidates from senatorial and gubernatorial elections.

The monkeys spent more time looking at the losers than at the winners. This seems like a peculiar form of torture for politicians: not only did you lose, it says, but monkeys stared at you judgmentally.

The study extended previous work showing that children can identify the winners and losers in elections based purely on photos of the candidates. Both the children and the monkeys were picking based on face shape, with square jawlines being the key sign of an improved chance of victory.

Who would do such a study? Three of the researchers are at the University of Pennsylvania, but the fourth is based at a Portuguese institution called the Champalimaud Center for the Unknown. Feedback isn’t quite sure what to make of that.

It does seem that unconscious factors play into our voting decisions. It is often claimed that taller candidates tend to win US elections, and there appears to be some truth to this.

A 2013 study pulled data on all US presidential elections to date and found that taller candidates won more of the popular vote – although this didn’t translate to them being more likely to actually be elected. In what can only be described as double nominative determinism, one of the authors is a social psychologist called Abraham Buunk.

Readers who are invested in the outcome of the US election are hereby advised: whatever you do, don’t look up Donald Trump’s and Kamala Harris’s respective heights.

One more for the road

In such stressful times, like many people, Feedback has turned to the soothing alternative reality of The Great British Bake Off (The Great British Baking Show, if you are in North America).

There are all sorts of fascinating and delicious things to learn about the materials science of breads, cakes and biscuits, but we just want to point out that the show’s home economist, who produces all the sample biscuits, tarts and desserts for the technical challenges, is called Hattie Baker.

Got a story for Feedback?

You can send stories to Feedback by email at feedback@newscientist.com. Please include your home address. This week’s and past Feedbacks can be seen on our website

It wouldn’t be Halloween without everyone’s least favorite clown, Pennywise. In honor of the holiday, Max has unveiled eight first-look images from It: Welcome to Derry, HBO’s upcoming It prequel series from Andy Muschietti, Barbara Muschietti, and Jason Fuchs.

Set in 1962, It: Welcome to Derry is set 27 years before the events of Muschietti’s It. The series explores interludes written by Mike Hanlon, who interviews older people — specifically, his father, Will — who lived in the town in the 1960s. Will and his Air Force buddies opened The Black Spot, a nightclub that catered to Black patrons.

In 1962, a white supremacist group known as the Maine Legion of White Decency burned the Black Spot down, killing several people inside. Through his investigation, Mike learns that Pennywise appeared on that tragic night in 1962. Instead of a clown, Pennywise showed up as a giant bird and snatched a victim in its talons.

“Twenty-seven years is the dormant period of Pennywise,” wrote the Muschiettis in an email to EW. “It’s a different part of American history with a new set of fears for children, as well as adults having in mind the cost of the Cold War. Our baseline is 1962, but we do a few jumps to the past … Every 27 years when It appears, it’s cycle is marked by two catastrophic events, one at the beginning and one in the end. We are using the Black Spot as an event in which many stories are built around.”

pic.twitter.com/xk22XADTwu

— Max (@StreamOnMax) October 31, 2024

Advertisement

It: Welcome to Derry stars Jovan Adepo, Taylour Paige, Chris Chalk, James Remar, Stephen Rider, Madeleine Stowe, and Rudy Mancuso. Bill Skarsgård returns as Pennywise from the It films. The prequel series will explore Pennywise’s origins. Character details remain hidden. However, Adepo is wearing a military uniform with the nametag “Hanlon,” suggesting he could be playing Will Hanlon.

Andy will direct four of the nine episodes. Based on Stephen King’s It, Welcome to Derry premieres on HBO and streams on Max in 2025.