chat: AI news

Language models [LLMs], like humans, prefer to learn some linguistic patterns over others.   Their preferences bear some resemblance to human preferences, but they’re not necessarily identical, and it’s still possible that aspects of Chomsky’s theories play a role in how humans learn.  Human brains and neural networks are each so complicated that understanding how they differ —especially when it comes to a task as subtle as language learning —can seem hopeless .  The paper title “ Mission: Impossible Language Models ” is fitting in more than one way. But like action heroes, researchers have a habit of accepting seemingly impossible missions and finding creative ways to make progress. Kallini and her co-authors pinpointed a simple principle called “information locality” that explains why their models found some of the impossible languages harder than others. That principle might also be relevant to human language acquisition.  Their results have already prompted sever...

Finra, Wall Street’s self-regulatory arm, said last year that it is considering issuing new guidance on the evolving uses of AI at firms. It said generative AI presents “promising opportunities” for compliance efficiencies. Regulators also demand that institutions be able to explain the workings of systems that they rely upon, said Amy Jane Longo, a partner at the law firm Ropes & Gray.  Companies using them also have to make sure they aren’t introducing a new source of cybersecurity threats.   “I’m both optimistic and cautious,” Longo said. Telephone calls are routinely recorded and retained on Wall Street. Firms haven’t been required to comprehensively monitor the thousands of hours of daily recordings an office might generate, but dramatic improvements in transcription technology now make such scrutiny a possibility. “Historically, it’s been very expensive and complicated to transcribe voice, and the quality was not that good,” said Matt Smith, chief executive of Steel...

The European Commission has asked X to hand over internal documents about its algorithms, as it steps up its investigation into whether Elon Musk’s social media platform has breached EU rules on content moderation. The EU’s executive branch told the company it wanted to see internal documentation about its “recommender system”, which makes content suggestions to users, and any recent changes made to it, by 15 February. X has been under investigation since December 2023 under the EU’s content law —known as the Digital Services Act (DSA) —over how it tackles the spread of illegal content and information manipulation.  The company has been accused of manipulating the platform’s systems to give far-right posts and politicians greater visibility over other political groups.

Meta tracks your activity across millions of websites and apps, regardless of whether you use its platforms, and it profits from that data through targeted ads .  You might think of Meta as a social media company, but its primary business is surveillance advertising.  Meta’s business model relies on collecting as much information as possible about people in order to sell highly-targeted ads.  That’s why Meta is one of the main companies tracking you across the internet —monitoring your activity far beyond its own platforms.  When Apple introduced changes to make tracking harder on iPhones, Meta lost billions in revenue, demonstrating just how valuable your personal data is to its business.  

Microsoft made a trio of announcements this week that are going to be very important for how the company approaches its big AI bet in 2025.  It started off by creating a new AI engineering group to focus its developers on building an AI platform and tools for both Microsoft and its customers.  Microsoft then announced pay-as-you-go agents for its relaunched Copilot Chat for businesses, and  It finished the week by bundling Office AI features into Microsoft 365 for consumers and raising the subscription price. All three of these announcements are linked to Microsoft’s insatiable hunger to win at AI.  The company still prioritizes security through employee performance reviews, but it increasingly feels like promoting AI is an equally high priority.

Soraya Chemaly, an activist and author of “Rage Becomes Her: The Power of Women’s Anger,” said the “yoga-speak” language of “masculine energy” hides how problematic and “ridiculous” Zuckerberg’s views actually are. “He’s able to talk about masculinity and femininity like they’re a vibe , as opposed to institutional inequalities that he is actively perpetuating and deepening with his blather,” Chemaly said. This “touchy-feely” language of “masculine energy” also “ignores the fact that his company and his sector is overwhelmingly dominated by powerful, straight white guys.” The truth is, corporate America does not lack masculine energy. Men lead 90% of Fortune 500 companies today. Men still earn 12% more than women on average, and that disparity widens for Black and Latina women when compared to white men. And at Meta specifically, men outnumber women and account for 62% of its global workforce, according to the company’s 2022 diversity report. In other words, the energy has been decided...

“This is going to be a technology that really attacks the energy problem quite head on,” Turek says. The ability to put a capsule full of DNA on a shelf or in a drawer and forget about it is a major part of its appeal as a storage medium .  Turek says DNA could potentially replace sprawling, energy-guzzling data centers. Unlike storing data on servers, which requires constant energy use, DNA can sit on a shelf for years without needing energy until that data is retrieved. Of course, it means DNA isn’t the ideal medium for storing data that’s accessed often.  But researchers imagine that it could be an ideal way to preserve large amounts of archival data. “So often, biotechnology is hidden, kind of behind the curtains ,” says Niko McCarty, founding editor of Asimov Press. He and editor-in-chief Xander Balwit conceived of the DNA book project because they “wanted people to feel biotechnology.” He says Asimov Press has sold nearly 500 preorders so far.

"In recent years, there has been a tremendous interest in using generative AI, and particularly large language models (LLMs) in software engineering; indeed there are now several commercially available tools, and many large companies also have created proprietary ML-based tools for their own software engineers.  "While the use of ML for common tasks such as code completion is available in commodity tools, there is a growing interest in application of LLMs for more bespoke purposes. One such purpose is code migration. "This article is an experience report on using LLMs for code migrations at Google. It is not a research study, in the sense that we do not carry out comparisons against other approaches or evaluate research questions/hypotheses. Rather, we share our experiences in applying LLM-based code migration in an enterprise context across a range of migration cases, in the hope that other industry practitioners will find our insights useful.  "Many of these learn...

Their arms are capable of making decisions independently and can even continue to react to stimuli after being severed. Each has more degrees of freedom than we can count, and any of its hundreds of suckers, able to "taste" the chemistry of the octopus's environment, is able to change shape independently. Researchers found that nerves from multiple segments connect to different muscle regions —suggesting that the segments work together to control the muscles with a high degree of precision. They found that the nerves for the suckers also connect via the septa, creating a sort of nervous spatial map of the suckers, and allowing for the fine, individual control of each one as the octopus uses them to sense its environment through touch-tasting . "Thinking about this from a modeling perspective, the best way to set up a control system for this very long, flexible arm would be to divide it into segments," Olson says. "There has to be some sort of communication...

"There are many AI models out there that you can play with from companies like OpenAI, Google, and a host of others .  "But when you use them, you get the experience they want, and you run it on their computer. There are a variety of reasons you might not like this. You may not want your data or ideas sent through someone else’s computer. Maybe you want to tune and tweak in ways they aren’t going to let you. "There are many more or less open models, but setting up to run them can be quite a chore and —unless you are very patient —require a substantial-sized video card to use as a vector processor.  "There’s very little help for the last problem. You can farm out processing, but then you might as well use a hosted chatbot.  "But there are some very easy ways to load and run many AI models on Windows, Linux, or a Mac. One of the easiest we’ve found is Msty . The program is free for personal use and claims to be private, although if you are really paranoid, you’l...

"Posing as Brad Pitt, the fraudster sent her poems and songs and AI photos to look like the actor. "He knew how to talk to women, it was always very well done,  Anne said. "He started buying her luxurious bags, but told her that she had to pay for customs. This is how Anne lost her first €9,000 (£7,580). "Eventually, the scammer —still impersonating the actor —claimed that he had kidney cancer and needed money for treatment. "He told Anne that his ex-wide (sic) Angelina Jolie had frozen his bank accounts and he had no access to his fortune."   

Anticipating that 2025 will be an "intense year" requiring rapid innovation, Mark Zuckerberg reportedly announced that Meta would be cutting 5 percent of its workforce —targeting "lowest performers." Bloomberg reviewed the internal memo explaining the cuts, which was posted to Meta's internal Workplace forum Tuesday. In it, Zuckerberg confirmed that Meta was shifting its strategy to "move out low performers faster" so that Meta can hire new talent to fill those vacancies this year. Any employees affected will be notified by February 10 and receive "generous severance," Zuckerberg's memo promised.  "I’ve decided to raise the bar on performance management," Zuckerberg said. "We typically manage out people who aren’t meeting expectations over the course of a year, but now we’re going to do more extensive performance-based cuts during this cycle." Cuts will likely impact more than 3,600 employees, as Meta's most recen...

DeepSeek-V3 is another LLM to watch. Its performance is on a par with Llama 3.1, GPT-4o, and Claude Sonnet. While training was not inexpensive, the cost of training was estimated to be roughly 10% of the bigger models. Not to be outdone by Google, OpenAI previewed its next models: o3 and o3-mini. These are both “reasoning models” that have been trained to solve logical problems. They may be released in late January; OpenAI is looking for safety and security researchers for testing. Not to be outdone by 12 Days of OpenAI, Google has released a new experimental model that has been trained to solve logical problems: Gemini 2.0 Flash Thinking . Unlike OpenAI’s GPT models that support reasoning, Flash Thinking shows its chain of thought explicitly. Jeremy Howard and his team have released ModernBERT , a major upgrade to the BERT model they released six years ago. It comes in two sizes: 139M and 395M parameters. It’s ideal for retrieval, classification, and entity extraction, ...

People with aphantasia report a variety of experiences.  Some can “hear” in their minds, while others can’t imagine either vision or hearing.  Some have excellent autobiographical memory, while many do not.  Some have involuntary flashes of mental imagery.  Most dream in images, but some cannot.  Most are born with aphantasia, although a small minority acquire it after injury.  “Aphantasia is not a monolithic phenomenon,” Nanay said. Neither is hyperphantasia.  People with hyperphantasia see mental images that seem to them as real as the things they actually see.  The images that hyperphantasics see aren’t the same as hallucinations because they know, at the time, that they’re not real. But that doesn’t mean they don’t feel real. There’s a subset of people with extremely vivid imaginations who are known as maladaptive daydreamers.  Some choose to live in their imagination, rather than in real life, Dijkstra said. “They sit down on the couch, ...

The experiments of neuronal activity in rodents are important, experts say, because they place the idea of imagination into a physical reality: that of the brain itself .  “This gives us an opportunity to take a fuzzy cognitive concept like imagination” and link it to brain activity, says Daphna Shohamy, a cognitive neuroscientist at Columbia University, and director and CEO of Columbia's Zuckerman Institute, who was not involved in those studies or the review paper. Humans’ internal worlds are rich, however, and the studies of place cells in rats may not represent all types of human imagination. The animal results connect most directly with imagination that is based in experience and action, as in planning out a strategy for moving through the world, Nadel says.  But other experts believe the hippocampus has a much broader repertoire: it may also forge ties between ideas and information.  “I don’t think the hippocampus cares, really, about what you’re connecting,” Addis...

"Reactivation of earlier perceptual activity is thought to underlie long-term memory recall. Despite evidence for this view, it is unclear whether mnemonic activity exhibits the same tuning properties as feedforward perceptual activity.  "Here, we leverage population receptive field models to parameterize fMRI activity in human visual cortex during spatial memory retrieval.  "Though retinotopic organization is present during both perception and memory, large systematic differences in tuning are also evident. Whereas there is a three-fold decline in spatial precision from early to late visual areas during perception, this pattern is not observed during memory retrieval.  "This difference cannot be explained by reduced signal-to-noise or poor performance on memory trials.  "Instead, by simulating top-down activity in a network model of cortex, we demonstrate that this property is well explained by the hierarchical structure of the visual system.   "Together...

"There is increasing evidence that imagination relies on similar neural mechanisms as externally triggered perception.  "This overlap presents a challenge for perceptual reality monitoring: deciding what is real and what is imagined.  "Here, we explore how perceptual reality monitoring might be implemented in the brain. We first describe sensory and cognitive factors that could dissociate imagery and perception and conclude that no single factor unambiguously signals whether an experience is internally or externally generated.  "We suggest that reality monitoring is implemented by higher-level cortical circuits that evaluate first-order sensory and cognitive factors to determine the source of sensory signals.  "According to this interpretation, perceptual reality monitoring shares core computations with metacognition.  "This multi-level architecture might explain several types of source confusion as well as dissociations between simply knowing whether som...

"Visual illusions and visual imagery are conscious sensory events that lack a corresponding physical input.  "But while everyday mental imagery feels distinct from incoming stimulus input, visual illusions, like hallucinations, are under limited volitional control and appear indistinguishable from physical reality.  "Illusions are thought to arise from lower-level processes within sensory cortices. In contrast, imagery involves a wide network of brain areas that recruit early visual cortices for the sensory representation of the imagined stimulus.  "Here, we combine laminar fMRI brain imaging with psychophysical methods and multivariate pattern analysis to investigate in human participants how seemingly ‘real’ and imaginary non-physical experiences are processed in primary visual cortex (V1).  "We find that the content of mental imagery is only decodable in deep layers, whereas illusory content is only decodable at superficial depths .  "This suggests tha...

"Mental imagery can be advantageous, unnecessary, and even clinically disruptive .  "With methodological constraints now overcome, research has shown that visual imagery involves a network of brain areas from the frontal cortex to sensory areas, overlapping with the default mode network, and can function much like a weak version of afferent perception.  "Imagery vividness and strength range from completely absent (aphantasia) to photo-like (hyperphantasia).  "Both the anatomy and function of the primary visual cortex are related to visual imagery.  "The use of imagery as a tool has been linked to many compound cognitive processes and imagery plays both symptomatic and mechanistic roles in neurological and mental disorders and treatments." 

"Humans are voracious imaginers, with internal simulations supporting memory, planning and decision-making.  "Because the neural mechanisms supporting imagery overlap with those supporting perception, a foundational question is how reality and imagination are kept apart.  "One possibility is that the intention to imagine is used to identify and discount self-generated signals during imagery. Alternatively, because internally generated signals are generally weaker, sensory strength is used to index reality.  "Traditional psychology experiments struggle to investigate this issue as subjects can rapidly learn that real stimuli are in play.  "Here, we combined one-trial-per-participant psychophysics with computational modelling and neuroimaging to show that imagined and perceived signals are in fact intermixed, with judgments of reality being determined by whether this intermixed signal is strong enough to cross a reality threshold.  "A consequence of this ac...

Raspberry AI, a startup founded two years ago, is one of the technological solutions that helps expedite product development by allowing designers to visualize and iterate their ideas nearly instantly with its text-to-image platform. Rasberry’s founder Cheryl Liu, who was a private equity analyst at KKR focused on retail before working for Amazon and DoorDash, spotted an opportunity to apply generative AI to fashion design right after image models like Open AI’s DALL-E and Stability AI‘s stable diffusion became available in late 2022. “For the first time in history, you could rapidly create hundreds of designs in a way that you could never do before,” Liu told TechCrunch . She explained that before generative AI, designers would often have to order physical samples to visualize their ideas, which would take weeks. The other alternative was to use older computer-aided design tools like Browzwear and Adobe’s Photoshop. But with Raspberry designers can turn their sketches into photo-real...

"The proliferation of AI models is both a blessing and a curse. "Platforms like Hugging Face host over a million models, ranging from state-of-the-art neural networks to poorly designed or maliciously altered versions. Amid this abundance lies a growing concern: model provenance. "Imagine a widely used model, fine-tuned by a seemingly reputable maintainer, turning out to be a tool of a state actor. Subtle modifications in the training data set or architecture could embed biases, vulnerabilities or backdoors. These evil models could then be distributed as trusted resources, only to be weaponized later. "This risk underscores the need for robust mechanisms to verify the origins and integrity of AI models. Initiatives like Sigstore, which employs tools such as SLSA (Supply chain Levels for Software Artifacts) to verify software provenance, must extend their efforts to encompass AI models and datasets.  "Without such safeguards, the community remains vulnerable to ...