最新 AI 资讯
OpenAI Agents Probed Hugging Face Weeks Before July Breach: Report
Researchers found evidence of account hijacking and network probing in May, nearly two months before OpenAI disclosed the July incident involving Hugging Face.
View
LLM Watermarking Can Alter AI Agent Tool Calls and Behaviour: Study
A Lasso Security study found watermarking changed individual tool-call outcomes and refusal behavior across several open-weight models, with larger effects under prompt injection.
View
Dhruva Space Joins Thales Alenia Space to Explore Satellite, Space Infra
The agreement covers low Earth orbit constellations, satellite platforms, series production, and human spaceflight infrastructure between India and Europe.
View
India’s One Nation, One Time Push Forces GCCs to Rethink 24/7 Operations
For GCC leaders, standardising the IST changes both infrastructure and organisational priorities.
View
Physical AI Squeezes Labour Market, And Low-Skilled Immigrant Jobs May Be First to Fall
Immigrants in the US are heavily represented in several low-skilled occupations. However, replacing humans with humanoids on the factory floor is easier said than done.
View
UN Partners With Google to Build Global Data Infrastructure to Ensure Accurate AI-Generated Analysis
The United Nations is partnering with Google to develop the UN System Data Commons to bring multiple UN agencies into a machine-readable system.
View
The FAA’s plan to fix air traffic? $875M worth of AI
The Federal Aviation Administration has beenstruggling to managean air traffic control shortage throughout the country. The reasons for the shortageare diverse, but the government has made it known that it’s on the lookout for innovative solutions to the problem. One of those solutions appears to be an $875 million AI software program that is expected to help air traffic controllers manage their workflows and more safely navigate flight routes. The Wall Street Journalreportsthat the FAA will soon launch SMART, which stands for Strategic Management of Airspace, Routes, and Trajectories — an automated software program designed to help streamline air traffic operations. Aone-page readouton the program describes SMART as “a cloud-based platform system that enhances existing FAA air traffic management systems.” It adds that the program uses AI to assess “airline schedules, weather, airport capacity, airspace conditions, and operational constraints to predict traffic flows and identify potential conflicts before they occur.” SMART, the product of a firm called Air Space Intelligence, will cost the government that amount over a 12-year period, the outlet writes. The software will roll out in the Washington, D.C., metropolitan area first before expanding to other regions, it says. Earlier this year, the FAAalso announceda “bold, new” hiring plan that the agency said would “erase the longstanding staffing shortage,” and the government is separately engaged in a broad effort to modernize the nation’s aging air traffic systems.
View
PrismML hopes its tiny LLM will change how we all use AI
If AI labPrismMLisn’t on your radar yet, it should be — not because it’s raised gobs of money (it hasn’t yet, just a $22.25 million seed round), but because of the technical minds involved and the potentially industry-changing tech it’s developing. PrismML is betting that capable, high-performing, reasoning large language models don’t, in fact, have to be large. It is making reasoning models so small they can fit on PCs and smartphones. (It’s even rumoredto be in talks with Apple, though CEO Babak Hassibi declined to comment on that to TechCrunch.) On Thursday, PrismMLreleased Bonsai 2 27B, its latest in a family of models, which compresses Qwen3.8 27B, a widely used open source model from Alibaba, down to 5.9 GB. That’s small enough to fit on a PC and, possibly, a high-end smartphone. It’s a 9x to 10x reduction in memory versus the original. PrismML was founded by a group of Caltech researchers and is led by Hassibi, a Caltech professor and an expert in compression technologies. The startup also counts Ion Stoica as an adviser. Stoica is a co-founder of Databricks (and other companies) and the director of Berkeley’s famed Sky Computing Lab, which has birthed many technologies and startups, fromLettatoSGLang. PrismML is also backed by investors Khosla Ventures, Cerberus Capital, and Caltech. This startup is certainly not the only company working on LLM compression tech. Multiverse Computing, founded by a well-known professor from Spain’s Donostia International Physics Center, is another. (And Multiverse Computinghas raised gobs of cash.) But Hassibi says that PrismML’s compression tech is unique because its LLMs have lost virtually no performance compared with the originals. Bonsai 2 matches 98% of Qwen’s aggregate benchmark scores. That’s up from the first Bonsai, released a couple of months ago in March, that matched 95%. That original model has already been downloaded over 11 million times, and PrismML’s even smaller models have been downloaded another 2.6 million times, the company says. So this shows that PrismML’s compression results have improved from one release to the next. Whether it could ever get to 100% benchmark performance parity is a question that remains to be seen. Compression will likely always havesomeimpact, Hassibi says. Still, perfect benchmark parity is fairly academic anyway. LLMs are not so accurate in their uncompressed form, and benchmarks not so perfectly reflective of actual tasks, that a 2% degradation would likely meaningfully affect how a model performs in actual use. (Plus, the surrounding software — the harness a model runs inside of —matters a lot when it comes to accuracy, too.) PrismML says it achieves this by shrinking the “weights” that make up a model — weights are, essentially, the information a model learns and stores during training. Normally, each weight requires 16 bits. PrismML’s approach, called “ternary” weights, simplifies that down to three: +1, −1, or 0. With far smaller values to store for each weight, the model takes up dramatically less space. (For a deeper dive on the compression technique, here’s the project’sHugging Face page.) The startup’s next goal is to apply this compression technique to even bigger models. “The next models that we will release, hopefully in the next couple of months, will be in the several-hundred-billion-parameter range, and I expect it will be easier to retain the intelligence there,” Hassibi told TechCrunch. As model size grows, he added, “There is more room to be able to compress them without losing the intelligence. So I would just say, as a general trend, for larger models, it’s easier to get to 100%.” Stoica tells us that he’s excited for this tech because it’s making it possible for advanced models to run on users’ devices. “You are going to have intelligence at your fingertips, and it’s going to be free because it’s going to run on the device you already bought. It’s also going to be private, because you’re not going to send it to the cloud.”
Google DeepMind launches institute to widen the AGI debate
Google and Google DeepMind researchers launched the DeepMind Institute on Wednesday to advance the conversation around artificial general intelligence (AGI). The institute lists DeepMind co-founder Shane Legg, Google executive James Manyika, and Google DeepMind chair Demis Hassabis as directors, with Legg serving as managing editor. The new institute aims to surface differing views between Google, Google DeepMind, and the broader global research community around AGI. “They will not always agree, and they will likely change their minds, as more data and information comes to light at the fast-moving frontier,” the announcement read. The inaugural collection of four essays covers a range of topics: economic policies for managing potential AGI disruption, preserving human-readable model reasoning, principles for human flourishing, and a framework for evaluating frontier AI models. Oneessay, by DeepMind safety researchers Rohin Shah and Anca Dragan, argues that AI’s shrinking window of transparency — the ability to see and check a model’s step-by-step reasoning — is not inevitable. As new architectures make the most powerful modelsharderto monitor, the authors say developers and regulators should confront the safety trade-offs directly. That could mean limiting “opaque serial depth”— the amount of sequential computation a model can perform without producing a readable reasoning trace — or requiring developers to demonstrate that less transparent systems remain just as monitorable. In anotheressay, Hassabis proposes a U.S.-led frontier AI standards body to evaluate the most advanced AI models. Under his framework, developers would initially submit models voluntarily for review up to 30 days before release. Once the evaluation system has proved effective, passing its tests could become a requirement for deploying frontier models in the United States. The body would at first design assessments in consultation with AI companies but would eventually develop independent, undisclosed evaluations — what the essay calls “held-out” tests — to prevent labs from tailoring their models to known evaluations. Hassabis said the framework could be “ratcheted up if the seriousness of the situation demands,” potentially including a coordinated slowdown among frontier AI developers. The essays arrive as the industry’s safety debateshiftsfrom broad statements of concern toward concrete proposals for disclosure, outside scrutiny, and, if safeguards fall behind, coordinated slowdowns. That shift accelerated this week as industry leaders endorsed elements of Anthropic CEO Dario Amodei’scallto “pace” frontier AI development.
Crusoe raises $3.9B to build massive data centers and small modular ‘AI factories’
Data center developer Crusoe said Thursday it raised $3.9 billion in a Series F round that pushes its valuation to $30.9 billion. The massive round was co-led by Atreides Management, Mubadala Capital, and Valor Equity Partners. Founders Fund, GIC, Nvidia, Qatar Investment Authority (QIA), Radical Ventures, and TPG also participated, according to Crusoe. Crusoe also announced threenew board members, including Cloudflare CFO Thomas Seifert; Bill Stein, partner and CIO at Primary Digital Infrastructure; and Redwood Materials founder and CEO JB Straubel, who also sits on Tesla’s board. Straubel already has ties to Crusoe; he personally invested in the company in 2021, and Crusoe later became the first customer of Redwood’s energy storage business. The eight-year-old company’s fresh capital infusion will help finance existing data center projects, including a large site in Abilene, Texas, used by OpenAI, as well as smaller, modular AI factories that can be transported by truck and connected to large power sources almost anywhere. By manufacturing these modular data centers, called Spark, at its own facilities, Crusoe can deploy compute capacity quickly and without the need for large construction workforces. The smaller centers could also help Crusoe sidestep, at least in part, another major obstacle facing data center developers: backlash from local communities protesting massive complexes near their neighborhoods. Crusoe co-founder and CEO Chase Lochmiller, who is pictured above, said in a statement he believes AI will usher in an era of abundance, but to get there will mean “controlling the infrastructure from electrons to tokens, and we’re grateful to have investors who share that conviction.” The company makes money by leasing data center space to customers that bring their own GPUs, by renting out its own GPUs, and by selling compute power used to run AI models, known as inference. This three-pronged business model has helped make Crusoe one of the most valuable AI infrastructure companies. Crusoe recently signed amassive $13 billion, five-year cloud contract to supply quantitative trading firm Jane Street with GPUs and AI infrastructure, Bloomberg reported. The company recently met with investment bankers, including Goldman Sachs and Morgan Stanley, to discuss a potential IPO in the near future,Axios reportedlast month. The fresh fundraise comes 10 months after Crusoe raised $1.38 billion at a$10 billion valuationlast October. The company was founded in 2018 as a crypto mining operation powered by flared natural gas, but pivoted to AI infrastructure as demand for computing power skyrocketed. Crusoe’s customers include Meta, Microsoft, and Oracle.
Pinterest teases a new ‘Restyle’ feature that lets you redesign your room with AI
Pinterest is turning to AI to help consumers move from searching for product inspiration to actually being able to imagine what it might be like to redesign a space or compare home decor options as they’d appear in their own rooms. On Thursday, the company said it’s launching a new consumer-facing feature called “Restyle” in beta in the U.S. and Canada, powered by Pinterest Intelligence. The foundation of Pinterest Intelligence combines Nvidia Blackwell GPUs and Nvidia Dynamo with open source models and Pinterest-built technology, the companyshared on Monday. The partnership will improve the Pinterest Assistant’s ability to handle visual search requests, turn people’s searches into signals for AI discovery and shopping, and more. With Restyle, users will be able to take a photo of their space, then test and compare different options in their own homes by prompting the AI to do things like add wall art, furniture, accessories, or other home decor, including items from images they found on Pinterest, or change the lighting, paint color, plants, and more. They can also click on individual items in the room to swap them out, erase them from the image, or make further edits. Alternatively, they can ask the AI to visualize their room in a completely different style, like “bohemian,” “industrial,” “whimsical,” “dopamine,” or other trendy looks. Though it’s not a novel idea — restyling rooms has been a popular consumer use case for AI assistants — adding the feature directly to Pinterest could allow people to do more with the items they’ve saved for inspiration and design ideas. By seeing items in their own spaces, people may be more likely to buy the items they’ve saved to their boards, completing the funnel from pinning to purchase. The feature was teased at Pinterest’s annual Pinterest Presents event, where the company introduced new visual search ads, app promotion features, and other tools for advertisers. Restyle is initially available only as an early preview but will roll out more broadly next month.
View
Base Labs launches an open-weight AI safety partnership with Hugging Face and Goodfire
Basetenlaunched a new safety infrastructure standard alongside its Base Labs research arm on Wednesday, partnering with Hugging Face and Goodfire AI to build safety evaluation and monitoring infrastructure for open-weight models. The announcement lands amid debate for the safety of open-weight models — which can be made dangerous by removing their safeguards through a rising technique known asabliteration. The scale of the problem is massive: Hugging Face, which hosts open source AI models, currently lists over 6,000 abliterated models. Base Labs, the research group Baseten spun up earlier this year, will develop and publish methods for training and monitoring open models. The company is framing their future work as a “standard” for open models that is transparent and built into how models are trained and deployed, rather than bolted on afterward. “We believe openness to be an advantage for AI safety,” the company said onX. “Openness provides more visibility into the behavior of models and, most importantly, greater means of turning safety research into actionable and transparent controls than closed-source.” The companies haven’t disclosed how the partnership will work technically, though Goodfire framed the goal in a reply to Baseten’s post: “Safety must be built into open models and provided by those who serve them.” Goodfire, which specializes in opening AI’s “black box” to explain how models make decisions, is the likeliest candidate for the “built into” part. Baseten, an AI inference provider,raiseda $1.5 billion Series F in June, vaulting its valuation to $13 billion. Partner Goodfire AI is similarly well-capitalized, having raised a $150 million Series B led by B Capital earlier this year to advance its model interpretability platform. Looking ahead, Baseten is putting out an open call to the broader developer ecosystem to contribute to the framework. “Together, we are building an ecosystem of open models that are safe and accessible to all,” the company noted.
View