最新 AI 资讯
A ‘pound of flesh’ from data centers: one senator’s answer to AI job losses
The signs that AI could lead to mass job displacement are already piling up: entry-level job postings in the U.S. havesunk 35% since 2023,masslayoffshavesweptacross Big Tech, and evenAI leaders themselvesare warning about what’s coming. Backstage at the Axios AI Summit in Washington on Wednesday, Sen. Mark Warner (D-VA) said a venture capitalist recently told him he’s writing software investments down to zero in large part due to the strides of Anthropic’s Claude, and a major law firm told him it’s not hiring first-year associates because AI can now handle much of the work once assigned to junior lawyers. Warner says the fear of AI-related job loss is “palpable,” even asdata from one AI companysuggests AI hasn’t yet started taking jobs. As those fears grow, they’re bleeding over into a different fight, which is who should foot the bill. Warner has a proposal: tax the data centers powering the AI boom and use that revenue to help workers through the transition. He hasn’t introduced legislation yet, but the idea is gaining urgency as public anger toward AI and data centers grows. Across the U.S., there’s beenpushback on data centers, including a bill on Wednesday introduced by Sen. Bernie Sanders (D-VT) and Rep. Alexandria Ocasio-Cortez (D-NY), calling for adata center moratorium. The loudest concerns are about noise, pollution, and rising electricity costs. But there’s a bubbling resentment underneath those concerns, a resistance to suffering the potential ill effects of having a data center in your backyard that powers the technology some fear will replace workers. Warner doesn’t plan to support his colleagues’ bill. On stage at the event, he said: “A data center moratorium simply means China is gonna move quicker, and this is one where we can’t lose.” There’s no stuffing the genie back into the bottle when it comes to AI and data centers, he added. And while Warner believes in strict requirements that ensure data centers don’t pass their water and power costs to residents, he told TechCrunch he thinks there’s another way for communities to extract their “pound of flesh” in a way that addresses the underlying job loss fears. “I’ve thought for a long time there’s an obligation from the industry to help figure this out and help pay for it, but one of the questions I was asking was, Who should pay?” Warner told TechCrunch. “Should it be the chip makers, Jensen [Huang, Nvidia’s CEO]? Should it be the large language model companies? Should it be the Goldman Sachs of the world who are using these tools to cut back on a number of first-year associates?” Ultimately, he said, he thinks the “easiest place to extract the pound of flesh is probably going to be from the data centers.” That could look like putting data center tax revenue toward training for new nurses or funding AI upskilling programs — so long as there’s a “tangible benefit to communities” as they navigate this economic transition AI companies have foisted on them. Warner sees it as a way to balance the need to build data centers with some obligation to the communities bearing their costs The idea is not without precedent. Warner pointed to Henrico County, Virginia which used the tax revenue from a local data center tokickstart a new affordable housingproject. Finding a way to connect data centers to a tangible benefit to the community will be essential, he says, because otherwise, “the pitchforks are coming out.” The public mood suggests he could be on to something. According to a recentNBC News poll, AI has a lower public approval rating than Immigration and Customs Enforcement (ICE), with 46% of registered voters viewing AI negatively compared to only 26% viewing it positively. In Virginia, that is playing out in a proposal torepeal the state’s tax breaksfor data center buildouts, which cost the state and localities nearly $2 billion a year in lost tax revenue in one of the world’s largest data center markets. Warner says other states might follow suit. AI and data centers, he said, are “easy to demonize.”
View
Cohere launches an open-source voice model specifically for transcription
Enterprise AI company Cohere on Thursday launched its first voice model: Transcribe is an open-source automatic speech recognition model that can be used for tasks like note-taking and speech analysis. Relatively light at just 2 billion parameters, the model is meant for use with consumer-grade GPUs for those who want to self-host it. It currently supports 14 languages: English, French, German, Italian, Spanish, Portuguese, Greek, Dutch, Polish, Chinese, Japanese, Korean, Vietnamese and Arabic. Cohere says Transcribe beats models such as Zoom Scribe v1, IBM Granite 4.0 1B, ElevenLabs Scribe v2, and Qwen3-ASR-1.7B Speech onthe Hugging Face Open ASR leaderboard, achieving an average word error rate (WER) of 5.42, lower than any other model on the benchmark. The company claims Transcribe had an average win rate of 61% over other models when human evaluators assessed its transcriptions for accuracy, coherence and usability. However, the model fell behind its rivals when it had to transcribe Portuguese, German and Spanish. Cohere says Transcribe can process 525 minutes of audio in a minute, which is high for its class of model. The company is planning to integrate Transcribe into its enterprise agent orchestration platform,North, and is making the model available through itsAPIfor free. The model will also be available onModel Vault, Cohere’s managed inference platform. Speech recognition models are growing increasingly popular as demand grows for note-taking and dictation apps like Granola andWispr Flow. Earlier this year, Cohere reportedlytoldinvestors that it was generating annual recurring revenue of $240 million in 2025, and its CEO, Aidan Gomez, was cited as saying that the startupmay go public “soon”.
View
Conntour raises $7M from General Catalyst, YC to build an AI search engine for security video systems
The surveillance tech industry today is in the spotlight, but not for the best reasons. With controversy around the U.S. Immigration and Customs Enforcementtapping into Flock’s camera networkto surveil people, and home camera maker Ringdrawing criticismfor building new features that would enable law enforcement to ask homeowners for footage of their neighborhoods, there’s currently a broad debate around safety, privacy, and who gets to watch whom. But controversy doesn’t erase markets, and the continued improvement of vision-language models has only blown more wind in the sails of companies building new ways to help companies monitor what goes on in their premises. According to Matan Goldner, co-founder and CEO of video surveillance startupConntour, the ethics around this topic are important enough that he says his company is quite picky about which clients to sell to. That may not come off as sound business sense for a startup barely two years in, but Goldner says he can afford to do this because Conntour already has several large government and publicly-listed customers, one of which is Singapore’s Central Narcotics Bureau. “The fact that we have such big customers allows us to select them and to stay in control […] We’re really in control of who is using it, what is the use case, and we can select what we think is moral and, of course, legal. We use all our judgment, and we make decisions based on specific customers that we’re okay [to work with] because we know how they will use it,” Goldner told TechCrunch in an exclusive interview. That traction has helped Conntour with more than being selective. Investors have taken note: The startup recently raised a $7 million seed round from General Catalyst, Y Combinator, SV Angel, and Liquid 2 Ventures. Goldner said the round closed within 72 hours. “I think I scheduled around 90 meetings in like eight days, and just after three days — we started on Monday and by Wednesday afternoon, we were done,” he said. Regardless, Conntour may be right in being picky, especially given how powerful AI tools in this space have become. The company’s own video platform uses AI models to let security personnel query camera feeds using natural language to find any object, person, or situation in the footage, in real-time — a Google-like search engine made specifically for security video feeds. It can also monitor and detect threats on its own based on preset rules, and surface alerts automatically. Unlike legacy systems that depend on preset definitions or parameters to detect specific objects, motion patterns or behaviors, Conntour claims its system uses natural and vision language models, which lends it a high degree of flexibility and usability. A user may ask, “Find instances of someone in sneakers passing a bag in the lobby,” and Conntour’s system will quickly search all the recorded footage or live video feeds to return relevant results. And because the platform bakes in AI models, users can simply ask questions about the footage and get answers in text, accompanied by the relevant video feeds, as well as generate incident reports. The company’s selling point, however, is its scalability. Goldner explained that the platform mainly differs from other AI video search services because it is designed to efficiently scale to systems comprising thousands of camera feeds. In fact, he said, Conntour’s system can monitor up to 50 camera feeds off a single consumer GPU like Nvidia’s RTX 4090. The company does this by using multiple models and logic systems, and then identifying which models and systems the algorithm should use for each query to require the lowest amount of computing power to give users the best results. Conntour claims its system can be deployed fully on premises, completely on the cloud, or a mix of both. It can plug into most security systems already in use, or can serve as a full surveillance platform on its own. But there’s been a long-running problem in the video surveillance industry: The quality of surveillance is only as good as the footage captured. It’s hard to make out details from the footage of a poorly-lit parking lot that was recorded by a low-resolution camera with a dirty lens, for example. Goldner says Conntour hedges for this inevitability by providing a confidence score along with its search results. If the source of a camera feed isn’t good enough quality, the system will return results with low confidence levels. Going forwards, Goldner says the biggest technical problem to solve is bringing the full level of LLM capability to its system while maintaining its efficiency. “We have two things that we want to do at the same time, and they contradict each other. One one hand, we want to provide full natural language flexibility, LLM-style, to let you ask anything. And on the other hand there’s efficiency, so we want to make it use very few resources, because again, processing [thousands] of feeds is just insane. This contradiction is the biggest technical barrier and technical problem in our space, and what we’re working really, really hard to solve.”
View
ByteDance’s new AI video generation model, Dreamina Seedance 2.0, comes to CapCut
OpenAI may be dialing back its efforts in the video generation market with theshutdown of its Sora app, but ByteDance on Thursday confirmed that its new audio and video model,Dreamina Seedance 2.0, is now rolling out in its editing platform,CapCut. ByteDance says the model allows creators to draft, edit, and sync video and audio content by using prompts, images, or reference videos. The phased rollout will begin with CapCut users in Brazil, Indonesia, Malaysia, Mexico, the Philippines, Thailand, and Vietnam, with more markets added over time. The news of the launch in CapCut follows a recent report that said themodel’s global rollout would be paused, while it worked to address intellectual property issues that drewcriticism from Hollywoodoveralleged copyright infringement.That likely explains the limited number of markets where the model is currently available within CapCut. In China, the model is available to users of ByteDance’s Jianying app. The video generation model works without reference images, even if the creator only uses a few words to describe the scene they have in mind, ByteDance says in itsannouncement. CapCut is also good at rendering realistic textures, movement, and lighting across a range of visual perspectives and angles, which the company notes could be used to edit, enhance, or correct creators’ own footage. Another use case would be allowing creators to test potential ideas based on early concepts or sketches before filming the real video. In addition, Dreamina Seedance 2.0 can be used for a wide range of content, including cooking recipes, fitness tutorials, business or product overviews, and videos with motion or action-focused content, where AI video models have historically faced challenges, the company explains. At launch, the model supports clips of up to 15 seconds long across six aspect ratios. In CapCut, the model will roll out across different areas, including editing features such as AI Video and generation tools like Video Studio. It will also come to ByteDance’s AI generation platform, Dreamina, and its marketing platform, Pippit. Given its ability to create realistic content, ByteDance says it has added safety restrictions, so the model won’t have the ability to make videos from images or videos that contain real faces. CapCut will also block the use of unauthorized generation of intellectual property. (However, if the restrictions were working properly, the model would be available now in the United States. Likely, more tweaks are still being made.) The content produced by Dreamina Seedance 2.0 will also include an invisible watermark, which will help to identify content made with the model when it’s shared off-platform, ByteDance added. This could aid in things like takedown requests from rights holders in the event that the model allowed copyright content through. ByteDance says it will partner with experts and creative communities as the model rolls out to iterate and improve upon the model’s capabilities.
View
Mistral releases a new open-source model for speech generation
French AI company Mistral released a new open-source text-to-speech model on Thursday that can be used by voice AI assistants or in enterprise use cases like customer support. The model, which lets enterprises build voice agents for sales and customer engagement, puts Mistral in direct competition with the likes of ElevenLabs, Deepgram, and OpenAI. The new model, called Voxtral TTS, supports nine languages, including English, French, German, Spanish, Dutch, Portuguese, Italian, Hindi, and Arabic. “Our customers have been asking for a speech model. So we built a small-sized speech model that can fit on a smartwatch, a smartphone, a laptop, or other edge devices. The cost of it is a fraction of anything else on the market, but it offers state-of-the-art performance,” Pierre Stock, vp of science operations at Mistral AI, told TechCrunch during a phone interview. Mistral said the new model can adapt a custom voice with a sample of less than five seconds, and also capture characteristics like subtle accents, inflections, intonations, and irregularities in the flow of speech. The model, based onMinistral 3B, can switch between languages easily without losing the characteristics of the voice, which is useful for use cases like dubbing or real-time translation. Stock said the company wanted the model to sound human and not robotic. The model has been built for real-time performance, according to the company. It has a time-to-first-audio (TTFA) — a measure of when the model starts ‘speaking’ after receiving input — of 90ms for a 10-second sample of 500 characters. The model also has a real-time factor (RTF) of 6x, which means it can render a 10-second clip in roughly 1.6 seconds. Earlier this year, Mistral launcheda pair of transcription models, one for large batch processing and the other for real-time use cases with low latency. With the new speech model, the company is likely aiming to provide a full suite of voice products to enterprises. “We plan to have an end-to-end platform that can handle multimodal streams of input, including audio, text, and image and output as well. The main benefit of that is you get way more information with an end-to-end agentic system that supports audio as an input or output,” Stock said. Mistral’s positioning is that its open source and customization bit will help enterprises adopt its voice models over competitors, as they can tune it the way they want.
View
Accenture, Anthropic Launch Cyber.AI to Expedite Cybersecurity Operations
The solution automates complex cybersecurity processes, protecting expansive digital environments without adding manual effort.
View
Humanoid Escorts Melania Trump, Greets Leaders in Bengali at White House Summit
In her address, the US First Lady urged governments and educators to adopt AI in education.
View
Namma Yatri Parent Moving Tech Acquires Automicle to Expand Zero-Commission Mobility Model into Europe
The deal extends MTI’s community-led mobility model internationally, reinforcing its vision of open, city-first infrastructure for sustainable urban transport and driver dignity.
View
Agentic AI Could Change Software Engineering Forever
Instead of developers writing every line of code, they now provide high-level specifications. AI agents take over from there.
View
Apple Has 'Complete Access' to Google's Gemini Model; Can Create Smaller Models via Distillation: Report
Apple has been granted full access to Google's Gemini model, which allows the iPhone maker to do more with the AI model used on Android smartphones, according to a report. The Cupertino company will be able to use the Gemini AI model for distillation in its own data centres, which means it can create smaller models that can be used for specific purposes. These models could be more efficient, would run on a user's device, and would not require access to the internet.
View
Yann LeCun Builds a World Model That Runs on a Single GPU
LeWorldModel can plan up to 48 times faster than some existing world models while maintaining competitive performance.
View
Infosys Announces FY26’s Largest Acquisition at $465 Mn; Total Deals Reach 5
The company has also acquired Stratus, extending the spree that includes Versent, MRE Consulting and The Missing Link.
View
