最新 AI 资讯
Amazon makes its AI-powered Alexa+ free on Fire TV, no Prime required
AI is coming to your TV, whether you want it or not. On Wednesday, Amazon said its AI assistant, Alexa+, will be rolled out to all compatible Fire TV devices in the U.S. for free, whether or not the customer has a Prime subscription. The update brings conversational search, smart home controls, and AI-powered recommendations, the company says. Previously,Alexa+cost $19.99 per month for anyone who didn’t have an Amazon Prime membership, and was initially made available to thenew Fire TV devicesthe company announced last fall. Now,Amazon saysthat everyone will be upgraded to Alexa+ automatically. They won’t need to download an app or sign up for a subscription. Compatible devices include all the current-generation Amazon Fire TV Sticks, the Fire TV Cube, Amazon Ember smart TVs, and other smart TVs that have Alexa+ built in, including Hisense and Panasonic. The move follows an industry-wide push to make AI services available on more consumer electronic devices, often through non-optional upgrades like this. Google, for instance, rolled outGemini to its Google TV platform earlier this year,replacing simple search features with AI-powered conversational modes. Rokuupgraded its voice assistant to AIlast year, too. The companies point to metrics like time spent with the features to suggest positive consumer adoption trends. For instance, Amazon says that Alexa+ customers now have nearly twice as many conversations on Fire TV as they did with the original Alexa, which apparently suggests that customers with Alexa+ are no longer using their TV only as a lean-back source of entertainment, and are instead engaging with the AI, too. Whether or not that’s a good thing is debatable. Amazon says that with Alexa+, users don’t have to ask for shows by title, but can instead ask for suggestions based on other factors such as theme, age or popularity — for example, “a top-rated thriller” or “a historical drama with a strong female lead.” The AI bot can also help customers manage and control their smart home, including displaying their Ring camera feeds on the TV.
View
TerraPower’s nuclear reactor has a secret weapon for powering AI data centers
Nuclear power startups have been pitching themselves as the antidote to what ails AI data centers: power that’s always available. Bill Gates-foundedTerraPoweris one of the latest to throw its hat that into that ring, with Bloombergreportingthat the startup plans to announce its first data center project this year. TerraPower did not say who the customer will be, though in January, it announced that Meta hadagreed to buyeight of its Natrium power plants. The data center project, expected to break ground in 2027, would be the company’s second power plant, with its first alreadyunder constructionin Wyoming. Not every nuclear reactor is suited to data center duty, but TerraPower possesses one key advantage — energy storage — that promises to give it an edge over competitors. And it’s all thanks to renewable power sources like wind and solar. Nuclear reactors, TerraPower’s included, work best when they’re running at full tilt. Of all the different types of power plants, nuclear reactors have the highest capacity factor —92.5% of the time, they generate at maximum power in the U.S. But in a way, they need to be. Existing reactors are slow to ramp up and down, capable of increasing or decreasing only about 5% of their total rated output per minute,accordingto the National Laboratory of the Rockies. New small modular reactors (SMRs), which many startups are pursuing, can react faster, about 10% of their rated output per minute, per NRL. But running at reduced capacity isn’t ideal — it’s hard to make money when you’re not generating electrons. That’s true of any power plant, but it’s especially true of nuclear, which has thehighest capital expendituresof any generating technology. Startups are hoping that mass manufacturing of SMRs will bring capex down, but that has yet to be proven. And if it does work, it could take adecade or moreto reap the benefits. Every startup acknowledges that its early power plants will be expensive, so it makes sense to operate it them at peak capacity as often as possible. For data centers, especially those that rely on behind-the-meter power, that poses a challenge. Their loads, especially when training AI or responding to prompts, can sink and soar quickly as GPUs respond to the tasks. The swings are so demanding that natural gas turbineshave been breaking under the stress. To smooth the curve, they need to use large banks of batteries, which increase costs further. TerraPower designed its 345-megawatt molten salt-cooled reactor to work around those challenges. One of the key considerations was ensuring the reactor could complement intermittent sources of electricity like wind and solar — the power plant needed to ramp up and down quickly. While TerraPower had renewable power, not data centers, in mind when it sketched its plans, the two are similarly intermittent, just on different sides of the equation. To ramp quickly, TerraPower doesn’t increase or decrease the power output of its reactor. Rather, it keeps on splitting atoms, and the extra heat gets stored in a giant vat of molten sodium. When power demand spikes, the power plant can tap that reservoir to generate more steam to spin the turbines. The expensive equipment keeps working even when demand is low, allowing the company to amortize its investment over more operational hours. The approach takes the best of nuclear power — high capacity factor — and pairs it with an energy storage technology that allows TerraPower to play nicely on a renewable-heavy grid or when connected to an data center. It’s a flexible approach that could give the startup an advantage in the race to power AI.
View
Bengaluru’s AlgoFET Raises ₹15 Cr to Scale Autonomous Drone Infrastructure
The startup will expand manufacturing, proprietary technology, and deployments across defence, enterprise, and global markets.
View
IBM Connects Cryogenic Modules for Next-Gen Quantum Computers
The company says the system could link hundreds of quantum chips as it works towards its planned 2029 fault-tolerant machine.
View
India is Building an AI Fibre Network. But Can It Survive a Cable Cut?
India is quickly linking AI data centres with new terrestrial and subsea fibre optic cables, but concentrated landing points and reliance on foreign repairs may jeopardise its AI ambitions.
View
Snowflake Adds Dynamic Model Routing to Cut Enterprise AI Costs
The update also brings DeepSeek-V4-Flash 0731 and GLM-5.3 to Cortex AI, alongside new controls for tracking usage, setting quotas and managing spending.
View
GitHub Has a New Problem Called Cursor Origin
Cursor is trying to change how developers manage code in an agent-driven world, but GitHub still has the advantage of habit and scale.
View
Murf AI Says Falcon 2 Voice Model Beats OpenAI, ElevenLabs on Naturalness
Murf AI plans to price Falcon 1 and Falcon 2 similarly. It also teased that Falcon 3 could arrive in the next couple of months.
View
Google Says Gboard Rambler Switches to Offline Mode After Usage Limit: Report
Google's Gboard Rambler, which uses Gemini to turn natural speech into polished text, comes with usage limits. The feature is designed to go beyond conventional voice typing by cleaning up spoken input and allowing users to make edits with voice commands. Google has now clarified how Rambler behaves when users reach its usage cap, providing more details about the feature's offline capabilities and the functions that remain available after the limit is reached.
View
Relativity Networks raises $22 million to bring a faster kind of fiber to data centers
Data center developers are expected to spend as much as $4 trillion by the end of the decade — and they’re already heavily constrained by both political and power-grid considerations in where they can build. But while most treat the speed of fiber as a given, one company is betting that faster fiber could change the geographical math behind the data center buildout. On Tuesday, Relativity Networks announced $22 million in SAFE note funding drawn by Rhapsody Venture Partners, Bell Ventures Inc., and Faster Than Glass LLC, among others. A SAFE note, in which an investment transfers into a specific numbers of shares once the company raises its first priced round, is a standard method used for pre-seed and seed rounds. The company also secured a $40 million follow-on order from a leading hyperscaler that declined to be named for this piece. Relativity Networks deals in hollow-core fiber, a rarely deployed technology that allows data to be transmitted 30% faster than conventional fiber. Where traditional fiber transmits light through fiber-optic glass, hollow-core fiber transmits the same light through a vacuum chamber in the center of the line, bringing it far closer to the theoretical limit of light speed. The difference is a matter of microseconds. CEO Jason Eisenholz estimates that a signal takes roughly five microseconds to travel one kilometer in conventional fiber. By switching to hollow-core, that figure can be reduced to only three and a half microseconds. When AI compute occurred across a single rack of GPUs, the fiber latency was easy to ignore — but as scale has grown, so has the physical distance between GPUs. Now, it’s common for a data center campus to sprawl across hundreds of acres and dozens of buildings. Eisenholz sees a particular opportunity for multi-campus deployments, in which pre-existing data centers are knit together to operate as a single unit. “The largest systems are distributing the compute across multiple campuses to reach the power that exists,” he tells TechCrunch. “They’re moving to where the warm shell is, but they still need to operate as one synchronized machine.” The result is a way to partially alleviate the harsh spatial logic that has restrained many ongoing data center buildouts. In latency terms, reducing time by 30% is giving developers an opportunity to span 30% larger distances before latency becomes a problem. As compute projects scale ever larger, Einholz thinks it could be a major shift for the industry. “The first era of AI optimized for compute,” he said. “It was GPU, GPU, GPU. The second era optimized the networking inside the data center to take advantage of that compute. The third era that we see coming is optimizing the geography.”
View
AI isn’t close to curing cancer. This startup says it knows what it will take.
A biotech startup calledVivodynesays the AI drug-discovery industry has a data problem, and that it has built a machine to fix it. HIVE, modular robotic labs built by the company, can grow 20 kinds of human tissue, then autonomously dose and monitor them, generating the kind of causal biological data that today’s AI models are missing — data that today mostly comes from animal testing, or studies of single cells or proteins, not living tissue. “Absent human testing, what are these [AI] models going to do?” asks Andrei Georgescu, Vivodyne’s CEO and co-founder. “They’re going to cure cancer in mice.” Even Anthropic CEO Dario Amodeiwroteover the weekend that claims that AI will cure cancer have become more cliche than credible — “the thing that will work isactually curing cancer,” as he put it. To be fair, the idea that AI will cure cancer is something Amodei himself has tossed out in previous essays; Sam Altman has repeatedly cited curing cancer as a justification for OpenAI’s push toward AGI and ever-larger compute buildouts; and Google DeepMind’s Demis Hassabissaidlast year that AI could potentially cure all disease within a decade. The actual results remain tepid. A handful of AI-designed drugs have proceeded into human trials — one as far as Phase III, widespread human testing — but the reality is that the roadblocks aren’t necessarily ones that AI can solve today. Nobel-prize winning Alphafold was a big advance for understanding the building blocks of life, but it has yet to actually produce a new drug. Isomorphic Labs, founded to build on Alphafold, is expecting its first trials, originally planned for 2025, by the end of this year. In February, the companywrotethat true drug discovery will require “highly accurate predictive models, across an expansive range of biochemical properties and interactions.” Georgescu says the space needs “a sanity check”— that existing models don’t have the data to capture the complexity of human biology. It’s a challenge already facing the pharmaceutical industry, where 90% of drugs that are effective in animal testing to enter clinical trials don’t receive regulatory approval for humans. Vivodyne’s plan is different. Vivodyne was spun out of the University of Pennsylvania in 2021, after Georgescu received a PhD in bioengineering there. The company says its tissues closely match the behavior of real human organs — that its liver cells have 94% predictive accuracy compared to human trials that test for toxicity, its airway tissue matches the behavior of real human tissue 96% of the time, and its bone marrow has achieved 100% concordance in tests of 20 different chemotherapy drugs. Last week, the company, which has raised just under $80 million across two rounds led by Khosla Ventures, opened what it calls the world’s largest “human data center” just outside of San Francisco, and Georgescu says his team is already achieving twice the throughput of all the animal trials being held in the US. The idea is to accelerate the path of drug candidates by having a better idea of what will work before going through the expense of a clinical trial, which typically costs tens of millions of dollars. Though it won’t name its partners publicly, Vivodyne says it is working with multiple major pharma companies to solve a problem that Georgescu compares to automotive crash tests: An automaker is typically confident its car will pass NHTSA requirements before testing it, but drugmakers rarely have that same confidence going into a clinical trial, where the vast majority of drugs fail to win FDA approval. But there is a larger vision: Georgescu sees his autonomous biology labs as key to generating the kind of causal data that can be used to train new models on human biology. He points to studies likethis one, published in Nature Methods last month, that find no clear data scaling laws when training generative AI models on existing cellular data. “All the training is done on static snapshots of these cells, and the models are not conditioned at all by thehowa cell got to that state,” Georgescu told TechCrunch. “In other words, the model learns ‘this is cell state A,’ ‘this is cell state B,’ but never ‘cell state B is the effect of inflaming cell state A.’” Vivodyne’s HIVE machines, however, are tracking hundreds of thousands of ongoing experiments where diseased tissue is exposed to some stimulus, which Georgescu expects to provide the kind of reinforcement learning that will produce AI models that understand human biology enough to make more meaningful progress in healthcare. Georgescu believes that will be key not just for today’s medicine challenges, but also for a future where complex diseases require drugs that, unlike the majority of those available today, target multiple pathways. “If we want combination therapies, the space that has to be searched explodes—it can’t be an experimental approach,” he told TechCrunch. “You have to say, ‘I want this effect to happen, so what cause should I invoke?’ Establishing causality in human biology is the basis of all of this.”
View
India Can Build Indigenous GPU by 2029, Says C-DAC Bengaluru Chief
C-DAC Bengaluru is developing a broader homegrown computing stack spanning CPUs, AI accelerators, RISC-V servers, and chiplets, with a fully indigenous HPC system targeted for 2030.
View
