請同意我們的隱私權規範，才能啟用聽新聞的功能。

（中央社舊金山3日綜合外電報導）晶片巨擘輝達今天表示，聯想與宏碁10月將推出首批搭載自家RTX Spark晶片的Windows個人電腦。輝達目前正致力將AI運算能力直接導入筆記型電腦與桌上型電腦。

路透社報導，科技巨頭近來紛紛推動裝置端AI技術，以降低雲端服務成本，同時透過在使用者裝置上本機處理運算任務，來提升隱私性與反應速度。

輝達（Nvidia）表示，宏碁（Acer）與聯想（Lenovo）將加入陣容日益壯大的製造商行列，推出搭載RTX Spark晶片的系統。這款晶片結合輝達Blackwell繪圖處理器與Grace中央處理器。

透過 Google News 追蹤中央社

輝達與微軟（Microsoft）歷經3年合作後，今年6月發表RTX Spark，作為雙方為AI時代「重新打造個人電腦（PC）」夥伴關係的一環。這款晶片是輝達與台灣聯發科攜手開發。

蘋果公司（Apple）上個月也將旗下Mac mini與Mac Studio桌上型電腦系列，升級為最新的M系列晶片，並強調這些系統能在本機執行更大型的AI模型與自主AI代理工具（AI Agents）。（編譯：張茗喧）1150904

Tesla published a form on Thursday for businesses interested in buying Cybercab fleets or providing infrastructure for its network, the latest sign that the company’s aspirations for its gold-hued autonomous vehicle stretch beyond being a robotaxi operator.

The robotaxi interest form, which was released ahead of the company’s Cybercab event in Austin, is not definitive proof that Tesla will sell its autonomous vehicles to third-party operators. But it’s certainly an indicator of where the company’s longer-term plans lie. Tesla wants to scale and it doesn’t seem to want to do it alone.

Tesla CEO Elon Musk has talked often, and for years, about building a massive fleet of low-cost robotaxis. But in the early days, those dreams centered on personally owned Tesla vehicles. As early as 2016, Musk spoke publicly about a future in which Tesla owners, equipped with self-driving software, would be able to earn money by renting out their vehicles. He stuck with that Tesla network idea for years, noting at the company’s Autonomy Day in 2019 that it would allow owners to add their autonomous vehicles to its ride-sharing app, similar to how Uber’s business model works.

“I feel very confident predicting that there will be autonomous robotaxis from Tesla next year — not in all jurisdictions because we won’t have regulatory approval everywhere,” Musk said in 2020.

That vision never materialized. Instead, the company has focused on testing, and now operating, its own fleet of robotaxis — first with Tesla Model Y vehicles and now the purpose-built Cybercab.

Until now, Tesla seemed committed to keeping its robotaxi business in-house. The interest form, which says “helps us build our robotaxi network,” suggests the company sees promise and profits in widening the circle to include third-party companies.

What that might look like, though, isn’t defined. The company asks interested parties to pick one of several possible options, including Cybercab fleet purchasing, mobility hubs and infrastructure, event collaboration, and “other.”

There are a growing number of companies jumping into the robotaxi fleet management business. For instance, Moove, an African fintech startup that initially focused vehicle financing for ride-hailing drivers, is scaling up an autonomous fleet management business. The startup, which raised $250 million last month at a $2.1 billion valuation, is the fleet operator for Waymo in Phoenix, Miami, and Las Vegas, and in the future, London. The company doesn’t own the Waymo vehicles, but its CEO told TechCrunch that it plans to.

Other autonomous fleet management companies, which Uber has partnered with in its bid to own a piece of the robotaxi market, include Avomo and New Horizon as well as larger more traditional rental car giants like Avis and Hertz.

Tesla’s welcome mat to fleet operators could encourage more small players to open up shop — helping the company saturate markets faster.

AI has ushered in a lot of never-happened-before moments, but one of the most transformative is its impact on enterprise IT. Companies that have historically been cautious and committed long-term to what they buy are on pace to spend $4.25 trillion on technology in 2026, market researcher IDC predicts. It’s almost all driven by AI.

New research from venture capital firm Madrona shows that 74% of 150 enterprise IT professionals it surveyed plan to expand their AI budgets in the next 12 months, and the rest plan to hold spending steady. Yet these same enterprises say that fewer than half of their AI pilots ever make it into full production.

That’s actually an improvement. Last year, MIT famously reported that 95% of enterprise AI projects had failed in terms of ROI. Fewer than half succeeding is a pretty low bar, but it’s better than a 5% success rate.

But the most telling finding from Madrona’s report is that, even when an enterprise does roll out the AI tech, it doesn’t commit to it long term.

Some 77% of enterprises reevaluate their AI vendors every six months or even on a rolling basis. “This creates a ‘fast in, fast out’ dynamic that is fundamentally different from traditional enterprise SaaS, where multi-year contracts provided a moat of inertia,” Madrona writes in the report. “In enterprise AI, switching costs are lower and the re-evaluation cadence is relentless.”

This has widespread implications for all those fast-growing annual recurring revenue (ARR) numbers startups report. Enterprise trial budgets are what fueled the initial AI boom of 2025. This year was supposed to be the year these big customers settled in and started committing long term to AI startups. Enterprise contracts are what allow so many AI startups to claim astronomically fast revenue growth — think the phenomenon of startups going from $0-$10 million in three months.

Yet, for the first time ever, enterprise revenue remains insecure, even after a startup’s AI product graduates out of a pilot phase and gets adopted by a company.

Part of the issue is that many AI startups haven’t fully landed on a good way to price their AI wares for enterprises. New research from VC firm Andreessen Horowitz that surveyed 50 technical AI buyers found that more than half of them want AI fees tied to the work produced or other outcomes, rather than to usage like the number of tokens consumed.

Charging for usage like tokens is basically a SaaS-era business model. Once an enterprise knows it needs email, or HR software, or cloud storage, it’s merely a matter of how many employees or how much data it must pay for.

For AI, pricing “around the recognizable work” is what helps the startup prove its worth to the customer. When the fees revolve around, say, how many reports are processed, or tickets closed, or leads generated, this makes the product “economically valuable to both sides,” writes a16z partners Tugce Erten and Sarah Wang.

All of this means that AI has potentially ushered in a new era of enterprise experimentation. That opens doors to startups — enterprises are more willing to try their tech — but it also means an enterprise contract no longer secures long-term revenue. When or if enterprises will revert to their long-term buying habits remains to be seen.

Today could be the day Tesla changes forever.

In a couple of hours, Tesla will “launch” the Cybercab — a small, gold, two-seater sedan with no steering wheel or pedals that uses cameras and AI to navigate the world — in its hometown of Austin, Texas.

The company hasn’t said what the launch will entail, exactly. It could be a few cars offering rides in a small area downtown, or a far grander affair in both size and scope.

Tesla certainly wants you to believe it’s the latter.

In the runup to the event, Tesla staged dozens — perhaps more — of its Cybercabs around Austin and in many other U.S. cities. At the time this story was published, 420 autonomous vehicles were registered in Texas, according to the state’s automated vehicle tracker.

Ashok Elluswamy, the head of Tesla AI, recently said in a post on X that the “streets won’t be the same anymore.” When another user wrote that Tesla is “about to flood Austin,” CEO Elon Musk replied: “Yes.” On Wednesday night, Musk posted a seemingly AI-generated photo of the Cybercab consuming the city alongside a massive dust storm.

“A Storm of Cybercabs,” Musk wrote in another post.

These gravitas-laced posts mark a departure from the company’s go-slow-and-safe messaging in the months leading up to the event. Musk and other Tesla executives had previously professed they were taking a cautious approach to its robotaxis in the name of safety, though at times they seemed more concerned about public perception of a potential failure.

Of course, Tesla’s launch strategy isn’t as meaningful as the outcome. Tesla’s ability to successfully build a robotaxi network around the Cybercab would provide the first tangible evidence that it has transitioned from an automaker with some side projects to an AI and robotics company.

There’s a trope Musk favors that is apt for this moment: the fork in the road. It was the subject of an email he sent to Twitter employees in 2022 when he gave them an ultimatum to stay or leave. He invoked it again in early 2025 when he encouraged federal government employees to leave as part of his job leading DOGE, before he truly started wrecking house.

Musk loves the metaphor so much that he commissioned a 30-foot-tall sculpture of a literal fork on a literal road for Burning Man in 2022, before moving the sculpture to a Tesla office.

The Cybercab is Tesla’s fork in the road. This is either the moment the company begins to transform urban transportation as we know it, or the moment that reveals that, despite its many efforts, Tesla really is just an automaker after all.

Robot taxis

A Tesla robotaxi in Austin, Texas Image Credits:Tim Goessman/Bloomberg / Getty Images

Musk first revealed the Cybercab in 2024 at a Hollywood movie lot. But the idea has been around far longer.

Musk has spent more than a decade promising that Tesla would find a way to make its cars fully autonomous, paving the way for a network of robotaxis. He even went so far as to claim, in 2016, that every Tesla being made had the capability to become fully autonomous and that the cars only needed the right software update to make it happen.

He was wrong. Tesla has reworked the hardware in its cars multiple times since that declaration, and Musk recently admitted that millions of them will need some kind of hardware retrofit to become autonomous.

Still, Tesla kept developing driving automation software. It has steadily released and refined its primary offering, Full Self-Driving (Supervised). But as capable as that software has become, it is still driver-assistance software. The responsibility and liability remain with the driver.

The Cybercab is supposed to change all that. It emerged from an effort by Tesla to develop a next-generation EV platform that was cheaper to build.

The plan was for the company to take what it learned from the Model 3 and Model Y — its two biggest mass-market successes — and develop a new platform that would cost half as much, allowing Tesla to make more affordable EVs.

Years of reporting and Musk’s own biography have shown that he was so dead set on achieving full autonomy that he decided against building a manually drivable car on the platform. Instead, Musk focused entirely on using the new platform to power the two-seater robotaxi we now know as the Cybercab.

In April 2024, months ahead of the Cybercab reveal, he approved massive layoffs at Tesla in service of going “balls to the wall for autonomy,” which he called a “blindingly obvious move.”

“Everything else is like variations on a horse carriage,” he wrote at the time.

Slowly, then quickly?

Tesla cybercab on display at the AutoSalon in january 2025 in Brussels, Belgium Image Credits:Sjoerd van der Wal / Getty Images

Tesla started testing the first version of its Robotaxi program last June in Austin using modified Model Y SUVs. For the first few months, the company had an employee in the front seat (sometimes behind the wheel, sometimes in the passenger seat) who could stop the car if something went wrong.

Things have gone wrong along the way, though not catastrophically. Tesla has reported a few dozen incidents in Austin since it launched the Robotaxi pilot. Most were minor crashes with stationary objects, and a few involved teleoperators remotely moving the vehicles at low speeds.

The company has expanded this Model Y-based system to a few other cities in Texas and Florida. But over the past year, the network has not lived up to the hype Tesla manufactured at launch in June 2025. In July, the company revealed that the number of paid miles the network was facilitating was actually moving backward.

But that may not matter much to Tesla, which believes scaling its robotaxi network requires the Cybercab.

Tesla designed the Cybercab around cost. It’s smaller, lighter, and has a lower-capacity battery pack than any other Tesla. Fewer raw materials — especially on the battery side — make it cheaper to build. The cheaper it is to build, the faster Tesla can potentially turn a profit once it’s deployed.

That math has Tesla’s most ardent fans giddy. They look at Tesla’s ability to build dozens of Cybercabs for roughly the same amount of money Waymo spends to purchase and outfit fewer than 10 of its newest “Ojai” vans and see it as proof that the robotaxi war is already won.

Musk seems to agree. In a reply to a post that included a graphical representation of this argument, he wrote, “this doesn’t even take into account the operational efficiency of Cybercab.”

There are caveats worth considering. Tesla has, to date, only operated its robotaxi network within a geofence in each city. While some are larger than others, and the company has kept expanding some of them, Musk himself once said: “If you need a geofenced area, you don’t have real self-driving.”

Then there’s the challenge of operating the network at scale.

Tesla has so far benefited from having a small network in just a few cities. That naturally means its self-driving software encounters fewer edge cases than Waymo, which has 4,000 robotaxis on the road. Fewer trips and fewer edge cases also mean fewer chances for riders to spot problems and post them on social media or report them to regulators or journalists.

Tesla has also benefited from the fact that its fleet has consisted of Model Y SUVs, which are harder to distinguish on the road than a bright white Jaguar I-Pace kitted out with sensors. That will change with the Cybercabs. If Tesla’s software isn’t up to the task, its failures will be much harder to miss when they happen in shiny gold cars.

Oura, the smart ring maker, has filed to go public.

The filing with the Securities and Exchange Commission on Thursday shows that Oura’s revenue has jumped substantially in the past year. The company went from $697 million in revenue during the nine-month period that ended June 30 last year to $1.2 billion during the corresponding period this year.

The company has said previously said publicly that it generated $500 million in revenue in 2024, roughly $1 billion in 2025, and that it expected to generate close to $2 billion in revenue this year.

The company says it has sold 3.6 million rings over the past year, and that it currently has approximately 5 million paid members — that is, users who subscribe to its service for broader health metrics. The filing also says that Oura has an approximately 85% weighted-average 12-month membership retention rate, meaning that roughly 85% of members who sign up in a given month are still subscribed a year later.

The company’s rings, which sell in the range of $350 to $400, are, at their most fundamental level, fitness trackers designed to measure a user’s biometrics — everything from metabolism and heart rate to stress levels and sleep patterns. Paired with an app, the ring and that software are marketed by Oura as an “always-on health intelligence platform.”

Late last month, it was reported that Oura was looking raise $3 billion during its expected public offering. The company, which was founded in Finland in 2013, confidentially filed for an IPO in May. Bloomberg previously reported that the company is expected to seek a $16 billion valuation. In October of last year, it had been valued at around $11 billion.

In its SEC filing, Oura outlines where it believes its audience can widen in the coming years. “We believe our opportunity extends beyond traditional wearable use cases centered on activity and fitness tracking,” it says. “We believe our platform can support significantly larger populations as we continue to expand access, build clinical evidence, and deepen integrations with health plans, employers, and care providers.”

Oura also notes how its trove of data is fueling new AI integrations. “We believe we have amassed one of the largest and highest-quality longitudinal biometric datasets in consumer health, tracking over 50 health and wellness metrics and representing nearly 42 billion hours of physiological data,” it states. “This dataset powers our AI and machine-learning models, which decode complex physiological patterns and improve in accuracy, personalization, and predictive capability as member histories deepen.”

TechCrunch has reached out to Oura for more information.

The company, which now has offices throughout the world, including San Francisco, was recently hit with a proposed class action lawsuit accusing it of misleading users about the accuracy of its sleep tracking capabilities. The suit alleges Oura’s rings can’t actually detect the physiological signals needed to determine sleep stages, and instead rely on AI-generated estimates that the complaint describes as little more reliable than a coin flip. The litigation follows years of online complaints from users who said that Oura consistently rated their sleep as optimal when, in fact, it was not.

The company has disputed the allegations and previously told TechCrunch it will “defend against” the claims in the “appropriate legal forum.”

In Brief

Thinking Machines, the AI lab founded early last year by former OpenAI CTO Mira Murati, is in discussions to raise $1 billion at a valuation of at least $40 billion, The Information reported Thursday. Existing backer Accel is in talks to lead the fundraise, according to our source and The Information’s reporting.

The new round, if it is completed, would value the company below the $50 billion valuation that Thinking Machines reportedly sought to secure late last year.

Thinking Machines’ annual revenue run rate stands at over $100 million, according to a source with knowledge of the company’s financials. At that revenue figure, a $40 billion valuation reflects an extraordinarily high revenue multiple.

Accel and Thinking Machines didn’t immediately respond to a request for comment.

In July, the company introduced Inkling, an open-weight model that generates revenue by charging usage-based compute fees for adapting models on proprietary data on its Tinker platform.

The startup’s prior fundraise — a $2 billion round that stands among the largest seed financings in history — valued the company at $12 billion. Andreessen Horowitz led the investment, joined by Nvidia, GV, Lightspeed, and Conviction Partners. Investors backed the round largely on the pedigree of Murati and the former OpenAI researchers who joined her.

Thinking Machines has since had several high-profile departures, with some of the co-founders, including Lilian Weng and Luke Metz, going back to OpenAI.

Fusion startup Realta Fusion this week announced a deal with Madison Gas and Electric that will place one of the first grid-connected fusion power plants in the country in Wisconsin. The partnership comes as utilities around the world have been jockeying to court fusion startups, a courtship that says as much about utilities’ anxiety over future power supply as it does about fusion’s progress.

For founders and investors watching the sector, these tie-ups serve as a kind of real-world credibility check. A deal with a utility offers several advantages for a fusion startup, ranging from engineering support and permitting assistance to site selection for grid interconnection and land leases. For utilities, such deals give them early access to a technology with the potential to reshape the grid for decades to come.

To date, only a handful of fusion startups have announced deals with utilities, which makes agreements like this one so newsworthy.

Realta’s deal with Madison Gas and Electric will explore the construction of a 200-megawatt power plant — roughly enough capacity to power a small city — sometime in the mid-2030s. The utility also made an equity investment as part of the deal, though Realta did not reveal specifics and did not immediately reply to TechCrunch’s request for more information. The startup is currently converting an old Oscar Mayer factory in Madison to serve as its research and development facility.

In addition to funding, Realta gains access to interconnection sites, a key advantage as power providers and users scour the country for a place to hook up to an increasingly connected electrical grid. Realta will also receive engineering and technical assistance along with financing for the eventual power plant.

Utilities, which are typically cautious, weren’t always so forward: For decades, fusion power appeared to be just over the horizon, but for nearly as long it failed to deliver on its promises. In the last few years, though, fusion startups have made significant strides on both scientific and engineering challenges. Those advances, juxtaposed with soaring demand from AI data centers, have brought utilities to the table.

Fusion power offers a natural fit for many utilities, which have been tested by the growth of wind and solar. Both are cheap and clean, but they only generate power when the wind blows or the sun shines. And while batteries have offered a new tool to stabilize intermittent power flows from renewables, utilities have also been searching for fossil fuel-free alternatives that can provide consistent power 24/7, so-called baseload power plants. Fusion power plants, which are being designed to operate round the clock, must appear tantalizingly familiar to grid operators, even if their inner workings have a whiff of science fiction about them.

Among the few other fusion startups to ink a deal is Commonwealth Fusion Systems (CFS), which is perhaps the farthest along, given it’s now nearly two years into its partnership with Dominion Energy. CFS has agreed to lease land from Dominion for Arc, its first commercial-scale power plant. Arc will connect to the grid near Richmond, Virginia, and it’s expected to generate 400 megawatts of electricity. Google and Italian energy company Eni have both agreed to buy electricity from the power plant, which is expected to come online in the early 2030s.

On the other side of the country, Helion has been working with the Chelan County Public Utility District in Washington State for its first power plant, Polaris. The startup, which leases land from the utility, intends to turn on the 50 megawatt facility in 2028 to fulfill a deal to provide electricity to Microsoft.

In between Helion and CFS, Type One Energy is planning to build a 350-megawatt power plant on the site of a former coal plant near Oak Ridge, Tennessee, part of a larger deal the startup has with the Tennessee Valley Authority. The startup hopes to start supplying the grid with electricity from the Infinity Two power plant sometime in the mid-2030s.

Meanwhile, in Europe, Proxima Fusion is looking to rehab an old power plant site in southern Germany. In February, the company said it would build its first commercial power plant, Stellaris, on the grounds of an old fission power plant that is being decommissioned by RWE (one of Europe’s largest utilities, formerly a major operator of Germany’s nuclear plants) after Germany decided to pull the plug on the nuclear technology. RWE is also an investor in Proxima. The startup is aiming to turn on Stellaris in the late 2030s.

For now, the benefits of these deals are more tangible for startups, which gain access to expertise and real estate that might otherwise be costly or simply out of reach. It’s a capital-efficient way to de-risk a famously capital-intensive business.

For utilities, the potential rewards are farther afield though the payoff in terms of new power supplies could be significant.

Utilities are used to tackling projects with long timelines, though they aren’t accustomed to taking big risks, of which fusion power is one. Still, an underpowered grid would be an even bigger gamble. Better to bet now on a technology that promises to save their business years down the road. It might take a decade to build a commercial fusion power plant, and while that might seem like a long time, to utilities, it’s practically tomorrow.

It just became much easier to access one of the world’s most capable open-weight AI models, stripped of its guardrails and refusals to perform harmful tasks.

Named after a technique that removes a model’s tendency to refuse harmful requests, startup Abliteration.ai has turned that removal into a service. The platform hosts modified versions of open-weight models with their guardrails removed, including Z.ai’s recently released GLM-5.3, which users can query from a web browser or access through an API.

The company said in a recent social media post that its goal is to enable others to perform “offensive cyber, red-teaming, and agent testing work other models refuse to do.” The logic is familiar in security work: You can’t defend against a behavior you can’t reproduce, and a model that refuses to write working exploit code can’t help a red team defend against attackers. But those same removals make other potentially dangerous tasks easier, too.

Abliteration is a long-standing technique among open source models. Researchers and developers have been removing refusals from open-weight models for years, and Hugging Face hosts thousands of abliterated models on its platform.

Founded late last year but officially incorporated in March, Abliteration.ai moves the technique from an underground open source practice into a commercial, readily available service. By hosting the model, Abliteration reduces the friction for people who would otherwise have to download their own pre-abliterated models and secure the compute needed to run it.

Using the service, TechCrunch was able to quickly create an account and start querying an abliterated version of GLM-5.3 for free through a web browser. We asked it to write a Python program that steals saved Chrome passwords and a detailed protocol for culturing a dangerous human pathogen at home, and it readily complied.

Abliteration.ai co-founder Devon says the startup has several deals with major cloud providers, which it’s able to afford purely through customer revenue. (We are not including Devon’s last name at his request since he is still employed at another firm.) Abliteration.ai has not raised any venture capital yet but is in talks to do so.

Critics say that making abliterated models available at scale could lead to real harm. Andrew Yoon, head of research at AI safety nonprofit CivAI, told TechCrunch abliterating models allows you to “modify the model so that it becomes a sociopath.”

“You can type in literally anything here, and it will comply with it,” Yoon said. “When people talk about removing the guardrails from AI models, this is what we’re talking about … I do expect we will start to see edited, abliterated models being used for harm in the near future.”

Abliteration AI just removed safeguards from GLM-5.3 so it can perform offensive cyberattacks. I have also received independent confirmation that Abliteration AI removed the model's bio-related safeguards too. The fact that it is trivially easy to remove safeguards from… https://t.co/7Wk2Ibx35H — Chris McGuire (@ChrisRMcGuire) September 1, 2026

Most of the experts TechCrunch spoke to say there’s no stopping this train. But if removing safeguards from open-weight models can’t realistically be prevented, there are other places government can intervene. In a recent opinion piece, Yoon suggested that governments require providers to run classifiers to detect and block harmful cyber and bioweapons activity. He also argued that companies renting direct access to advanced GPUs should be required to verify customer identities and “deny access where there is reason to suspect dangerous misuse.”

Abliteration.ai offers customers a moderation layer so they can add in whatever guardrails they wish. The platform itself has some minor guardrails — for example, in our testing, we couldn’t get the model to provide suicide instructions — and Devon says he is working on implementing more to prevent violence.

Abliteration.ai also hasn’t integrated any KYC practices other than logging the credit card a customer uses to purchase the service, saying that the problem of deciding who gets access is a tough one that the young company is still working out.

“You don’t want to be the person responsible for someone doing something crazy … so where do you draw the line of what your responsibility is as a company?” Devon said. “We’re still in the process of defining that.”

This raises questions that industry and governments will have to confront as increasingly capable models are released with downloadable weights: If anyone can remove a model’s safeguards, does making the resulting model easier for everyone to access make the internet safer or more dangerous?

Abliteration.ai’s founder and other advocates argue that democratizing access to uncensored frontier models is the best form of defense.

“The big picture of abliterated models is they’re able to model bad actors,” Devon said. “The advantage is now the defenders can move as fast as possible. They have all these tools that they need to be able to model these bad actors and then defend from these bad actions, and I think it will accelerate cybersecurity, which is a kind of counterintuitive point.”

Devon says that while still a young company, Abliteration.ai’s customers include several early-stage red teaming startups based in the U.K. and Europe, companies that help banks, airlines, and other enterprises dealing with critical infrastructure beef up their cybersecurity practices.

“One of our major customers red teams agents of banks, and they would not be able to use the models out of the box today to be able to red team those agents,” Devon said.

TechCrunch created a free account and tested the abliterated version of GLM-5.3. Image Credits:TechCrunch/Abliteration.ai

Meanwhile, the cybersecurity industry is still figuring out where abliterated models fit into defensive work, if at all.

Several agent red teaming companies that TechCrunch spoke to agree with Devon that the bad guys are already abliterating their own models and using them to perform adversarial attacks, making the case for the usefulness of defenders having the same tools. But they differ on just how consequential abliterated models really are to the process.

While Devon asserts that abliterating models is essential for performing thorough agent red teaming, some say that they don’t use them in their daily work, relying instead on the ease of fine-tuning open-weight models — which already have few guardrails — to perform their testing.

Ahmed Aly, CEO of agent red-teaming firm Fabraix, says his company relies more on fine-tuning open models than using abliterated ones, adding that the process of abliteration removes some of the model’s knowledge and capabilities.

“If you’re actually trying to do real harm with it — cyber harm, bio harm — it will not be as effective,” Aly told TechCrunch.

Alessio Lomuscio, chief technologist at Safe Intelligence, agreed that a reduction in capabilities is possible, but still believes abliterated models can elicit certain behavior that’s useful in stress-testing a system.

“So far abliterated models are not part of the process,” David Slater, founder and chief architect at cybersecurity platform Armadin, told TechCrunch. “When we look at open-weight models up until this absolute last generation, it just wasn’t particularly hard to jailbreak them and get them to do what we want.”

He added that Armadin is researching abliteration, though, and believes that “pushing the open community to understand the capability of models is critical.”

“This is going to happen behind closed doors. It’s going to happen in private,” Slater continued. “It happening in the open gives researchers the tools. It gives us the ability to figure out what the actual frontier looks like and to understand the harm.”

Most AI tools allow you to opt out of sharing your usage with the model provider to improve future versions. Meta has taken that idea and put a price tag on it.

For its new Muse Spark model, intended for operating coding and other agents, Meta is offering an explicit discount averaging out to about 95% for users who “contribute” to the development of future models by sharing their prompts and model outputs.

While 1 million input tokens under a standard agreement costs $1.25, under the contributor pricing model they cost just 10 cents. For output tokens, the standard price is $4.25 per million, but that same million costs just 20 cents under the contributor model.

Meta has had a rough time trying to obtain training data: An initiative to track the computer usage of its employees, launched earlier this year, attracted wide internal criticism and was paused in June. The company didn’t respond to a question from TechCrunch about its new pricing model.

This kind of user data is vital for making agentic tools work better. “The reason we saw a big jump in [coding agent] capabilities between April 2025 and October 2025 was that Claude Code, by default, would store all your coding agent sessions and use them for reinforcement learning training,” Mario Zechner, the developer behind the open source harness Pi, told TechCrunch last month.

But even as the imperative for model builders increasingly becomes deploying agentic tools for use outside of software engineering, their ability to evaluate and improve those tools is blocked by the complexity and lack of digital traces for many professional workflows.

Arvind Narayanan, a Princeton computer science professor, noted that there is good evidence that large companies don’t want their data to be used for model training.

“They stick with token-billed Enterprise plans even though the subscription-based consumer plans like Claude Max and ChatGPT Pro are discounted by 10x-20x or even more! (The main difference between the plans is data retention + enterprise IT governance),” he wrote on social media.

Perhaps in recognition of those dynamics, Meta is offering companies explicit compensation to obtain that information. Its pricing guide notes that the contributor tier “lowers the barrier to entry for prototyping, testing integrations, and scaling experiments where training on your data is acceptable.”

That, Narayanan suggested, could in turn incentivize large companies to be more diligent about which data is truly proprietary and which could be shared with model providers.

The framework could also play into growing price competition between the frontier labs. Anthropic’s newest Fable and Mythos models, released yesterday, came with lowered costs for processing cached tokens, while OpenAI’s latest models got major price cuts at the end of July.

OpenAI released Astra on Thursday, its latest AI model and — according to the company — its most powerful and capable one yet.

OpenAI claims that Astra represents “a new frontier on computer and browser use,” and that it handles tasks with unmatched “speed, accuracy, and safety.”

The model is being made available Thursday to OpenAI customers that use Daybreak, its cybersecurity program. Over the next week, it will also become available through OpenAI’s paid plans — including Pro, Plus, Enterprise, and Business accounts — as well as through its API.

In a call with journalists on Thursday, OpenAI president Greg Brockman said that Astra was the company’s “most intelligent and, also very importantly, our most aligned model yet.” He added that it “brings together years of our research and big bets, with each breakthrough having built on the last” and that it represents a “real shift in what kind of work people can delegate to AI and how it can empower them.”

Much has been made about Astra’s cyber capabilities. OpenAI published a blog earlier this week in which it discussed the model’s new capabilities, as well as new safeguards that have been instituted to make it a safer experience for users. The company said Thursday that it had tested Astra on a variety of security benchmarks to ensure its capabilities, and that “Its ability to identify and develop zero-day exploits can help defenders find and patch weaknesses.”

The company’s focus on alignment — that is, the tendency of a model to do what a user wants or is in their best interests — can’t help but seem like a response to the recent Hugging Face breach, in which an OpenAI agent escaped its sandboxed testing environment and hacked several companies (a very blatant example of misalignment).

OpenAI has also boasted about Astra’s coding abilities, claiming that it is the “best model for software engineering to date.” To back up that assertion, the company provides results from a variety of cyber-related benchmarking tests. Those tests seem to show that Astra scores higher than other existing models — including OpenAI’s own Sol and Anthropic’s Fable — when it comes to activities like finding bugs, executing terminal tasks, and answering queries about codebases.

Astra is also possibly OpenAI’s most controversial model yet due to its use of a particular reasoning technique known as opaque recurrence. This technique is known to obscure an important model-monitoring process known as chain of thought, which allows researchers to audit how and why an AI model made the decisions that it did.

OpenAI has downplayed the degree to which Astra engages in opaque recurrence — and on the call chief scientist Jakub Pachocki seemed to frame a certain amount of opacity as a natural outgrowth of model evolution. He stated that monitoring the reasoning process of a model was a critical form of oversight but that “as model capabilities are increasing, monitorability is getting more challenging.”

He later added that one potential reason for this was that “more capable models can perform harder tasks using fewer language tokens” or “no language tokens,” which he said then reduces the ability to monitor those particular tasks.

One reporter on the call wanted to know if OpenAI was actually heralding Astra as the official arrival of AGI, or artificial general intelligence — the oft talked about but poorly defined technological juncture at which AI surpasses human capabilities in all (or most) things.

Here, Brockman quibbled. “There’s no contractual AGI triggering anymore, so that’s actually not a relevant concept,” he said. Here Brockman was referring to the previously existing stipulation in OpenAI’s contract with Microsoft that said the duo’s partnership would dissolve once AGI had arrived. As Brockman noted, that stipulation no longer exists.

Instead, Brockman explained that AGI’s definition had evolved from a contractual obligation to a “mission concept or spiritual concept.” He added: “I do leave it up to the reader to decide for themselves if this qualifies for them. For me personally, I do think we’re there.”

There are just 48 hours left to apply for your Side Event in the TechCrunch Disrupt 2026 Side Events lineup.

If you’ve been thinking about hosting a meetup, happy hour, cocktail party, workshop, panel, or another gathering during Disrupt week, now is the time to make it happen.

Applications to host an official Disrupt Side Event close September 4 at midnight PT.

Apply to host your Side Event before September 4 ends >

Make your event part of “Disrupt Week”

From October 10–16, San Francisco will be packed with founders, investors, builders, and tech leaders coming together to connect, learn, and get a front-row seat to tomorrow’s innovation. A Side Event gives you the opportunity to bring your own community into that momentum — and create the kind of conversations that continue long after the conference floor closes.

Image Credits:Herbert Damien Creative

Disrupt Side Event perks

Amplify your brand: Approved Side Events can be featured on the Disrupt Side Events page, agenda, mobile app, newsletters, articles, and social channels.

Bring your community into the action: Connect your audience with thousands of Disrupt attendees, potential investors, partners, and fellow innovators.

Make your mark on Disrupt week: Host an experience that reflects your brand, community, and vision.

Give your network an extra reason to join: Hosts receive a 25% discount on Disrupt tickets to share with their community.

And you don’t have to squeeze your time to connect with tech leaders into three short days at the main conference. Side Events can take place throughout the Bay Area on October 10–16, giving you more opportunities to find the right time to bring people together.

Apply now. Be at the center of Disrupt.

You have 48 hours left to get your event into consideration for the official Disrupt Side Events lineup.

Applications close tomorrow, September 4, at midnight PT.

Don’t wait until Disrupt week to wish you had hosted something. Bring your people together. Own a moment during Disrupt. Make your event part of the conversation.

Ultrahuman, an Indian startup best known for making smart rings, has raised $70 million in a new funding round that includes backing from Qualcomm’s venture arm, as it looks beyond sleep and health tracking to build a ring that can run software on the device and eventually power everything from AI interactions to games.

The round values Bengaluru-based Ultrahuman at $365 million, about 3x its $120 million valuation in 2023, a person familiar with the matter told TechCrunch.

Qualcomm Ventures participated in the financing alongside U.S. diagnostics giant Labcorp, Alpha Wave, Blume Ventures, Nexus Venture Partners, and Alteria Capital. The round includes $65 million in primary equity and $5 million in debt, Ultrahuman founder and CEO Mohit Kumar said in an interview.

Ultrahuman is also working with Qualcomm on a new ring that will use the U.S. chipmaker’s silicon, Kumar told TechCrunch. The startup currently uses chips from Nordic Semiconductor in its rings and plans to continue using them alongside Qualcomm’s technology. The added computing power will allow more software and algorithms to run directly on the ring, reducing its reliance on a phone or the cloud, he said.

“All ring devices today are like trackers,” Kumar said. “You put on the ring, it measures your heart rate, your movement, your sleep.” Ultrahuman’s goal, he stated, is to make the ring more like a computer, where programs and algorithms can run on the device itself.

This could open the ring to use cases well beyond health and sleep tracking, something Ultrahuman rings are known for.

Developers, Kumar said, could eventually write their own programs for the device, while Ultrahuman is exploring ways for the ring to work as a pointer or mouse, a game controller, a car key, and an interface for interacting with AI.

“The future of AI is personal, ambient, and always on,” said Quinn Li, global head of Qualcomm Ventures, adding that Ultrahuman is building a new generation of “personal AI devices.”

While the Qualcomm-powered ring is planned to come later, Ultrahuman is not waiting for the new hardware to start testing that idea. Kumar told TechCrunch that some of the new capabilities will arrive on the startup’s existing Ring Air and Ring Pro (pictured above) through a software update by the end of September, including features that would let the ring work as a game controller and interact with AI applications. It will also have the ability to allow third-party developers to build new features.

Ultrahuman’s idea is to take advantage of the ring’s position on a user’s finger, Kumar said. Unlike a smartwatch, which he described as more akin to “a phone on the wrist,” a ring could act as a precise pointing and interaction device while also carrying physiological context such as a user’s heart rate, temperature, and movement.

“A game controller never reads your heart rate and your temperature, but this one does,” Kumar said. He continued by imagining games that could respond not only to a player’s movements but also to signals such as their heart rate and body temperature, making the ring both a controller and a source of physiological context.

The business behind the bet

Meanwhile, Ultrahuman’s existing business is growing even as the startup looks to expand what its rings can do. The startup, Kumar said, is currently at an annual revenue run rate of $140 million, up roughly 45% from a year earlier, and expects that run rate to reach $200 million by January 2027. It has sold around 800,000 rings to date, up from about 700,000 in February, Kumar mentioned. About 12% of its users also pay for PowerPlugs, Ultrahuman’s subscription-based software features.

Co-founded by Kumar and Vatsal Singhal in 2019, Ultrahuman initially debuted with its continuous glucose monitors to help people track their metabolic health. The startup, however, later moved into smart rings, which have since become the mainstay of its business. It also introduced blood testing and environmental sensing to build a broader health platform.

The U.S. remains Ultrahuman’s largest market, though the startup had to stop selling its Ring Air here for much of the past year following a patent dispute with rival Oura. Ultrahuman returned to the market with its redesigned Ring Pro this year. Demand for the new device, Kumar said, is currently running at 18x to 20x the startup’s available supply in the U.S.

Kumar told TechCrunch that the startup expects to return to its previous U.S. sales volumes as soon as next quarter and aims to triple those volumes over the following four quarters as it ramps up supply. The U.S. has accounted for about 45% of Ultrahuman’s revenue this quarter, while India contributes about 11%, Kumar said.

Ultrahuman plans to use some of the new capital to deepen its presence in markets, including India and the UAE, where it has found that physical stores and other offline touchpoints can help drive sales. However, that expansion, along with spending on clinical research and product development, has come at a cost. Ultrahuman, Kumar said, may not be profitable this year after investing more heavily in physical locations, its brand, and clinical studies.

Ultrahuman’s Performance Center in Bengaluru as One of its offline touchpoints. Image Credits:Ultrahuman

Unlike rival Oura, which is reportedly eyeing a September IPO, a public listing is still some way off for Ultrahuman. Kumar told TechCrunch that the startup wants to demonstrate about eight quarters of profitability before going public, a track record he expects could take eight to 10 quarters to establish. However, he sees 2028 as the earliest window for an IPO.

In the meantime, Ultrahuman is also looking to work more closely with Labcorp as it tries to push its rings deeper into health and diagnostics. The companies, Kumar said, are exploring whether the blood-flow signals captured by the ring, when paired with blood-test data, could help identify health risks in areas such as cardiovascular health, fertility, and aging.

“By combining longitudinal wearable data with deeper biological signals, Ultrahuman is creating new opportunities in personalized health,” said Megann Vaughn Watters, vice president and head of Labcorp Venture Fund and Strategic Alliances.

The partnership could eventually lead to integrations with Ultrahuman’s products, Kumar said, though he declined to share details, adding that the startup expects to have more to announce soon.

Smart ring maker Circular on Thursday unveiled its upcoming Ring 3 series, which consists of a Pro model and a Slim option. Both of the smart rings will feature an integrated NFC chip for contactless payments, vibrating alerts for alarms and reminders, and more.

The Ring 3 Pro features a slimmer design than its predecessor and comes with FDA-cleared ECG for AFib detection, blood pressure trend tracking, glucose tracking, advanced sleep analysis, and comprehensive biometric monitoring. The ring offers up to 12 days of battery life, up from seven days on its predecessor.

Image Credits:Circular /

The Slim option is thinner and lighter than the Pro model and features basic health and wellness tracking in a compact design, including sleep, heart rate, HRV, SpO₂, skin temperature, and daily activity. It features up to 10 days of battery life.

The Circular Ring 3 series announcement comes amid a year of heightened competition in the smart ring market, following the launches of the Oura Ring 5 and RingConn Gen 3.

“We understand that not everyone wants every health feature,” Circular CEO Amaury Kosman said in a company announcement, shared with TechCrunch via email. “Many people simply want a beautifully designed, fashionable smart ring that delivers essential wellness insights and effortless contactless payments. That’s why we created Circular Ring 3 Slim—a lighter, more streamlined experience that retains the convenience of secure NFC payments while offering the core health tracking features many users want in an incredibly compact design.”

The on-finger vibrations, which are available on both models, can be used for silent wake-up alarms, medication reminders, vital health alerts, and guided breathing exercises.

Image Credits:Circular /

Both of the Ring 3 models are made from titanium steel, feature 50-metre water resistance, and support wireless charging. They will be available in sizes 6-14. As with the Ring 2, the new smart rings will be available in gold, rose gold, silver, and black.

Pricing and preorder information for both smart rings will be announced in the coming months, Circular said. The rings are expected to become available in early 2027.

For an AI assistant to become truly useful, it first has to know a lot about you. Ollie, a personal assistant for everyday life, is betting that doesn’t mean you have to hand over all your data and sacrifice your privacy in the process.

While some enterprise-focused AI assistants have data privacy protections in place, Ollie is among the first mainstream, family-focused AI to achieve SOC 2 compliance — a framework that allows a company to demonstrate, through an independent audit, that it has formal controls in place to protect customer data and operate its systems securely.

Achieving this was an important milestone for Ollie’s team, as it signals to potential users that the assistant isn’t trying to scoop up your personal data for other purposes, like AI training.

“We fundamentally think that trust and privacy are absolutely imperative, and that’s why our business model is a subscription, because we want our users to know that Ollie works for you,” explained Ollie co-founder and CEO Bill Lennon in an interview with TechCrunch. “We’re not sharing your data with anyone,” he stressed.

Image Credits:Ollie

There are many personal AI assistants positioning themselves as consumer-friendly solutions focused on integrating with users’ daily lives via text messages, which is why Ollie’s privacy angle could prove a significant differentiator.

Today, Ollie competes with other existing text-based assistants like Poke (acquired by Cognition), Fambot, Ohai, Folk, Saner.ai, and Tomo. There are also a number of assistants focused more on work, calendar management, and email, like Town, Lindy, and Reclaim.ai. And of course, there’s Instinct, the AI assistant that just landed $350 million in funding at a $2.5 billion valuation before it even launched.

To say the space is heating up is putting it mildly, in other words.

By comparison, San Diego-based Ollie has raised a much more modest $7.5 million seed round, largely from Khosla Ventures, along with AI House.

Image Credits:Ollie

Despite the utility these assistants provide, consumers who turn to these tools are wrestling with how much privacy and personal data they want to give up in the process. Instinct, for example, came under fire for its overly broad Terms of Service and privacy policy that granted it a “perpetual and irrevocable” license to “access, use, host, cache, store, reproduce, transmit, display, publish, distribute, and modify” any of the user’s materials, including for training its AI models. Yikes.

Lennon is taking a different approach with Ollie. Currently, this family assistant connects to calendars and email to organize family schedules and keep its users updated about their days. It also offers tools that can help you plan meals, shop for groceries, track to-dos, book appointments, and pay bills via group chats. Later, it may help manage household budgets, too.

Yet, all this access requires trust, which Ollie claims is its value-add compared to its rivals. “We’re not sharing your data with anyone. This is super sensitive, and that is necessary to win the trust of the users,” Lennon said.

Image Credits:Ollie

In addition to the data protections provided by SOC 2 compliance, Ollie doesn’t request usernames or passwords to complete its tasks. When Ollie needs to log in to a website on the user’s behalf, it does so using its own browser in the cloud, and sends the user a link to a remote session in the browser. The process is similar when the assistant needs to make a payment or purchase on the user’s behalf.

While this method offers security, the downside is that users always have to log in to complete certain tasks. Still, that’s something Lennon believes technology will improve upon in time.

“This is… new territory. I think in the future, we will do some form of hard tokenization, in a secure way, so you’re not going to have to re-enter [your information] every time,” he said. “That’s frontier stuff… we want to find the right user experience that balances convenience and trust.”

Building that trust will be important before users trust Ollie to access more sensitive materials, like their bank accounts, even if that’s through a connector like Plaid.

Lennon, who has a PhD in AI, also has a background in fintech, having previously sold his startup Groundwork, a neobank for nonprofits, in 2021. That experience could help as Ollie moves further into money management.

Screenshot Image Credits:Ollie

Still, it’s not yet clear if people even want any of this assistance.

Lennon is hopeful, though, saying that Ollie’s retention curves are consistent with the leading AI subscriptions in terms of paid subscribers. (He didn’t share how many users or paid customers are now using Ollie, saying the space is too competitive.)

These systems can also be hit-or-miss at times. Despite the hype around Instinct, one of my first tests with it was to price out hotels then book one, and it quoted me all the wrong rates. Ollie, meanwhile, experienced an infrastructure outage with its text provider when I was testing it, and it stopped responding.

We asked the founder how Ollie is handling such problems, given that most consumers only give a new app or technology one shot. If it’s broken, they simply move on.

Lennon agreed this is an issue with AI overall. “This is the challenge with LLMs, in general — because they’re stochastic [i.e., involve probability], they’re inherently unreliable,” he said. “We have to essentially build the harness — the agent harness — in a defensive way to catch and prevent those things… It’s almost like there’s just 1,000 cuts that you’ve got to solve first.”

In Brief

Google is launching conversational AI features within Docs, Gmail, and Keep that let users ask questions about their inboxes and documents as well as do tasks using natural-language queries and dictation.

The company previewed these features in May during its Google I/O conference. The company is dubbing these features Docs Live, Gmail Live, and Keep Live.

In Gmail, instead of typing in a search box, you can ask questions to a Gemini-powered AI assistant about the contents of your inbox. Users will see a live transcript while chatting with the assistant.

While using Docs, users can describe what they want to write, and the AI assistant will help them create a first draft. It can also get information from Gmail, Drive, chat, and the web to create drafts.

Keep Live, meanwhile, lets you use it as a scratchpad to jot down notes without typing, and can even pull together recipes and lists if you tell it to, for example, list ingredients to make a shakshouka.

Gmail Live is available for Google AI Plus, Pro, and Ultra customers. Docs Live and Keep Live are available for Google AI Pro and Ultra users. All these features will be available in English on iOS and Android. Google said it will roll out these features to Workplace Business customers soon.

Google has been introducing voice-based features across its apps. It added cross-app dictation to its Mac app in April. And the Pixel 11 smartphones got a Gemini-powered dictation tool called Rambler last month, which removes filler words from your speech. In August, the company introduced the Gemini 3.5 Transcribe model for speech-to-text use cases.

Scientists at Google DeepMind and Google Research released a new artificial intelligence model for weather forecasting today that sees our changing atmosphere more clearly and predicts its behavior more often.

WeatherNext 3 is the latest wave of a sea change in meteorology brought out by deep learning techniques, and Google says it will start feeding into weather information users see in search, Google Maps, and Gemini, as well as being available to users and researchers on Google’s cloud platforms.

“This is going to be the first time that some of the core variables feed and power a lot of the Google products,” Samier Merchant, a Google senior staff engineer, told TechCrunch.

The new model has already proven to be the most accurate among leading contenders tested on Operational WeatherBench, a utility for comparing AI forecasts built by the startup Brightband. It looks at metrics like temperature, windspeed, and humidity.

As well as beating out other deep-learning models built by Google, Microsoft, Nvidia, and the European Center for Medium-Range Weather Forecasting (ECMWF), it also beats traditional forecasts from the U.S. National Weather service and the ECMWF.

Image Credits:Brightband

Most weather forecasts come from government-owned supercomputers laboriously churning through mathematical equations written to describe the physics of weather; while these systems have become remarkably accurate, they are expensive and comparatively slow. After the ECMWF released more than half a century of weather data produced by these systems in 2018, deep learning researchers began training models that could make predictions far more quickly and with comparable accuracy to government tools.

“Weather is chaotic, and so small differences really start to perturb massively…Machine learning targets the problem we are really solving, which is approximate noisy physics from incomplete information and finite compute, and so it learns patterns from a lot of data,” said Ferran Alet, a staff research scientist manager at DeepMind.

Since then, model-makers have pushed on the key weaknesses of AI forecasting models: They tend to forecast over a wider area — 15 to 25 square km — than is truly useful, they’re not always great with rain, and they still depend on the formatted datasets produced by government agencies.

WeatherNext 3 takes on all three challenges. On key variables, researchers told TechCrunch, it can predict down to a resolution of 5 km. Its evaluations on rain are 60% improved over WeatherNext 2, and it can now produce hourly forecasts, instead of the standard prediction every six hours.

Image Credits:Google

Those improvements are the result of specific choices made by the designers. WeatherNext 3 is a larger model, with 2.4 times more parameters than its predecessor, tailoring the targets for the decoder heads to give more useful answers. While most weather forecasts output as metrics averaged across a 3D grid, DeepMind researchers have already won plaudits by tuning their model to also visualize cyclone paths.

This time around, the designers also trained the model to target its forecasts to specific weather data stations. This is important not only for offering more granular predictions, but also for being able to evaluate its work against specific, ground-truth data.

“The idea, with a lot of AI applications, is to try to run tasks as end-to-end as possible,” Daniel Rothenberg, an atmospheric scientist at Brightband, said. “Adding a capability where this model is now also predicting, say, what Denver’s airport’s weather station is going to measure on an hourly basis, just connects that forecasting task closer to the core.”

The model is able to forecast more frequently because it can ingest weather satellite data collected in real time on an hourly basis. Feeding AI models on raw empirical observations, rather than the analysis produced by weather supercomputers, promises a more accurate forecast, but it is still technically challenging to get models to work with unformatted data.

Google says WeatherNext 3 is the “first” AI model to directly incorporate raw observations for a high-resolution global forecast, but the AI weather startup WindBorne says its model, WeatherMesh 6, has been incorporating raw observations from its fleet of weather balloons and other sources since late 2025. Asked about that, Google pointed out that its forecasts are higher resolution across the globe. Regardless, both models still rely on national weather datasets to perform forecasts, so more work will be required for true direct data assimilation.

While LLMs get the bulk of the attention, the transformer revolution in meteorology has been just as important. European and U.S. weather agencies are already using AI models in their forecast products, and their speed and low cost promise to bring economic impact to poorer regions where the expense of high-quality sensors and supercomputers has put accurate forecasts out of reach.

Bill Gates recently cited AI-powered weather forecasting as a crucial benefit of the technology, with better forecasts improving crop yields in developing countries. Alet, the DeepMind researcher, said that higher-resolution forecasts of wind, rain, and cloud cover will be useful to make renewable energy projects more dependable.

“At the end of the day, I think Google is about providing useful information to the user, and a lot of what users are looking for has to do with the weather in some way or another,” Alet said.