Rivian is, by every measure, a maker and seller of EVs. But in 2025, it was the company’s software and services that helped its annual revenue grow by 8%.

Rivian reported Thursday $5.38 billion in total revenue in 2025, up from $4.97 billion from the prior year. That rosy picture dulls a bit when looking just at its automotive revenue, which fell 15% to $3.8 billion in 2025. The fall was fueled by a $134-million drop in regulatory credit sales and lower vehicle deliveries, which were partially offset by higher average selling prices, according to Rivian.

Meanwhile, software and services revenue grew more than threefold to $1.55 billion for the year. And the joint venture with Volkswagen Group was behind most of that growth, according to Rivian. The “services” component of this line item, which Rivian doesn’t break out, includes a variety of items including vehicle repair, vehicle trade-ins, and maintenance services. The remainder, and the bulk of the revenue, is from software, and specifically due to the joint venture with VW Group.

VW and Rivian formed a technology joint venture in 2024 that is worth up to $5.8 billion. The joint venture is milestone-based and in 2025 Rivian hit the mark, which meant a $1 billion payout in the form of a share sale. Under the terms of the JV, Rivian will supply VW Group with its existing electrical architecture and software technology stack.

Rivian received an initial $1 billion convertible note in 2024 and another $1 billion payment in July 2025.

Rivian is expected to continue to receive payments from VW Group through 2027. Rivian is expected to receive an additional $2 billion of capital as part of the joint venture in 2026, CFO Claire McDonough said Thursday on the company earnings call. About $1 billion of that is subject to the successful completion of winter testing, which is underway. The remaining $1 billion is non recourse debt, which is expected to be received in October.

And while the funds provide a hefty stopgap, Rivian’s financial success in 2026 will hinge largely on the rollout of its next EV, the R2.

Techcrunch event TechCrunch Founder Summit 2026: Tickets Live On June 23 in Boston, more than 1,100 founders come together at TechCrunch Founder Summit 2026 for a full day focused on growth, execution, and real-world scaling. Learn from founders and investors who have shaped the industry. Connect with peers navigating similar growth stages. Walk away with tactics you can apply immediately



Save up to $300 on your pass or save up to 30% with group tickets for teams of four or more. TechCrunch Founder Summit: Tickets Live On June 23 in Boston, more than 1,100 founders come together at TechCrunch Founder Summit 2026 for a full day focused on growth, execution, and real-world scaling. Learn from founders and investors who have shaped the industry. Connect with peers navigating similar growth stages. Walk away with tactics you can apply immediately



Save up to $300 on your pass or save up to 30% with group tickets for teams of four or more. Boston, MA | REGISTER NOW

Rivian confirmed in its earnings report Thursday that the R2 SUV, which is designed to be cheaper to build and less expensive for customers, will come to market by June 2026. That “cheaper to build” line item is particularly crucial for Rivian, which has historically lost money on every vehicle it makes.

Rivian has spent years trying to push cost of goods sold figure down. And it has made progress with the rollout of its second-generation flagship R1T truck and R1S SUV. For instance, McDonough said “in the fourth quarter it was able to deliver $92,000 of cogs per unit and that was about a $4,000 per unit improvement relative to the third quarter.” Rivian’s cogs per unit was $99,000 in the fourth quarter of 2024.

The company saw its total automotive cost of revenue decrease year-over-year from $1.4 billion in fourth quarter of 2024 to $898 million in the same quarter in 2025. Notably, the company’s cost of revenue for software steadily inched up throughout 2025.

The R2 SUV, which will initially launch as a dual motor all-wheel drive model, is an opportunity to further reduce costs. The company is expected to release more information about the R2, including final specs, on March 12.

Rivian’s guidance for 2026 suggests that it is banking on demand for R2 and its ability to ramp up production. The company said Thursday it expects to deliver between 62,000 and 67,000 vehicles in 2026 — which could provide up to a 59% bump from last year. Rivian delivered 42,247 vehicles in 2025, which includes its two R1 consumer vehicles and the electric delivery van (EDV).

Rivian CEO RJ Scaringe noted that the company expects some growth in EDV sales in 2026. Rivian plans to an all-wheel drive version and a larger battery pack variant of the EDV, for which Amazon is its primary customer.

“Both of those are to help unlock specific use cases within the Amazon network,” Scaringe said. “We’re working really closely with Amazon in defining the requirements of those and excited to get those launched.”

The company isn’t signaling profitability — on an adjusted basis — just yet. But it is offering up considerable improvement on that front. Rivian reported a $3.6 billion net loss in 2025; it expects an adjusted net loss of between $1.8 billion and $2.1 billion for 2026. Rivian also projects capital expenditures will be between $1.95 billion and $2.05 billion this year.

Tim Spencer realized just how complicated manufacturing procurement can be while running Markai, an e-commerce startup in Asia, during the pandemic.

“We had thousands of suppliers, and we were distributing products into dozens of countries around the world,” Spencer (pictured left) told TechCrunch. His staff was overwhelmed by the manual complexity of sourcing suppliers, negotiating pricing, tracking orders, and managing payments.

“I found myself running this big team that was not really set up for success,” he said. He sold Markai in 2023, just as it was becoming clear that generative AI could streamline the most time-consuming procurement hurdles for manufacturers and distributors.

Later that year, Spencer launched Didero with Lorenz Pallhuber (pictured center), a veteran of McKinsey’s procurement practice, and Tom Petit, the former technical co-founder of Landis.

Didero, whose mission is to automate many of the complexities of global procurement, just raised a $30 million Series A co-led by Chemistry and Headline, with participation from Microsoft’s venture fund M12.

“Global trade runs on natural language communication,” Spencer said. “It’s emails, WeChat, phone calls, purchase orders, and packing lists.”

Until the advent of generative AI, these fragmented pieces had to be tracked by humans who spent their days chasing suppliers and manually updating systems of record. Didero claims its platform can ingest that communication, putting a significant portion of the procurement workflow on autopilot.

Techcrunch event TechCrunch Founder Summit 2026: Tickets Live On June 23 in Boston, more than 1,100 founders come together at TechCrunch Founder Summit 2026 for a full day focused on growth, execution, and real-world scaling. Learn from founders and investors who have shaped the industry. Connect with peers navigating similar growth stages. Walk away with tactics you can apply immediately



Save up to $300 on your pass or save up to 30% with group tickets for teams of four or more. TechCrunch Founder Summit: Tickets Live On June 23 in Boston, more than 1,100 founders come together at TechCrunch Founder Summit 2026 for a full day focused on growth, execution, and real-world scaling. Learn from founders and investors who have shaped the industry. Connect with peers navigating similar growth stages. Walk away with tactics you can apply immediately



Save up to $300 on your pass or save up to 30% with group tickets for teams of four or more. Boston, MA | REGISTER NOW

Didero functions as an agentic AI layer that sits on top of a company’s existing ERP, acting as a coordinator that reads incoming communications and automatically executes the necessary updates and tasks.

“The goal is to go from ‘I need a good’ to payment without having to lift a finger,” Spencer said.

Unlike Levelpath, Zip, or Oro Labs, which use AI to streamline corporate purchasing, Didero focuses on the supply chain. Its platform is designed for manufacturers and distributors who need to source raw materials and inputs required to build or sell their products.

Didero has a few smaller competitors that can handle some of the tasks that the company does. For instance, Cavela and Pietra help brands source and negotiate pricing with manufacturers, but according to Spencer, these companies serve small and medium-sized companies and don’t handle the full procurement process, from the first quote to the final payment.

Didero has dozens of customers but named only one, Footprint, a provider of sustainable, plant-based packaging.

In Brief

Anthropic has just concluded a $30 billion Series G fundraising round, the company announced on Thursday. The company’s new value is $380 billion — a huge jump from its previous Series F valuation of $183 billion.

Some details of the round were reported earlier this week by Bloomberg.

The new round was led by Singaporean wealth fund GIC and investment management firm Coatue, with a number of other prominent firms — including D. E. Shaw Ventures, Peter Thiel’s Founders Fund, and Abu Dhabi’s MGX — co-leading the round. Other significant investors listed include Accel, General Catalyst, Jane Street, and the Qatar Investment Authority, among many others.

The infusion of funding for the AI startup takes place as it is vying for customers and cultural attention with its competitor, OpenAI. OpenAI also recently made it known that it is seeking to clinch an additional $100 billion in funding, which, were it to secure the funds, would balloon the company’s valuation to some $830 billion.

“Whether it is entrepreneurs, startups, or the world’s largest enterprises, the message from our customers is the same: Claude is increasingly becoming more critical to how businesses work,” said Krishna Rao, Anthropic’s chief financial officer, in a press release. “This fundraising reflects the incredible demand we are seeing from these customers, and we will use this investment to continue building the enterprise-grade products and models they have come to depend on.”

After two years of initial hesitation, YouTube has finally unveiled its dedicated app for the Apple Vision Pro.

When Apple’s headset first launched, YouTube opted for a web-based approach instead of developing a dedicated app, directing viewers to use Safari to access content. Until now, anyone wanting to watch YouTube on the headset has missed out on helpful features like offline downloads — making it much less useful for travel or offline viewing.

To fill the gap, third-party options, such as Juno, briefly appeared on visionOS before being removed for violating YouTube’s Terms of Service.

Today’s announcement changes the game, allowing users to enjoy their favorite content — including all standard videos and YouTube Shorts —on a theater-sized virtual screen in a fully immersive environment.

Standout features include a Spatial tab that lets users discover spatial videos, including 3D, VR180, and 360-degree formats. For consumers with the latest Apple Vision Pro models powered by the M5 chip, the app supports 8K playback, a notable upgrade.

Additionally, users can interact with the new visionOS app using gesture controls, such as resizing windows or scrubbing through videos.

The timing of this app’s release is interesting, as YouTube initially hesitated to invest in a dedicated visionOS experience, possibly waiting to gauge the headset’s traction. Most major streaming services have offered native experiences on the platform since launch, including Disney+, Amazon Prime Video, Paramount, and Peacock, among others.

Techcrunch event TechCrunch Founder Summit 2026: Tickets Live On June 23 in Boston, more than 1,100 founders come together at TechCrunch Founder Summit 2026 for a full day focused on growth, execution, and real-world scaling. Learn from founders and investors who have shaped the industry. Connect with peers navigating similar growth stages. Walk away with tactics you can apply immediately



Save up to $300 on your pass or save up to 30% with group tickets for teams of four or more. TechCrunch Founder Summit: Tickets Live On June 23 in Boston, more than 1,100 founders come together at TechCrunch Founder Summit 2026 for a full day focused on growth, execution, and real-world scaling. Learn from founders and investors who have shaped the industry. Connect with peers navigating similar growth stages. Walk away with tactics you can apply immediately



Save up to $300 on your pass or save up to 30% with group tickets for teams of four or more. Boston, MA | REGISTER NOW

Another reason why YouTube’s decision to launch the app now is interesting is that the initial hype surrounding the headset appears to have diminished.

Despite the Apple Vision Pro introducing Apple Intelligence-powered features almost a year ago, user engagement with the headset has reportedly been lukewarm. Recent estimates indicate that sales have significantly declined, with only about 45,000 new units shipped during the fourth quarter of 2025. The Financial Times reported that production was halted due to weak demand, resulting in a sharp reduction in marketing efforts across key markets.

The app is now available in the visionOS App Store and is compatible with both the M2 and M5 chip models.

Vincenzo Iozzo, a renowned hacker linked to convicted sex offender Jeffrey Epstein, is no longer listed on the website of Black Hat, one of the largest cybersecurity conferences in the world, nor on the Japanese security conference Code Blue.

As of Thursday, Iozzo does not appear on the official review board pages of Black Hat or Code Blue. He was still listed on both pages as of last week. Iozzo had been on the Black Hat review board since 2011, according to his LinkedIn profile.

In a statement shared with TechCrunch through a spokesperson, Iozzo said he told Black Hat that he “will not willingly resign” and welcomed “a full investigation.”

Spokespeople for Black Hat did not respond to requests for comment.

Iozzo, currently the founder and chief executive of cybersecurity startup SlashID, has had a long career in the industry. Iozzo authored one of the first manuals for hackers researching Apple’s mobile software, and in 2015, founded cybersecurity startup IperLane, which was later bought by CrowdStrike, leading him to serve as a senior director at the company for almost four years.

Iozzo’s name appears in more than 2,300 documents, some of which contain several emails, released on January 30 as part of the Department of Justice’s legally required effort to publish materials from its investigation into the late financier and sex trafficker.

Iozzo’s interactions with Epstein span from October 2014 until December 2018. In late 2018, the Miami Herald published news stories detailing allegations that Epstein abused more than 60 women, some of them teenage girls.

After these stories were published, newly released emails show Iozzo was trying to meet with Epstein at his New York town house.

Contact Us Do you have more information about Epstein’s connections in the cybersecurity world?? From a non-work device, you can contact Lorenzo Franceschi-Bicchierai securely on Signal at +1 917 257 1382, or via Telegram, Keybase and Wire @lorenzofb, or Do you have more information about Epstein’s connections in the cybersecurity world?? From a non-work device, you can contact Lorenzo Franceschi-Bicchierai securely on Signal at +1 917 257 1382, or via Telegram, Keybase and Wire @lorenzofb, or by email .

Among the new material published by the Justice Department, beyond the more than 2,300 documents mentioning Iozzo, there was also a report from an FBI informant who claimed Epstein had a “personal hacker.” The document is redacted and does not name the alleged hacker. However, some of the identifying details included in the document strongly suggest that the informant believed Iozzo was Epstein’s hacker. Italian newspaper Il Corriere della Sera reported on the emails earlier this month and named Iozzo as the person likely redacted in the informant document.

It’s important to note that the claims and allegations by the FBI informant were not confirmed by the FBI and may be partially wrong. Furthermore, there is no evidence in the emails that suggests Iozzo did anything unlawful for Epstein.

Iozzo said in his statement to TechCrunch that he “knew Epstein for professional reasons” and that he wished he had not, but he denied claims that he was Epstein’s hacker or did any hacking for him.

“We were introduced in 2014 when I was a 25-year-old at MIT fundraising for my startup, by people whom I trusted and admired. Because of this, I failed to ask the right questions that, in retrospect, seem obvious,” read the statement, sent by his spokesperson Joan Vollero. “I foolishly accepted the narrative that was presented to me by others that greatly minimized the magnitude of his horrific actions. I regret the past association and take full responsibility for not exercising greater judgment at the time.”

“My interactions with Epstein were limited to business opportunities that never materialized, as well as discussions of the markets and emerging technologies. I never observed nor participated in any illegal activity or behavior,” Iozzo added.

In 2008, Epstein pleaded guilty to soliciting sex from girls as young as 14 and registered as a sex offender in Florida and New York. In 2018, new reports emerged that Epstein was allegedly a serial sex abuser and had trafficked underage girls at his private island. After these new reports, the Justice Department formally charged Epstein in 2019 of trafficking, exploiting, and abusing dozens of underage girls. Epstein later died in jail.

Neither Iozzo’s spokesperson Vollero, nor his attorney Emma Spiro, explained why Iozzo was removed from Black Hat’s website, but did not dispute the removal.

“Mr. Iozzo welcomed an independent investigation from Black Hat, rather than a knee-jerk removal decision, because he is confident that he would be cleared through that process,” said Vollero.

Code Blue spokesperson Ken-ichi Saito confirmed to TechCrunch that the conference removed Iozzo’s name from its review board. Saito said the conference had been “preparing for this update for several months” to remove Iozzo and two other review board members “who had not been active” and that the “the timing of our website update coincidentally overlapped with the public release of the Epstein documents.”

In Brief

After months of telegraphing the move, the Trump administration today officially repealed the EPA’s 2009 “endangerment finding,” which found that greenhouse gases such as carbon dioxide and methane were a threat to human health and welfare.

The finding had so far supported the agency’s regulation of greenhouse gases under the Clean Air Act. Currently, the new rule only applies to car and truck tailpipe emissions, but it’s expected to be the first of several similar changes to federal air pollution regulations.

Before the endangerment finding can be fully repealed, though, the EPA must go through a lengthy process. The original finding took two years to put in place.

The move by Lee Zeldin, the EPA administrator, will slow the decline in emissions by about 10%, according to Axios. That’s a significant amount, but not enough to reverse the trend, partly because cheap renewables have dominated new electricity-generation capacity in recent years.

“This action will only lead to more pollution, and that will lead to higher costs and real harms for American families,” Fred Krupp, president of the Environmental Defense Fund, told TechCrunch in a statement.

Unabated climate change is expected to raise mortality rates by around 2% in the U.S., and it could reduce global GDP by 17% by 2050, or about $38 trillion.

Web hosting company WP Engine has filed an amended complaint with brow-raising new allegations in its ongoing legal battle with WordPress co-founder Matt Mullenweg and his company Automattic (WordPress.com’s parent company). The company now claims that Mullenweg intended to target 10 different hosting companies with royalty payments for their use of the WordPress trademark and tried to get payment processor Stripe to cancel its contract with WP Engine.

At the heart of the dispute, Mullenweg believes WP Engine is profiting from the open source WordPress project without contributing back to the community, and demanded the hosting company pay 8% of its monthly gross revenues as a royalty fee for using the WordPress brand.

The suit, originally filed by WP Engine in October 2024, accused Automattic and Mullenweg of defamation and abuse of power. Automattic filed its counterclaims last year, alleging that the hosting company has been abusing the WordPress trademark and engaging in deceptive marketing practices.

In this latest filing, WP Engine has amended its complaint for the third time after gaining access to information that was uncovered during the discovery process. The information, now unredacted, had been previously sealed at Automattic’s request.

Notably, one of the new claims accuses Automattic of allegedly planning to target 10 other competitors with royalty claims similar to those aimed at WP Engine.

The complaint also states that Newfold, a company whose portfolio includes hosting providers like Bluehost and HostGator, among others, is already paying Automattic for use of its trademarks and that Automattic is in conversations with others. (Names of the other hosts were redacted in the complaint, which referenced email conversations between those companies and Mullenweg.)

WP Engine is also alleging that Mullenweg reached out to a Stripe executive via email to pressure the company to cancel WP Engine’s contract. This occurred after WP Engine had filed its lawsuit against Automattic, the complaint states.

The filing also challenges the 8% rate for the royalty payments as being somewhat arbitrary. Referencing Mullenweg’s comments at TechCrunch Disrupt 2024, WP Engine claims that the Automattic founder essentially came up with the rate because it was what he thought WP Engine “could afford to pay.”

At Disrupt, Mullenweg had responded to a question about how he settled on the 8% fee by saying it was based on a “business analysis.”

“If you estimate that would be about $32 million — they still would have been free cash-flow positive, and based on our estimates, and the negotiations over the past 18 months, we felt like that was a fair amount,” he said at the event.

Other new claims in the filing point to aggressive language Mullenweg allegedly used against WP Engine, like threats that if the web hosting provider didn’t comply, he would start stealing its customers. “If they don’t take the carrot, we’ll give them the stick,” was one of the quotes cited from internal correspondence at Automattic, for instance.

The complaint also includes allegations that Mullenweg used the term “nuclear war” to describe his approach to WP Engine’s defiance.

Asked about the new filing, Automattic shared the following statement with TechCrunch: “There is nothing new here. This is the same narrative WP Engine has been pushing for over a year, and the Court has already dismissed many of its central claims. The lawsuit is going nowhere. This latest filing simply repackages the same tired allegations in an effort to keep the story alive. WP Engine is trying to recast robust competition as something nefarious. We are confident the courts will continue to reject that theory.”

Updated after publication with Automattic’s statement.

“Join xAI if the idea of mass drivers on the Moon appeals to you,” CEO Elon Musk proclaimed yesterday following a restructuring that saw a stream of former executives exit the AI lab.

This is an interesting recruitment strategy after the company’s merger with Musk’s rocket maker, SpaceX, and the combined company’s anticipated IPO. You might think that xAI employees ought to be fascinated with achieving AGI, using deep learning models to disrupt traditional software companies, or simply bad wordplay like “Macrohard.” But instead, Elon is going to the moon.

After outlining plans to build AI data centers in orbit, the primary synergy between the two companies, Musk took the idea further. “What if you want to go beyond a mere terawatt per year?” Musk asked. “To do that, you have to go to the moon…I really want to see a mass driver on the moon that is shooting AI satellites into deep space.”

New year, new dream Image Credits:SpaceX

In Musk’s telling, the step beyond data centers orbiting Earth is even larger computers in deep space. And furthermore, Musk says the best way to achieve that is to build a city on the moon to manufacture space computers and hurl them into the solar system using a big maglev train.

If that all feels a bit much, veteran Musk watchers know there’s a clue about where the discussion appears in a video of an all-hands meeting xAI shared with the public. The slide describing the moon base comes at the end of the presentation deck, where, during SpaceX pep talks, Musk typically shares renderings of SpaceX rockets landing on Mars and waxes rhapsodic about the future of multi-planetary humanity.

Notably, the moon base comes just after SpaceX has publicly backed away from its long-held goal of colonizing Mars. Now, with xAI in the corporate fold, Musk needs a new science fiction metaphor for the future: In this case, the Kardashev Scale, a theoretical measure of galactic civilizations coined by the eponymous Soviet astronomer in the 1960s. The idea is climbing the scale of energy usage — early civilizations figure out how to leverage all the power sources on their planets, and then (hypothetically) go to space and build infrastructure to capture the energy of the sun.

With the moon base, Musk says the company could harness “maybe even a few percent of the sun’s energy” to train and operate AI models. “It’s difficult to imagine what an intelligence of that scale would think about,” he told his staff, “but it’s going to be incredibly exciting to see it happen.”

Techcrunch event TechCrunch Founder Summit 2026: Tickets Live On June 23 in Boston, more than 1,100 founders come together at TechCrunch Founder Summit 2026 for a full day focused on growth, execution, and real-world scaling. Learn from founders and investors who have shaped the industry. Connect with peers navigating similar growth stages. Walk away with tactics you can apply immediately



Save up to $300 on your pass or save up to 30% with group tickets for teams of four or more. TechCrunch Founder Summit: Tickets Live On June 23 in Boston, more than 1,100 founders come together at TechCrunch Founder Summit 2026 for a full day focused on growth, execution, and real-world scaling. Learn from founders and investors who have shaped the industry. Connect with peers navigating similar growth stages. Walk away with tactics you can apply immediately



Save up to $300 on your pass or save up to 30% with group tickets for teams of four or more. Boston, MA | REGISTER NOW

In the nine years since Musk unveiled his plan for Martian exploration and colonization, the vision has been an effective hiring tool for SpaceX: The founding tale of Musk’s interest in the Red Planet offered a long-term vision that united the company’s various development efforts, and signaled the company’s ambition among other space contractors that settled for incremental work on government priorities. “Occupy Mars” t-shirts offered a visible symbol of SpaceX’s aspirations.

That’s where the hypothetical moon base fits in — part of a long history of Musk wrapping his companies in a powerful narrative. It’s one million people living on Mars, but now catering to a future where AI is the most interesting thing. Martian mission creep became apparent less in Musk’s May 2025 Starship update, when the presentation ended with a now-cancelled vision of Tesla Optimus robots clomping across the Red Planet.

Poor robot Image Credits:SpaceX

There was just one problem with SpaceX and Mars: No one wanted to pay them to go there. Plans announced in 2016 to repurpose the company’s Dragon spacecraft as a Mars lander were abandoned the next year after the technical challenges became too costly. And since Musk unveiled the vehicle that would become Starship in 2016, its capabilities, initially intended for Mars colonization, have been scaled back to focus on two more remunerative tasks: launching satellites for the Starlink comms network and $4 billion worth of contracts to land astronauts on the moon for NASA.

Unlike a multi-planetary civilization, there may be some logic in having SpaceX purchase a money-burning AI and social media to build data centers in Earth orbit, particularly if forecasts of rising demand and costs on the ground come true. Experts suggest that it might be possible in the 2030s.

Hypothetically, building satellites on the moon would require a lot more of Musk’s other dreams coming true first. Scientists and startups are experimenting with building chips and other precision components in space. But mass-producing many tons of advanced computers on the moon means we’re living in a universe where it is dramatically cheaper to get to space, which is the central requirement for those technologies, getting all the raw materials for such an effort to the moon, plus whatever is required for a “self-sustaining city.”

In a sense, that’s the point: This is the, uh, stretch goal. If meme-happy retail investors buy the argument, they could turn SpaceX shares into the next Tesla. The engineers, AI or aerospace, who Musk needs to achieve his goals may find the shift jarring. But the vision is one way to explain what xAI is about, other than an LLM perhaps best known as a pervert. As one of the company’s departing executives said on his way out the door, “all AI labs are building the exact same thing, and it’s boring.”

Mass-producing a solar system-scale supercomputer on the moon is many things (I’m going to get emails for not using the word “insane”), but it is not the exact same thing, and it is not boring.

On Thursday, OpenAI announced the release of a light-weight version of its agentic coding tool Codex, the latest model of which OpenAI launched earlier this month. GPT-5.3-Codex-Spark is described by the company as a “smaller version” of that model, one that is designed for faster inference. To power that inference, OpenAI has brought in a dedicated chip from its hardware partner Cerebras, marking a new level of integration in the company’s physical infrastructure.

The partnership between Cerebras and OpenAI was announced last month, when OpenAI said that it had reached a multi-year agreement with the firm worth over $10 billion. “Integrating Cerebras into our mix of compute solutions is all about making our AI respond much faster,” the company said at the time. Now, OpenAI calls Spark the “first milestone” in that relationship.

Spark, which OpenAI says is designed for swift, real-time collaboration and “rapid iteration,” will be powered by Cerebras’ Wafer Scale Engine 3. The WSE-3 is Cerebras’ third-generation waferscale megachip, decked out with 4 trillion transistors. OpenAI describes the new lightweight tool as a “daily productivity driver, helping users with rapid prototyping” rather than the longer, heavier tasks that the original 5.3 is designed for. Spark is currently enjoying a research preview for ChatGPT Pro users in the Codex app.

In a tweet in advance of the announcement, CEO Sam Altman seemed to hint at the new model. “We have a special thing launching to Codex users on the Pro plan later today,” Altman tweeted. “It sparks joy for me.”

In its official statement, OpenAI emphasized Spark as designed for the lowest possible latency on Codex. “Codex-Spark is the first step toward a Codex that works in two complementary modes: real-time collaboration when you want rapid iteration, and long-running tasks when you need deeper reasoning and execution,” OpenAI shared. The company added that Cerebras’ chips excel at assisting “workflows that demand extremely low latency.”

Cerebras has been around for over a decade but, in the AI era, it has enjoyed an increasingly prominent role in the tech industry. Just last week, the company announced that it had raised $1 billion in fresh capital at a valuation of $23 billion. The company has previously announced its intentions to pursue an IPO.

“What excites us most about GPT-5.3-Codex-Spark is partnering with OpenAI and the developer community to discover what fast inference makes possible — new interaction patterns, new use cases, and a fundamentally different model experience,” Sean Lie, CTO and co-founder of Cerebras, said in a statement. “This preview is just the beginning.”

Aurora’s self-driving trucks can now travel nonstop on a 1,000-mile route between Fort Worth and Phoenix — exceeding what a human driver can legally accomplish.

The distance, and the time it takes to travel it, offers up positive financial implications for Aurora — and any other company hoping to commercialize self-driving semitrucks.

It takes Aurora about 15 hours to carry freight in its driverless trucks on the 1,000-mile journey, according to the company. Human truck drivers take much longer to complete the same distance due to federal regulations that limit how long they can be behind the wheel. For instance, truck drivers must stop for a 30-minute break after eight hours and can operate a semitruck for a maximum of 11 hours at a time, according to federal regulations. Once drivers hit that threshold, they cannot get behind the wheel for another 10 hours.

“This represents more than a technological achievement,” Aurora co-founder and CEO Chris Urmson said during the company’s earnings call Wednesday afternoon. “It is the dawn of a superhuman future for freight.”

It also offers compelling economics to its customers, which includes Uber Freight, Werner, FedEx, and Schneider. The company said that eventually it can cut transit times nearly in half, a stat that has won over companies like Hirschbach, an early customer on the Fort Worth-to-Phoenix route.

Aurora said in a letter to shareholders that it is poised to expand across the Sun Belt of the United States. Today, the company operates driverless trucks — some with a human observer still in the cab — on routes between Dallas and Houston, Fort Worth and El Paso, El Paso and Phoenix, Fort Worth and Phoenix, and Laredo and Dallas.

The expansion has helped Aurora transition from a developer of autonomous trucks to a commercial operator who is earning money on its driverless routes.

Techcrunch event TechCrunch Founder Summit 2026: Tickets Live On June 23 in Boston, more than 1,100 founders come together at TechCrunch Founder Summit 2026 for a full day focused on growth, execution, and real-world scaling. Learn from founders and investors who have shaped the industry. Connect with peers navigating similar growth stages. Walk away with tactics you can apply immediately



Save up to $300 on your pass or save up to 30% with group tickets for teams of four or more. TechCrunch Founder Summit: Tickets Live On June 23 in Boston, more than 1,100 founders come together at TechCrunch Founder Summit 2026 for a full day focused on growth, execution, and real-world scaling. Learn from founders and investors who have shaped the industry. Connect with peers navigating similar growth stages. Walk away with tactics you can apply immediately



Save up to $300 on your pass or save up to 30% with group tickets for teams of four or more. Boston, MA | REGISTER NOW

Aurora has brought in revenue since April 2025 when it first deployed driverless, heavy-duty trucks for commercial use on public roads. Aurora reported $1 million in the fourth quarter and $3 million for the year, according to a report filed with the U.S. Securities and Exchange Commission. The company’s CFO, David Maday, said total year adjusted revenue, which includes money earned through pilot programs early last year, was $4 million.

That’s a tiny figure, especially when compared to its expenses. Aurora reported a net income loss of $816 million in 2025, up 9% from the year before as it focuses on scaling its operations. Still, it shows measurable progress from 2024, when it didn’t recognize any revenue.

Revenue is projected to continue as the company adds more trucks and driverless routes to its network. Today, the company has 30 trucks in the fleet, 10 of which are operating driverlessly. That fleet is expected to grow to more than 200 trucks by the end of the year. Urmson said the company’s trucks have racked up 250,000 driverless miles as of January 2026 with a perfect safety record.

In the second quarter, Aurora plans to deploy a fleet of driverless International Motors LT trucks, which will not have a human observer on board. Aurora’s driverless operations that use Paccar trucks currently have a human safety observer in the cab as requested by the truck manufacturer.

Urmson took a bullish view of Aurora’s future, buoyed by advancements in its self-driving software, an impending second-generation hardware kit that will lower costs, and expansion of its driverless routes. The expansion of its driverless trucking routes has been propelled by a new software release, its fourth since launching commercial service in April 2025.

The first release validated initial driverless operations between Dallas and Houston, the second validated operations at night, and the third validated El Paso, according to Aurora. The company said this latest software release will give its self-driving system the capabilities to navigate the diverse geography and climate of the southern United States.

“Just as the last two years brought robotaxis into the mainstream, we expect 2026 to mark the inflection point where the market recognizes that self-driving trucks have arrived and are quickly becoming a permanent fixture in our transportation landscape,” Urmson said on the company’s earnings call. “If you’re in the Sun Belt in 2026, you won’t just read about the Aurora driver. You’ll see it every day.”

Aurora currently operates driverless routes through Texas, New Mexico, and Arizona, and has future driverless operations planned for Nevada, Oklahoma, Arkansas, Louisiana, Kentucky, Mississippi, Alabama, North Carolina, South Carolina, Georgia, and Florida, the company said.

Coupang’s massive data breach in South Korea has now become a geopolitical flashpoint as a growing number of the company’s U.S. investors take legal action against the South Korean government.

What began as a regulatory investigation into data security failures has expanded into a broader dispute over alleged unfair treatment of the U.S.-headquartered company.

While Coupang — which operates in South Korea, Taiwan, and Japan — is often referred to as the “Amazon of South Korea,” its worldwide headquarters are actually in Seattle, Washington.

The company’s investors are now seeking international arbitration under the U.S.-Korea Free Trade Agreement (FTA). On January 23, 2026, U.S. investment firms Greenoaks and Altimeter filed a notice with South Korea’s Ministry of Justice, saying they suffered losses from what they characterized as the government’s discriminatory investigation into the data breach. They said they plan to pursue investor–state dispute settlement (ISDS) arbitration under the U.S.-Korea FTA.

South Korea’s Ministry of Justice said Thursday that three more investors, including Abrams Capital, Durable Capital Partners, and Foxhaven Asset Management have now joined the case. They are alleging the government acted unlawfully toward the e-commerce company.

To recap the incident: In December, Coupang disclosed that nearly 34 million Korean customers’ personal information had been leaked in a data breach that had been going on for more than five months. The breach involved customer names, email addresses, phone numbers, shipping addresses, and certain order histories, the company said.

While other tech breaches in Korea resulted in less severe penalties, Coupang has faced extraordinary government pressure. The government reportedly threatened massive fines, suspension of operations, and travel bans for executives while Coupang’s investors allege it also sought to block public communication and repeatedly misrepresented the scope of the breach.

Techcrunch event TechCrunch Founder Summit 2026: Tickets Live On June 23 in Boston, more than 1,100 founders come together at TechCrunch Founder Summit 2026 for a full day focused on growth, execution, and real-world scaling. Learn from founders and investors who have shaped the industry. Connect with peers navigating similar growth stages. Walk away with tactics you can apply immediately



Save up to $300 on your pass or save up to 30% with group tickets for teams of four or more. TechCrunch Founder Summit: Tickets Live On June 23 in Boston, more than 1,100 founders come together at TechCrunch Founder Summit 2026 for a full day focused on growth, execution, and real-world scaling. Learn from founders and investors who have shaped the industry. Connect with peers navigating similar growth stages. Walk away with tactics you can apply immediately



Save up to $300 on your pass or save up to 30% with group tickets for teams of four or more. Boston, MA | REGISTER NOW

Korea’s Personal Information Protection Commission (PIPC) said that more than 30 million Coupang accounts were exposed — but the facts point to just 3,000 affected accounts, according to Coupang’s investors.

In December, South Korea’s government and the PIPC said the Coupang breach was serious enough to justify higher fines. Under current law, penalties are capped at 3% of revenue, more than $800 million for Coupang, according to the U.S. investors, but some lawmakers have proposed raising the limit to 10% and applying it retroactively.

Even if the new law passes, it wouldn’t apply to Coupang, since the breach occurred before the rules changed. But a Democratic Party lawmaker in the country suggested imposing punitive fines, through either new legislation or a special parliamentary act, and PIPC backed the idea, per news reports. South Korean president Lee Jae Myung also publicly called for heavy penalties, suggesting the company had not faced sufficient consequences.

Based on the notice of intent filing released by the investors’ legal adviser, the investors argue that the South Korean government’s actions constitute an “unprecedented assault” on Coupang. In the filing, they argue:

The Government’s unprecedented assault on a U.S. company to benefit its Korean and Chinese competitors is an egregious violation of the Treaty, principles of international law, and the historic partnership between Korea and the United States … The Government’s shocking conduct has left the U.S. investors with no choice. If the Government does not immediately cease its attacks against Coupang, fully restore the company’s ability to operate its business, and permanently end its longstanding campaign of discrimination against the company, then the U.S. investors will be forced to seek billions of dollars in damages from Korea to protect their investments in Coupang and remedy the Government’s ongoing Treaty violations, including attempted expropriation.

The filing is a preliminary, pre-litigation step. South Korea’s Ministry of Justice is now reviewing the notice of intent, which kicks off a mandatory 90-day consultation period before formal arbitration can begin.

Coupang, Abrams Capital, and Foxhaven Asset Management did not respond to TechCrunch’s request for comment. Durable Capital Partners could not be reached.

According to the investors’ filing, South Korea’s handling of data breaches has been inconsistent, specifically citing other recent data breaches in South Korea, including KakaoPay, SK Telecom, Upbit, and Alibaba’s AliExpress.

KakaoPay reportedly transferred 54 billion customer records to Alipay Singapore, yet faced only a $10 million fine and a CEO warning, while SK Telecom was fined $91 million after a massive SIM card breach. Upbit and AliExpress also saw minimal government action. The investors say these examples underscore the stark contrast with the government’s response to Coupang.

South Korea’s Ministry of Science and ICT said Wednesday that the Coupang data breach was carried out by a former employee who had worked on the company’s authentication systems and was aware of vulnerabilities in both the authentication framework and key management system.

The Ministry alleges that Coupang failed to report the breach to the Korea Internet & Security Agency (KISA) within 24 hours and did not fully implement a November 2025 data preservation order, leading to the deletion of key web and app access logs. The ministry has referred the matter to investigators and ordered Coupang to submit a prevention plan by February 2026, with compliance monitored through July.

Coupang released a statement, saying that the employee, a Chinese national, accessed data from over 33 million accounts but retained only about 3,000 before deleting it, and that no sensitive info such as payment data, passwords, or government IDs was accessed.

Coupang also replaced its CEO, Park Dae-jun, with Harold Rogers, its U.S. parent’s top lawyer, in December.

Adam Farrar, senior associate at CSIS and senior geoeconomics analyst for APAC at Bloomberg, said on Tuesday’s Impossible State podcast that what began as a major data breach involving Coupang has grown into a broader issue between the United States and South Korea.

Farrar said that the case is amplifying broader U.S. claims of unfair treatment toward American technology firms, raising trade and tariff risks for South Korea as the U.S. Congress becomes increasingly engaged.

“The massive data breach [by Coupang] led to a series of investigations in the National Assembly and some very combative back and forth with Coupang and a series of executives over the past several months,” Farrar said in the podcast. “The additional dynamic here is that Coupang, while driving almost all of its earnings from Korea, is now a U.S.-based company that adds to the dynamic on both sides, impacting how they’re perceived and seen.”

The issue extends beyond Coupang, raising broader questions about whether South Korea is unfairly targeting U.S. companies, Farrar continued.

Critics point to digital policies they say favor domestic firms, including network usage fees on content providers like Netflix, Apple’s App Store and Google Play payment rules, and data localization requirements that limit services like Google Maps on national security grounds.

In Brief

Apple has acquired the IP and all rights to its hit show “Severance” from its original studio, Fifth Season, as first reported by Deadline. Under the deal, which was worth just under $70 million, Apple’s in-house studio will be producing the show’s future seasons. Fifth Season will remain as an executive producer.

The deal is similar to one Apple struck with AMC Studios for sci-fi show “Silo” after season one.

“Severance” will now be one of Apple’s marquee titles, joining popular series such as “Your Friends and Neighbors” and “Stick.”

According to Deadline, the production costs for “Severance” had exceeded what Fifth Season could afford. The studio had already requested advances from Apple and was considering relocating production from New York to Canada for bigger and quicker tax rebates. Since Apple definitely has the funds to back the show, the company decided to take full control.

“Severance” is important to the streamer, as the show’s second season became Apple’s most-watched series at the time. It also had the highest number of nominations at the 2025 Emmy Awards.

Deadline reports that the show is expected to run for four seasons, with the possibility of spin-offs, a prequel, and foreign versions.

The U.S. Federal Trade Commission (FTC) has raised concerns over allegations that Apple is censoring conservative content on the Apple News app.

In a letter to Apple CEO Tim Cook, FTC chair Andrew Ferguson cited reports from Media Research Center, a right-leaning think tank, which accused Apple of excluding right-leaning outlets from the top 20 articles in the Apple News feed.

“These reports raise serious questions about whether Apple News is acting in accordance with its terms of service and its representations to consumers […] I abhor and condemn any attempt to censor content for ideological reasons,” Ferguson’s letter reads.

Ferguson, a Big Tech critic who Trump appointed to lead the competition regulator, noted the FTC doesn’t have any powers to require Apple to take ideological or political positions when curating news, but he said that if the company’s practices are “inconsistent” its terms of service or “reasonable expectations of consumers,” they may be in violation of the FTC Act.

Brendan Carr, the chairman of the Federal Communications Commission (another Trump appointee critical of Big Tech), supported Ferguson’s stance, writing, “Apple has no right to suppress conservative viewpoints in violation of the FTC Act.”

Ferguson has urged Apple to conduct a “comprehensive review” of its terms of service and ensure that the content curated on Apple News is consistent with its policies, and “take corrective action swiftly” if the curation isn’t in line.

The letter comes a day after President Donald Trump shared the report by Media Research Center on his social media platform, Truth Social. Trump has repeatedly accused Big Tech companies of censoring right-leaning content, though many major platforms have rolled back several measures to curb fake news and disinformation they had imposed in the years prior to his second stint at the White House.

Techcrunch event TechCrunch Founder Summit 2026: Tickets Live On June 23 in Boston, more than 1,100 founders come together at TechCrunch Founder Summit 2026 for a full day focused on growth, execution, and real-world scaling. Learn from founders and investors who have shaped the industry. Connect with peers navigating similar growth stages. Walk away with tactics you can apply immediately



Save up to $300 on your pass or save up to 30% with group tickets for teams of four or more. TechCrunch Founder Summit: Tickets Live On June 23 in Boston, more than 1,100 founders come together at TechCrunch Founder Summit 2026 for a full day focused on growth, execution, and real-world scaling. Learn from founders and investors who have shaped the industry. Connect with peers navigating similar growth stages. Walk away with tactics you can apply immediately



Save up to $300 on your pass or save up to 30% with group tickets for teams of four or more. Boston, MA | REGISTER NOW

Apple’s relationship with the Trump administration has oscillated between warm and cold over the past year. Trump has criticized Big Tech, especially Apple, for manufacturing its devices in China, but after Cook promised to spend more than $600 billion over the next four years stateside and moved to mend fences, relations between the administration and the company have improved. Apple also dodged planned tariffs on smartphones made overseas and imported into the U.S.

The FTC last year also launched an investigation into “censorship by tech platforms,” seeking input from the public who felt they were silenced due to their political ideologies or affiliations. “Tech firms should not be bullying their users,” Ferguson said at the time. “This inquiry will help the FTC better understand how these firms may have violated the law by silencing and intimidating Americans for speaking their minds.”

Apple did not immediately return a request for comment.

Has AI coding reached a tipping point? That seems to be the case for Spotify at least, which shared this week during its fourth-quarter earnings call that the best developers at the company “have not written a single line of code since December.” That statement, from Spotify co-CEO Gustav Söderström, came alongside other comments about how the company is using AI to accelerate development.

Of note, Spotify pointed out it shipped more than 50 new features and changes to its streaming app throughout 2025. And, most recently, it has rolled out more features, like AI-powered Prompted Playlists, Page Match for audiobooks, and About This Song, which all launched within the past few weeks.

At Spotify, engineers are using an internal system called “Honk” to speed up coding and product velocity, the company told analysts on the call. This system allows for things like remote, real-time code deployment using generative AI, and specifically Claude Code.

“As a concrete example, an engineer at Spotify on their morning commute from Slack on their cell phone can tell Claude to fix a bug or add a new feature to the iOS app,” Söderström said. “And once Claude finishes that work, the engineer then gets a new version of the app, pushed to them on Slack on their phone, so that he can then merge it to production, all before they even arrive at the office.”

Spotify credited the system in helping to speed up coding and deployment “tremendously.”

“We foresee this not being the end of the line in terms of AI development, just the beginning,” Söderström said.

The exec also touted Spotify’s ability to build a unique dataset that other LLMs could not commoditize, the way they could other online resources, like Wikipedia. That’s because there’s not always a factual answer for music-related questions, he said.

Techcrunch event TechCrunch Founder Summit 2026: Tickets Live On June 23 in Boston, more than 1,100 founders come together at TechCrunch Founder Summit 2026 for a full day focused on growth, execution, and real-world scaling. Learn from founders and investors who have shaped the industry. Connect with peers navigating similar growth stages. Walk away with tactics you can apply immediately



Save up to $300 on your pass or save up to 30% with group tickets for teams of four or more. TechCrunch Founder Summit: Tickets Live On June 23 in Boston, more than 1,100 founders come together at TechCrunch Founder Summit 2026 for a full day focused on growth, execution, and real-world scaling. Learn from founders and investors who have shaped the industry. Connect with peers navigating similar growth stages. Walk away with tactics you can apply immediately



Save up to $300 on your pass or save up to 30% with group tickets for teams of four or more. Boston, MA | REGISTER NOW

For instance, if you asked what workout music is, you’d get different answers from different people, sometimes based on their geography. Americans tend to prefer hip-hop overall, though millions prefer death metal. And while a number of Europeans would work out to EDM, many Scandinavians like heavy metal.

“This is a dataset that we are building right now that no one else is really building. It does not exist at this scale. And we see it improving every time we retrain our models,” Söderström noted.

Analysts on the call also asked about Spotify’s approach to AI-generated music. The company explained that it’s allowing artists and labels to indicate in a track’s metadata how the song was made but is still policing the platform for spam.

重點新聞（0206～0212）

GPT Transformer 簡化

200行Python程式碼重現GPT核心，AI也可以很極簡

前OpenAI研究員、Eureka Labs創辦人Andrej Karpathy最近發起一個實驗專案microGPT，只用243行純Python程式碼、不依賴任何機器學習框架，就完整重現GPT語言模型的訓練與推論流程，把GPT的核心運作邏輯完整寫出來。他形容，GPT真正「不可或缺」的演算法就是這些，其他程式碼只是為了讓模型跑得更快、更有效率。他後來又再優化架構，進一步把程式碼精簡到200行。

Andrej Karpathy沒有用PyTorch、TensorFlow或NumPy等任何深度學習框架，而是用最基本的數學運算，把GPT需要的所有核心機制一個個寫出來。從Transformer架構和自注意力機制，到前向與反向傳播、梯度更新，再到一個字一個字生成的自回歸流程，都完整寫在這幾百行程式碼裡。只不過，模型在訓練時的梯度計算，由他自製的micrograd引擎負責，並搭配Adam演算法更新參數。

他後來精簡過的版本，把反向傳播寫得更直觀：每個運算步驟只需說明「如果輸出改變，自己會怎麼改變」，整體誤差的連鎖傳遞由系統統一處理。每個模組只負責自己的部分，程式結構因此變得更乾淨。

新版程式也排成三欄形式：資料與自動微分、GPT模型本體，以及訓練和推論流程。這個專案就像是把大型語言模型（LLM）的核心，從龐大框架和黑盒子中抽出來，回到最基本的數學與演算法本質。（詳全文）

OpenAI Codex 程式開發

OpenAI推GPT-5.3-Codex，AI開始協助打造AI了

OpenAI再推新版代理型程式開發模型GPT-5.3-Codex，不只強化開發能力，更首度公開表示，在訓練與部署過程中，新模型的早期版本還能使用工具、協助開發後來的正式版本，就像是自己開發自己一樣，揭示「AI輔助AI開發」開始走向實務。

GPT-5.3-Codex延續400K的上下文長度，並優化長時間、多步驟任務的穩定性和執行速度，與前代相比，整體效能提升約25%。這套模型專為Nvidia GB200 NVL72架構設計，在SWE-Bench Pro與Terminal-Bench 2.0等基準測試中，表現也比前代模型好，部分成績也高於Claude Opus 4.6。

OpenAI強調，GPT-5.3-Codex可支援從寫程式、除錯、部署到文件記錄和測試等完整軟體生命周期。在整體開發能力上，OpenAI強調GPT‑5.3-Codex具備更強的編碼、美學和壓縮優化能力，能在數天內從零開始到建立複雜的遊戲和應用程式。這波更新不只是新模型發布，更顯示AI從輔助寫程式，開始參與下一代模型開發，意味著代理型AI競賽進入新階段。（詳全文）

OpenAI 代理 Frontier

OpenAI推代理管理平臺Frontier，瞄準企業級AI協作

OpenAI日前推出代理人管理平臺Frontier，讓企業可開發、部署並統一管理各類AI代理。目前已開放少數客戶使用，未來將逐步擴大。

OpenAI指出，企業要真正提升生產力，關鍵不只是導入AI，而是建立跨系統的AI代理管理機制。Frontier可串接資料倉儲、CRM系統和內部應用，建立共用的「企業脈絡」，讓不同代理理解相同資料和任務背景。企業可透過ChatGPT Enterprise等介面操作Frontier，並同時管理自建、OpenAI提供的或第三方AI代理。

功能上，Frontier涵蓋四大面向：一是整合資料與工具，讓代理人能在既有流程中運作；二是任務規劃與資源調度，可依需求排程模型運算；三是品質回饋與記憶機制，協助代理持續優化；四是企業級安全與治理，為每個代理人設定身份、權限與護欄。目前HP、Oracle和Uber等企業已導入或測試，OpenAI也派出前導部署工程師協助落地。（詳全文）

Anthropic 金融分析 Office

Anthropic新模型專攻金融分析和Office整合

Anthropic在2月5日發表新模型Claude Opus 4.6，主打強化金融研究和分析能力，並同步升級Claude in Excel，還推出研究預覽版Claude in PowerPoint與桌面工具Cowork，把生成式AI進一步帶進金融專業人員的日常工作。

Anthropic表示，新版在推理、多工處理與長時間多步驟任務的穩定性上，都有明顯進步。在其內部Real-World Finance評測中，Claude Opus 4.6在50項投資與財務分析情境下，表現比幾個月前推出的Claude Sonnet 4.5高出23個百分點。在外部測試中，在Vals AI的Finance Agent評測達60.7%，TaxEval稅務測試則是76%。

產品面上，Claude in Excel可直接協助規畫分析流程、編輯樞紐分析表和圖表，降低跨檔案操作負擔。研究預覽版Claude in PowerPoint則可讀取簡報版型和母片，自動生成或修改投影片。桌面工具Cowork可在授權下，存取指定資料夾，同時處理多份文件與任務，並支援外掛整合財務流程。市場反應上，多家傳統企業軟體股在這項消息公布後出現下跌，顯示AI對既有工作軟體版圖的衝擊正在擴大。（詳全文）

Mistral 語音轉文字 會議

Mistral揭露新一代語音轉文字模型

法國AI新創Mistral推出語音轉文字系列模型Voxtral Transcribe 2，鎖定會議逐字稿、客服通話和即時語音互動等需求。這次發布兩款模型：適合大量錄音處理的Voxtral Mini Transcribe V2，以及主打即時應用的Voxtral Realtime。官方表示，即時模型的轉錄延遲可設定到200毫秒以下。

Voxtral Realtime採用串流架構，音訊一進來就同步產生文字，而不是先切段再處理。開發者可依需求，調整速度和準確度。模型權重採Apache 2.0授權，企業可自行部署在本地或邊緣設備。

批次模型則新增說話者分離、詞級時間戳記與語境偏移功能，可加入最多100個關鍵詞，幫助辨識人名或專業術語，方便整理多人會議。官方公布，模型在FLEURS測試中詞錯誤率約4%，API定價為每分鐘0.003美元。Mistral也在Mistral Studio新增音訊試用介面，讓團隊可上傳檔案、即時檢視轉錄效果。（詳全文）

GitHub 基礎設施 AI代理

GitHub把AI代理變成開發基礎設施

GitHub旗下工具Agent HQ開始支援Claude和Codex模型（公開預覽版），意味著GitHub正把多家AI模型納入核心開發流程，而不是只當作外部輔助工具。Copilot Pro+和Copilot Enterprise用戶可在GitHub網站、行動版與VS Code啟動代理工作階段。

這次的更新關鍵，在於讓不同代理直接參與既有協作流程。開發者可將議題指派給代理，或要求其在Pull Request中分析、提出修改建議，所有對話、變更和日誌都會保留在儲存庫內，納入原本的審查與追蹤機制，減少工具切換的麻煩。

VS Code 1.109以上版本可支援本機互動與雲端自動化兩種模式，長時間任務可交由GitHub端執行。企業管理者則能集中設定模型存取權限和安全政策。GitHub也預告，接下來將引進更多AI供應商，來打造多代理並存的開發環境。（詳全文）

104人力銀行 推薦 履歷

104人力銀行揭3項AI應用，推薦與履歷健檢加速工作媒合

104人力銀行日前公布旗下3項AI亮點應用，包括AI工作／人才推薦、AI訊息分類和AI履歷健檢。這些功能結合自研演算法與開源、商用大型語言模型（LLM），目標是要提高求職者與企業間的雙向媒合效率。

在推薦機制方面，AI會根據求職者瀏覽行為、履歷內容與偏好，自動推送高關聯職缺；企業端則可每日收到符合職缺條件的人才建議，並透過回饋持續優化模型。104統計，透過AI推薦獲得面試邀約的機率，是自行搜尋的3.2倍，企業端邀約同意率也約為傳統方式的3倍。

為降低溝通落差，104也推出AI訊息分類，將企業面試邀約分為高關聯的「專屬」與其他「一般」訊息，協助求職者快速判斷優先順序。此外，AI履歷健檢可在約3秒內給出客製化建議，引導補強關鍵經歷與成果，而非直接代寫。近4成使用者依建議更新履歷，整體面試機會提升約20%。（詳全文）

Databricks LLM 評審

Databricks發表MemAlign，用雙記憶機制強化LLM評審能力

Databricks旗下Mosaic AI Research團隊發表MemAlign，是一套以「雙記憶系統」累積專家回饋的大型語言模型（LLM）評審對齊框架，並已整合進開源模型管理平臺MLflow。。

企業常讓LLM擔任評審，檢查客服機器人或代理程式輸出是否合規，但通用模型的判斷往往和領域專家標準有落差。Mosaic指出，單靠提示詞，容易前後不一致，若靠微調，則成本高，因此改採記憶機制累積專家自然語言回饋。

MemAlign把可重複使用的原則存入「語意記憶」，把具體案例保留在「情節記憶」，新任務評估時會同時檢索兩類記憶作為參考，不需更新模型權重。在Prometheus-eval測試中，僅用最多50筆回饋樣本，即可在約40秒內完成對齊，成本約0.03美元。不過推論時需進行向量搜尋，單筆評估可能增加約1秒延遲。目前，MemAlign已可在MLflow與Databricks平臺使用。（詳全文）

圖片來源／OpenAI、Anthropic、104人力銀行、Databricks

AI趨勢近期新聞

1. Google將Data Commons MCP改為雲端託管，免本機部署即可查詢公開統計資料

資料來源：iThome整理，2026年2月