Latest AI News
Everyone is navigating AI security in real time — even Google
I recently had the opportunity to sit down with Francis de Souza, COO of Google Cloud, backstage at aneventin Los Angeles. Amid the din around us, de Souza, who speaks in the calm, measured manner of a university professor, offered useful advice for companies navigating the AI security moment we’re all living through, noting that “there’ll be a transition period, and then I think we get to this better place.” He wasn’t speaking about Google at that moment, but it’s clear that even Google is still figuring things out. De Souza’s core message was one security professionals have been trying to get executives to internalize for years, now made urgent by AI: security can’t be an afterthought. “As companies embark on this AI journey, they need to take a platform approach,” he said. “Security is not something you can bolt on later, and it’s not something you can leave up to employees to do on their own.” He warned specifically about “shadow AI” — employees reaching for consumer tools without organizational oversight — and argued that companies need to demand security, governance, and auditability from their platforms from the start. “There’s no such thing as an AI strategy without a data strategy and a security strategy. They need to go hand in hand.” Worth noting: he wasn’t pitching Google Cloud alone. When I observed that his advice sounded like a Google advertisement, he pushed back. Google, he said, is committed to a multicloud approach, and he made the case that companies that think they’re operating on a single cloud almost certainly aren’t. “Even if they pick a single cloud, they’re relying on SaaS applications, there are business partners that may be using different clouds,” he said. “It’s important for companies to have a security posture that is consistent across clouds, across models.” He also made the case that the threat landscape has changed so fundamentally that old defensive models are too slow. He noted that the average time between an initial breach and the handoff to the next stage of an attack has dropped from eight hours to 22 seconds, and that the attack surface has expanded well beyond the traditional network perimeter. “In addition to your usual estate, you have models now. You have data pipelines used to train the models. You have agents, you have prompts. All of this needs to be protected.” One threat de Souza flagged that doesn’t get enough attention: agents moving through a company’s internal systems can surface forgotten data repositories that nobody has thought about in years. “A lot of organizations have old SharePoint servers [and access controls] they haven’t really updated, but it didn’t matter because nobody really knew where they were. But agents roaming your enterprise will find those data assets and will expose the data on them.” The answer, in his view, is to meet machine speed with machine speed. “We’re now seeing the emergence of an AI-native, fully agentic defense where organizations can run agents driving their defense,” he said. “Instead of having a human-led defense or even a human in the loop, you can now have humans overseeing a fully agentic defense.” He added that this has become a leadership issue, not just a technology one. “This is a board-level issue and an executive team issue. It’s not just a security team’s issue.” But even as AI takes on more of the defensive workload, the people qualified to oversee it are in short supply — and the vulnerabilities that AI itself is introducing are multiplying faster than security teams can address them. “We’re going to need people to deal with the bug-pocalypse,” LinkedIn’s chief information security officer Lea Kissnertold the New York Timesthis week, adding that she doesn’t expect the industry to understand AI security in any sustainable long-term way for at least several years. Which brings us back to the platform providers themselves. The Register has published a series of reports over the past several weeks documenting a wave of Google Cloud developers hit with five-figure bills following unauthorized API calls to Gemini models — services many of them had never used or intentionally enabled. The cases followed a familiar pattern: API keys originally deployed for Google Maps, placed publicly per Google’s own instructions, had quietly become capable of accessing Gemini after Google expanded their scope without clearly disclosing the change. Rod Danan, CEO of interview-prep platform Prentus, said his bill hit$10,138 in roughly 30 minutesafter attackers exploited his compromised API key. Isuru Fonseka, a Sydney-based developer whose account was similarly compromised, woke up to charges of roughly AUD $17,000 despite believing he had a $250 spending cap in place. What neither knew was that Google’s automated systems had upgraded their billing tiers based on account history, raising their effective ceilings to as high as $100,000 without explicit consent. Google refunded both after The Register published its initial report. Still, Google told The Register it has no plans to change its automatic tier-upgrade policy, saying it prioritizes preventing service outages over enforcing users’ stated budget preferences. In the meantime, there is the separate question of what happens when a developer tries to shut things down. The Registerreported this weekon research by security firm Aikido finding that even developers who catch a compromised key and immediately delete it may not be safe. According to Aikido’s findings, attackers can apparently continue using that key for up to 23 minutes because Google’s revocation propagates gradually across its infrastructure. Aikido researcher Joseph Leon told The Register that during that window, success rates are unpredictable — in some minutes over 90% of requests still authenticated — and attackers can use the time to exfiltrate files and cached conversation data from Gemini. Leon also noted that Google’s own newer credential formats don’t appear to have the same problem: service account API credentials revoke in about five seconds, and Gemini’s newer AQ-prefixed key format takes about a minute. “Both run at Google scale,” he wrote in Aikido’s related paper. “Both suggest this is technically solvable for Google API keys, too.” In short, according to Leon, the 23-minute window isn’t an engineering constraint but a matter of priorities for the company. That’s worth considering when reading de Souza’s advice, which is sound and should be taken very seriously. He’s not wrong, but there is currently a gap between the platforms are prescribing and how fast they are themselves adapating, and it’s good to be aware of this, too.
View
I tried Amazon’s Bee wearable and am both intrigued and slightly creeped out
I recently had the opportunity to test out a wearable fromBee, the AI wrist gadget that Amazonacquiredlast year and has sinceupdatedwith a number of new features. Like other AI wearables, Bee is designed as a kind personal assistant: it records, transcribes, and summarizes the user’s conversations throughout the day, providing an ongoing note-taking capability that’s useful if you’re forgetful or just want to be more organized about your life. If you sync it with your calendar, it can also send you alerts and reminders about things you’re supposed to do throughout the day. TechCrunch haswritten about Beebefore, and the way it works is pretty simple: the user powers it up, puts it on, syncs it with the Bee mobile app, and enters some basic personal information. Bee has a built-in recorder that can be turned on and off by clicking the wearable’s button. When Bee is recording, a green light flashes. When it’s not, that green light goes off. After a conversation has been recorded, the app will create an automated summary that is easy to read, as well as an entire transcription of the conversation. Your mileage may vary on how exciting (or not) this whole conceit is. The problem for me is that I am something of a privacy enthusiast. In a world where the average person is beset from all sides by constant digital surveillance, I appreciate any opportunity I can get to not be recorded. Therefore, the idea of walking around with an eavesdropping gizmo strapped to my wrist 24/7 was not particularly appealing. Yet, even I have to admit that — in the right context — Bee could have a lot of potential to help organize your life. Bee really comes through in the context of professional engagements. If your day is full of meetings and you have trouble keeping it all straight, Bee could be a moderately competent assistant. During a business-related phone call this week, I activated Bee after getting confirmation that I could record our meeting. Afterward, the app faithfully regurgitated a summary of the conversation, helpfully breaking down each segment of our talk so that I could review it later without having to re-listen to our entire conversation. This was undeniably helpful, although it should be noted that this isn’t something that’s markedly different than those offered by other transcription services,like OtterorGranolaand others, which also offer transcriptions and auto-generated summaries. That said, you could envision a situation in which a professional who has to navigate between various meetings throughout the day would be well-served by this device. You could just keep Bee running throughout the day and, later, review the conversation summaries for anything you’re not clear about. Bee does a relatively good job at summarizing conversations, but the actual transcripts offered by the wearable can be a bit of a mess. Previous critics have noted that you usually have to manually enter the names of other speakers, as Bee doesn’t always know who is talking. During my conversation, I noticed that it had also omitted certain sections of our chat — nothing huge, but it wasn’t a complete account of everything that had been said. I also took Bee to my semi-weekly movie night with my friends and left it running throughout the night. Given the fact that we watchedReservoir Dogs, I was mildly afraid that the wearable would mistake all of the vulgar carnage for real-life bloodshed and potentially trigger some sort of internal alarm. However, Bee knew — basically — what was happening. The wearable figured out that we were watching a movie and, in the summary of events afterward, the wearable labeled the conversation “Tarantino Film Scene Analysis.” While Bee shows early promise as a professional tool, I would not want this thing recording me in my personal life. Weirdly enough, Bee has largely been marketed as a product for personal use. To be comfortable with that, you have to be comfortable with Bee having access to a majority of both your offline and digital life. Indeed, to work well, Bee needs expansive mobile permissions — including access to your location, photos, phone contacts, calendar, and mobile notifications. You can also share your health data with it — should you, for whatever reason, want it to know about your sleep patterns or your resting heart rate. The large accumulation of data Bee collects is stored in the cloud, which — again, for the digital privacy enthusiast — presents its own concerns. In a message to tech YouTuber Becca Farsace, Bee apparentlyunveiled a demoof the device running entirely locally. Were the company able to produce such a device, I would be thoroughly impressed — and might even consider buying one. That said, Amazon hasn’t offered any update on those plans. As for Bee’sdigital privacy protections, the company says that it offers encryption to protect user data — both at rest and in-transit. In its privacy policy, the company states that it has “implemented technical and organizational security measures designed to protect the security of any personal information” that the company processes. Bee also claims that it undergoes “rigorous third-party security audits” and employs continuous security monitoring. That all sounds quite good, although it’s worth noting that Amazon — like many large tech companies — has been subject to the occasionaldata security issue or two(not exactly surprising for a company that governs as much of the global cloud environment as it does, but still). In short, Bee is a curious piece of hardware that, given some time and some tweaking, could have some promising professional applications further down the road. As a digital assistant for your personal life, however, it might prove to be a little too invasive for some users.
View
A Bionic Arm That Helped India’s Amputees Face The World Again
India has an estimated five lakh upper-limb amputees, according to rehabilitation studies. Access to prosthetics, however, remains sharply unequal.
View
Ferrari is using IBM’s AI to create F1 superfans
Two years ago, IBM realized there was one glaring omission in its roster of sports partnerships: Formula One. Formula One has become one of the world’s most popular sports, especially in the U.S., where Netflix’s “Drive to Survive” documented the working lives of F1 drivers and turned them into mainstream celebrities. The tech-centric sport has also become ahot ticket for tech companieslike AWS, Oracle, and Anthropic, which partner with teams for sponsorship visibility and to provide data analytics and AI tools that can deliver a competitive edge. So when IBM went looking for its next major sports partnership, it’s no wonder the company picked F1 and one of its most iconic teams,Scuderia Ferrari HP. “They’re the winningest team in history,” Kameryn Stanhouse, IBM’s Vice President of Sports and Entertainment Partnerships, told TechCrunch. At the heart of this partnership, however, is what has led other teams to start working with tech giants: access to more sophisticated tech solutions that can help them make the most of, especially, artificial intelligence. In fact, one of the best parts of sports, Stanhouse said, is how much data is available and can be used to help people get comfortable with AI. “They actually see how it serves them,” she said of how AI is used in sports storytelling. The IBM-Ferrari partnership centers on that idea of storytelling, enhancing fan engagement by overhauling the technology powering the Ferrari fan app. To help with this, Ferrari hired Stefano Pallard in the newly titled role “head of fan development,” who said the challenge the team wanted to tackle was not just reaching fans, but “making each of them feel like we know them.” “That starts with taking the data we get from the track and turning it into content that is easy to follow and engaging,” he told TechCrunch. Teamsprocess millions of data pointsper second during each race, capturing every movement of the driver and the car. Turning this into content that fans can engage with is just one way that advanced enterprise AI can help businesses better interact with their consumers. Among the 11 teams, Ferrari is one of the few (alongside the likes of McLaren and Williams) to have a standalone fan app strategy rather than lean on social media or the official F1 platforms instead, showing how the sport is slowly starting to capitalize on its growing global fandom. Some of the changes to the Ferrari app were simple, like offering it in Italian. Even though Ferrari is an Italian company and many of its fans are Italian, their fan app was not available in Italian until the IBM partnership. Stanhouse said the old Ferrari fan app was a place where people went to find race details and then leave. This new app has games where fans can play with others in the app, new AI-written race summaries, more behind-the-scenes stories about the team and the drivers, a place to make predictions, and an AI companion for fans to ask questions. “There are two drivers, but did you know it takes 24 people working simultaneously in two seconds to change a tire?” Stanhouse said, adding that storytelling helps fans feel closer to the team. Unlike other sports apps IBM has built, Stanhouse said the Ferrari app’s main focus is storytelling because it wants fans to stay engaged with it all year long, rather than for a few weeks a year, as with tournaments like the Masters. Engagement data for the app has been trending upward since IBM came into the scene, Stanhouse said, citing a 62% increase in engagement over race weekends as an example. Pallard said the team then uses AI to analyze engagement signals in the app, such as which content people like to read and the sentiment of the messages fans send. “That helps us understand what resonates most with the Tifosi [the fan nickname for Ferrari] and it directly informs how we shape our storytelling and how we deliver content,” he said. The team hopes to dive deeper into personalization and create more immersive fan experiences. The app developers also took into account Ferrari’s fanbase, which is much more diverse than it was even five years ago. F1 releasedstats last year showingthat 75% of new fans were women, many of whom were Gen Z. A particular draw for women is the F1 Academy, an all-female racing series that aims to develop the next generation of women drivers. But these new fans, much like the old, are after one thing — more. “They are asking for more data, more insight, more features, and we have to be able to deliver that,” Pallard said. “With IBM, the vision for the next five years is to make every fan feel like the experience was built for them, whether they have been with us for 30 years or 30 days. That is how you build loyalty that lasts.”
View
Elon Musk has given up on solar power (on Earth)
Has Elon Musk given up on Tesla’s Master Plans, on the electrified economy, on solar power as we know it? From theSpaceX IPO filingreleased yesterday, it sure seems like it. A recap for those not enmeshed in the Musk-verse: Tesla has releasedfour Master Plansover the years, and while details have varied, the through line has been electrification of the economy. Musk put it best in his first edition: “the overarching purpose of Tesla motors…is to help expedite the move from a mine-and-burn hydrocarbon economy towards a solar electric economy.” But recently, one of Musk’s companies, xAI, has embraced the mine-and-burn hydrocarbon economy, usingdozens of unregulated natural gas turbinesto power its data centers with plans tobuy $2.8 billion more, effectively cementing the fossil fuel’s role in the company’s AI operations. It’s a curious turn for a businessman who built his empire on clean energy — and who has no qualms directing his companies to buy from one another. SpaceX spent $131 million on 1,279 Cybertrucks, and xAI has spent $697 million in the last two years on Tesla Megapacks, it’s grid-scale battery storage systems that the company will use to manage peak loads. But so far, xAI hasn’t bought a materially significant number of solar panels from Tesla. Solar power isn’t missing in the SpaceX filing, it’s just all concentrated on space, which the company touts as the future of data center power. Terrestrial solar garners a few mentions — not as a power source for xAI data centers but instead to show how much better SpaceX thinks space-based solar will be. It’s no secret that Musk and other Silicon Valley executives have become obsessed with space-based solar power. SpaceX says that space-based solar arrays can generate “more than five-times the energy” of terrestrial ones thanks to 24/7 illumination. As AI data centers have run into opposition here on Earth, CEOs like Musk have started mulling big server racks in space powered by that 24/7 sunshine. Hammer, meet nail. Even if SpaceX is able to bring down the cost of boosting a data center into orbit, the economics arechallenging at best. Power prices for Starlink satellites are multiples higher than what a terrestrial data center typically spends, and protecting chips from the rigors of space won’t be easy or cheap. It’s also not clear whether AI training can be distributed across multiple satellites, leaving a significant chunk of AI work earthbound. It’s not just one problem that SpaceX needs to solve, but many. It’s likely that Musk considers xAI’s current data centers as stopgaps, that once SpaceX is able to loft gigawatts worth of servers into orbit — probably just a few years away, in his mind — he’ll scrap what’s here on the ground, natural gas turbines included and not have to think about NIMBYs anymore. The risk, of course, is that he’s wrong. It’s not just NIMBYs that Musk is worried about, though. He’s clearly concerned that computing demands from AI will quickly outstrip what we can provide here on Earth. Sprinkled throughout the SEC filing are references to “terawatt-scale annual AI compute growth,” which will require power to match. That’s a stunning figure when you consider that all the world’s data centers use around40 gigawatts today. This is Musk’s “first principles” thinking in action. At some point, he assumed the world will need an additional terawatt worth of compute every year, and he worked back from there. “We believe that third-party estimates on data center demand are constrained by the practical supply limitations that exist in a terrestrial context and the power shortage may be far greater than what research estimates suggest,” the company argues. Possible? Sure, I suppose. But consider that humanity today usesabout 35,000 terawatt-hoursof energy annually, or about 4 terawatts on a continuous basis. Energy demand has risen lately, and for AI, it probably is in an phase of exponential growth, which could either continue or level off. We have no way of knowing at this point, but if there’s one thing Musk is good at, it’s spotting a trend at its inflection point and extrapolating wildly. Here’s where Musk’s problems settle back down to Earth. I’m no rocket scientist, but I suspect that shipping solar panels on a flatbed truck uses less energy than sending them into orbit. Plus, space-ready solar panels will need to be manufactured at unprecedented scale. Not insurmountable problems, but also maybe a distraction. We’ve barely scratched solar’s potential here on Earth, for example. The perfect doesn’t have to be the enemy of the good. There’s plenty of room to improve things here on Earth even while we chase after our dreams in the stars. Just three years ago, Musk and his colleagues at Tesla released the “Master Plan Part 3,” which thoughtfully outlined a “plan to eliminate fossil fuels.” A good starting point might be xAI’s data centers.
View
'India Cannot Build Deeptech on Hype Alone'
While India looks to be a leader in AI and semiconductors, investor Aditya Vuchi suggests focusing on practical applications and real customer needs.
View
AI is being used to resurrect the voices of dead pilots
In the latest sign of these AI-heavy times, the National Transportation Safety Board temporarily removed access to its docket system after discovering that voices of pilots who were killed in a UPS plane crash last year had been re-created using AI and were circulating on the internet. The NTSB is prohibited by federal law from including cockpit audio recordings in its docket system, which otherwise contains troves of data on investigations and has historically been open to the public. But the accident docket for this flight included a spectrogram file of the voice recorder. A spectrogram uses a mathematical process to turn sound signals, including low and high frequencies, into an image. Scott Manley, a popular YouTuber whose channel combines physics, astronomy, and video games,noted on Xthat it could be possible to reconstruct audio from the megabytes of data encoded in that image. And that’s what happened. People took the spectrogram, along with the publicly available transcript, to create approximations of the cockpit voice recorder audio from UPS Flight 2976 in Louisville, Kentucky,according to the NTSB. They used AI tools like Codex, according to posts on social media. The agencyrestoredpublic access to the docket system on Friday but kept 42 investigations closed pending review — including the one related to Flight 2976.
View
SpaceX files to go public, and the math requires a little faith
Loading the player… The SpaceX S-1 is finally here, and the story it tells goes way further than rockets. The filing runs to 36 pages of risk factors alone, and the numbers inside match the ambition: a $28 trillion total addressable market, a pay package tied to establishing a Mars colony, and a valuation target that would make it the largest IPO in American history. Watch asEquitypodcast hosts Kirsten Korosec, Anthony Ha, and Sean O’Kane dig into what the filing actually says, what it leaves out, and whether any of this math connects to reality. The team also coversNanoCo turning down a $20M buyoutto raise a $12M seed for its secure Nano Claw alternative, Anthropic’s$300M acquisition of SDK startup Stainless, and the Google I/O announcement thatpromises to change search as we know it. Subscribe to Equity onYouTube,Apple Podcasts,Overcast,Spotifyand all the casts. You also can follow Equity onXandThreads, at @EquityPod.
View
Elon Musk can’t hear you over the sound of his $1.75 trillion IPO
The SpaceX S-1 is finally here, and the story it tells goes way further than rockets. The filing runs to 36 pages of risk factors alone, and the numbers inside match the ambition: a $28 trillion total addressable market, a pay package tied to establishing a Mars colony, and a valuation target that would make it the largest IPO in American history. On this episode of TechCrunch’sEquitypodcast, Kirsten Korosec, Anthony Ha, and Sean O’Kane dig into what the filing actually says, what it leaves out, and whether any of this math connects to reality. Listen to the full episode to hear about: Subscribe to Equity onYouTube,Apple Podcasts,Overcast,Spotifyand all the casts. You also can follow Equity onXandThreads, at @EquityPod.
View
How VCs and founders use inflated ‘ARR’ to crown AI startups
Last month, Scott Stevenson, co-founder and CEO of the legal AI startup Spellbook, took to X in an effort to expose what he called a “huge scam” among AI startups: inflation of the revenue figures that they announce publicly. “The reason many AI startups are crushing revenue records is because they are using a dishonest metric. The biggest funds in the world are supporting this and misleading journalists for PR coverage,” he wrote in his tweet. Stevenson isn’t the first to claim that annual recurring revenue (ARR) — a metric historically used to sum up annual revenue of active customers under contract — is being manipulated by some AI companies beyond recognition. Certain aspects of ARR shenanigans have been the subject of multipleother newsreportsandsocialmedia posts. However, Stevenson’s tweet seemed to have struck a particular nerve within the AI startup community, drawing over 200 reshares and comments fromhigh-profile investors, manyfounders, anda fewheadlines. “Scott at Spellbook did a great job of highlighting some of what you might describe as bad behavior on the part of some companies,” Jack Newton, co-founder and CEO of legal startup Clio, told TechCrunch, adding that the post brought much-needed awareness to the topic, referring to anexplanatory postfrom YC’s Garry Tan about proper revenue metrics. TechCrunch spoke with over a dozen founders, investors, and startup finance professionals to assess whether the ARR inflation is as pervasive as Stevenson suggests. Indeed, our sources, many of whom spoke on the condition of anonymity, confirmed that fudged ARR in public declarations is a common occurrence among startups, and how, in many cases, investors are aware of the exaggerations. The main obfuscation tactic is substituting “contracted ARR,” sometimes referred to as “committed ARR” (CARR), and simply calling it ARR. “For sure they are reporting CARR” as ARR, one investor said. “When one startup does it in a category, it is hard not to do it yourself just to keep up.” ARR is a metric established and trusted since the cloud era to indicate total sales of products where usage, and therefore payments, is metered out over time. Accountants don’t formally audit or sign off on ARR primarily because generally accepted accounting principles (GAAP) focus on historical, already-collected revenue, rather than future revenue. ARR was intended to show the total value of signed-and-sealed sales, typically multiyear contracts. (Today, this concept tends to go by another name: remaining performance obligations.) Meanwhile, the term “revenue” is typically reserved for money already collected. CARR is supposed to be another way to track growth. But it’s a much squishier metric than ARR because it counts revenue from signed customers that aren’t onboarded yet. One VC told TechCrunch that he has seen companies where CARR is 70% higher than ARR, even though a significant chunk of that contracted revenue will never actually materialize. CARR “builds on the ARR concept by adding committed but not yet live contract values to total ARR,” Bessemer Venture Partners (BVP)wrote in a blog postback in 2021. Critically, though, BVP says, the startup is supposed to adjust CARR to take into account expected customer churn (how many customers leave) and “downsell” (those who decide to buy less). The main problem with CARR is counting revenue before a startup’s product is implemented. If implementation is lengthy or goes awry, clients might cancel during the trial before all — or any — of the contracted revenue has been collected. Several investors told TechCrunch that they directly know of at least one high-profile enterprise startup that reported it surpassed $100 million in ARR, when only a fraction of that revenue came from currently paying customers. The rest was from contracts that hadn’t been deployed yet and in some cases may take a long time to implement the technology. One former employee at a startup that routinely reported CARR as ARR told TechCrunch that the company counted at least one substantial, yearlong free pilot as ARR. The company’s board, including a VC from a large fund, was aware that the revenue from the eventual paying part of the contract had been counted in ARR during the lengthy pilot program, the person said. The board was also aware that the customer could cancel before paying the full contract amount. The obvious problem with using CARR and calling it ARR is that it is far more susceptible to being “gamed” than traditional ARR. If a startup doesn’t account realistically for churn and downsell, CARR could be inflated. For instance, a startup could offer big discounts for the first two years of a three-year contract and count the whole three years as CARR (or ARR), even though customers may not stick around to pay the higher prices in year three. “I think Scott [Stevenson] is right. I’ve heard all sorts of anecdotes as well,” Ross McNairn, co-founder and CEO of legal AI startup Wordsmith told TechCrunch about ARR misrepresentations. “I speak to VCs all the time. They’re like, ‘There are some choppy, choppy standards out.’” Most cases are slightly less extreme. For instance, an employee at another startup described a discrepancy where marketing materials claimed $50 million in ARR, while the actual figure was $42 million. However, this person claimed that investors had access to the company’s books, which accurately reflected the lower amount. The source said some startups and their investors are comfortable playing fast and loose with their public metrics in part because AI startups are growing so quickly that an $8 million gap is viewed as a rounding error they’ll grow into quickly. There’s another issue surrounding all those public ARR declarations. Sometimes founders use another measurement with the same “ARR” acronym and a similar name: annualized run-rate revenue. This ARR is also controversial because it extrapolates current revenue over the next 12 months based on a given period’s haul (e.g., a quarter, month, week, or even a day). Since many AI companies charge based on usage or outcomes, that method of calculating annualized run-rate ARR can be misleading because revenue is no longer locked into predictable contracts. Most people interviewed for this story said that ARR overstatements of all kinds are hardly a novel phenomenon, but startups have become far more aggressive amid the AI hype. “The valuations have gotten higher, and so the incentives are stronger to do it,” Michael Marks, a founding managing partner at Celesta Capital, told TechCrunch. In the age of AI, startups are expected to grow much faster than ever before. “Going from 1 to 3 to 9 to 27 is not interesting,” Hemant Taneja, CEO and managing director of General Catalyst, said on the20VC podcastlast September, referring to the millions in ARR a startup is traditionally projected to hit each year. “You got to go like 1 to 20 to 100.” The pressure to show rapid growth is prompting some VCs to support, or at least overlook, startups presenting inflated ARR figures to the public. “There are definitely VCs in on this because they’re incentivized to create a narrative that they have runaway winners. They’re incentivized to get press coverage for their companies,” Stevenson told TechCrunch. Newton, whose legal AI startup Clio was valued at$5 billionlast fall, also alleges that VCs are often aware but silent about ARR misrepresentations. “We see some investors looking the other way when their own companies are inflating numbers because it makes them look good from the outside in,” he told TechCrunch. Other investors who spoke with TechCrunch say there is no reason for VCs to expose the overstatements. By turning a blind eye to public pronouncements of inflated ARR, VCs are effectivelyhelping to kingmaketheir own portfolio companies. When a startup publicly reports high revenue, it is more likely to attract the best talent and customers who believe the company is the undisputed winner in its category. “Investors can’t call it out,” a VC told TechCrunch. “Everyone has a company monetizing CARR as ARR.” Still, anyone intimately familiar with the industry’s intricacies has a hard time believing that some of these startups actually reached $100 million in ARR within a few years of launch. “To everyone who’s inside, it just feels fake,” said Alex Cohen, co-founder and CEO of health AI startup Hello Patient. “You read the headlines and you’re like, ‘I don’t believe it.’” However, not all startups feel comfortable representing growth by reporting CARR instead of ARR. They prefer to be clean and clear about their numbers in part because they understand that public markets measure software companies on ARR rather than CARR. These founders prioritize transparency. Wordsmith’s McNairn, who remembers the struggle startups faced justifying high valuations after the 2022 market correction, said he doesn’t want to create an even higher hurdle by exaggerating his startup’s revenue. “I think it is short-sighted, and I think that when you do things like that for a short-term gain, you’re overinflating already crazy high multiples,” he said. “I think it’s super bad hygiene, and it’s going to come back and bite you.”
View
Google goes for the glitter with disco-ball icons: ‘Are y’all sure you still want this?’
So bad, it’s good? Google on Friday joined in the disco ball icon fun taking place on home screens everywhere. After Spotify’stemporarynew disco ball app icon, released to celebrate the company’s 20th anniversary, drewextensive online backlash(and a bit of praise for those who like a little kitsch!), Google decided to get in on the joke and rolled out a custom set of Android app icons sporting a similar disco ball theme. On X, Android ecosystem head Sameer Samat posted, “Your wish is our command. Disco icons available on Pixel as of today…Are y’all sure you still want this?” Your wish is our command. Disco icons available on Pixel as of today.… Are y'all sure you still want this ?? 😅@DurvidImel@RaceJohnsonhttps://t.co/S9dwLZRtHlpic.twitter.com/nvevL7fTSb His post included a screenshot of a Pixel phone fully decked out with sparkly, disco-ball-inspired icons, which looks just as terrible (incredible??) as it sounds. The new icons are available through Pixel’s relatively new custom icons feature, which allows users to choose from different AI-generated styles for their app icons. Before this, users could only customize their icons by changing their colors to match the phone’s wallpaper and theme. The custom icons feature rolled out inMarch’s Pixel Drop— Google’s term for its periodic feature updates to Pixel phones — introducing app icon templates like a hand-drawn “Scribbles” aesthetic, a gold look called “Treasure,” a colorful, painted style dubbed “Easel,” and others. Earlier this week, Samat had jokinglytweeted, “Should we make this icon pack happen on Android?” alongside a Chrome icon turned into a disco ball. Should we make this icon pack happen on Android??@RaceJohnson😅https://t.co/Xbd5xlIVzhpic.twitter.com/nZDhAlGHfL Silly as it may be, Google actually made it happen. Many people had complained about the Spotify icon, calling it ugly, prompting the company toremindthem it was just a temporary sitch. “Alright, we know glitter is not for everyone,” the streamer wrote. Google, seemingly, disagrees. As off-brand as its disco-themed icons are, there’s also something whimsical about turning your whole homescreen into a sparkly landscape of little apps. (And, in case you missed it,the Zillennials are really into whimsyright now, The New York Times reports, describing their “playful response to a difficult world.”) Upon seeing Google’s release, X user and former Pixly co-founder Race Johsnonquipped, “When your home screen gets bottle service.”Said another, “Omg it’s awful. I’ll take it!”
View
We tried Google’s AI glasses and they’re almost there
At the Google I/O developer conference this week, we had the opportunity for a brief hands-on with Google’s upcoming AI-powered glasses — not theaudio-only glassesthat the company said will begin shipping this fall but rather the glasses that offer a combined audio and visual experience. Firstannouncedat last year’s event, these Android XR glasses offer an in-lens display that puts helpful information in front of you, overlaid on top of the real world. This includes widgets that could display things like the weather, walking directions, Uber pickup details, live translation, and more — even widgets you designed yourself using AI. The glasses will also pair with iOS and Android phones, the company noted, both in the audio-only format and in the future display version. The eyewear with the display is meant to be the next step beyond the first generation of audio glasses coming out later this year. The glasses were developed in partnership with Warby Parker, Gentle Monster, and Samsung, blending Google’s technology with their brands’ design aesthetics. The glasses we tested, meanwhile, were still very much a prototype, although one polished enough to now be tested externally. The reps demoing the XR glasses explained that the prototype allowed Google to not worry about some of the cosmetic details related to different styles and shapes, so it could instead focus on experimenting with the display technology more freely and its impacts on battery life. That means these spectacles are very different from any future shipping version of glasses, in terms of fit, shape, dimensions, and attention to detail. Rather, it’s more like being able to experiment with the “insides” of the glasses, while still in a basic, comfortable frame. The shipping version of the glasses will be able to detect when the glasses are placed on your head and taken off, but the ones we tried didn’t have this feature. To activate Gemini, you perform a two-second press on the right side of the glasses’ frame. A startup chime sounds, letting you know that Gemini is on and listening. In the demo version, starting Gemini also starts the camera at the same time, but the shipping version will allow the user to configure whether they want to turn on the camera when Gemini starts. In an initial test, we played music via the glasses by asking Gemini to play a favorite artist. The venue was too noisy to evaluate the sound quality, unfortunately, as the music was dialed up to the maximum volume and was still relatively hard to hear crisply and in detail. But the initial impression from this limited experience was that the glasses would not be a great substitute for higher-quality earbuds, though they would do if you just wanted some music while you were outside, walking, hiking, or doing chores around the house. The advantage of not having earbuds in is that you can more easily hear someone talking, compared with the transparency mode experiences on devices like Apple’s AirPods. To turn the music off, tap once on the side of the frame, around the middle, as if tapping on your temple. In the second test, we pressed the photo capture button to take a photo of a person. The display was off, so the picture was transferred to our phone and watch. (You’ll later be able to capture video with a long press, but this option was not available to test with the prototype. In the case of video, you would see a video thumbnail preview instead of a photo.) You can also simply ask Gemini to take a photo without having to press the photo button, and perform some sort of AI manipulation on the result. For instance, you can say something like, “Take a photo and turn the person into an anime character.” The photo is sent to the phone, then to the Gemini and Nano Banana servers, and then returns in its edited version. At the Google I/O venue, where Wi-Fi was under a heavy load, the round-trip took around 45 seconds. With the display enabled, you’ll see a simple home screen appear in your field of view. The demo version had some widgets preloaded that showed the weather and a countdown to Google’s I/O event. You could also build quick launchers into specific apps, like Google Maps or Translate, if those were among your main use cases for the glasses. The prototype had just one display over the right eye, but the platform can support both single and dual displays, as well as audio-only glasses. The image itself was a little fuzzy, but we chalked this up to our prescription contacts, which involve wearing one lens optimized for distance on one side, and one optimized for near vision on the other. When we closed one eye, the image came into better focus, but the experience almost immediately left us with some eye strain above the right eye, and it’s unclear if the prescription was entirely to blame. One of the best demos was of the language translation experience on the glasses, which is backed by the Google Translate app on the phone. One of the demonstrators spoke rapid Spanish, and the glasses automatically detected the language and showed the text in English on the display, while Gemini spoke English in our ear. We could see world travelers buying the glasses for this experience alone. We should note that Translate will work on the audio-only glasses, too, just without the text being displayed on the glasses. Instead, you could see the transcription on the phone, if needed, in addition to the real-time audio feedback. Another demo involved using the glasses to navigate. While obviously we couldn’t go out on a walk and leave the venue to test its accuracy, we could get an idea of how it works. You can start the Google Maps experience by asking Gemini to navigate to a destination — which can even be as vague as something like “the nearest coffee shop.” Gemini will activate Google Maps on the phone, and after a brief delay while the experience loads, the glasses will display turn-by-turn directions. When you are looking forward, your next turn information will be displayed. But if you need to get oriented in space, look down at the ground to see your blue dot on a map. You can also turn to the left and right to rotate in space, just like you would try to get the blue dot to point the right way on your phone. Then when you look up again, you can keep walking without the map being in your way. Because the experience is tied to Google Maps on your phone, saved destinations like “home” and “work” will already be available. We were also able to briefly use the glasses to identify a variety of objects in our view and ask questions about them. The glasses initially struggled to identify the replica of a Monet painting on a shelf in front of us, but that’s because the prototype didn’t automatically enable the camera — it had to be turned on again from the app. Still, it took a couple of questions before Gemini said that it looked like a Monet even after we moved in closer to focus on the Monet signature in the bottom left. Other tests were smoother, as the glasses immediately identified the plant on the shelf and answered questions about different recipes in a book. Still, these were things you could do today with Google Lens (or other AI models integrated in chatbot apps), though we suppose it’s interesting to be able to do them without having to pull your phone out. Google says it will have more to share about its Android XR display glasses later this year, when it expands its trusted tester program. In the meantime, the company believes that audio will suffice for some users’ needs, which is perhaps a smart way to spin the fact that it doesn’t have its display glasses ready, despite the competition from Meta and Snap on this front. Like the display version, the audio glasses also provide access to Google’s Gemini AI, which you hear privately through the glasses’ frame speakers. You can do things like listen to music through the glasses, press a button to take a photo, make a call, or tap into your phone apps from these glasses, as you can on the future display versions. Tapping into other third-party apps wasn’t among the items we demoed, but the glasses will allow users to tell Gemini to do things like “take the ingredients from this recipe and add them to my shopping list.” In another example that Google showcased during the event’s keynote, the glasses could see a meal that the wearer was cooking on the stove and offer feedback about the meal, like whether the meat was fully done yet.
View