Latest AI News
Glean’s top line crosses $300M as AI budget-cutting becomes its major selling point
Glean, a company often described as the Google for enterprise, said it has reached $300 million in annual recurring revenue (ARR), a three-fold increase from the $100 million milestone it reached just 15 months ago. While many AI startups are growing at a blistering pace, Glean’s progress is particularly remarkable. After years of essentially being the only player in the category, the seven-year-old startup is accelerating its growth as tech giants enter the enterprise AI search market with rival products. “The first four or five years of our existence, we had no competition,” Glean CEO Arvind Jain told TechCrunch. “Given how important search is to make AI work in the enterprise, every single company in the world wants to be in this space.” Tech heavyweights building Glean-like tools include Google, Microsoft, OpenAI, Anthropic, Salesforce, and Atlassian. Jain maintains that there’s value in being a first mover in the space, but that it’s also equally important to offer a better product. What Glean does better than its competition, according to Jain, comes down to the deep understanding that its AI tools have of customers’ business needs. Glean’s AI achievesthis knowledge— a concept captured by the new, popular term “context graph” — by connecting to and learning from enterprises’ internal software systems. Jain claims that Glean’s context graph also helps enterprises cut AI computing costs. “If you connect your AI to Glean, it gives you all the information that you need to do your work, and that results in AI consuming far fewer tokens compared to if you unleash AI onto your systems directly,” Jain said. That’s because with Glean, AI ends up performing fewer operations, he added. At a time when many companies are blowing through their AI budgets, those token cost savings have become a major selling point for the company. “One of the things you know our customers really like about Glean is the fact that we can reduce your AI bill significantly,” he said. The company, which was last valued at $7.2 billion when it raised a $150 million Series F last June, offers various pricing structures to its customers, which include Databricks, Reddit, Pinterest, and Samsung. According to Jain, Glean offers both a consumption-based model, where clients pay per use, and a hybrid model that combines a fixed monthly fee for active users with separate usage fees for model consumption. Glean is definitely not the first company to do this, but it’s worth pointing out that the company’s $300 million milestone cannot be fully described as traditional ARR, because a consumption model by definition doesn’t have a strictly recurring component. Pure consumption pricing models depend on fluctuating user activity rather than predictable subscription renewals, therefore a portion of Glean’s topline is more accurately described as anannualized revenue run rate. Glean did not immediately respond to a request for comment; this post will be updated if the company replies.
View
Just like gold and oil, we’ll soon be able to trade AI token futures
The most important market of the future could be in LLM tokens — and financial groups are rushing to build new infrastructure for them. China’s Shanghai Futures Exchange is currently designing a derivatives market for AI tokens, Reutersreports. The news comes as major derivatives exchangeCME Groupand theIntercontinental Exchange(the owner of the NYSE) have separately said they’re working on launching futures contracts for renting GPUs. GPU markets are still maturing, but given the wide range of companies using, selling, and renting GPUs, there’s a robust market for spot prices on GPU rental, typically charged by the hour. According to data fromAI Mining Co., which tracks daily GPU rental pricing across 28 marketplaces and cloud providers, median prices for Nvidia H100 GPUs ranged from $1.40 to $4.27 per hour across 13 marketplaces, while the average price for H200 GPUs were between $2.34 and $5 per hour across 10 marketplaces. And just over the past seven days, average H100 prices ranged from $2.79 to $3.33. But while mature markets exist for GPUs, there’s less infrastructure around tokens themselves — the fundamental building blocks of contemporary AI models. Enterprise plans for major AI companies are commonly denominated in tokens: OpenAI, for example, charges $5 per million input tokens, and $30 per million output tokens if you want to use the API for its latest GPT-5.5 model. Even cloud providers are increasingly offering the opportunity to charge per token, as inAmazon’s Bedrock system. The effort comes amid an unprecedented buildout of AI infrastructure. Cloud service providers, private equity firms, and infrastructure players alike have poured hundreds of billions into building data centers, anticipating that demand for GPUs and compute will continue to rise. An emerging crop of globalneocloud companiesis also vying for a piece of this demand. Some of these new entrants are specializing, focusing on inference, while others are competing with cloud giants like Oracle, AWS, and Google Cloud to offer their services to AI companies. By targeting AI tokens, the Shanghai exchange’s derivative product would be tied to how AI companies price their services, giving businesses, investors, and data center operators a way to hedge against the cost of compute.
View
Anthropic raises $65 billion, nears $1T valuation ahead of IPO
Anthropic has snagged$65 billion in fundingat a $965 billion post-money valuation in its latest funding round, marking what could be the AI startup’s last private fundraising before debuting on the public markets. The Series H round was co-led by Altimeter Capital, Dragoneer, Greenoaks, Sequoia Capital, Capital Group, Coatue, D1 Capital Partners, and others. Institutional investors including Baillie Gifford, Blackstone, Brookfield, D.E. Shaw Ventures, DST Global, and Fidelity Management & Research participated in the round. Strategic infrastructure partners, including Samsung, SK Hynix, and Micron, also joined the round. A portion of the round — $15 billion — is also made up of previously committed investments from hyperscalers, including$5 billion from Amazonannounced in April. TechCrunchreported last monththat Anthropic was close to closing a $50 billion round, with investors clamoring to get on the cap table. One institutional investor had even pledged as much as $5 billion just to get a meeting with Anthropic CFO Krishna Rao. Anthropic plans to use the new funds to “advance our safety and interpretability research, expand compute to meet growing demand for Claude, and scale the products and partnerships our customers rely on.” The round comes the same day that Anthropic released its newClaude Opus 4.8 model,which touts better capabilities in agentic tasks, advanced coding, and focus on honesty and self-correction. The AI startup is alsoreportedlyplanning to more widely launch models that are on par with its powerful cybersecurity model Mythos, which it has only released in limited fashion due to potential safety concerns. The company has seen increased growth since its last funding round, particularly among enterprise customers that rely on Claude Code. The company said its run rate revenue crossed $47 billion earlier this month, andThe Wall Street Journal recently reportedthat the startup expects a 130% revenue surge to bring it to its first operating profit. “Claude’s latest advancements have driven large-scale adoption among the world’s most demanding organizations. This momentum positions Anthropic to lead the next phase of AI innovation and capture the enormous opportunity ahead,” said Brad Gerstner, founder and CEO of Altimeter Capital. Anthropic has been in tight competition with OpenAI for fundraising and user growth in advance of their respective IPOs. OpenAI last raised awhopping $122 billion roundin March at an $852 billion post-money valuation. Elon Musk’s SpaceX — whichmerged with xAIearlier this year — is targeting a $2 trillion valuation in itspending IPO, and seeking to raise more than $75 billion.
View
Asana acquires no-code agent-builder StackAI
Asana has acquired the workflow automation company StackAI for $75 million, part of a larger effort to position itself as an AI-native workplace platform. StackAI’s founders, Tony Rosinol and Bernard Aceituno, will join Asana as part of the acquisition. Asana framed the acquisition as part of its broader AI pivot, in which it seeks to build its platform into “the operating system for human-agent teams.” The announcement was announced Thursday afternoon to coincide with Asana’s earnings and investor call. Builtas an AI workflow-automation system, StackAI designs agents to operate within existing business systems, pulling in data from systems like Salesforce, Slack, and Gsuite. Part of Y Combinator’s Winter ’23 cohort, the company has faced fierce competition from automation tools like Zapier as well as AI labs like OpenAI and Anthropic. StackAI had raised just under $20 million, according to PitchBook data, with most of it coming in a recent $16 million Series A round.That roundincluded funding from Gradient, Epaklon Capital, Lobby VC, LifeX Ventures, and Vercel CEO Guillermo Rauch. While users are likely most familiar with Asana’s work management system, the company has released a number of AI-oriented products in recent years, most notably the AI Studio agent builder andAI Teammatesseries of pre-built automations. While equivalent tools are available from major labs, Asana sees its deep integration into existing corporate workflows as a key advantage, allowing it to distill context and training data that would otherwise be unavailable. Asana has struggled on public markets during the AI era, losing more than half its market cap value since the introduction of ChatGPT — a spiral that grew worse with the departure of founder Dustin Moskovitz as CEO last March. But revenue has continued to grow steadily, and the new leadership is confident that its human-agent products will enable it to rebound. “This acquisition accelerates our roadmap and takes us into the next phase of human-agent work,” said CEO Dan Rogers in a statement. “We’re already seeing real momentum with AI Teammates and AI Studio … StackAI now lets them go further, agentifying the most complex business processes end-to-end.”
View
The internet is being rebuilt for machines
Cloud infrastructure has long been designed around humans who search, click, scroll, and stream in a steady and predictable fashion. AI agents behave differently. They can unleash a swell of activity, spinning up multiple sub-agents that query hundreds of databases, search documents, and call APIs in seconds and then disappear as quickly as they arrived. Under that premise, Amazon is redesigning a core piece of its cloud infrastructure. On Thursday, AWSlaunched its next generation of OpenSearch Serverless, a fully managed search and vector database — essentially a system for storing and retrieving information at scale — that’s designed specifically for agentic workloads. AWS says the new system can instantly scale up when agents trigger tasks and scale back down to zero when idle. The launch reflects a growing realization across the tech industry: infrastructure originally designed for a human-driven internet doesn’t work as well in a world increasingly populated by agents. While AI agents still represent a relatively small portion of internet activity, machine-generated traffic is already significant, and poised to grow. Cloudflare says bots accounted for 31% of overall HTTP traffic over the last six months. AI crawlers, search engines, and assistants made up roughly a quarter of all bot requests during that period. “Non-human traffic will exceed human traffic sometime in the first half of 2027,” saidLi Yi Ohlsen, senior product manager at Cloudflare, to TechCrunch. At Google’s I/O developer conference last week, the company said users will be able to startdelegating tasksto AI systems, like researching purchases, booking travel, browsing the web, and interacting with apps. But the buck doesn’t stop at consumer-focused AI agents. Enterprises are increasingly deploying agents internally and for their customers, creating new kinds of machine-generated traffic behind the scenes. As a result, cloud providers and infrastructure companies have been reckoning with how to adapt systems built for humans to a world of agents that are constantly and autonomously retrieving information, invoking tools, and generating machine-to-machine traffic. That’s where AWS’s new OpenSearch Serverless comes in. “The timing is straightforward. Agents are moving from experimentation into production, and they create traffic patterns that previous infrastructure simply wasn’t designed for,” Tia White, general manager for Amazon OpenSearch Service, told TechCrunch. “They spike without warning, they go idle without notice, and enterprise needs search that keeps up without paying for empty or idle compute.” The key technical change with this new generation is that it decouples compute from storage, allowing compute to scale up in seconds to accommodate agent traffic bursts and to scale down to zero, so customers pay $0 when agents are idle. “Previously, even in our prior Serverless version, you had to have at least one instance operational and running because storage and compute were coupled,” White said. “You couldn’t just automatically spin up [compute] at the rate you needed to, so you always had idle compute reserved for your workload, whether you were using it or not.” Think of it like always paying for a parking space, even when you’re not using it. With AWS’s upgraded Serverless, it’s more like paying for a metered parking spot. At launch, OpenSearch Serverless will integrate natively with AI development platforms like Vercel and Kiro, so developers can deploy production-ready search and vector backends for agents without managing infrastructure. The shift is emerging across the cloud industry. Databricks andSnowflakeare repositioning themselves as AI memory and retrieval systems for enterprise data. Microsoft has rolled outupdates to Azuredesigned to handle AI agent bursts and share memory between agents. Cloudflare, in a similar vein to Amazon,last month introducedinfrastructure aimed at giving agents persistent environments and instant scalability. The more companies deploy AI agents, the more pressure there will be to redesign infrastructure around machine-generated workloads, which in turn could make agents cheaper and easier to deploy at larger scales. Loading the player…
View
2 days left: Lock in ticket savings of up to $410 to TechCrunch Disrupt 2026
Early Bird pricing ends tomorrow, May 29, at 11:59 p.m. PT. After that, prices forTechCrunch Disrupt 2026go up. Miss this, and you’ll be paying more for the same access to one of the most anticipated tech epicenters of the year.Register nowto secure discounts of up to $410 on your pass, or up to 30% ongroup passes. If you want to raise capital, hire top talent, launch your startup, or discover your next portfolio company, missingDisruptfrom October 13–15 at San Francisco’s Moscone West is not an option. Here’s what you’ll gain by attending: Founder Pass: Accelerate growth with the right insights, tools, and connections. Meet investors aligned with your startup. Investor Pass: Discover standout startups and expand your portfolio with curated access. Use matchmaking tools to make every conversation count. This window to the lowest ticket rates of the year is closing at 11:59 p.m. PT tomorrow, May 29.Register nowto secure your ticket with up to a $410 discount. Or save up to 30% withcommunity passesof 4+.
View
Visa invests in Replit to power agentic payments for developers
Visa has announced an undisclosed investment in AI coding platform Replit. The two companies are also exploring how to integrate Visa’s payment products into Replit, so that developers — and the AI agents they build — can accept payments directly from customers without leaving the platform. Visa added that more than 1,000 of its employees have been using Replit for prototyping and development. As part of the partnership, the companies are exploring how developers on Replit can use Visa’s suite for AI-powered payments, called Visa Intelligent Commerce, as well as Visa’s Trusted Agent Protocol — a system that allows AI agents to securely identify themselves by sharing information like their intent and relevant customer details, so that payments made by agents can be verified and trusted. All of these projects are in an exploratory stage, and the companies haven’t formally announced any joint products. The investment reflects a broader race to establish the infrastructure for so-called agentic payments — a world in which AI agents buy and sell things on users’ behalf. Besides Replit and Visa, other tech companies are also moving quickly in this space. Retail investing platform Robinhood now wants people touse agents to trade, while Google wants users to deployagents for shopping. “Over the last few months, our enterprise traction has been growing, and Visa coming on board underscores our mission of making coding available to anyone in a secure and robust manner,” Amjad Masad, CEO and founder of Replit, said in a statement. Replit is also launching self-serve enterprise access, allowing companies to sign contracts worth up to $200,000 without talking to a salesperson. The tier offers enterprise-grade compliance and controls, including SSO — single sign-on, a system that lets employees access multiple tools with one set of credentials — audit logs, and advanced permissions. “Our continued customer and partner additions in the enterprise, coupled with our new self-serve program, bring us closer to a world where any team can go from idea to production-ready software quickly and securely,” Masad added. As demand for so-called vibe-coding platforms has shot up, valuations of startups like Replit, Cursor, and Lovable have risen rapidly, along with investor interest. In September of last year, Replit hit the$3 billionvaluation mark. Six months later, in March, the company raised $400 million in a Series D led by Georgian Partners at a$9 billionvaluation — tripling its valuation in under six months. In May at TechCrunch’sStrictlyVC event in San Francisco, Masad said that Replit’s churn is very low, and customers are sticking around. “Churn is very, very low, and net retention is incredibly high — 300% in some cases. What we actually hear from customers is that when engineers get nervous and try to rebuild an app into their own stack, they often make it worse. Once enterprises get comfortable with the full Replit stack — especially when we set up a single-tenant environment for them — they keep the apps on Replit,” he said.
View
YouTube adds new podcast features, including an AI recommendation tool and ‘Auto speed’
YouTubeannouncedon Thursday it’s introducing new podcast features for Premium users, including an AI-powered recommendation tool, an “Auto speed” setting, and a new on-the-go listening mode. The update signals YouTube’s ongoing efforts to compete with other platforms for podcast audiences, especially as streaming giantNetflixis investing heavily in video podcasts. Additionally, by focusing on personalized discovery and hands-free listening features, the company also appears to be targeting users who consume podcasts on audio-first apps like Spotify and Apple Podcasts. The Google-owned platform’s “Ask Music” feature already lets Premium users generate personalized radio stations and playlists, and now users will be able to get podcast recommendations based on genres, their current mood, or shows they already enjoy. Users will also get access to a new “Auto speed” feature designed to make listening more efficient by intelligently adjusting playback speed during slower speech or information-dense segments, creating a more streamlined experience without sacrificing comprehension. While you can already adjust your playback speed, it can be inconsistent if the hosts are speaking at different speeds or changing their tone. With this new feature, listeners will be able to listen to content at a speed that addresses these changes throughout the conversation. The new on-the-go mode gives Premium users access to listener-friendly controls designed for activities like running, commuting, or multitasking. Users will get access to quick controls like skipping forward or backward, or jumping to the next episode. YouTube says the feature is designed to make it easier to get the most out of background playback. Auto speed and on-the-go mode are now available for Premium users on Android and are coming to iOS in the coming months. YouTube says Premium users watched over 800 million hours of podcasts in April 2026, and that YouTube Podcasts has over 1 billion monthly active users. With these new features, the company is looking to both retain and grow these numbers.
View
At TechCrunch Disrupt 2026: Databricks’ co-founder on what kills enterprise AI deals
Enterprise organizations are not rejecting AI. They are rejecting operational instability. That is the shift many founders still misunderstand — and it is becoming one of the defining realities separating enterprise AI companies that scale from the ones that stall after early momentum. For the last several years, AI startups benefited from a market driven by experimentation. A strong demo, an impressive model, and a powerful vision were often enough to generate enterprise interest, pilot programs, and investor enthusiasm. But enterprise AI is entering a different phase now, one where enterprises are no longer evaluating whether AI is exciting. They are evaluating whether it is safe to deploy broadly. AtTechCrunch Disrupt 2026,taking place October 13–15 at Moscone West in San Francisco,Arsalan Tavakoli-Shiraji, co-founder and SVP of field engineering at Databricks, will unpack that shift during his AI Stage session, “The Enterprise Isn’t Broken. Your Assumptions About It Are.” Disrupt will bring together 10,000+ founders, investors, and operators to explore the technologies and operational pressures changing how companies are built and scaled. The three-day event will feature 250+ sessions across six stages, led by tech leaders directing the industry today. Explore the sessions appearing on the Disrupt AI Stage.Ticket savings of up to $410 end on May 29 at 11:59 p.m. PT.Register here. The enterprise AI market is full of successful pilots that never became real deployments. Not because the technology failed. But because the organization could not absorb the operational consequences of adopting it. Now the reality founders need to face is that startup AI deals rarely die because the model underperformed. They die because the enterprise lost confidence in what the deployment would require. That is the gap Tavakoli-Shiraji’s session is designed to explore. Most enterprises are not simply evaluating whether an AI product works. They are evaluating: An AI product can perform exceptionally well in a controlled environment and still fail commercially if its deployment creates instability within the business. That distinction is important to founders because many AI startups are still optimizing for the wrong outcome. They are building for initial excitement rather than long-term operational adoption. And enterprises are becoming far more disciplined about recognizing the difference. Register for Disrupt to hear how enterprise AI leaders evaluate what actually survives beyond the pilot phase.Lock in your ticket savings of up to $410when you register by May 29 at 11:59 p.m. PT. The AI startups gaining traction inside large organizations increasingly share one thing in common: They reduce uncertainty. They integrate more cleanly into existing systems. They create less workflow friction. They are easier to govern, easier to explain internally, and easier for organizations to trust over time. That sounds less exciting than breakthrough demos or model benchmarks. But it is quickly becoming the difference between AI startups that generate attention and those that generate durable revenue. The market is maturing. Enterprise buyers are asking different questions now: Those concerns are no longer secondary. In many organizations, they have become core to the buying decision itself. For AI founders selling into the enterprise, this session breaks down what actually drives adoption after the pilot phase ends.Check out the session detailsandget your $410 ticket savingsto learn what to prioritize to gain traction with enterprise AI deals. Tavakoli-Shiraji brings an unusually relevant perspective to this conversation because his background spans both enterprise strategy and deeply technical systems architecture. Before joiningDatabricks, he was an associate principal at McKinsey & Company, advising enterprises, technology vendors, and public-sector organizations on cloud computing, next-generation IT, and enterprise transformation strategy. He also earned a PhD in computer science from UC Berkeley, focused on networking and distributed systems. That lens is valuable to startups because enterprise AI success increasingly depends on more than strong engineering alone. Founders now need to understand how technical systems interact with organizational behavior, infrastructure realities, procurement processes, governance concerns, and operational risk. The startups that succeed in enterprise AI over the next several years may not necessarily be the ones with the most advanced models. They may be the ones that best understand how enterprises actually absorb change. That is the kind of operational pressure that Tavakoli-Shiraji and other speakers on theAI Stage at Disruptwill explore. Presented by Google Cloud, the stage examines how AI agents and generative AI are reshaping SaaS, enterprise adoption, software economics, security, and operational infrastructure — including Tavakoli-Shiraji’s session on why enterprise AI success increasingly depends on operational trust rather than simply technical performance. Across the stage, founders will learn how and why the focus is shifting away from AI novelty and toward the real-world challenges of deploying, governing, and scaling AI systems inside real organizations. Explore the Disrupt agendaand learn how founders, investors, and enterprise operators are managing the next phase of AI adoption.Register by May 29 at 11:59 p.m. PT to save up to $410 on your passes.
View
RSI is the new AGI — and it’s just as hard to pin down
The word “recursion” is the latest buzzword in AI circles. Two separate startups have taken on the name, and many more have started referencing Recursive self-improvement (RSI) in their roadmaps. Like AGI before it, RSI has become a three-letter byword for a cataclysmic AI takeoff – even if there’s still a little disagreement about exactly what it means. In basic terms, RSI refers to an AI system that can continuously upgrade itself. Once AI systems can manage the upgrade cycle better than humans, the process can become a closed loop, limited only by the compute power they can access, and humans no longer necessary or even helpful. Scary or not, that’s a vision that a lot of AI labs are eager to chase. Earlier this month, well-known AI researcher, Richard Socher, launched the aptly named Recursive Superintelligence launched with RSI as an explicit goal. “Our main focus is to build truly recursive, self-improving superintelligence at scale,” Sochertold TechCrunch at launch, “which means that the entire process of ideation, implementation, and validation of research ideas would be automatic.” A number of other prominent researchers are already chasing that same goal, hoping for a breakthrough that will make recursive self-improvement possible. One of the most prominent is Alex Karpathy, a legendary figure from Tesla and OpenAI, who is using agent swarms to train LLMs on simple tasks for a project he callsAuto-Research. Karpathy has been unusually open about the project,tweeting about milestones regularlyand making the building blocks available through a public GitHub repo. So far, the work has mostly been confined to making minor improvements on a GPT-2 scale model – as Karpathy noted in March, “It’s not novel, ground-breaking ‘research’ (yet)” – but it’s been enough to convince lots of other researchers to follow the RSI dream. And with Karpathy now workingon pre-training at Anthropic, he will have plenty of opportunity to apply the idea at a larger scale. Adaption – founded by Cohere and Google alum Sara Hooker – recentlylaunched a similar tool called AutoScientistin an effort to automate frontier training. Like Karpathy’s auto-researchers, the system trains agents to make incremental improvements – but for Adaption, the goal is to make it easier to train a full-scale frontier model. If those same researchers start to push the frontier forward, the system could quickly spiral into something very much like RSI. Disarray founder Doris Xin drew more specific RSI interest when her self-trained machine learning agenttook home 28 medals in a recent Kaggle competition, beating out many human-trained agents. As she sees it, the major challenge is reliability. “I would argue, given infinite compute and infinite time horizon, we are already there,” Xin told me. “I want to make an argument that this is not a creative endeavor, really. It’s just a lot of meat-and-potatoes engineering.” There’s also plenty of evidence that the AI industry isn’t very close to recursive systems in any meaningful way — and is still grappling with talking to a wary public about its progress. So Google CEO Sundar Pichai basically admitted ina recent podcast interview. “It’s a continuum, and we are all definitely making progress,” Pichai said. “But in the way people describe R.S.I., that would represent a next level of acceleration and would have a lot of implications, but we aren’t quite there yet.” But the continuum includes an awful lot of self-improving AI systems.In January,one of Anthropic’s lead programmers for Claude Code estimated that “close to 100%” of his team’s code was written by the tool – a frank admission that Claude Code was literally writing itself. Just because engineers are using an AI tool doesn’t mean the tool can replace them – but Anthropic seems to be getting close to replacing engineers too. In a recent surveytied to the Mythos preview, five out of 18 Anthropic engineers believed that, with harness improvements, this version of Mythos could soon substitute for an L4 engineer – a mid-level programmer who can take on involved projects without supervision. Still, there were some of the same weaknesses you might expect. “Some of Claude’s major reported weaknesses compared to an L4 include: self-managing week-long ambiguous tasks, understanding org priorities, taste, verification, instruction-following, and epistemics,” the report reads. In other words, its weaknesses are everything involved with self-direction, which is the cornerstone for RSI. But sure, for everything else, Claude is ready to step right in. Just like the AGI term before it, the AI industry also can’t tell us how far away it is from showcasing a meaningful recursive system. When Georgetown’s Center for Security and Emerging Technologyassembled a group of experts to study RSIlast year, the group found a major split in assessments – some expecting an imminent “superintelligence” style explosion while others expected slower progress and an eventual plateau. But all agreed that recursion made the future especially difficult to predict. Helen Toner, director of CSET and a former board member at OpenAI, told TechCrunch that simply using AI tools to do AI research isn’t enough to qualify as RSI. “They’re just using AI for as much as they can,” Toner tells TechCrunch. “And I think that is different from the classic definition of RSI, which is really that there are no humans needed.” Toner points toa recent post by METR’s Ayeja Cotra, which distinguishes different milestones on the path to the AI research takeover. One step, which Cotra calls “adequacy,” would come when the system can still perform research after all humans are removed – even if the resulting research isn’t as valuable or efficient. “Parity” comes when an AI-only system is as good at research as a human-only system. “Supremacy,” the final stage, comes when an AI-only system outperforms a collaborative system between humans and AI. Ultimately, Cotra concludes that AI is very close to the adequacy threshold of being able to produce some work on its own – similar to the incremental changes made by Karpathy’s Auto-Research system. “I wouldn’t be totally shocked if you told me this milestone had already passed, and I expect it to happen in the next couple years,” Cotra writes. She’s less clear on when parity will come, but once it does, she thinks it would “massively accelerate the pace of AI progress, leading to AI research supremacy within another year.” With so much of AI built on scaling laws, there’s a strong tendency to think RSI will follow the same curve. Toner thinks that many of those pursuing AI research and development via RSI “ think of it as a pretty smooth ladder, where you can just keep scaling up.” But even if AI researchers are able to make incremental improvements like Karpathy’s auto-researchers, there will be larger challenges in handing off the whole process of research. Toner puts it in terms of the history of computing, which sees human beings handing off more and more of the process while still directing things from the top. “We went from machine languages to assembly language and compiled languages; you’re getting further and further from the guts of the computer,” Toner says. “But the human is still, in some intuitive sense, running the show.” Moving beyond that paradigm will take significant challenges, both in engineering and alignment. But even with the massive investments happening, there’s no infinite compute available – and the basic tradeoff between human labor and machine intelligence will be hard to overcome. As for a total recursive AI system of apocalyptic visions? The only thing researchers essentially agree on is that, like AGI, it’s not here yet.
View
Sneak peek at new Siri app reveals Apple’s plans to take on ChatGPT and more
Just ahead of Apple’s Worldwide Developers Conference (WWDC) in June,Bloomberghas published leaked renders of what Apple’s planned AI upgrade could look like on iPhone — including a brand-new Siri app meant to rival ChatGPT and other AI chatbots — as well as how Siri’s new capabilities will be woven throughout the operating system. The images were produced by Bloomberg based on what it saw and learned from sources. While you’ll still be able to press a button in iOS 27 to trigger Siri, the animation and response will now emerge from the iPhone’sDynamic Island— that’s the black pill-shaped area at the top of the screen that today houses Live Activities, the real-time updates and interactive displays from apps that appear directly on the phone’s Home Screen. This mode will work best for quick voice queries or searches, much like how people use Siri currently. A new mode, however, will put Siri-powered search within easy reach, capitalizing on people’s muscle memory for swiping down on their screen to access Spotlight Search — a built-in way to find information from both your phone and the web in one place. The swipe-down gesture will still open search, but now those searches will draw on the AI-powered Siri, which includes a rebuilt AI model thatuses Google’s Gemini AI technology under the hoodfor added intelligence. From here, iPhone users can search, launch apps, start messages, ask about the weather, add calendar appointments, search their notes, and trigger app shortcuts, Bloomberg reports, with results displayed in formatted text in a card-style interface that also emerges from the Dynamic Island. Apple’s approach to AI is strikingly similar to its earliermulti-billion dollarpartnership with Google that made Google the default search engine on iPhone. Just as building a search engine from scratch was never in Apple’s playbook, AI presents a similar calculus — it’s too expensive and complex to go it alone, at least right now. So Apple is working with outside partners for AI technology that users want today, while simultaneously building out its own models,including local AI, that runs on local devices rather than the cloud — an approach that allows Apple to lean into it privacy brand without needing to catch up. Bloomberg also notes there will be a new standalone Siri app — as previously reported — designed to compete directly with chatbots like ChatGPT, Claude, Gemini, and others. The app will surface your past chat history and allow you to upload documents and photos, in addition to text. Scale, as ever, is Apple’s advantage. While ChatGPT now has900 millionweekly active users, Apple’s install base (all devices, not just iPhone) is2.5 billion— meaning the company has an unmatched runway to introduce AI to people who haven’t yet adopted standalone AI tools.
View
Sesame, the conversational AI startup from Oculus founders, launches its iOS app
On Thursday, the AI startupSesame, co-founded byOculus’s foundersand others from theVR company that sold to Meta, released a public preview of the conversational AI agents it’s been developing for over a year. With itsnew iOS app, Sesame is rethinking the traditional AI chatbot experience popularized by apps like ChatGPT, creating one where conversation flows, even if the AI needs time to think. As the company explains in its launchannouncement, “There’s an inherent tension between replying quickly and taking the time to compose thoughtful responses. A slower response is usually more correct, but it can also feel unnatural if it takes too long.” To address this challenge, Sesame claims to have built fast search and retrieval systems, so the AI can have up-to-date information, as well as technology that allows it to run multiple parallel searches while speaking, weaving those results into its responses as it talks. That means the AI will talk more like a human, even pivoting mid-sentence if need be, as it taps into newer information — as a human might when remembering another key fact or point they want to add. The app offers four distinct AI agents called Maya, Miles, Simone, and Charlie, each of which have their own distinct voice, personality, point of view, and memory. Maya and Miles werepreviously availablein Sesame’s Research Preview of its technology, where they were soon accessed by over a million people within the first few weeks,said Sesame investor Sequoiaat the time. (The company had then just raised its $250 million Series B from Sequoia and others, and was opening up a beta.) During the beta, Sesame learned from user feedback and rolled out features including search cards with image results for visualizing concepts, notes for capturing takeaways, a texting mode for those times when speaking aloud is not an option, and support for deep dives where you can get more in-depth results. There’s also a new incognito mode for private conversations, which allows the agents access to prior context, but saves nothing to memory. The app, however, is only the first step towards Sesame’s bigger plans for AI involving intelligent eyewear, which the team expects to launch in 2027. Before that, the agents will also learn to do more than just think with you, Sesame hints, suggesting they’ll later be able to take action on your behalf — hence why they’re called “agents” in the first place, instead of just chatbots. That is potentially even more interesting, as working with agentic tools or apps today requires being able to prompt for what you need and have a specific idea of what you want to happen, and sometimes, even how it should happen. A conversational agent that you could talk to naturally could help you take the next steps, without you having to perfect the command you’re giving it. TheiOS appis out today in 39 countries, and the full experience is free for the time being. However, there still may be a short waitlist at sign-up. AnAndroid previewis coming in the future, the company says.
View