Elon Musk is contemplating decommissioning SpaceX's Dragon spacecraft, responding to President Donald Trump's apparent intent to terminate government subsidies and contracts with the billionaire's companies. It looks like the feud between the former allies has quickly turned vicious.SpaceX's CEO initially announced that the company would retire its Dragon spacecraft in an X post on Thursday, with Musk sharing a screenshot of a post published on Trump's Truth Social account earlier in the day."The easiest way to save money in our Budget, Billions and Billions of Dollars, is to terminate Elon’s Governmental Subsidies and Contracts," said Trump in the screenshotted post. "I was always surprised that Biden didn’t do it!" You May Also Like "In light of the President’s statement about cancellation of my government contracts, @SpaceX will begin decommissioning its Dragon spacecraft immediately," Musk wrote on X.SpaceX's Dragon spacecraft are a family of vehicles designed to carry passengers and cargo. The National Aeronautics and Space Administration (NASA) has previously relied upon them to transport astronauts to and from the International Space Station (ISS). Mere hours prior to Musk's announcement, SpaceX posted on X that it was preparing to launch a Dragon next Tuesday.For a few hours, it seemed reasonable to assume that this launch would now not go ahead. However, Musk then appeared to quickly walk back his decision. Responding to an X user advising him to "cool off and take a step back for a couple days," the billionaire subsequently stated that Dragon will not be decommissioned after all.It's unclear whether Musk's initial announcement was sincere, or whether his apparent about-face might be sarcastic. Musk has a history of making flippant comments online with no apparent regard to their consequences. What is clear is that Musk and Trump's relationship is well past the honeymoon phase, and now looks much more like an ugly divorce.If Trump does terminate government contracts with Musk's companies, it would deal a significant blow to the billionaire. According to a Washington Post investigation, NASA has invested over $15 billion in SpaceX alone. When put together with Musk's other companies such as EV automaker Tesla, his various businesses have received at least $38 billion in government contracts, loans, subsidies, and tax credits. Mashable Trend Report: Coming Soon! Decode what’s viral, what’s next, and what it all means. Sign up for Mashable’s weekly Trend Report newsletter. By clicking Sign Me Up, you confirm you are 16+ and agree to our Terms of Use and Privacy Policy. Thanks for signing up! Musk and Trump go through messy public breakup Credit: Roberto Schmidt / AFP via Getty Images Musk's relationship with Trump has significantly deteriorated in recent days. The billionaire announced that he was leaving his position as de facto head of the Department of Government Efficiency (DOGE) last Wednesday, just one day after he criticised Trump's tax bill as undermining its work. The split was presented as amicable at the time, with Trump presenting Musk with a golden key and words of praise. However, their love affair has quickly turned sour.Musk continued to lambast Trump's bill after his departure from DOGE, arguing that it will increase government debt by trillions of dollars. Strongly disagreeing with the president's characterisation of the proposed legislation as a "Big Beautiful Bill," Musk labelled it a "disgusting abomination" and has been calling for lawmakers to crush it.For his part, Trump has claimed that Musk is simply throwing a tantrum because the bill supposedly cut an alleged "EV mandate." The president stated on Thursday that he had asked the billionaire to leave his administration, and that Musk had been "wearing thin.""I took away his EV Mandate that forced everyone to buy Electric Cars that nobody else wanted (that he knew for months I was going to do!), and he just went CRAZY!" Trump claimed. Despite Trump's assertions, he did not abolish any EV mandate as there has never been any U.S. law which makes switching to an electric car mandatory. However, Trump has taken several anti-EV measures since his inauguration, including abolishing incentives encouraging EV adoption, pausing $3 billion in funding for a U.S. charging network, and introducing a $250 annual fee for EV users in his recent tax bill. Related Stories Trump's claim about Musk is an interesting contrast to his statements in March, when he praised the billionaire for not complaining about the supposed end of the non-existent EV mandate. The president made the comments while he and Musk co-hosted a Tesla ad on the White House lawn in an effort to boost the company's cratering stock prices.Tesla's struggling share value has now fallen again amidst Musk's feud with Trump, plummeting more than 14 percent on Thursday to wipe out over $150 billion in value."I don’t mind Elon turning against me, but he should have done so months ago," Trump wrote on Thursday.Meanwhile, Musk went all-in attacking Trump on Thursday, claiming that the president is linked to child sex offender Jeffrey Epstein and sharing posts calling for him to be impeached. Musk has also hit out at Trump's tariffs on international trade, predicting that they will "cause a recession in the second half of the year.""Without me, Trump would have lost the election," Musk alleged on X. "Such ingratitude."

One of the key use cases for generative AI involves answering questions over private datasets, with retrieval-augmented generation (RAG) as the go-to framework. As new RAG techniques emerge, there’s a growing need to benchmark their performance across diverse datasets and metrics.  To meet this need, we’re introducing BenchmarkQED, a new suite of tools that automates RAG benchmarking at scale, available on GitHub (opens in new tab). It includes components for query generation, evaluation, and dataset preparation, each designed to support rigorous, reproducible testing.   BenchmarkQED complements the RAG methods in our open-source GraphRAG library, enabling users to run a GraphRAG-style evaluation across models, metrics, and datasets. GraphRAG uses a large language model (LLM) to generate and summarize entity-based knowledge graphs, producing more comprehensive and diverse answers than standard RAG for large-scale tasks.  In this post, we walk through the core components of BenchmarkQED that contribute to the overall benchmarking process. We also share some of the latest benchmark results comparing our LazyGraphRAG system to competing methods, including a vector-based RAG with a 1M-token context window, where the leading LazyGraphRAG configuration showed significant win rates across all combinations of quality metrics and query classes. In the paper, we distinguish between local queries, where answers are found in a small number of text regions, and sometimes even a single region, and global queries, which require reasoning over large portions of or even the entire dataset.  Conventional vector-based RAG excels at local queries because the regions containing the answer to the query resemble the query itself and can be retrieved as the nearest neighbor in the vector space of text embeddings. However, it struggles with global questions, such as, “What are the main themes of the dataset?” which require understanding dataset qualities not explicitly stated in the text.   AutoQ: Automated query synthesis This limitation motivated the development of GraphRAG a system designed to answer global queries. GraphRAG’s evaluation requirements subsequently led to the creation of AutoQ, a method for synthesizing these global queries for any dataset. AutoQ extends this approach by generating synthetic queries across the spectrum of queries, from local to global. It defines four distinct classes based on the source and scope of the query (Figure 1, top) forming a logical progression along the spectrum (Figure 1, bottom). Figure 1. Construction of a 2×2 design space for synthetic query generation with AutoQ, showing how the four resulting query classes map onto the local-global query spectrum.  AutoQ can be configured to generate any number and distribution of synthetic queries along these classes, enabling consistent benchmarking across datasets without requiring user customization. Figure 2 shows the synthesis process and sample queries from each class, using an AP News dataset. Figure 2. Synthesis process and example query for each of the four AutoQ query classes.  About Microsoft Research Advancing science and technology to benefit humanity View our story Opens in a new tab AutoE: Automated evaluation framework  Our evaluation of GraphRAG focused on analyzing key qualities of answers to global questions. The following qualities were used for the current evaluation: Comprehensiveness: Does the answer address all relevant aspects of the question?  Diversity: Does it present varied perspectives or insights?  Empowerment: Does it help the reader understand and make informed judgments?  Relevance: Does it address what the question is specifically asking?   The AutoE component scales evaluation of these qualities using the LLM-as-a-Judge method. It presents pairs of answers to an LLM, along with the query and target metric, in counterbalanced order. The model determines whether the first answer wins, loses, or ties with the second. Over a set of queries, whether from AutoQ or elsewhere, this produces win rates between competing methods. When ground truth is available, AutoE can also score answers on correctness, completeness, and related metrics. An illustrative evaluation is shown in Figure 3. Using a dataset of 1,397 AP News articles on health and healthcare, AutoQ generated 50 queries per class (200 total). AutoE then compared LazyGraphRAG to a competing RAG method, running six trials per query across four metrics, using GPT-4.1 as a judge. These trial-level results were aggregated using metric-based win rates, where each trial is scored 1 for a win, 0.5 for a tie, and 0 for a loss, and then averaged to calculate the overall win rate for each RAG method. Figure 3. Win rates of four LazyGraphRAG (LGR) configurations across methods, broken down by the AutoQ query class and averaged across AutoE’s four metrics: comprehensiveness, diversity, empowerment, and relevance. LazyGraphRAG outperforms comparison conditions where the bar is above 50%. The four LazyGraphRAG conditions (LGR_b200_c200, LGR_b50_c200, LGR_b50_c600, LGR_b200_c200_mini) differ by query budget (b50, b200) and chunk size (c200, c600). All used GPT-4o mini for relevance tests and GPT-4o for query expansion (to five subqueries) and answer generation, except for LGR_b200_c200_mini, which used GPT-4o mini throughout. Comparison systems were GraphRAG (Local, Global, and Drift Search), Vector RAG with 8k- and 120k-token windows, and three published methods: LightRAG (opens in new tab), RAPTOR (opens in new tab), and TREX (opens in new tab). All methods were limited to the same 8k tokens for answer generation. GraphRAG Global Search used level 2 of the community hierarchy. LazyGraphRAG outperformed every comparison condition using the same generative model (GPT-4o), winning all 96 comparisons, with all but one reaching statistical significance. The best overall performance came from the larger budget, smaller chunk size configuration (LGR_b200_c200). For DataLocal queries, the smaller budget (LGR_b50_c200) performed slightly better, likely because fewer chunks were relevant. For ActivityLocal queries, the larger chunk size (LGR_b50_c600) had a slight edge, likely because longer chunks provide a more coherent context. Competing methods performed relatively better on the query classes for which they were designed: GraphRAG Global for global queries, Vector RAG for local queries, and GraphRAG Drift Search, which combines both strategies, posed the strongest challenge overall. Increasing Vector RAG’s context window from 8k to 120k tokens did not improve its performance compared to LazyGraphRAG. This raised the question of how LazyGraphRAG would perform against Vector RAG with 1-million token context window containing most of the dataset. Figure 4 shows the follow-up experiment comparing LazyGraphRAG to Vector RAG using GPT-4.1 that enabled this comparison. Even against the 1M-token window, LazyGraphRAG achieved higher win rates across all comparisons, failing to reach significance only for the relevance of answers to DataLocal queries. These queries tend to benefit most from Vector RAG’s ranking of directly relevant chunks, making it hard for LazyGraphRAG to generate answers that have greater relevance to the query, even though these answers may be dramatically more comprehensive, diverse, and empowering overall. Figure 4. Win rates of LazyGraphRAG (LGR) over Vector RAG across different context window sizes, broken down by the four AutoQ query classes and four AutoE metrics: comprehensiveness, diversity, empowerment, and relevance. Bars above 50% indicate that LazyGraphRAG outperformed the comparison condition.  AutoD: Automated data sampling and summarization Text datasets have an underlying topical structure, but the depth, breadth, and connectivity of that structure can vary widely. This variability makes it difficult to evaluate RAG systems consistently, as results may reflect the idiosyncrasies of the dataset rather than the system’s general capabilities. The AutoD component addresses this by sampling datasets to meet a target specification, defined by the number of topic clusters (breadth) and the number of samples per cluster (depth). This creates consistency across datasets, enabling more meaningful comparisons, as structurally aligned datasets lead to comparable AutoQ queries, which in turn support consistent AutoE evaluations. AutoD also includes tools for summarizing input or output datasets in a way that reflects their topical coverage. These summaries play an important role in the AutoQ query synthesis process, but they can also be used more broadly, such as in prompts where context space is limited. Since the release of the GraphRAG paper, we’ve received many requests to share the dataset of the Behind the Tech (opens in new tab) podcast transcripts we used in our evaluation. An updated version of this dataset is now available in the BenchmarkQED repository (opens in new tab), alongside the AP News dataset containing 1,397 health-related articles, licensed for open release.   We hope these datasets, together with the BenchmarkQED tools (opens in new tab), help accelerate benchmark-driven development of RAG systems and AI question-answering. We invite the community to try them on GitHub (opens in new tab).  Opens in a new tab

Last year, China saw a boom in foundation models, the do-everything large language models that underpin the AI revolution. This year, the focus has shifted to AI agents—systems that are less about responding to users’ queries and more about autonomously accomplishing things for them.  There are now a host of Chinese startups building these general-purpose digital tools, which can answer emails, browse the internet to plan vacations, and even design an interactive website. Many of these have emerged in just the last two months, following in the footsteps of Manus—a general AI agent that sparked weeks of social media frenzy for invite codes after its limited-release launch in early March.  These emerging AI agents aren’t large language models themselves. Instead, they’re built on top of them, using a workflow-based structure designed to get things done. A lot of these systems also introduce a different way of interacting with AI. Rather than just chatting back and forth with users, they are optimized for managing and executing multistep tasks—booking flights, managing schedules, conducting research—by using external tools and remembering instructions.  China could take the lead on building these kinds of agents. The country’s tightly integrated app ecosystems, rapid product cycles, and digitally fluent user base could provide a favorable environment for embedding AI into daily life.  For now, its leading AI agent startups are focusing their attention on the global market, because the best Western models don’t operate inside China’s firewalls. But that could change soon: Tech giants like ByteDance and Tencent are preparing their own AI agents that could bake automation directly into their native super-apps, pulling data from their vast ecosystem of programs that dominate many aspects of daily life in the country.  As the race to define what a useful AI agent looks like unfolds, a mix of ambitious startups and entrenched tech giants are now testing how these tools might actually work in practice—and for whom. Set the standard It’s been a whirlwind few months for Manus, which was developed by the Wuhan-based startup Butterfly Effect. The company raised $75 million in a funding round led by the US venture capital firm Benchmark, took the product on an ambitious global roadshow, and hired dozens of new employees.  Even before registration opened to the public in May, Manus had become a reference point for what a broad, consumer‑oriented AI agent should accomplish. Rather than handling narrow chores for businesses, this “general” agent is designed to be able to help with everyday tasks like trip planning, stock comparison, or your kid’s school project.  Unlike previous AI agents, Manus uses a browser-based sandbox that lets users supervise the agent like an intern, watching in real time as it scrolls through web pages, reads articles, or codes actions. It also proactively asks clarifying questions, supports long-term memory that would serve as context for future tasks. "Manus represents a promising product experience for AI agents," says Ang Li, cofounder and CEO of Simular, a startup based in Palo Alto, California, that’s building computer use agents, AI agents that control a virtual computer. “I believe Chinese startups have a huge advantage when it comes to designing consumer products, thanks to cutthroat domestic competition that leads to fast execution and greater attention to product details.” In the case of Manus, the competition is moving fast. Two of the most buzzy follow‑ups, Genspark and Flowith, for example, are already boasting benchmark scores that match or edge past Manus’s.  Genspark, led by former Baidu executives Eric Jing and Kay Zhu, links many small “super agents” through what it calls multi‑component prompting. The agent can switch among several large language models, accepts both images and text, and carries out tasks from making slide decks to placing phone calls. Whereas Manus relies heavily on Browser Use, a popular open-source product that lets agents operate a web browser in a virtual window like a human, Genspark directly integrates with a wide array of tools and APIs. Launched in April, the company says that it already has over 5 million users and over $36 million in yearly revenue. Flowith, the work of a young team that first grabbed public attention in April 2025 at a developer event hosted by the popular social media app Xiaohongshu, takes a different tack. Marketed as an “infinite agent,” it opens on a blank canvas where each question becomes a node on a branching map. Users can backtrack, take new branches, and store results in personal or sharable “knowledge gardens”—a design that feels more like project management software (think Notion) than a typical chat interface. Every inquiry or task builds its own mind-map-like graph, encouraging a more nonlinear and creative interaction with AI. Flowith’s core agent, NEO, runs in the cloud and can perform scheduled tasks like sending emails and compiling files. The founders want the app to be a “knowledge marketbase”, and aims to tap into the social aspect of AI with the aspiration of becoming “the OnlyFans of AI knowledge creators”. What they also share with Manus is the global ambition. Both Genspark and Flowith have stated that their primary focus is the international market. A global address Startups like Manus, Genspark, and Flowith—though founded by Chinese entrepreneurs—could blend seamlessly into the global tech scene and compete effectively abroad. Founders, investors, and analysts that MIT Technology Review has spoken to believe Chinese companies are moving fast, executing well, and quickly coming up with new products.  Money reinforces the pull to launch overseas. Customers there pay more, and there are plenty to go around. “You can price in USD, and with the exchange rate that’s a sevenfold multiplier,” Manus cofounder Xiao Hong quipped on a podcast. “Even if we’re only operating at 10% power because of cultural differences overseas, we’ll still make more than in China.” But creating the same functionality in China is a challenge. Major US AI companies including OpenAI and Anthropic have opted out of mainland China because of geopolitical risks and challenges with regulatory compliance. Their absence initially created a black market as users resorted to VPNs and third-party mirrors to access tools like ChatGPT and Claude. That vacuum has since been filled by a new wave of Chinese chatbots—DeepSeek, Doubao, Kimi—but the appetite for foreign models hasn’t gone away.  Manus, for example, uses Anthropic’s Claude Sonnet—widely considered the top model for agentic tasks. Manus cofounder Zhang Tao has repeatedly praised Claude’s ability to juggle tools, remember contexts, and hold multi‑round conversations—all crucial for turning chatty software into an effective executive assistant. But the company’s use of Sonnet has made its agent functionally unusable inside China without a VPN. If you open Manus from a mainland IP address, you’ll see a notice explaining that the team is “working on integrating Qwen’s model,” a special local version that is built on top of Alibaba’s open-source model.  An engineer overseeing ByteDance’s work on developing an agent, who spoke to MIT Technology Review anonymously to avoid sanction, said that the absence of Claude Sonnet models “limits everything we do in China.” DeepSeek’s open models, he added, still hallucinate too often and lack training on real‑world workflows. Developers we spoke with rank Alibaba’s Qwen series as the best domestic alternative, yet most say that switching to Qwen knocks performance down a notch. Jiaxin Pei, a postdoctoral researcher at Stanford’s Institute for Human‑Centered AI, thinks that gap will close: “Building agentic capabilities in base LLMs has become a key focus for many LLM builders, and once people realize the value of this, it will only be a matter of time." For now, Manus is doubling down on audiences it can already serve. In a written response, the company said its “primary focus is overseas expansion,” noting that new offices in San Francisco, Singapore, and Tokyo have opened in the past month. A super‑app approach Although the concept of AI agents is still relatively new, the consumer-facing AI app market in China is already crowded with major tech players. DeepSeek remains the most widely used, while ByteDance’s Doubao and Moonshot’s Kimi have also become household names. However, most of these apps are still optimized for chat and entertainment rather than task execution. This gap in the local market has pushed China’s big tech firms to roll out their own user-facing agents, though early versions remain uneven in quality and rough around the edges.  ByteDance is testing Coze Space, an AI agent based on its own Doubao model family that lets users toggle between “plan” and “execute” modes, so they can either directly guide the agent’s actions or step back and watch it work autonomously. It connects up to 14 popular apps, including GitHub, Notion, and the company’s own Lark office suite. Early reviews say the tool can feel clunky and has a high failure rate, but it clearly aims to match what Manus offers. Meanwhile, Zhipu AI has released a free agent called AutoGLM Rumination, built on its proprietary ChatGLM models. Shanghai‑based Minimax has launched Minimax Agent. Both products look almost identical to Manus and demo basic tasks such as building a simple website, planning a trip, making a small Flash game, or running quick data analysis. Despite the limited usability of most general AI agents launched within China, big companies have plans to change that. During a May 15 earnings call, Tencent president Liu Zhiping teased an agent that would weave automation directly into China’s most ubiquitous app, WeChat.  Considered the original super-app, WeChat already handles messaging, mobile payments, news, and millions of mini‑programs that act like embedded apps. These programs give Tencent, its developer, access to data from millions of services that pervade everyday life in China, an advantage most competitors can only envy. Historically, China’s consumer internet has splintered into competing walled gardens—share a Taobao link in WeChat and it resolves as plaintext, not a preview card. Unlike the more interoperable Western internet, China’s tech giants have long resisted integration with one another, choosing to wage platform war at the expense of a seamless user experience. But the use of mini‑programs has given WeChat unprecedented reach across services that once resisted interoperability, from gym bookings to grocery orders. An agent able to roam that ecosystem could bypass the integration headaches dogging independent startups. Alibaba, the e-commerce giant behind the Qwen model series, has been a front-runner in China's AI race but has been slower to release consumer-facing products. Even though Qwen was the most downloaded open-source model on Hugging Face in 2024, it didn’t power a dedicated chatbot app until early 2025. In March, Alibaba rebranded its cloud storage and search app Quark into an all-in-one AI search tool. By June, Quark had introduced DeepResearch—a new mode that marks its most agent-like effort to date.  ByteDance and Alibaba did not reply to MIT Technology Review’s request for comments. “Historically, Chinese tech products tend to pursue the all-in-one, super-app approach, and the latest Chinese AI agents reflect just that,” says Li of Simular, who previously worked at Google DeepMind on AI-enabled work automation. “In contrast, AI agents in the US are more focused on serving specific verticals.” Pei, the researcher at Stanford, says that existing tech giants could have a huge advantage in bringing the vision of general AI agents to life—especially those with built-in integration across services. “The customer-facing AI agent market is still very early, with tons of problems like authentication and liability,” he says. “But companies that already operate across a wide range of services have a natural advantage in deploying agents at scale.”

AGI will be embedded into humanoid robots, which makes white-collar and blue-collar jobs a target ... More for walking/talking automation.getty In today’s column, I examine the highly worrisome qualms expressed that the advent of artificial general intelligence (AGI) is likely to usurp white-collar jobs. The stated concern is that since AGI will be on par with human intellect, any job that relies principally on intellectual pursuits such as typical white-collar work will be taken over via the use of AGI. Employers will realize that rather than dealing with human white-collar workers, they can more readily get the job done via AGI. This, in turn, has led to a rising call that people should aim toward blue-collar jobs, doing so because (presumably) those forms of employment will not be undercut via AGI. Sorry to say, that misses the bigger picture, namely that AGI when combined with humanoid robots is coming not only for white-collar jobs but also blue-collar jobs too. It is a proverbial double-whammy when it comes to the attainment of AGI. Let’s talk about it. This analysis of an innovative AI breakthrough is part of my ongoing Forbes column coverage on the latest in AI, including identifying and explaining various impactful AI complexities (see the link here). Heading Toward AGI And ASI First, some fundamentals are required to set the stage for this weighty discussion. There is a great deal of research going on to further advance AI. The general goal is to either reach artificial general intelligence (AGI) or maybe even the outstretched possibility of achieving artificial superintelligence (ASI). AGI is AI that is considered on par with human intellect and can seemingly match our intelligence. ASI is AI that has gone beyond human intellect and would be superior in many if not all feasible ways. The idea is that ASI would be able to run circles around humans by outthinking us at every turn. For more details on the nature of conventional AI versus AGI and ASI, see my analysis at the link here. We have not yet attained AGI. In fact, it is unknown as to whether we will reach AGI, or that maybe AGI will be achievable in decades or perhaps centuries from now. The AGI attainment dates that are floating around are wildly varying and wildly unsubstantiated by any credible evidence or ironclad logic. ASI is even more beyond the pale when it comes to where we are currently with conventional AI. AGI Problem Only Half Seen Before launching into the primary matter at hand in this discussion, let’s contemplate a famous quote attributed to Charles Kettering, a legendary inventor, who said, “A problem well-stated is a problem half-solved.” I bring this up because those loud clamors right now about the assumption that AGI will replace white-collar workers are only seeing half of the problem. The problem as they see it is that since AGI is intellectually on par with humans, and since white-collar workers mainly use intellect in their work endeavors, AGI is going to be used in place of humans for white-collar work. I will in a moment explain why that’s only half of the problem and there is a demonstrative need to more carefully and fully articulate the nature of the problem. Will AGI Axiomatically Take White-Collar Jobs On a related facet, the belief that AGI will axiomatically replace white-collar labor makes a number of other related key assumptions. I shall briefly explore those and then come back to why the problem itself is only half-baked. The cost of using AGI for doing white-collar work will need to be presumably a better ROI choice over human workers. If not, then an employer would be wiser to stick with humans rather than employing AGI. There seems to often be an unstated belief that AGI is necessarily going to be a less costly route than employing humans. We don’t know yet what the cost of using AGI will be. It could be highly expensive. Indeed, some are worried that the world will divide into the AGI haves and AGI have-nots, partially due to the exorbitant cost that AGI might involve. If AGI is free to use, well, that would seem to be the nail in the coffin related to using human workers for the same capacity. Another angle is that AGI is relatively inexpensive in comparison to human labor. In that case, the use of AGI is likely to win over human labor usage. But if the cost of AGI is nearer to the cost of human labor (all in), or more so, then employers would rationally need to weigh the use of one versus the other. Note that when referring to the cost of human labor, there is more to that calculation than simply the dollar-hour labor rate per se. There are lots of other less apparent costs, such as the cost to manage human labor, the cost of dealing with HR-related issues, and many other factors that come into the weighty matter. Thus, an AGI versus human labor ROI will be more complex than it might seem at an initial glance. In addition, keep in mind that AGI would seemingly be readily switched on and off, and have other capacities that human labor would not equally tend to allow. The Other Half Is Coming Too Assume that by and large the advent of AGI will decimate the need for white-collar human labor. The refrain right now is that people should begin tilting toward blue-collar jobs as an alternative to white-collar jobs. This is a logical form of thinking in the sense that AGI as an intellectual mechanism would be unable to compete in jobs that involve hands-on work. A plumber needs to come to your house and do hands-on work to fix your plumbing. This is a physicality that entails arriving at your physical home, physically bringing and using tools, and physically repairing your faulty home plumbing. A truck driver likewise needs to sit in the cab of a truck and drive the vehicle. These are physically based tasks. There is no getting around the fact that these are hands-on activities. Aha, yes, those are physical tasks, but that doesn’t necessarily mean that only human hands can perform them. The gradual emergence of humanoid robots will provide an alternative to human hands. A humanoid robot is a type of robot that is built to resemble a human in form and function. You’ve undoubtedly seen those types of robots in the many online video recordings showing them walking, jumping, grasping at objects, and so on. A tremendous amount of active research and development is taking place to devise humanoid robots. They look comical right now. You watch those videos and laugh when the robot trips over a mere stick lying on the ground, something that a human would seldom trip over. You scoff when a robot tries to grasp a coffee cup and inadvertently spills most of the coffee. It all seems humorous and a silly pursuit. Keep in mind that we are all observing the development process while it is still taking place. At some point, those guffaws of the humanoid robots will lessen. Humanoid robots will be as smooth and graceful as humans. This will continue to be honed. Eventually, humanoid robots will be less prone to physical errors that humans make. In a sense, the physicality of a humanoid robot will be on par with humans, if not better, due to its mechanical properties. Do not discount the coming era of quite physically capable humanoid robots. AGI And Humanoid Robots Pair Up You might remember that in The Wonderful Wizard of Oz, the fictional character known as The Strawman lacked a brain. Without seeming to anthropomorphize humanoid robots, the current situation is that those robots typically use a form of AI that is below the sophistication level of modern generative AI. That’s fine for now due to the need to first ensure that the physical movements of the robots get refined. I have discussed that a said-to-be realm of Physical AI is going to be a huge breakthrough with incredible ramifications, see my analysis at the link here. The idea underlying Physical AI is that the AI of today is being uplifted by doing data training on the physical world. This also tends to include the use of World Models, consisting of broad constructions about how the physical world works, such as that we are bound to operate under conditions of gravity, and other physical laws of nature, see the link here. The bottom line here is that there will be a close pairing of robust AI with humanoid robots. Imagine what a humanoid robot can accomplish if it is paired with AGI. I’ll break the suspense and point out that AGI paired with humanoid robots means that those robots readily enter the blue-collar worker realm. Suppose your plumbing needs fixing. No worries, a humanoid robot that encompasses AGI will be sent to your home. The AGI is astute enough to carry on conversations with you, and the AGI also fully operates the robot to undertake the plumbing tasks. How did the AGI-paired humanoid robot get to your home? Easy-peasy, it drove a car or truck to get there. I’ve previously predicted that all the work on devising autonomous vehicles and self-driving cars will get shaken up once we have suitable humanoid robots devised. There won’t be a need for a vehicle to contain self-driving capabilities. A humanoid robot will simply sit in the driver’s seat and drive the vehicle. This is a much more open-ended solution than having to craft components that go into and onto a vehicle to enable self-driving. See my coverage at the link here. Timing Is Notable One of the reasons that many do not give much thought to the pairing of AGI with humanoid robots is that today’s humanoid robots seem extraordinarily rudimentary and incapable of performing physical dexterity tasks on par with human capabilities. Meanwhile, there is brazen talk that AGI is just around the corner. AGI is said to be within our grasp. Let’s give the timing considerations a bit of scrutiny. There are three primary timing angles: Option 1: AGI first, then humanoid robots. AGI is attained before humanoid robots are sufficiently devised. Option 2: Humanoid robots first, then AGI. Humanoid robots are physically fluently adept before AGI is attained. Option 3: AGI and humanoid robots arrive about at the same time. AGI is attained and at the same time, it turns out that humanoid robots are fluently adept too, mainly by coincidence and not due to any cross-mixing. A skeptic would insist that there is a fourth possibility, consisting of the possibility that we never achieve AGI and/or we fail to achieve sufficiently physically capable humanoid robots. I am going to reject that possibility. Perhaps I am overly optimistic, but it seems to me that we will eventually attain AGI, and we will eventually attain physically capable humanoid robots. I shall next respectively consider each of the three genuinely reasonable possibilities. Option 1: AGI First, Then Humanoid Robots What if we manage to attain AGI before we manage to achieve physically fluent humanoid robots? That’s just fine. We would indubitably put AGI to work as a partner with humans in figuring out how we can push along the budding humanoid robot development process. It seems nearly obvious that with AGI’s capable assistance, we would overcome any bottlenecks and soon enough arrive at top-notch physically adept humanoid robots. At that juncture, we would then toss AGI into the humanoid robots and have ourselves quite an amazing combination. Option 2: Humanoid Robots First, Then AGI Suppose that we devise very physically adept humanoid robots but have not yet arrived at AGI. Are we in a pickle? Nope. We could use conventional advanced AI inside those humanoid robots. The combination would certainly be good enough for a wide variety of tasks. The odds are that we would need to be cautious about where such robots are utilized. Nonetheless, we would have essentially walking, talking, and productive humanoid robots. If AGI never happens, oh well, we end up with pretty good humanoid robots. On the other hand, once we arrive at AGI, those humanoid robots will be stellar. It’s just a matter of time. Option 3: AGI And Humanoid Robots At The Same Time Let’s consider the potential of AGI and humanoid robots perchance being attained around the same time. Assume that this timing isn’t due to an outright cross-mixing with each other. They just so happen to advance on a similar timeline. I tend to believe that’s the most likely of the three scenarios. Here’s why. First, despite all the hubris about AGI being within earshot, perhaps in the next year or two, which is a popular pronouncement by many AI luminaries, I tend to side with recent surveys of AI developers that put the date around the year 2040 (see my coverage at the link here). Some AI luminaires sneakily play with the definition of AGI in hopes of making their predictions come true sooner, akin to moving the goalposts to easily score points. For my coverage on Sam Altman’s efforts of moving the cheese regarding AGI attainment, see the link here. Second, if you are willing to entertain the year 2040 as a potential date for achieving AGI, that’s about 15 years from now. In my estimation, the advancements being made in humanoid robots will readily progress such that by 2040 they will be very physically adept. Probably be sooner, but let’s go with the year 2040 for ease of contemplation. In my view, we will likely have humanoid robots doing well enough that they will be put into use prior to arriving at AGI. The pinnacle of robust humanoid robots and the attainment of AGI will roughly coincide with each other. Two peas in a pod.Impact Of Enormous Consequences In an upcoming column posting, I will examine the enormous consequences of having AGI paired with fully physically capable humanoid robots. As noted above, this will have a humongous impact on white-collar work and blue-collar work. There will be gargantuan economic impacts, societal impacts, cultural impacts, and so on. Some final thoughts for now. A single whammy is already being hotly debated. The debates currently tend to be preoccupied with the loss of white-collar jobs due to the attainment of AGI. A saving grace seems to be that at least blue-collar jobs are going to be around and thriving, even once AGI is attained. The world doesn’t seem overly gloomy if you can cling to the upbeat posture that blue-collar tasks remain intact. The double whammy is a lot more to take in. But the double whammy is the truth. The truth needs to be faced. If you are having doubts as a human about the future, just remember the famous words of Vince Lombardi: “Winners never quit, and quitters never win.” Humankind can handle the double whammy. Stay tuned for my upcoming coverage of what this entails.

Last year, China saw a boom in foundation models, the do-everything large language models that underpin the AI revolution. This year, the focus has shifted to AI agents—systems that are less about responding to users’ queries and more about autonomously accomplishing things for them.  There are now a host of Chinese startups building these general-purpose digital tools, which can answer emails, browse the internet to plan vacations, and even design an interactive website. Many of these have emerged in just the last two months, following in the footsteps of Manus—a general AI agent that sparked weeks of social media frenzy for invite codes after its limited-release launch in early March.  These emerging AI agents aren’t large language models themselves. Instead, they’re built on top of them, using a workflow-based structure designed to get things done. A lot of these systems also introduce a different way of interacting with AI. Rather than just chatting back and forth with users, they are optimized for managing and executing multistep tasks—booking flights, managing schedules, conducting research—by using external tools and remembering instructions.  China could take the lead on building these kinds of agents. The country’s tightly integrated app ecosystems, rapid product cycles, and digitally fluent user base could provide a favorable environment for embedding AI into daily life.  For now, its leading AI agent startups are focusing their attention on the global market, because the best Western models don’t operate inside China’s firewalls. But that could change soon: Tech giants like ByteDance and Tencent are preparing their own AI agents that could bake automation directly into their native super-apps, pulling data from their vast ecosystem of programs that dominate many aspects of daily life in the country.  As the race to define what a useful AI agent looks like unfolds, a mix of ambitious startups and entrenched tech giants are now testing how these tools might actually work in practice—and for whom. Set the standard It’s been a whirlwind few months for Manus, which was developed by the Wuhan-based startup Butterfly Effect. The company raised $75 million in a funding round led by the US venture capital firm Benchmark, took the product on an ambitious global roadshow, and hired dozens of new employees.  Even before registration opened to the public in May, Manus had become a reference point for what a broad, consumer‑oriented AI agent should accomplish. Rather than handling narrow chores for businesses, this “general” agent is designed to be able to help with everyday tasks like trip planning, stock comparison, or your kid’s school project.  Unlike previous AI agents, Manus uses a browser-based sandbox that lets users supervise the agent like an intern, watching in real time as it scrolls through web pages, reads articles, or codes actions. It also proactively asks clarifying questions, supports long-term memory that would serve as context for future tasks. "Manus represents a promising product experience for AI agents," says Ang Li, cofounder and CEO of Simular, a startup based in Palo Alto, California, that’s building computer use agents, AI agents that control a virtual computer. “I believe Chinese startups have a huge advantage when it comes to designing consumer products, thanks to cutthroat domestic competition that leads to fast execution and greater attention to product details.” In the case of Manus, the competition is moving fast. Two of the most buzzy follow‑ups, Genspark and Flowith, for example, are already boasting benchmark scores that match or edge past Manus’s.  Genspark, led by former Baidu executives Eric Jing and Kay Zhu, links many small “super agents” through what it calls multi‑component prompting. The agent can switch among several large language models, accepts both images and text, and carries out tasks from making slide decks to placing phone calls. Whereas Manus relies heavily on Browser Use, a popular open-source product that lets agents operate a web browser in a virtual window like a human, Genspark directly integrates with a wide array of tools and APIs. Launched in April, the company says that it already has over 5 million users and over $36 million in yearly revenue. Flowith, the work of a young team that first grabbed public attention in April 2025 at a developer event hosted by the popular social media app Xiaohongshu, takes a different tack. Marketed as an “infinite agent,” it opens on a blank canvas where each question becomes a node on a branching map. Users can backtrack, take new branches, and store results in personal or sharable “knowledge gardens”—a design that feels more like project management software (think Notion) than a typical chat interface. Every inquiry or task builds its own mind-map-like graph, encouraging a more nonlinear and creative interaction with AI. Flowith’s core agent, NEO, runs in the cloud and can perform scheduled tasks like sending emails and compiling files. The founders want the app to be a “knowledge marketbase”, and aims to tap into the social aspect of AI with the aspiration of becoming “the OnlyFans of AI knowledge creators”. What they also share with Manus is the global ambition. Both Genspark and Flowith have stated that their primary focus is the international market. A global address Startups like Manus, Genspark, and Flowith—though founded by Chinese entrepreneurs—could blend seamlessly into the global tech scene and compete effectively abroad. Founders, investors, and analysts that MIT Technology Review has spoken to believe Chinese companies are moving fast, executing well, and quickly coming up with new products.  Money reinforces the pull to launch overseas. Customers there pay more, and there are plenty to go around. “You can price in USD, and with the exchange rate that’s a sevenfold multiplier,” Manus cofounder Xiao Hong quipped on a podcast. “Even if we’re only operating at 10% power because of cultural differences overseas, we’ll still make more than in China.” But creating the same functionality in China is a challenge. Major US AI companies including OpenAI and Anthropic have opted out of mainland China because of geopolitical risks and challenges with regulatory compliance. Their absence initially created a black market as users resorted to VPNs and third-party mirrors to access tools like ChatGPT and Claude. That vacuum has since been filled by a new wave of Chinese chatbots—DeepSeek, Doubao, Kimi—but the appetite for foreign models hasn’t gone away.  Manus, for example, uses Anthropic’s Claude Sonnet—widely considered the top model for agentic tasks. Manus cofounder Zhang Tao has repeatedly praised Claude’s ability to juggle tools, remember contexts, and hold multi‑round conversations—all crucial for turning chatty software into an effective executive assistant. But the company’s use of Sonnet has made its agent functionally unusable inside China without a VPN. If you open Manus from a mainland IP address, you’ll see a notice explaining that the team is “working on integrating Qwen’s model,” a special local version that is built on top of Alibaba’s open-source model.  An engineer overseeing ByteDance’s work on developing an agent, who spoke to MIT Technology Review anonymously to avoid sanction, said that the absence of Claude Sonnet models “limits everything we do in China.” DeepSeek’s open models, he added, still hallucinate too often and lack training on real‑world workflows. Developers we spoke with rank Alibaba’s Qwen series as the best domestic alternative, yet most say that switching to Qwen knocks performance down a notch. Jiaxin Pei, a postdoctoral researcher at Stanford’s Institute for Human‑Centered AI, thinks that gap will close: “Building agentic capabilities in base LLMs has become a key focus for many LLM builders, and once people realize the value of this, it will only be a matter of time." For now, Manus is doubling down on audiences it can already serve. In a written response, the company said its “primary focus is overseas expansion,” noting that new offices in San Francisco, Singapore, and Tokyo have opened in the past month. A super‑app approach Although the concept of AI agents is still relatively new, the consumer-facing AI app market in China is already crowded with major tech players. DeepSeek remains the most widely used, while ByteDance’s Doubao and Moonshot’s Kimi have also become household names. However, most of these apps are still optimized for chat and entertainment rather than task execution. This gap in the local market has pushed China’s big tech firms to roll out their own user-facing agents, though early versions remain uneven in quality and rough around the edges.  ByteDance is testing Coze Space, an AI agent based on its own Doubao model family that lets users toggle between “plan” and “execute” modes, so they can either directly guide the agent’s actions or step back and watch it work autonomously. It connects up to 14 popular apps, including GitHub, Notion, and the company’s own Lark office suite. Early reviews say the tool can feel clunky and has a high failure rate, but it clearly aims to match what Manus offers. Meanwhile, Zhipu AI has released a free agent called AutoGLM Rumination, built on its proprietary ChatGLM models. Shanghai‑based Minimax has launched Minimax Agent. Both products look almost identical to Manus and demo basic tasks such as building a simple website, planning a trip, making a small Flash game, or running quick data analysis. Despite the limited usability of most general AI agents launched within China, big companies have plans to change that. During a May 15 earnings call, Tencent president Liu Zhiping teased an agent that would weave automation directly into China’s most ubiquitous app, WeChat.  Considered the original super-app, WeChat already handles messaging, mobile payments, news, and millions of mini‑programs that act like embedded apps. These programs give Tencent, its developer, access to data from millions of services that pervade everyday life in China, an advantage most competitors can only envy. Historically, China’s consumer internet has splintered into competing walled gardens—share a Taobao link in WeChat and it resolves as plaintext, not a preview card. Unlike the more interoperable Western internet, China’s tech giants have long resisted integration with one another, choosing to wage platform war at the expense of a seamless user experience. But the use of mini‑programs has given WeChat unprecedented reach across services that once resisted interoperability, from gym bookings to grocery orders. An agent able to roam that ecosystem could bypass the integration headaches dogging independent startups. Alibaba, the e-commerce giant behind the Qwen model series, has been a front-runner in China's AI race but has been slower to release consumer-facing products. Even though Qwen was the most downloaded open-source model on Hugging Face in 2024, it didn’t power a dedicated chatbot app until early 2025. In March, Alibaba rebranded its cloud storage and search app Quark into an all-in-one AI search tool. By June, Quark had introduced DeepResearch—a new mode that marks its most agent-like effort to date.  ByteDance and Alibaba did not reply to MIT Technology Review’s request for comments. “Historically, Chinese tech products tend to pursue the all-in-one, super-app approach, and the latest Chinese AI agents reflect just that,” says Li of Simular, who previously worked at Google DeepMind on AI-enabled work automation. “In contrast, AI agents in the US are more focused on serving specific verticals.” Pei, the researcher at Stanford, says that existing tech giants could have a huge advantage in bringing the vision of general AI agents to life—especially those with built-in integration across services. “The customer-facing AI agent market is still very early, with tons of problems like authentication and liability,” he says. “But companies that already operate across a wide range of services have a natural advantage in deploying agents at scale.”

Larry Fabbroni is an architect, strategic advisor, and Chief Innovation Officer for Practice of Architecture. Throughout his career, he has led efforts to reform studio culture and innovate practice. He earned his MBA from the University of Chicago’s Booth School of Business. In 2017, as leaders in the AIA’s Young Architects Forum (YAF), we led the launch of the Practice Innovation Lab (PIL) and hosted a symposium that imagined new architectural practice models. At that time, we already felt that practice innovation was overdue in a profession that has not seen scaled disruption to its business model in over a century. Today, we are confident that there has never been a more critical time for the profession to embrace innovation. Redefining Innovation Henley Hall: Institute for Energy Efficiency by KieranTimberlake, Santa Barbara, California | KieranTimberlake’s research expertise creates value beyond a baseline labor model.  Currently, artificial intelligence dominates strategy conversations, but just as we saw back in 2017, larger patterns prompt calls for innovation. Talent attraction is increasingly challenging, disruptive technology continues to emerge, and actors from outside our industry show growing interest in the space. While incremental innovation has long been a part of the profession, relatively few firms have adopted new practices that create value beyond a baseline labor model. Firms such as KieranTimberlake have shown that research expertise can do this. MASS Design has pioneered a mission-driven approach. BIG has taken on the role of architect-as-developer. Snøhetta houses a product design division. We could continue to list great firms that have pushed the boundaries of practice, but they represent exceptions that have yet to be recognized as new standards. Indeed, the confluence of those factors that led to the original PIL continues to make the case that the time for scaled innovation is now. A Melting Iceberg: Incremental Changes Depleting the Profession Powerhouse Telemark by Snøhetta, Vestfold og Telemark, Norway | Photo by Ivar Kvaal | Snøhetta houses a product design division, innovatively presenting a alternative business model for firms.  One of the dangers of operating in a slow-moving industry is that change is difficult to detect and even more challenging to comprehend. If an iceberg loses 1% of mass per year, it’s tough to take notice, but the end result is catastrophic. This is what is happening to our profession. For newcomers, if it feels like there are increasingly more attractive opportunities elsewhere, that’s because there are. For seasoned professionals, if it feels like it’s become more challenging to maintain the same levels of prosperity, that’s because it has. Less(er) Talent In some ways, the shift towards companies recognizing “talent” as their most excellent resource has bewildered architects: we have always relied on talent. However, the patterns of talent leaving our profession are concerning. We say “feel” because there is no significant data. We spoke to Kendall A. Nicholson, Senior Director of Research at the Association of Collegiate Schools of Architecture (ACSA), who confirmed that aggregated data on graduate placement does not exist. So we inquired about what placement looks like at several programs around the country. Omar Khan, Head of the Carnegie Mellon University School of Architecture, informed us that approximately 90% of students pursue a minor to expand their horizons, and that in 2022, nearly one in three graduates entered the tech sector. Khan stated that these opportunities aren’t just student-driven — large innovative companies increasingly seek the value that graduates of architecture schools will provide. This increasing difficulty in capturing the talent that architecture schools are producing results in a shrinking and diluted talent pool. For a profession so reliant on human resources, this presents extreme risk. Pay Gaps In an increasingly expensive world, we are not able to compete for the best talent with emerging industries. It’s easy to understand why a popular career pivot for architects has become UX design. Designing user experience for websites pays significantly better than designing the same for the built environment. According to Glassdoor, 2023 entry-level UX designers earned an average of $78K, while the AIA salary calculator suggests architecture grads can expect to earn an average of $59 K. The talent we do attract into the profession often loses interest when they experience low pay and long hours, all while most firms lack clear paths and criteria for advancement or compensation increases. A Smaller Piece of the Pie Examining data in isolation, one might conclude that the profession continues to grow; the number of architects has increased substantially over the last century, and this trend has persisted in recent years. The problem with this growth is that the estimated share of the US GDP for Architectural Services has shrunk over time. This is not a manageable number to measure before 1999, when NCARB first aggregated local jurisdictional data. Due to limitations in industry economic data, we’re only showing data since 2011 for the purposes of this article. In that time, the number of architects has grown, the market size for services has grown, but the share those services represent as a portion of the US GDP has declined — by 15% if we use US Census data to almost 30% if we use industry research data (we used IbisWorld.com, however we found data that suggested a worse and others that offered a slightly better picture). To put it another way, architecture is a stagnant industry with a shrinking share of the economy. It’s challenging to examine this data and emerge feeling confident about the profession, but there is a silver lining. The biggest impediment to innovation for architects is not a lack of talent, but rather the business model. Design thinking has been widely adopted throughout the world as a key component of innovation processes; however, the problem is that we operate in the realm of professional services, which inherently is not well-suited to promoting innovation. Reliance on that formula is causing our iceberg to melt. The Tsunami: The AI Tidal Wave is Here The Rwanda Institute for Conservation Agriculture by MASS Design Group, Rwanda | MASS Design has pioneered a mission-driven approach that creates value beyond a baseline labor model.  As we confront the exodus of talent, it is easy for both firm owners and clients to imagine AI bringing efficiencies and replacing “CAD-monkeys” with machines. However, any firm that wants to operate — and win — as anything more than a low-cost provider will need a strategy to increase value, not just cut costs. AI is merely part of the toolbox required to confront a perfect storm of forces. Jobs will Disappear Goldman Sachs predicts that as much as 37% of our industry tasks will be replaced by AI. Many see this as a pathway to lower costs and increased profits. However, that is short-sighted. Markets will adjust quickly and demand lower costs for services; additional new value will need to be articulated and proven, and this will only happen through innovation. New Jobs will Emerge (but fewer of them) AI prophets often emphasize that technological innovation has historically led to net employment gains. Previous World Economic Forum estimates predicted losses of up to 85 million existing jobs worldwide, with parallel gains of as many as 97 million new jobs. However, these estimates were revised in the WEF 2023 Economic Outlook, which now anticipates a net loss of 14 million jobs. This stark outlook signals an even greater need for architects to become more innovative. The 2024 RIBA AI Report indicates that 41% of architecture firms were already utilizing AI, though current tools are indeed just the beginning. Marketing, business development and content creation will be standard areas of AI deployment moving forward. Still, revolutionary changes will come in how we learn, not only to use new tools, but also to collaborate with digital agents. How will this happen? We can theorize, but it is not possible to know for sure until it arrives, so we need to have a plan before we can see the tidal wave from land. The Alien Invasion: Outsiders Are Entering Our Orbit VIA 57 West by BIG – Bjarke Ingels Group, New York City, New York | BIG has pioneered a new model for practice by taking on the role of architect-as-developer. For years, we’ve heard cries that “architects gave away the role of master builder.” But how much did architects actually give, and how much was taken by innovative competition? This distinction is critical because the wagons are circling, and the AEC space has become ever more attractive to investors. Venture Capital and Private Equity Investment The numbers are often difficult to parse because architecture can impact so many verticals and does not operate as its own sector in the investment realm; however, the trends suggest a groundswell is underway. A 2023 McKinsey report shows that construction tech deals nearly doubled from 2019 to 2022, growing by 85%. At the same period, the number of deals increased by 30%, indicating that interest continues to grow. An increasing size of deals also suggests a maturity of the market. As interest in infrastructure investments has declined from its high in 2020, and along with real estate, has been blunted by high interest rates, institutional investors continue to see opportunities in the AEC space. Firm Acquisitions AEC firms that deliver predictable returns have proven to be attractive targets for PE firms. In the second quarter of 2024, private equity firms accounted for over one-third of AEC firm mergers and acquisitions. For M&A deals, the industry has seen an increase in attractiveness with expanded infrastructure spending as a catalyst. However, this interest can also be tied to the lack of innovation that has resulted in an industry ripe for consolidation. M&A orchestrators generate large amounts of profit by streamlining operations, eliminating redundancies, and then stamping out competition. An entire community has been built around this, with AEC Advisors hosting an annual “Private Equity Summit” that brings together CEOs of AEC firms with PE investors. Startups As an extension of the growing interest from venture capital in the space, there is an upward trend in the AEC space being targeted for disruption by entrepreneurs who see an industry that represents a significant portion of the global GDP. AEC Works, a project of e-verse that catalogs AEC startups and investors, lists nearly 800 startups from around the world, with almost 200 identified as “architecture-focused.” The signal is clear: startups are looking to figure out how to do what you do cheaper, better, or perhaps both. Combining this environment with depleted talent pools, a declining share of GDP, and revolutionary technology, it is a correct response to be alarmed. Significant change is inevitable. It is time for architects to see the same opportunities that investors and entrepreneurs see, and learn to navigate within these spaces. The Great Opportunity Throughout history, new actors have enjoyed a “leap-frog” effect and been able to surpass established incumbents to reshape industries, markets and economies. From climate change to pandemic ripple effects, to the housing crisis, to generational shifts in the workforce, there are many forces that directly impact the work of architects and call for innovation. The need for new ways of designing and delivering different components of the built environment is ever-present and will be solved by teams that either include — and might be led by — architects, or those that do not. Most end users will only care if the resulting product is superior. This time of tension is indeed a time of great opportunity. Architects who embrace innovation in pursuing new iterations of our dated business models may actually achieve what many of us have dreamed of from the start: to leave a positive mark on the world. We think the future of the profession depends on it. Top image: Powerhouse Telemark by Snøhetta, Vestfold og Telemark, Norway The post Architects, Your Real Competition Isn’t AI — It’s Business Complacency appeared first on Journal.

Last year, China saw a boom in foundation models, the do-everything large language models that underpin the AI revolution. This year, the focus has shifted to AI agents—systems that are less about responding to users’ queries and more about autonomously accomplishing things for them.  There are now a host of Chinese startups building these general-purpose digital tools, which can answer emails, browse the internet to plan vacations, and even design an interactive website. Many of these have emerged in just the last two months, following in the footsteps of Manus—a general AI agent that sparked weeks of social media frenzy for invite codes after its limited-release launch in early March.  These emerging AI agents aren’t large language models themselves. Instead, they’re built on top of them, using a workflow-based structure designed to get things done. A lot of these systems also introduce a different way of interacting with AI. Rather than just chatting back and forth with users, they are optimized for managing and executing multistep tasks—booking flights, managing schedules, conducting research—by using external tools and remembering instructions.  China could take the lead on building these kinds of agents. The country’s tightly integrated app ecosystems, rapid product cycles, and digitally fluent user base could provide a favorable environment for embedding AI into daily life.  For now, its leading AI agent startups are focusing their attention on the global market, because the best Western models don’t operate inside China’s firewalls. But that could change soon: Tech giants like ByteDance and Tencent are preparing their own AI agents that could bake automation directly into their native super-apps, pulling data from their vast ecosystem of programs that dominate many aspects of daily life in the country.  As the race to define what a useful AI agent looks like unfolds, a mix of ambitious startups and entrenched tech giants are now testing how these tools might actually work in practice—and for whom. Set the standard It’s been a whirlwind few months for Manus, which was developed by the Wuhan-based startup Butterfly Effect. The company raised $75 million in a funding round led by the US venture capital firm Benchmark, took the product on an ambitious global roadshow, and hired dozens of new employees.  Even before registration opened to the public in May, Manus had become a reference point for what a broad, consumer‑oriented AI agent should accomplish. Rather than handling narrow chores for businesses, this “general” agent is designed to be able to help with everyday tasks like trip planning, stock comparison, or your kid’s school project.  Unlike previous AI agents, Manus uses a browser-based sandbox that lets users supervise the agent like an intern, watching in real time as it scrolls through web pages, reads articles, or codes actions. It also proactively asks clarifying questions, supports long-term memory that would serve as context for future tasks. “Manus represents a promising product experience for AI agents,” says Ang Li, cofounder and CEO of Simular, a startup based in Palo Alto, California, that’s building computer use agents, AI agents that control a virtual computer. “I believe Chinese startups have a huge advantage when it comes to designing consumer products, thanks to cutthroat domestic competition that leads to fast execution and greater attention to product details.” In the case of Manus, the competition is moving fast. Two of the most buzzy follow‑ups, Genspark and Flowith, for example, are already boasting benchmark scores that match or edge past Manus’s.  Genspark, led by former Baidu executives Eric Jing and Kay Zhu, links many small “super agents” through what it calls multi‑component prompting. The agent can switch among several large language models, accepts both images and text, and carries out tasks from making slide decks to placing phone calls. Whereas Manus relies heavily on Browser Use, a popular open-source product that lets agents operate a web browser in a virtual window like a human, Genspark directly integrates with a wide array of tools and APIs. Launched in April, the company says that it already has over 5 million users and over $36 million in yearly revenue. Flowith, the work of a young team that first grabbed public attention in April 2025 at a developer event hosted by the popular social media app Xiaohongshu, takes a different tack. Marketed as an “infinite agent,” it opens on a blank canvas where each question becomes a node on a branching map. Users can backtrack, take new branches, and store results in personal or sharable “knowledge gardens”—a design that feels more like project management software (think Notion) than a typical chat interface. Every inquiry or task builds its own mind-map-like graph, encouraging a more nonlinear and creative interaction with AI. Flowith’s core agent, NEO, runs in the cloud and can perform scheduled tasks like sending emails and compiling files. The founders want the app to be a “knowledge marketbase”, and aims to tap into the social aspect of AI with the aspiration of becoming “the OnlyFans of AI knowledge creators”. What they also share with Manus is the global ambition. Both Genspark and Flowith have stated that their primary focus is the international market. A global address Startups like Manus, Genspark, and Flowith—though founded by Chinese entrepreneurs—could blend seamlessly into the global tech scene and compete effectively abroad. Founders, investors, and analysts that MIT Technology Review has spoken to believe Chinese companies are moving fast, executing well, and quickly coming up with new products.  Money reinforces the pull to launch overseas. Customers there pay more, and there are plenty to go around. “You can price in USD, and with the exchange rate that’s a sevenfold multiplier,” Manus cofounder Xiao Hong quipped on a podcast. “Even if we’re only operating at 10% power because of cultural differences overseas, we’ll still make more than in China.” But creating the same functionality in China is a challenge. Major US AI companies including OpenAI and Anthropic have opted out of mainland China because of geopolitical risks and challenges with regulatory compliance. Their absence initially created a black market as users resorted to VPNs and third-party mirrors to access tools like ChatGPT and Claude. That vacuum has since been filled by a new wave of Chinese chatbots—DeepSeek, Doubao, Kimi—but the appetite for foreign models hasn’t gone away.  Manus, for example, uses Anthropic’s Claude Sonnet—widely considered the top model for agentic tasks. Manus cofounder Zhang Tao has repeatedly praised Claude’s ability to juggle tools, remember contexts, and hold multi‑round conversations—all crucial for turning chatty software into an effective executive assistant. But the company’s use of Sonnet has made its agent functionally unusable inside China without a VPN. If you open Manus from a mainland IP address, you’ll see a notice explaining that the team is “working on integrating Qwen’s model,” a special local version that is built on top of Alibaba’s open-source model.  An engineer overseeing ByteDance’s work on developing an agent, who spoke to MIT Technology Review anonymously to avoid sanction, said that the absence of Claude Sonnet models “limits everything we do in China.” DeepSeek’s open models, he added, still hallucinate too often and lack training on real‑world workflows. Developers we spoke with rank Alibaba’s Qwen series as the best domestic alternative, yet most say that switching to Qwen knocks performance down a notch. Jiaxin Pei, a postdoctoral researcher at Stanford’s Institute for Human‑Centered AI, thinks that gap will close: “Building agentic capabilities in base LLMs has become a key focus for many LLM builders, and once people realize the value of this, it will only be a matter of time.” For now, Manus is doubling down on audiences it can already serve. In a written response, the company said its “primary focus is overseas expansion,” noting that new offices in San Francisco, Singapore, and Tokyo have opened in the past month. A super‑app approach Although the concept of AI agents is still relatively new, the consumer-facing AI app market in China is already crowded with major tech players. DeepSeek remains the most widely used, while ByteDance’s Doubao and Moonshot’s Kimi have also become household names. However, most of these apps are still optimized for chat and entertainment rather than task execution. This gap in the local market has pushed China’s big tech firms to roll out their own user-facing agents, though early versions remain uneven in quality and rough around the edges.  ByteDance is testing Coze Space, an AI agent based on its own Doubao model family that lets users toggle between “plan” and “execute” modes, so they can either directly guide the agent’s actions or step back and watch it work autonomously. It connects up to 14 popular apps, including GitHub, Notion, and the company’s own Lark office suite. Early reviews say the tool can feel clunky and has a high failure rate, but it clearly aims to match what Manus offers. Meanwhile, Zhipu AI has released a free agent called AutoGLM Rumination, built on its proprietary ChatGLM models. Shanghai‑based Minimax has launched Minimax Agent. Both products look almost identical to Manus and demo basic tasks such as building a simple website, planning a trip, making a small Flash game, or running quick data analysis. Despite the limited usability of most general AI agents launched within China, big companies have plans to change that. During a May 15 earnings call, Tencent president Liu Zhiping teased an agent that would weave automation directly into China’s most ubiquitous app, WeChat.  Considered the original super-app, WeChat already handles messaging, mobile payments, news, and millions of mini‑programs that act like embedded apps. These programs give Tencent, its developer, access to data from millions of services that pervade everyday life in China, an advantage most competitors can only envy. Historically, China’s consumer internet has splintered into competing walled gardens—share a Taobao link in WeChat and it resolves as plaintext, not a preview card. Unlike the more interoperable Western internet, China’s tech giants have long resisted integration with one another, choosing to wage platform war at the expense of a seamless user experience. But the use of mini‑programs has given WeChat unprecedented reach across services that once resisted interoperability, from gym bookings to grocery orders. An agent able to roam that ecosystem could bypass the integration headaches dogging independent startups. Alibaba, the e-commerce giant behind the Qwen model series, has been a front-runner in China’s AI race but has been slower to release consumer-facing products. Even though Qwen was the most downloaded open-source model on Hugging Face in 2024, it didn’t power a dedicated chatbot app until early 2025. In March, Alibaba rebranded its cloud storage and search app Quark into an all-in-one AI search tool. By June, Quark had introduced DeepResearch—a new mode that marks its most agent-like effort to date.  ByteDance and Alibaba did not reply to MIT Technology Review’s request for comments. “Historically, Chinese tech products tend to pursue the all-in-one, super-app approach, and the latest Chinese AI agents reflect just that,” says Li of Simular, who previously worked at Google DeepMind on AI-enabled work automation. “In contrast, AI agents in the US are more focused on serving specific verticals.” Pei, the researcher at Stanford, says that existing tech giants could have a huge advantage in bringing the vision of general AI agents to life—especially those with built-in integration across services. “The customer-facing AI agent market is still very early, with tons of problems like authentication and liability,” he says. “But companies that already operate across a wide range of services have a natural advantage in deploying agents at scale.”

The Woven Apartment, designed by Khang Concept, is where the curves do the talking. The architectural forms create a statement of their own, letting understated design have its moment! The color palette is dominated by calming neutrals. Simultaneously, organic textures come together to add coziness. Whether it’s the arched details, woven finishes, or the muted palette, this home proves that minimalism doesn’t have to feel cold… it can feel like a hug. In the living room, a sculptural cream-toned sofa curves gently along the wall. This is grounded by a layered rug and organic coffee table. Floor-to-ceiling sheer curtains filter in natural light. The built-in bench seating and abstract wall art add character without clutter. The kitchen in the Woven Apartment is clean-lined and calming. It is a blend of matte cabinetry, wood textures, and soft under-cabinet lighting. The palette sticks to soothing tones, while the oversized hood and sleek open shelving bring drama. Tucked into a cozy nook, the dining area is all about organic flow. A custom-built banquette curves around a round black pedestal table, paired with a mix of rattan and sculptural chairs. A pendant lamp with natural textures floats above. This corner invites long, lingering meals and easy conversations. This bedroom is simple yet sophisticated. A soft upholstered bed adds functionality, while the wooden cabinetry provides storage space. A nightstand holds a globe lamp and a plant, adding to the decor of the otherwise minimalist room. The bedroom is a tranquil retreat. Custom wardrobes with woven panels frame a niche styled with ambient lighting. To the side, a sculptural desk with a chunky black leg and floating vanity shelves makes the space multifunctional. Natural light filters through full-length curtains, brushing softly against neutral bedding and organic textures. Playful yet polished, this bedroom features a soft, rainbow-arched headboard that adds charm. The mint green desk and sculptural shelves introduce a fun pop of color. With built-in wardrobes and a cozy workspace, it’s perfect for comfort as well as creativity. This bathroom continues the use of curves, as seen in the mirror. A terracotta-hued shower cubicle is segregated with a glass partition, keeping the space airy. The second bathroom uses the same terracotta color, but this time for the area around the vanity. This time, a curved partition separates the shower area. Matte black fixtures add visual interest.

Cellebrite said the deal will help with the "accelerated identification of mobile vulnerabilities and exploits."

News Earth Small earthquakes can have a big impact on the movements of major faults The little quakes can disrupt the growth of slow-slip events along megathrust faults Megathrust faults can generate some of Earth’s largest quakes. One example is the 2011 Tohoku earthquake, which rocked Japan and fomented a tsunami that devastated coastal areas, as shown. These same faults can also release immense amounts of seismic energy over days or weeks in harmless slow-slip events. enase/Getty Images By Nikk Ogasa 15 hours ago Seismic symphonies of minor earthquakes may affect grand movements on major faults. Small and distant earthquakes can disrupt the growth of slow-slip events — gradual fault movements that can release tremendous amounts of energy at gentle tempos, a new analysis of seismic data suggests. Reported in the May 16 Science Advances, the research shows that the more frequently small earthquakes occur near a fault’s slow-slip zone, the less synchronized the slipping becomes. Megathrust faults — massive fractures where one tectonic plate pushes under another — are notorious for hatching Earth’s most devastating temblors. These same faults can also slide steadily for days or weeks in slow-slip events while emitting faint vibrations called tectonic tremor. First identified around two decades ago, these softly humming events can shift stress on faults and may influence the timing of large, destructive earthquakes. But it’s not clear why these enigmatic events can grow for hundreds of kilometers along some faults while remaining restricted to small parts of others. Sign up for our newsletter We summarize the week's scientific breakthroughs every Thursday.