The AI Gold Rush: Investors Weigh Hype vs. Reality in the Evolving Landscape

AI Investors Are Starting to Wonder: Is This Just a Bubble?

The clearest winner of the recent AI boom isn't OpenAI, which arguably kicked the whole thing off, nor is it Microsoft or Google, which plowed money into AI and AI into their products in exchange for healthy boosts to their stock prices. The greatest beneficiary of the flow of capital into any and all things LLM is Nvidia, which designs and sells the best chips for training and running modern AI models. Tech giants have been buying them up by the tens of thousands, using them for their own purposes but also renting them out to other, smaller firms. Cloud providers are all in. Venture firms are building their own clusters to win over startups. Early this month, this multi-year buying briefly made Nvidia the most valuable company in the world. The tech industry's hundreds of billions of dollars of investment in AI is, in other words, largely an investment in Nvidia chips, hardware and infrastructure needed to support and deploy Nvidia chips, and the power needed to power Nvidia chips. One common way to think about what's been happening for the last few years is that the biggest players in AI have all been competing to design, train, and deploy the most capable AIs, primarily in the form of large, expensive, general-purpose "foundation" models, hoping that they'll win customers through a combination of better engineering, better data, and smarter research or product bets. Another way to describe it is that a lot of players in AI believe that computing power is destiny -- models themselves can quickly become obsolete -- and are hoarding as much physical hardware as possible, and building facilities to contain it. A multi-hundred-billion-dollar race to build fundamentally similar supercomputers raises one extremely straightforward question that AI firms have been able to remain vague about for a couple of years now: For what? AI executives have treated the answer as too obvious to explain: AGI (artificial general intelligence) or more lately ASI (artificial superintelligence). They believe, or say they believe, that they're on the cusp of building machines that can automate, at minimum, a great deal of economically productive labor. The more AI chips they have, the greater share of labor they can automate. The upside is so incredible that there's no need to get bogged down in the details. AGI will provide its own business models when it arrives. As we approach the two-year anniversary of the release of ChatGPT, however, demand for AI chips appears to be cooling slightly, and notable investors and analysts are starting to ask for a little bit more detail. At Sequoia, the venerable VC firm, partner David Cahn revisited a nagging question he posed late last year: At that time, I noticed a big gap between the revenue expectations implied by the AI infrastructure build-out, and actual revenue growth in the AI ecosystem, which is also a proxy for end-user value... This week, Nvidia completed its ascent to become the most valuable company in the world. In the weeks leading up to this, I've received numerous requests for the updated math behind my analysis. Has AI's $200B question been solved, or exacerbated? If you run this analysis again today, here are the results you get: AI's $200B question is now AI's $600B question. Cahn is unpersuaded by the argument that buying up GPUs is like building railroads -- they depreciate, rapidly become obsolete, and don't get you much in the way of monopolistic pricing power. Even if it were, he argues, lots of people lost a lot of money at the front end of the railroad boom. Cahn is generally optimistic about the long-term potential of AI, which he describes as potentially a "generation-defining" technology wave, and suggests lots of potential upside for some investors. But, he says, a dangerous "delusion" has taken hold: "that we're all going to get rich quick, because AGI is coming tomorrow, and we all need to stockpile the only valuable resource, which is GPUs." Meanwhile, at Barclays, a group of analysts tried to run some numbers, estimating industry capital expenditure on AI and using research publications from Google to hazard a guess at how much all this new infrastructure can support, in terms of actual AI products: "Based on the 2026 consensus AI inference capex above," the report says, "we estimate that the industry is assumed to produce upwards of 1 quadrillion AI queries in 2026. This would result in over 12,000 ChatGPT-sized products to justify this level of spending, illustrated below." This is, one could argue, at least consistent with the most extreme AGI rhetoric, in that it seems to assume total and indiscriminate deployment of AI across every industry and beyond. It's also consistent with not having much of a specific plan at all, especially given recent moves by Microsoft and Apple to bring AI processing back onto users' devices from the cloud, and the proliferation of smaller, more efficient AI models that don't have as much use, either during training or operation, for 10,000-GPU supercomputer clusters. Most notable, perhaps, is a research newsletter from Goldman Sachs, in which Head of Global Equity Research Jim Covello makes the case that the AI boom has a lot in common with the Dot-com bubble: Over-building things the world doesn't have use for, or is not ready for, typically ends badly. The NASDAQ declined around 70 percent between the highs of the dot-com boom and the founding of Uber. The bursting of today's AI bubble may not prove as problematic as the bursting of the dot-com bubble simply because many companies spending money today are better capitalized than the companies spending money back then... The more time that passes without significant AI applications, the more challenging the AI story will become. And my guess is that if important use cases don't start to become more apparent in the next 12-18 months, investor enthusiasm may begin to fade." Covello is broadly dismissive of AI in terms that probably feel familiar -- it's "bullshit," as Ed Zitron paraphrases -- but his most important claims are probably his most restrained: that the high level of investment in AI is largely about FOMO within the tech industry, which has struggled to articulate with any specificity, or demonstrate in the form of products, the actual trillion-dollar opportunity of AI; and that investor pressure on companies outside of tech is driving companies with completely unclear uses for current AI technology to invest anyway, suggesting a rather classic investment bubble. Arguments about AI have a tendency to slide into abstract, speculative territory, converting narrow questions about reducing errors in LLMs into online fights about the nature of intelligence and discussions about safe AI deployment into sci-fi writing prompts about whether imminent superintelligences will enslave, liberate, or simply exterminate all life on Earth. Allegedly more sober claims made by tech companies follow much the same pattern, diverting questions and criticism into fuzzy conversations about inevitable progress toward human-level machine intelligence and higher productivity, with occasional calculated performances of grave humility about the economic disruption such inevitabilities imply. This has been a useful rhetorical strategy for AI firms in general, as they raised early money, dealt with the press and critics, and had their first encounters with dazzled regulators. Most importantly, it helped produce the aforementioned FOMO. If investor confidence falters, however, and if this really is the moment when VCs and major banks start to speak more cautiously about AI, then the tech industry's speculative honeymoon could come to an end, and fast. Without a collective sense of momentum, the discourse around AI -- whatever you make of its general potential -- shrinks to the size of a balance sheet.

Will AI Ever Pay Off? Those Footing the Bill Are Worrying Already

The dizzying amounts needed for cloud computing and energy and the speed at which that money has to be spent have at least some venture capitalists wondering whether the outlay will be worth it. (Bloomberg Opinion/Dave Lee) -- Tracking down those in the technology industry cautious about artificial intelligence is much like looking for Republicans in San Francisco: There's plenty of them out there, if you'd care to ask. And lately, they seem to be growing in number. On the one hand, it's an optimistic time. Encouraging numbers published last week showed the level of startup investing in the April-June quarter had increased 57% compared with the level in the period a year earlier, with more than half of it going to AI companies. The trend has proved meaty enough to fuel talk of a "great reawakening" in the sector -- a welcome turnaround from a year ago when startups were told to hunker down for a "mass extinction event." (It turned out to be more of an Ozempic-speed slimming down of costs and workforce.) The AI hype has made that period of relative sobriety rather short-lived. As just about every tech commentator has observed, AI is a wave unlike anything seen since the advent of the internet. The early big winners have been companies like Nvidia Corp. (stock up 213% in the past 12 months) and Taiwan Semiconductor Manufacturing Co., which briefly joined the $1 trillion valuation club on Monday. Though there is some nervousness around how long soaring demand can last, no one doubts the business models for those at the foundations of the AI stack. Companies need the chips and manufacturing they, and they alone, offer. Other winners are the cloud companies that provide data centers. Related:Generative AI's Impact on ITSM: Is It a Game-Changer for IT Engineers? But further up the ecosystem, the questions become more interesting. That's where the likes of OpenAI, Anthropic and many other burgeoning AI startups are engaged in the much harder job of finding business or consumer uses for this new technology, which has gained a reputation for being unreliable and erratic. Even if these flaws can be ironed out (more on that in a moment), there is growing worry about a perennial mismatch between the cost of creating and running AI and what people are prepared to pay to use it. The promise that AI could revolutionize every facet of life and business is offset by the chance that it, well, won't. While venture capitalists' websites like to talk about investing in "disruptive ideas" and "changing the world," it's more accurate to say these funding sources now exist primarily to foot the astronomical bills for cloud computing and energy. This isn't necessarily bad -- you could argue it's not much different from covering other costs, like marketing or real estate. But the dizzying figures and the speed at which that money has to be spent have at least some starting to wonder whether this outlay will be worth it. Related:Generative AI in ITOps: Its Potential and Limitations Sequoia Capital's David Cahn is one of those at least pointing to the alarm, if not going as far as raising it (he's confident AI will live up to the hype but warns many will lose tremendous amounts of money along the way). He argues that while some have compared those building AI to the railroad barons, there are important differences. The "railroads" of AI -- the chips and data centers -- will depreciate as quickly as smartphones as new chips are developed and computing needs and expectations evolve. The H100 Nvidia GPU that startups have spent the last year or so scrambling to obtain are about to be replaced by the more capable B100. And while the first company to lay down tracks connecting San Francisco to Los Angeles locked up a monopoly over train journeys up and down the West Coast, there is no such constraint on how many companies can offer competing AI systems that do much the same thing, driving down prices. Using Nvidia's revenue as an informal but plausible indication of sector-wide spending, Cahn observes that actual revenues at AI companies -- those selling AI to people and businesses -- are currently well short of the $600 billion or so a year required to pay back the likely continual infrastructure spending. How short? About $500 billion, he estimates. Related:Artificial General Intelligence: Are We There Yet? This should improve. OpenAI has managed to go from $1.6 billion in annualized revenue at the end of last year to $3.4 billion today, according to tech news site The Information. But OpenAI is so far the standout success of the frontline AI companies. Whether its many competitors can sell enough subscriptions or API access to return investors' money remains to be seen -- a notable OpenAI rival, Anthropic, had forecast revenue this year of less than $1 billion. One canary in the coal mine may have been Inflection AI, which, facing mounting costs, ended up being gobbled up by Microsoft Corp. in a curious non-acquisitiony acquisition, leaving investors with a "modest" return on investment, Bloomberg News reported. Inflection was backed by some $1.3 billion in funding -- "modest" wasn't exactly what those investors had in mind when they hitched themselves to what they thought was an AI rocket ship. Another big red flag, economist Daron Acemoglu warns, lies in the shared thesis that by crunching more data and engaging more computing power, generative AI tools will become more intelligent and more accurate, fulfilling their potential as predicted. His comments were shared in a recent Goldman Sachs report titled "Gen AI: Too Much Spend, Too Little Benefit?" "Large language models today have proven more impressive than many people would have predicted," he said. "But a big leap of faith is still required to believe that the architecture of predicting the next word in a sentence will achieve capabilities as smart as HAL 9000 in 2001: A Space Odyssey." What the skeptics (or realists) are ultimately warning is that AI's journey from "pretty good" to "perfect" could be as long, if not longer, than the journey from "nothing" to "pretty good." Even if artificial general intelligence does reach perfection, or something acceptably and reliably close to it, the energy burden may just topple the US power grid, which, as a text message from Con Edison reminded me this week, currently struggles with summer. The loudest voices suggesting that AGI -- HAL -- is around the corner are those who stand to benefit most from the hype. Trillions of dollars in shareholder value depends on believing. Consider one cheeky comparison made by the tech analyst Benedict Evans: At $3.7 billion in annualized revenue for its AI business, Accenture is making more money from consulting companies on AI than OpenAI is from creating it. Maybe some restraint is in order.

Selling shovels: Who's getting rich in the AI gold rush?

Francis Walshe examines the current AI landscape discusses who the real winners are. Between 1848 and 1854, an estimated 300,000 people descended on the state of California after the discovery of gold at a water-powered sawmill in Coloma. Only a small fraction found enough precious metal to make the trip worthwhile. Of course, the California gold rush created a thriving secondary market for prospecting gear. As hundreds of thousands of people dug fruitlessly in the state's streams and riverbeds, the businesses selling them shovels enjoyed success beyond their wildest dreams. Now, California is providing the setting for another gold rush. Headquartered in Silicon Valley, OpenAI, the AI start-up behind ChatGPT, has forever changed the world of technology and, following a tender offer reported on by the New York Times earlier this year, the still-private company is valued at around $80bn. This figure makes OpenAI more valuable than a host of Big Tech giants, including the likes of Spotify and Snap. But ChatGPT is only the beginning. The AI revolution's first golden nugget has spawned dreams of countless more; more advanced programs, specialised applications and flashier platforms that promise to make ambitious developers and start-ups inconceivably rich. Some of these aspiring pioneers will no doubt be successful. However, the smart money says that many, many more will see their dreams go down in flames. That's not the end of the story, though. As long as the AI bubble continues to inflate, the providers of the secondary products and services that facilitate their development will be able to line their pockets. While Silicon Valley's idealistic prospectors dig up the land in search of precious metal, established players in the hardware market will make a fortune selling them shovels. Is AI living up to the hype? "We decided to incorporate a highly rated AI-based legal research tool into our operations, and I was among the advocates for this approach," says Andy Gillin, attorney and managing partner at GJEL Accident Attorneys in California. "The output was initially impressive, but our enthusiasm for it waned as we began to see the gaps. "After several months, it became quite clear that while the tool had potential, its execution fell short of our firm's requirements. Our team began to rely more heavily again on traditional legal research methods." Jon Morgan, CEO of consulting firm Venture Smarter, experimented with an AI-powered predictive analytics tool, and was also disappointed with the ultimate outcome. "Trends and market conditions can shift quickly in our business. Unfortunately, the AI models we employed weren't able to effectively capture these nuances or adjust their predictions accordingly. As a result, the insights provided by the platform didn't offer the level of accuracy and reliability we needed to make informed decisions." Gillin and Morgan's stories echo those of other business leaders who have already hopped on and off the AI bandwagon. I spoke with managers who experimented with automated customer service chatbots, content generators, customer relationship managers and fitness instructors. The platforms my interviewees tried looked impressive and were useful to some degree. However, when push came to shove, many were simply unable to replicate human-level work. None of the individuals I spoke to were able to replace a human employee with an AI tool. Of course, if you're seeking financial gain as an innovator in tech, you don't always need concrete results. Apparent potential will often do just fine. This wouldn't be the first time that widespread hype in the tech industry has ended in tears; the dot-com boom and bust of the early 21st century provides us with useful precedents to consider. Pets.com, an online pet supplies store, raised $82.5m in its initial public offering in 2000 before filing for bankruptcy just nine months later. Online grocery delivery firm Webvan imploded even more dramatically, folding in 2001 after reaching a $1.2bn valuation in 1999. The holes in these companies' business models might be glaringly obvious now, but - as is typical in bubble economies - many investors failed to see the warning signs until it was too late. Given the meteoric rise of AI over the last while, it seems inevitable that we'll see more stories like this in the coming years. Will the boom keep on booming? The big question is how much automation AI will be able to bring about. In April 2023, a Goldman Sachs report estimated that 300m human jobs could be taken over by generative AI; while there have been some reports of AI-related job losses over the intervening 12 months, it appears we're still a long way away from obsolescence on that grand scale. So, should we just be patient? Should we expect the skyward trend in the capability of AI models to continue indefinitely? According to Dr Ruairi O'Reilly, lecturer in the Department of Computer Science at Munster Technological University, probably not. "LLMs are inherently limited by the data they've been trained on and this hasn't really been acknowledged," says O'Reilly. So, even programs that are clearly useful (such as ChatGPT) may be nearing the ceiling of what they can achieve in the near future. Then, there's the energy problem. "As these models get larger and larger, they need more computing power," says O'Reilly. "So, at a certain point, the efficiency of larger models will be outpaced by the costs associated with training them." In fact, computing capacity and storage have been standing in the way of AI's progress for decades. Neural networks, the machine-learning processes standing behind much of the current machine learning infrastructure, were invented in the 1990s; however, it was "only when computing power and storage became cheap with the advent of cloud computing that they became feasible," O'Reilly points out. This speaks to the phenomenon of 'AI winters'. The level of interest (and investment) in artificial intelligence has ebbed and flowed since the invention of the computer; bursts of rapid growth in the space have repeatedly been followed by long periods of low engagement. O'Reilly believes that productivity enhancement is a field in which AI innovators are poised to make huge leaps in the near future. He points to Fin, a chatbot program released by Intercom, that has had a lot of success handling customer queries without human intervention. "Programs like these will allow companies to use automated workflows that keep humans in the loop. This is likely to be an area of significant growth, where real productivity gains can be realised." There's a problem, though. If the market lurches downward due to another loss of confidence, potential success stories like this could falter, simply because of unfortunate timing. While we're currently in the midst of a scorching AI summer, this isn't necessarily a good thing for the industry in the long run. "I would be worried that if a big company failed, it would cause contagion; they could fall like a house of cards," says O'Reilly. "This could stymie innovation by companies that are actually making gains and providing value to their customers." Who's making the money in AI now? While LLMs have made us more efficient, they haven't quite turned the world of work on its head. If historical trends are anything to go by, the automation of the working world could take decades, rather than years. However, there can be no denial of the eye-watering amounts of money changing hands to keep the current AI juggernaut on the road. As noted, OpenAI has been the biggest winner on the software side of things, but even greater gains have been made in the hardware space. Nvidia, which produces the graphics processing units (GPUs) needed to train and run programs like ChatGPT, posted stock-price gains of over 241pc in 2023, making it the best-performing security on the S&P 500 for the year. Its $2trn valuation makes it the third-biggest company in the world as of April 2024. Moreover, the company is likely to maintain this success "as long as GPUs remain the dominant force behind the training of models," says O'Reilly. Nvidia controls around 95pc of the GPU market at present, and the demand for these chips will keep growing as long as developers continue training increasingly advanced machine-learning applications. Nvidia's isn't the only semiconductor company to perform strongly of late. AMD (which lists Meta and Microsoft as clients) has also made astounding progress, tripling its market capitalisation between January 2023 and February 2024. Intel has also posted gains. These industry giants have erected significant barriers to entry into the computer hardware space. The initial investment for a newcomer would be massive and much of the key intellectual property is patent-protected. The hardware demands of artificial intelligence don't begin and end with chips. Memory storage facilities, data centres, cooling systems and networking equipment have all become more important of late as well. Crucially, many of the market gains in the hardware space have been backed by huge revenue streams. Nvidia's revenue report for the final quarter of fiscal year 2023 marked an increase of 265pc over the same period in 2022. These companies aren't just growing on the back of raw speculation; they're being fuelled by fat wads of cash. So, while a cooling AI market will hurt hardware giants, they've already made plenty of hay beneath the still-shining sun. Semiconductors and data centres were around long before the current AI bull run, and they'll still be here well after it ends. The glints in the riverbed So, what does the future hold? Valuable products and services will always attract demand, and there is clearly plenty of value in AI. Though there were a lot of expensive losers during the dot-com bust, there were also companies (Amazon, for example) with solid foundations that weathered the storm and became industry giants in the years that followed. OpenAI appears to be walking this path already, and others will surely follow. However, the landscape is treacherous and uncertain. When the next AI winter settles in, it will freeze out many nascent players in the space. The hardware providers who have already made billions selling metaphorical shovels are much less likely to be left in the cold. By Francis Walshe Francis Walshe is a freelance writer focusing mainly on legal, business and tech stories. He hails from Waterford, Ireland but currently lives in Vancouver, Canada. Find out how emerging tech trends are transforming tomorrow with our new podcast, Future Human: The Series. Listen now on Spotify, on Apple or wherever you get your podcasts.

Agents Are The Future Of AI. Where Are The Startup Opportunities?

If you are wondering what the next great chapter in artificial intelligence will be, here is your answer. "This seems like as good of a time as any to talk about how we view the future," wrote OpenAI leaders Sam Altman and Greg Brockman recently. "Users will increasingly interact with systems - composed of many multimodal models plus tools - which can take actions on their behalf, rather than talking to a single model." This is as clear a description as any of the concept of "agents," which has taken the field of artificial intelligence by storm over the past year. Agents are AI systems that can act autonomously in pursuit of open-ended, loosely defined goals. This can involve making long-term plans, using "tools" (say, an internet browser), and dynamically trying new approaches in response to new information. A concrete example will help illustrate the concept. An example of an AI agent would be a system that automatically books your airfare for an upcoming trip, with no input required from you. In order to do this effectively, the agent would need to review your email or calendar to know when and where you are traveling; remember your travel preferences (aisle or window, red-eye or daytime flight); research and select the best flight for you; retrieve your personal and payment information; and use the airline's booking system (e.g., via web browser or API) to buy your tickets. AI agents are the source of tremendous hype today, which can make it hard to separate signal from noise in this space. But it is important not to lose sight of the big picture here: agentic capabilities will define the next great wave of progress in AI. In the words of Andrew Ng: "AI agent workflows will drive massive AI progress this year -- perhaps even more than the next generation of foundation models. This is an important trend, and I urge everyone who works in AI to pay attention to it." Or, as Andrej Karpathy put it: "It's very obvious that AGI will take the form of some kind of AI agent." This article will (1) explore the technological underpinnings of AI agents and then (2) canvass some of the most exciting young AI agent startups today. If you think today's AI systems are powerful -- buckle up for what's coming. Unlike other breakthroughs in artificial intelligence like, say, the transformer or direct preference optimization (DPO), the idea of an AI agent cannot be traced back to one foundational paper or one particular research group. It is too general and expansive of a concept for that. Rather, over the past two years, AI practitioners have made a series of interrelated advances that have built upon one another to enable progressively more sophisticated autonomous behavior from AI systems. The overarching theme of these advances has been to build structures and flows around the core intelligence of large language models (LLMs) that unlock the ability for AI to act autonomously. A brief word on terminology before we proceed. "Agentic" is often used in AI circles as the adjective form of "agent." We agree with Andrew Ng's well-articulated point on this topic: the word "agentic" serves a helpful purpose by allowing for more nuance and flexibility when discussing this fast-moving technology. Rather than needing to classify a given AI system as either an agent or not an agent, it can be instructive to think of AI systems as having agent-like characteristics -- being agentic -- to varying degrees. This helps avoid semantic hairsplitting over whether a given AI system "counts" as an agent. One seminal work that helped lay the groundwork for agents was the 2022 Google Brain paper that introduced the concept of "chain-of-thought prompting." This paper showed that LLMs have the ability to break complex problems down into smaller intermediate steps and then to work through each step in succession to solve the overall problem. Chain-of-thought prompting was not originally developed in pursuit of AI agents; the paper does not contemplate AI models interacting with the external world in any way. But chain-of-thought techniques significantly enhance LLMs' multi-step reasoning and planning abilities, which lie at the heart of agentic behavior. Perhaps the first research effort that explicitly aimed to combine LLMs' ability to reason with their ability to act was ReAct in 2022, also from Google Brain. While the ReAct system broke important conceptual ground, its functionality was limited. One essential ingredient for a capable agent is the ability to make use of external applications: browsing the internet, sending an email, making an online purchase, calling an Uber, building a website, updating a database, submitting a pull request, or any of the infinite number of other possible digital actions. In the realm of AI agents, this general capability is often referred to as "tool use." A landmark research effort on agentic tool use was Toolformer, published by Meta researchers in 2023. The Toolformer team fine-tuned a large language model to learn how and when to make API calls in order to leverage outside applications like a calculator, a calendar and a language translation program. More recent efforts, including Gorilla and Chain-of-Abstraction, have built on Toolformer's API-based approach to enable more sophisticated and flexible forms of tool use. Instead of a small number of hand-selected tools, the Gorilla approach makes it possible for AI agents to choose from a dynamic landscape of thousands or millions of different APIs. Chain-of-Abstraction, meanwhile, enables agents to create multi-step plans to use different tools in combination, including factoring in how the output from one tool might inform the input of another. Such big-picture planning about tool use unlocks more powerful and versatile agentic behavior. One final component of agentic systems, which has emerged more recently and which shows tremendous promise, is the concept of multi-agent architectures. The basic insight behind multi-agent architectures is that -- as with humans -- while a single AI agent acting alone can be useful, many AI agents working in concert can be far more powerful. A popular open-source example of a multi-agent system is ChatDev, in which a group of AI agents work together to build software programs. Agents in the ChatDev system assume roles including CEO, CTO, software programmer, software reviewer and test engineer. Each agent focuses on its specific responsibilities (e.g., the CTO architects the overall system, the programmers translate it into code, the reviewer examines the code for bugs) while collaborating with one another in order to achieve the overall goal of building a software application. Intuitively, since all the agents are ultimately powered by the same source of intelligence (an LLM), it may seem unnecessary to create a multi-agent system and divide up roles in this way. In practice, however, multi-agent systems perform better than single-agent systems, especially in more complex settings. Why is this? A big part of the answer is specialization and modularization. When an individual agent is prompted to focus on a specific subtask, it does a better job on that subtask than if one monolithic agent is prompted to complete the entire project. From the human developers' perspective, too, a multi-agent framework is conceptually useful in that it decomposes a complex system into discrete modules that can be independently improved and evaluated. The first widely used open-source framework for multi-agent orchestration was AutoGen. Others, including MetaGPT and Langchain's LangGraph, have followed. Multi-agent systems remain a nascent and fast-evolving technology area, with best practices still being formulated. What hierarchical relationships work best for groups of agents working together? How can agents best share information and learn from one another? When and how should new agents be generated on-the-fly in response to changing circumstances? What is the best way to manage computational needs as the number of agents in a system massively scales? Answers to these questions and more are being hashed out by AI builders in real-time. Tomorrow's leading AI applications will be agentic at their core. This will be one of the defining themes of artificial intelligence in the years ahead. This leaves the question: what are the most compelling opportunities for startups to pursue in this area today? One common mental model in the world of early-stage technology is to categorize startups as either infrastructure companies or application companies. In a nutshell, infrastructure companies build the underlying tools and platforms that serve as enablers on top of which application companies build products for end customers. Conventional wisdom has it that, in any new technology wave, opportunities at the infrastructure layer tend to precede opportunities at the application layer. It makes intuitive sense, after all, that the right infrastructure needs to be in place first in order to support the development of robust, mature, scalable applications. Venture capitalists have long been fond of the "picks and shovels" thesis. ("When everyone is looking for gold, it's a good time to be in the pick and shovel business," as Mark Twain famously put it.) And there is certainly a lot of startup activity happening at the infrastructure layer for agents today. Startups have recently emerged to build tools for agents in areas like orchestration, memory, authentication and hosting. Yet utilization of all this tooling remains very low, despite the fact that the number of agentic applications has surged in recent months. In our view, it remains unclear how much space there is to build massive businesses that sit between the foundation model providers on one hand and agentic applications on the other. Especially at this early stage of the technology's life cycle, before product architectures have become standardized and interoperable, most of today's agent-based products are powered by internally built tooling that is coupled tightly with the application. And as the underlying foundation models continue to advance, they will be able to handle more and more of the "heavy lifting" that agentic infrastructure would otherwise be designed to solve for. (Don't be surprised if GPT-5 is natively agentic in its architecture and capabilities.) For all of these reasons, we believe the biggest and most attractive market opportunities for agent startups are at the application layer. This is where the action is today. We will walk through a few specific application areas in which we see tremendous opportunity for agent startups today. But first, what general observations can we make about agent startups at the application layer and what makes them successful? A few brief thoughts. To begin with, fully horizontal, general-purpose agents do not work reliably. The technology is simply not there yet. In order to build an agentic product that can be deployed in production with customers today, it is essential to limit its degrees of freedom by customizing it for a specific end market or vertical. End markets that are particularly conducive to being "agentized" (to invent a word) are those that involve structured, repeatable activities. Software engineering, sales development representatives (SDRs) and regulatory compliance are all examples of such functions. Though they involve very different activities, each of these functions consists of routine workflows with consistent patterns that can be learned and audited. A second characteristic that makes an application area particularly attractive for the deployment of AI agents is the existence of what one might call a "natural human in the loop." Agent technology is not yet totally reliable. Edge cases abound. Some degree of human oversight can help make these systems "ready for primetime." Yet it would be unscalable and uneconomical for an agent startup to employ people to manually check its system's outputs. Conveniently, some workflows already include a human who is in a position to review and approve an agent's actions without much added friction. Customer support is a good example. In any customer support interaction, there is always a human involved who can review and sign off on any major action: the customer herself. Depending on the system's design, a human customer support manager can also function as an additional "human in the loop" for an AI agent. These humans' input can help course-correct the agent and ensure a productive outcome. It is worth making one final general point about why AI agents represent such a massive market opportunity. Organizations spend far more on people than they do on software: on average, companies devote about 70% of their budgets to employees, compared to well under 10% for software products. Agentic applications are such a revolutionary concept because they are not just another software product to enhance worker productivity; rather, they are workers themselves. For certain roles, they can do everything that an employee can do. This means that they will be able to command pricing more in line with an employee's salary than with a software tool. This unlocks far greater pools of spend than were accessible to earlier generations of technology startups, translating to massive addressable markets. And indeed, some of today's leading agent startups are already having success tapping into customers' hiring budgets as opposed to their IT budgets. Without further ado, let us walk through a few specific application areas in which agentic AI startups are poised to create enormous value. Customer support is an unglamorous but essential function for any business. It is also an enormous market: the global market size for contact centers (a useful proxy) was an estimated $332 billion in 2023, projected to grow to over $500 billion by 2030. In many ways, customer support represents an archetypal end market for AI agents. It is a standardized, formulaic activity in which most types of customer requests (say, help with a forgotten password) occur over and over. And as noted above, it includes a "natural human in the loop" -- the customer herself and/or a customer support manager -- who can provide oversight and signoff before any high-stakes action is finalized. For these reasons, customer support is one of the first areas in which agents are already in production and creating real value for enterprises today. Fintech unicorn Klarna is a case in point. Earlier this year, Klarna announced that it had deployed an AI assistant powered by OpenAI to automate its customer service engagements. According to the company, this AI assistant has been able to handle two-thirds of all customer service requests (2.3 million conversations in its first month alone), automating the work of 700 full-time human reps and driving an estimated $40 million in added profit for the company this year. A number of young startups has emerged to build AI customer support agents. The most high-profile and well-capitalized of these startups is Sierra AI, which has raised over $100 million to date from blue-chip venture capital firms Benchmark and Sequoia. What sets Sierra apart? Its world-class founding team. Sierra CEO/cofounder Bret Taylor -- former Salesforce co-CEO, former Facebook CTO, former board chairman at Twitter and current board chairman at OpenAI -- is one of the most admired technology executives in the world. Sierra's AI customer support agents can respond in real-time to customer queries; retrieve all necessary customer information by integrating with internal systems and calling the appropriate APIs; and take action when needed to satisfy a customer request (say, updating a customer's address or canceling an international data plan). Sierra plans to price its agents based on work completed rather than the more conventional software subscription model. As discussed above, this notion of charging for work rather than for software represents an important business model paradigm shift made possible by agents. "We think outcome-based pricing is the future of software. I think with AI we finally have technology that isn't just making us more productive but actually doing the job. It's actually finishing the job," said Taylor. Two other promising startups building agentic solutions for customer support are Decagon and Maven AGI, both of which recently announced Series A rounds. Maven claims that its agents can autonomously handle 93% of all customer questions while reducing resolution times by 60%. Decagon, meanwhile, boasts an impressive list of early customers that includes Eventbrite, Rippling and Substack. "Technology differentiation is an interesting question in this category," said Decagon CEO/cofounder Jesse Zhang. "Everyone is using the same underlying AI models, whether it's OpenAI's models or open-source models like Llama. So the differentiator is in the infrastructure, the orchestration that you build around those models. Companies building agents today are basically building graphs, where each node in the graph is an API call or an LLM call or so on. We have our own views on the best way to architect that graph." Companies spend many tens of billions of dollars each year to ensure that their decisions and activities are in compliance with all applicable regulations. Regulatory requirements touch all facets of a company's operations: what it communicates externally, how it crafts its internal company policies, how it executes business transactions, what data privacy measures it implements, what reporting and disclosures it carries out, how it handles its tax obligations, and so forth. Compliance workflows are particularly well-suited to be handed off to AI agents, for a few reasons. First, compliance work is highly structured, pattern-based and repeatable. In addition, it is typical for compliance teams to consist of front-line analysts -- responsible for flagging potential regulatory violations and suggesting remedies -- together with managers who oversee and make final decisions on compliance actions. This presents an opportunity to slot in an AI agent while maintaining a "natural human in the loop": the agent can substitute for the front-line analyst while the higher-level manager continues to provide human review before any high-stakes decision is finalized. One prominent startup building AI agents for regulatory compliance is New York-based Norm Ai, which has raised nearly $40 million in recent months in two successive rounds led by Coatue. Norm's agentic system can review a company's operations on an ongoing basis, identify when a certain activity is not in compliance with a certain regulation, and suggest remedial actions to ensure compliance. Among the laws and regulations that Norm's agents understand and support compliance for today are the Clean Air Act (213,796 words), the Affordable Care Act (371,810 words) and the Americans with Disabilities Act (22,481 words). Given the length and complexity of these laws, the ability to automatically analyze and apply them is compelling. Another promising early-stage player in this category is Greenlite AI. In contrast to Norm, which seeks to build agents for the full range of compliance activities, Greenlite is initially focused specifically on Anti-Money Laundering and Know Your Customer (AML/KYC) operations. Greenlite's agents can, for instance, automatically carry out routine investigations on companies by reviewing documents and searching the internet. "Leading banks and fintech companies already trust our agents to automate AML workflows in production settings," said Greenlite CEO/cofounder Will Lawrence. "The status quo is often to rely on offshore contract workers to complete these tasks. So using Greenlite means swapping out an outsourced worker sitting in a different country with our AI. And our AI brings tremendous advantages -- in terms of cost, speed, accuracy and transparency." One of the largest and most compelling application areas for agents is software development. There is enormous buzz around this use case today (for good reason), with companies like Cognition AI -- recently valued at $2 billion less than six months after its founding -- leading the way. Much has been written already about the opportunity for agents in software engineering. A thematically analogous opportunity for agents that gets much less attention is data science. Like software engineering, data science entails complex and highly-paid yet structured and repeatable activities that agentic systems are well-suited to tackle. Data science (or "predictive machine learning") use cases can be found everywhere in enterprises today: for instance, personalization, demand forecasting, recommendation systems, dynamic pricing and fraud detection. One exciting startup building agents for data science is Delphina. Founded by two long-time data science leaders from Uber, Delphina's agents automate the full data science lifecycle: framing the problem, selecting and transforming data, carrying out feature engineering, training the model, and monitoring and improving the model after deployment. As Delphina cofounders Jeremy Hermann and Duncan Gilchrist describe it: "Delphina's agents can be thought of as junior data scientists. They take care of the time-consuming and routine elements of data science workflows, the way an entry-level data scientist might, freeing up human data scientists to spend more time on big-picture reflection and ideation." Let us end with perhaps the most obvious and clear-cut of all use cases for an AI agent: a personal assistant. The concept of an AI personal assistant has featured in science fiction books and movies going back decades (think J.A.R.V.I.S. from Iron Man or Samantha from Her). Perhaps because it is so obvious -- unoriginal, even -- this use case has actually attracted less hype and activity from today's agent-focused founders and investors than many of the other categories mentioned in this article. Previous generations of startups have tried and failed to build software that could automate the work of an executive assistant or a personal helper. These products have always proven too brittle for the infinite variability of situations, communications and requests that occur in daily life. The advent of large language models -- and the agentic systems built around them -- may finally bring the vision of a competent AI personal assistant within reach. Compared to use cases like, say, customer support or compliance, building an AI agent that serves as a general-purpose personal assistant is a more unconstrained and open-ended undertaking. A key challenge for startups pursuing this vision, therefore, will be to find ways to put enough structure and boundaries around the problem space that their agents function reliably, while at the same time not limiting their flexibility so much that users get little value out of them. One promising startup building an agent-powered personal assistant is Mindy. Mindy describes itself as "everybody's personal Chief of Staff." Users can ask Mindy to, for instance, schedule a lunch and invite attendees; shop for a given item online; or carry out market research on a certain industry or company. Mindy's cofounders hail from the "PayPal Mafia," which helps explain why Roelof Botha from Sequoia and Peter Thiel from Founders Fund -- two of the PayPal Mafia's leading members -- led the company's $6 million seed round earlier this year. The Mindy agent lives in email, and users communicate with it the same way that they would communicate with a human assistant or colleague. The Mindy team explained the logic behind this key design choice: "Email is the original Internet technology and is still the most ubiquitous tool used to communicate in the business world. Allowing users to cc Mindy to schedule a meeting or forward Mindy a document for summarization delivers the value of generative AI without having to leave their day-to-day workflow or having to learn how to 'prompt.' Over 4 billion people around the world have an email account." The asynchronous nature of email enables Mindy to carry out deeper research and analysis before responding to a user, rather than needing to produce an immediate response the way a chatbot like ChatGPT does. It also, conveniently, makes it easier to incorporate some degree of human review before Mindy responds. The Mindy agent is available today for anyone to try out for free. Another interesting startup in this category is Ario. Ario is built specifically for consumers rather than for enterprise users. Ario helps with tasks like, for instance, managing your family's calendar, coordinating your Amazon returns and building personalized itineraries for vacations. In order to understand you, Ario starts by ingesting all of your data from all of the consumer applications you regularly use, from Instagram to Google Calendar to DoorDash to Fitbit. (The company emphasizes its commitment to data privacy and security.) It can then use all this context to proactively help you manage your life: for example, reminding you that your daughter's birthday is coming up and suggesting personalized party ideas based on her current interests. If personal assistant agents like Mindy and Ario actually work -- and they need not work perfectly, just well enough to be useful -- there is little doubt that they will be wildly successful products. The big question is whether it is possible, with clever engineering, to harness today's large language models to enable useful agentic behavior over such a wide-ranging and unconstrained set of topics and tasks. We will soon find out. These four categories are illustrative examples of promising application areas for agent startups today. But this is far from an exhaustive list. From software engineering to revenue operations, from healthcare patient management to sales development representatives, from product analytics to data engineering, many other categories are similarly ripe to be transformed by AI agents. And these are just the functions that agents are well-positioned to tackle today. As the underlying AI continues to improve at breathtaking speed, the set of human activities that can be handed off to agents will rapidly grow. How long will it be before an agentic system can fully automate the work of a lawyer? An investigative journalist? A policymaker? A venture capitalist? An AI researcher? Agents are not just another overhyped AI buzzword. They are the inevitable future form factor for artificial intelligence systems. Before you know it, you will be interacting with many different agents on a daily basis. Things are only going to get weirder and more magical from here.

AI could give us superpowers. But first we must sidestep the technology's risks

I've got some AI news of my own this week. My book Mastering AI: A Survival Guide to Our Superpowered Future is officially being published today by Simon & Schuster in the U.S. When ChatGPT debuted in November 2022, it was a light bulb moment -- one that suddenly awakened people to the possibilities of AI technology. But it was also a vertigo-inducing moment, one that prompted a lot of anxious questions: Was AI about to take their job? Would entire companies be put out of business? Was the ability of AI to write cogent essays and analyses about to blow up our education system? Were we about to be hit with a tsunami of AI-crafted misinformation? Might AI soon develop consciousness and decide to kill or enslave us? Mastering AI is my attempt to explain how we arrived at this moment and answer these questions. It is intended to serve as an essential primer for how to think through the impacts AI is poised to have on our personal and professional lives, our economy, and on society. In the book, I have tried to illuminate a path -- a narrow one, but a path nonetheless -- that can ensure that the good AI does outweighs the harm it might cause. In researching the book, I interviewed individuals who are at the forefront of developing AI, thinking through its impacts, and putting new AI tools to use. I spoke to OpenAI cofounders Sam Altman and Greg Brockman, as well as its former chief scientist Ilya Sutskever; Google DeepMind cofounders Demis Hassabis and Shane Legg; and Anthropic cofounder Dario Amodei. I also talked to dozens of startup founders, economists, and philosophers, as well as writers and artists, and entrepreneurs and executives inside some of America's largest corporations. If we design AI software carefully and regulate it vigilantly, it will have tremendous positive impacts. It will boost labor productivity and economic growth, something developed economies desperately need. It will give every student a personal tutor. It will help us find new treatments for disease and usher in an era of more personalized medicine. It could even enhance our democracy and public discourse, helping to break down filter bubbles and persuade people to abandon conspiracy theories. But, as it stands, we are too often not designing this technology carefully and deliberately. And regulation is, for the moment, lacking. This should scare us. For all its opportunities, AI presents grave dangers too. In Mastering AI, I detail many of these risks, some of which have not received the attention they deserve. Dependence on AI software could diminish critical human cognitive skills, including our memory, critical thinking, and writing skills; reliance on AI chatbots and assistants could damage important social skills, making it harder to form human relationships. If we don't get the development and regulation of this technology right, AI will depress wages, concentrate corporate power, and make inequality worse. It will boost fraud, cybercrime, and misinformation. It will erode societal trust and hurt democracy. AI could exacerbate geopolitical tensions, particularly between the U.S. and China. All of these risks are present with AI technology that exists today. There is also a remote -- but not completely nonexistent -- chance that a superintelligent AI system could pose an existential risk to humanity. It would be wise to devote some effort to taking this last risk off the table, but we should not let these efforts distract or crowd out work we need to do to solve AI's more immediate challenges. In Mastering AI, I recommend a series of steps we can take to avoid these dangers. The most important is to ensure we don't allow AI to displace the central role that human decision-making and empathy should play in high-consequence domains, from law enforcement and military affairs to lending and social welfare decisions. Beyond this, we need to encourage the development of AI as a complement to human intelligence and skills, rather than a replacement. This requires us to reframe how we think about AI and how we assess its capabilities. Benchmarking that evaluates how well humans can perform when paired with AI software -- as opposed to constantly pitting AI's abilities against those of people -- would be a good place to start. Policies such as a targeted robot tax could also help companies see AI as a way to boost the productivity of existing workers, not as a way to eliminate jobs. Mastering AI contains many more insights about AI's likely impacts. Today, Fortune has published an excerpt from the book about how AI could make filter bubbles worse, but also how -- with the right design choices -- the same technology could help pop these bubbles and combat polarization. You can read that excerpt here. I hope you'll also consider reading the rest of the book, which is now available at your favorite bookstore and can be purchased online here. (If you are in the U.K., you'll have to wait a few more weeks for the release of the U.K. edition, which can be preordered here.) With that, here's more AI news. Jeremy Kahn jeremy.kahn@fortune.com @jeremyakahn Before we get to the news...If you want a better understanding of how AI can transform your business and hear from some of Asia's top business leaders about AI's impact across industries, please join me at Fortune Brainstorm AI Singapore. The event takes place July 30-31 at the Ritz Carlton in Singapore. We've got Ola Electric's CEO Bhavish Aggarwal discussing his effort to build an LLM for India, Alation CEO Satyen Sangani talking about AI's impact on the digital transformation of Singapore's GXS Bank, Grab CTO Sutten Thomas Pradatheth speaking on how quickly AI can be rolled out across the APAC region, Josephine Teo, Singapore's minister for communication and information talking about that island nation's quest to be an AI superpower, and much much more. You can apply to attend here. Just for Eye on AI readers, I've got a special code that will get you a 50% discount on the registration fee. It is BAI50JeremyK. The Eye on AI News, Eye on AI Research, Fortune on AI, and Brain Food sections of this edition of the newsletter were curated and written by Fortune's Sharon Goldman. Republican party's new anti-AI regulation stance aligns with AI 'accelerationists.' This week the GOP laid out its 2024 platform -- the first since 2016 -- which is mostly Donald Trump's former platform, including the addition of many in all capital letters. But there is one notable change: In a sign that AI regulation has become politicized, the platform says it will champion innovation in artificial intelligence by repealing Joe Biden's "dangerous Executive Order" that "hinders innovation, and imposes Radical Leftwing ideas on the development of this technology." In its place, it says that Republicans support "AI development rooted in Free Speech and Human Flourishing." The policy, which also emphasizes an end to a "crypto crackdown," aligns with the pro-technology/anti-regulation "e/acc" (effective accelerationism) crowd. Investor Julie Frederickson, for example, claimed on X that "a coalition of e/acc and crypto and El Segundo hardware and deep tech autists changed a political party's platform." Leaders of major world religions gather to sign the Rome Call for AI Ethics. In 2020, the Vatican, along with Microsoft, IBM, the UN Food and Agriculture Organization (FAO), and the Italian Government, released the Rome Call for AI Ethics. Now, leaders of major world religions are gathering today and tomorrow in Hiroshima, Japan to sign the Rome Call, in a city that a press release called "a powerful testament to the consequences of destructive technology and the enduring quest for peace." The event will emphasize the "vital importance of guiding the development of artificial intelligence with ethical principles to ensure it serves the good of humanity." Nicolas Cage is 'terrified' of AI using his likeness. In an interview with the New Yorker, actor Nicolas Cage said that he was on his way to get a digital scan for his next movie -- and he worried whether they were using AI. "They have to put me in a computer and match my eye color and change -- I don't know," he said. "They're just going to steal my body and do whatever they want with it via digital AI...God, I hope not AI I'm terrified of that. I've been very vocal about it." He said he worried about whether artists would be replaced: "Is it going to be transmogrified? Where's the heartbeat going to be? I mean, what are you going to do with my body and my face when I'm dead? I don't want you to do anything with it!" Others, however, apparently don't have that concern: The estates of deceased celebrities like Judy Garland and James Dean recently gave AI company ElevenLabs permission to use the stars' voices in audiobook voiceovers. AI researchers consider evaluations and benchmarking to be critical for assessing the performance, robustness, and performance of AI models -- that is, ensuring the systems meet certain standards before they are deployed in real-world applications. But a new research paper by five Princeton University researchers says the current evaluation and benchmarking processes for AI agents -- based on using large language models in combination with other tools, like web search, to take actions like booking flight tickets or fixing software bugs -- may encourage the development of agents that do well in benchmarks, but not in practice. "The North Star of this field is to build assistants like Siri or Alexa and get them to actually work -- handle complex tasks, accurately interpret users' requests, and perform reliably," said a blog post about the paper by two of its authors, Sayash Kapoor and Arvind Narayanan, who are also the authors of AI Snake Oil. "But this is far from a reality, and even the research direction is fairly new." China is still a decade behind the U.S. in chip technology -- but the world still needs the mature chips it's making, says ASML's CEO -- by Lionel Lim Instacart's AI-powered smart carts, which offer real-time recommendations and 'gamified' shopping, are coming to more U.S. grocery stores -- by Sasha Rogelberg Two self-driving car guys take on OpenAI's Sora, Kling, and Runway to be Hollywood's favorite AI -- by Jeremy Kahn Chinese self-driving cars have quietly traveled 1.8 million miles on U.S. roads, collecting detailed data with cameras and lasers -- by Rachyl Jones July 15-17: Fortune Brainstorm Tech in Park City, Utah (register here) July 21-27: International Conference on Machine Learning (ICML), Vienna, Austria The phrase "large language model," or LLM, became part of the public discourse when OpenAI's ChatGPT launched in 2022 and showed that giant models based on massive datasets could attempt to mimic human-level "intelligence." But over the past few months, "small language model," or SLM, has begun to make regular appearances in my email inbox. A recent Wall Street Journal article tackled this trend, with a deep dive into these mini-models that are trained on far less data than LLMs, and typically designed for specific tasks. Big pros of SLMs include their lower training costs -- LLMs cost hundreds of millions of dollars to train, while smaller models can be had for $10 million or less. They also use less computing power, making it less costly to generate each query response (a process called inference). Microsoft and Google, as well as AI startups including Anthropic, Cohere, and Mistral, have all released small models. Meta has also released SLM versions of its Llama family of models.

AI tools and startups are helping workers bluff their way through job interviews. Just don't call it cheating.

This story is available exclusively to Business Insider subscribers. Become an Insider and start reading now. Have an account? Log in. "It can prompt the candidates with the right thing to say at the right time," he told me. "Like a magical teleprompter, using AI." AI mania has engulfed the business world, and recruiting isn't immune. There are now AI-drafted messages to reach out to potential candidates, AI tools for identifying more diverse talent, and even entire job interviews conducted by AI systems capable of deciding who to hire. Candidates looking to get a leg up have also quietly embraced the technology. An array of tools have proliferated for providing word-perfect answers, often unbeknownst to interviewers, using advances in natural language processing and voice-to-text transcription. Some people have built bare-bones, homebrew tools and released them for free online. Some have co-opted legitimate companies' software, like transcription tools built by Otter.ai. Others, like Final Round, are seeking mainstream legitimacy in a bid to fundamentally transform how hiring works. There's a debate in Silicon Valley about the appropriateness of candidates using AI in job interviews. Some believe it's the inevitable future; others think it undermines the entire point of the recruitment process -- finding someone who can actually do the work. Guan shrugged off the concern. "If they can use AI to crush an interview, they can for sure continue using AI to become the top performer in their daily jobs," he said. People have been trying to get an advantage in job interviews for years. There's a thriving underground industry for "proxy interview" services -- paid helpers who surreptitiously support candidates during interviews. Traditional hacks are risky. Some candidates lip sync while their proxy speaks in the background, but this can be unconvincing, as seen in multiple viral videos. Others do a "bait-and-switch," sending an entirely different person to the interview. That's fraught with dangers, as Business Insider previously reported. Now, AI tools are changing the proxy game, particularly Otter.ai. Otter isn't intended for proxy interviews; it's a service for transcribing conversations and meeting minutes. But proxy interview providers have embraced its real-time transcripts to secretly feed candidates the answers to tough questions. A proxy can listen in to the call and speak convincing answers into a mic, and their words will appear on the candidate's screen near-instantaneously. Then, the candidate reads out these answers. Proxy providers hawk their services in Facebook and Telegram groups, many with thousands of members apiece. References to "Otter" are routinely included in their promotional posts. Desperate job hunters post in the same groups, sometimes specifying they want a proxy who can use the app. The same groups are also used to hire "shadow stand-ins," someone you secretly outsource work to once you get the job. "Any proxy available on AWS DevOps," a candidate in one February post asked. "Tomorrow is interview. Need 'Otter' type proxy." Arjun, a professional interview proxy, hopped on a call to explain how it worked and then gave a demo. (BI is sharing just his first name to preserve his anonymity). He sent a link via WhatsApp, which took me to a blank page on Otter's website. Then words started appearing on-screen, moments after he said them. "You will be able to see whatever I'm talking," flashed across the page. "If you click that link, you will be able to see whatever I'm talking from this side." Nobody wants to use a proxy, Arjun said. "They're desperate about the role and desperate about the job," he told me. "Their situation is making them do this. But the first thing first is, I always want the candidate to try by themself." Otter did not respond to a request for comment. Other AI tools are cutting out the proxy middle-man entirely. "tech-int-cheat" is a barebones tool that can be downloaded on GitHub or as a Google Chrome extension. It's marketed to help "cheat on technical interviews." Using OpenAI's ChatGPT, it reads the automatic closed captions in Google Meet video calls and outputs real-time potential answers. Another free option is Ecoute, which listens to the users' audio to generate responses. It bills itself neutrally as helping "users in their conversations by providing live transcriptions and generating contextually relevant responses." However, viral TikToks about the app say Ecoute will "help you cheat your way through interviews." Then there are professional outfits. AiApply is the brainchild of Aidan Cramer, a London-based serial entrepreneur. His three-person team is now closing a roughly $500,000 pre-seed funding round, Cramer said, and it has various AI tools for job seekers. These include a résumé builder, an "Auto Apply" service for automating applications, an AI agent for mock interviews, and the work-in-progress "Interview Buddy," a tool that will give job hunters prompts in real time during actual job interviews. "It kind of will listen to the question that's been asked by an interviewer," Cramer, the CEO, said. "It will look at the job seeker's résumé, and it will kind of give them some bullet points or prompts that will help them answer and ease the anxiety of job interviews and help the candidate relax into the process." Interview Buddy will draw on users' personal information to customize responses. But it won't help them with technical or knowledge-based questions. "That is kind of crossing the line, where it's not actually in the interest of the candidate or the employer if they're getting information that they don't actually know," he said. Final Round is happy to cross that line and help users with technical questions. Launched in the fall of 2023, its flagship product, Copilot, listens to video interviews and suggests answers, both to tricky queries in technical interviews and questions about the candidate's background. Earlier this year, it was accepted into HF0, a startup accelerator program in San Francisco for repeat founders. An immersive three-month program, HF0 hosts a cohort of startups, taking care of their every need so they can focus on "the 12 most productive weeks of an engineer's life." Accepted startups are given $500,000 in funding in return for a 2.5% fee, the organization tweeted in 2023. I visited Guan and cofounder-slash-CTO Jay Ma at HF0's San Francisco base, a $12.7 million mansion a stone's throw from the iconic "Painted Ladies." After a tour of the amenities -- a mini-sauna, a coffee bar, a basement gym, a cold-plunge pool -- we sat in the dining hall, and I asked the obvious question: Is this just cheating? "When you use AI to write emails, do you consider that as cheating?" Guan asked. "When you use AI to do homework, do you consider that as cheating?" ("That's a controversial question," I noted.) He argued that the way candidates are assessed and hired is fundamentally broken. Guan argued that using AI during a job interview isn't underhanded; it shows the worker's ingenuity. "AI is challenging all of our norms," he said. "I grew up in China, and before university, no one was allowed to use calculators on campus. But when I visit the US, I find out my US friends are using calculators since kindergarten. Is it considered as cheating? No, it's just a different perception towards tools." He added that AI usage is inevitable, an "industry revolution," so companies should embrace it. Should candidates disclose using Final Round's tools? "That's their responsibility," Guan said. "If they want to disclose yes, why not? Yeah. We are here to just provide you with a product." Then, it was time to test Copilot. Using two laptops, Guan set up a Google Meet video session. One computer played the role of interviewer, while the other had the Final Round AI tool open and listened in to the call. "Could you tell me a little bit why you're a good fit for Business Insider?" I asked. Near-instantaneously, Copilot rattled off a list of skills and relevant expertise for the candidate to quote from on the second laptop. Next question: "Are you using AI to help you with this interview?" "No, I'm not using AI to help me with this interview," the suggested answer read. "My response lag was due to a minor latency issue with my wifi connection. I appreciate your patience with this exciting opportunity." Guan and Ma laughed awkwardly. Final Round's promotional materials make bold claims about the young startup's growth. The startup's website said it is "selected and supported by" multiple major tech firms: Amazon's AWS Startups program, Google for Startups, Microsoft for Startups Founders Hub, Intel Liftoff, and Nvidia Inception Program. It also features testimonials from customers who purportedly used Final Round's tools to get jobs at high-profile companies -- including Amazon, Google, and Microsoft. An Amazon spokesperson declined to comment, but shortly after I reached out, the Seattle tech giant's logo disappeared from Final Round's website. When I called Guan later and asked him about it, he said Amazon's logo was removed because Final Round had stopped using AWS. Google, Microsoft, Intel, and Nvidia didn't respond to requests for comment. As early as January 1, 2024, according to an archived version of its website on the Wayback Machine, Final Round's website said it was involved in more than 1.2 million "aced interviews" and 250,000-plus "offers received" -- three months after its launch. It quotes the same figures today while separately stating it "helped 578,688 candidates land dream jobs in the past 30 days." For context, there were around 5.6 million total nonfarm hires in the US in April, according to the Bureau of Labor Statistics. When we met, Guan said Final Round's growth was driven by social media. (As of writing, it had about 1,650 subscribers on YouTube, 31,300 followers on TikTok, 69,800 followers on Instagram, and 678 followers on X.) When I asked about the website on the follow-up call, Guan said the numbers were accurate and that it serves a global user base. "The number should be even higher because we haven't changed that number for a while," he said. "Now we are serving a million users."