AI Regulation Stalls as Chevron Collapses and Tech Giants Jockey for Position

This Week in AI: With Chevron's demise, AI regulation seems dead in the water

Hiya, folks, and welcome to TechCrunch's regular AI newsletter. This week in AI, the U.S. Supreme Court struck down "Chevron deference," a 40-year-old ruling on federal agencies' power that required courts to defer to agencies' interpretations of congressional laws. Chevron deference let agencies make their own rules when Congress left aspects of its statutes ambiguous. Now the courts will be expected to exercise their own legal judgment -- and the effects could be wide-reaching. Axios' Scott Rosenberg writes that Congress -- hardly the most functional body these days -- must now effectively attempt to predict the future with its legislation, as agencies can no longer apply basic rules to new enforcement circumstances. And that could kill attempts at nationwide AI regulation for good. Already, Congress was struggling to pass a basic AI policy framework -- to the point where state regulators on both sides of the aisle felt compelled to step in. Now any regulation it writes will have to be highly specific if it's to survive legal challenges -- a seemingly intractable task, given the speed and unpredictability with which the AI industry moves. Justice Elena Kagan brought up AI specifically during oral arguments: Courts will fill that gap now. Or federal lawmakers will consider the exercise futile and put their AI bills to rest. Whatever the outcome ends up being, regulating AI in the U.S. just became orders of magnitude harder. Google's environmental AI costs: Google has issued its 2024 Environmental Report, an 80-plus-page document describing the company's efforts to apply tech to environmental issues and mitigate its negative contributions. But it dodges the question of how much energy Google's AI is using, Devin writes. (AI is notoriously power hungry.) Figma disables design feature: Figma CEO Dylan Field says that Figma will temporarily disable its "Make Design" AI feature, which was said to be ripping off the designs of Apple's Weather app. Meta changes its AI label: After Meta started tagging photos with a "Made with AI" label in May, photographers complained that the company had been applying labels to real photos by mistake. Meta is now changing the tag to "AI info" across all of its apps in an attempt to placate critics, Ivan reports. Robot cats, dogs and birds: Brian writes about how New York state is distributing thousands of robot animals to the elderly amid an "epidemic of loneliness." Apple bringing AI to the Vision Pro: Apple plans go beyond the previously announced Apple Intelligence launches on the iPhone, iPad and Mac. According to Bloomberg's Mark Gurman, the company is also working to bring these features to its Vision Pro mixed-reality headsets. Text-generating models like OpenAI's GPT-4o have become table stakes in tech. Rare are the apps that don't make use of them these days, for tasks that range from completing emails to writing code. But despite the models' popularity, how these models "understand" and generate human-sounding text isn't settled science. In an effort to peel back the layers, researchers at Northeastern University looked at tokenization, or the process of breaking down text into units called tokens that the models can more easily work with. Today's text-generating models process text as a series of tokens drawn from a set "token vocabulary," where a token might correspond to a single word ("fish") or a piece of a larger word ("sal" and "mon" in "salmon"). The vocabulary of tokens available to a model is typically determined before training, based on the characteristics of the data used to train it. But the researchers found evidence that models also develop an implicit vocabulary that maps groups of tokens -- for instance, multi-token words like "northeastern" and the phrase "break a leg" -- to semantically meaningful "units." On the back of this evidence, the researchers developed a technique to "probe" any open model's implicit vocabulary. From Meta's Llama 2, they extracted phrases like "Lancaster," "World Cup players" and "Royal Navy," as well as more obscure terms like "Bundesliga players." The work hasn't been peer-reviewed, but the researchers believe it could be a first step toward understanding how lexical representations form in models -- and serve as a useful tool for uncovering what a given model "knows." A Meta research team has trained several models to create 3D assets (i.e., 3D shapes with textures) from text descriptions, fit for use in projects like apps and video games. While there's plenty of shape-generating models out there, Meta claims its are "state-of-the-art" and support physically based rending, which lets developers "relight" objects to give the appearance of one or more lighting sources. The researchers combined two models, AssetGen and TextureGen, inspired by Meta's Emu image generator into a single pipeline called 3DGen to generate shapes. AssetGen converts text prompts (e.g., "a t-rex wearing a green wool sweater") into a 3D mesh, while TextureGen ups the "quality" of the mesh and adds a texture to yield the final shape. The 3DGen, which can also be used to retexture existing shapes, takes about 50 seconds from start to finish to generate one new shape. "By combining [these models'] strengths, 3DGen achieves very-high-quality 3D object synthesis from textual prompts in less than a minute," the researchers wrote in a technical paper. "When assessed by professional 3D artists, the output of 3DGen is preferred a majority of time compared to industry alternatives, particularly for complex prompts." Meta appears poised to incorporate tools like 3DGen into its metaverse game development efforts. According to a job listing, the company is seeking to research and prototype VR, AR and mixed-reality games created with the help of generative AI tech -- including, presumably, custom shape generators. Apple could get an observer seat on OpenAI's board as a result of the two firms' partnership announced last month. Bloomberg reports that Phil Schiller, Apple's executive in charge of leading the App Store and Apple events, will join OpenAI's board of directors as its second observer after Microsoft's Dee Templeton. Should the move come to pass, it'll be a remarkable show of power on Apple's part, which plans to integrate OpenAI's AI-powered chatbot platform ChatGPT with many of its devices this year as part of a broader suite of AI features. Apple won't be paying OpenAI for the ChatGPT integration, reportedly having made the argument that the PR exposure is as valuable as -- or more valuable than -- cash. In fact, OpenAI might end up paying Apple; Apple is said to be mulling over a deal wherein it'd get a cut of revenue from any premium ChatGPT features OpenAI brings to Apple platforms. So, as my colleague Devin Coldewey pointed out, that puts OpenAI's close collaborator and major investor Microsoft in the awkward position of effectively subsidizing Apple's ChatGPT integration -- with little to show for it. What Apple wants, it gets, apparently -- even if that means contentiousness its partners have to smooth over.

This Week in AI: With Chevron's demise, AI regulation seems dead in the water | TechCrunch

Hiya, folks, and welcome to TechCrunch's regular AI newsletter. This week in AI, the U.S. Supreme Court struck down "Chevron deference," a 40-year-old ruling on federal agencies' power that required courts to defer to agencies' interpretations of congressional laws. Chevron deference let agencies make their own rules when Congress left aspects of its statutes ambiguous. Now the courts will be expected to exercise their own legal judgment -- and the effects could be wide-reaching. Axios' Scott Rosenberg writes that Congress -- hardly the most functional body these days -- must now effectively attempt to predict the future with its legislation, as agencies can no longer apply basic rules to new enforcement circumstances. And that could kill attempts at nationwide AI regulation for good. Already, Congress was struggling to pass a basic AI policy framework -- to the point where state regulators on both sides of the aisle felt compelled to step in. Now any regulation it writes will have to be highly specific if it's to survive legal challenges -- a seemingly intractable task, given the speed and unpredictability with which the AI industry moves. Justice Elena Kagan brought up AI specifically during oral arguments: Let's imagine that Congress enacts an artificial intelligence bill and it has all kinds of delegations. Just by the nature of things and especially the nature of the subject, there are going to be all kinds of places where, although there's not an explicit delegation, Congress has in effect left a gap. ... [D]o we want courts to fill that gap, or do we want an agency to fill that gap? Courts will fill that gap now. Or federal lawmakers will consider the exercise futile and put their AI bills to rest. Whatever the outcome ends up being, regulating AI in the U.S. just became orders of magnitude harder. Google's environmental AI costs: Google has issued its 2024 Environmental Report, an 80-plus-page document describing the company's efforts to apply tech to environmental issues and mitigate its negative contributions. But it dodges the question of how much energy Google's AI is using, Devin writes. (AI is notoriously power hungry.) Figma disables design feature: Figma CEO Dylan Field says that Figma will temporarily disable its "Make Design" AI feature, which was said to be ripping off the designs of Apple's Weather app. Meta changes its AI label: After Meta started tagging photos with a "Made with AI" label in May, photographers complained that the company had been applying labels to real photos by mistake. Meta is now changing the tag to "AI info" across all of its apps in an attempt to placate critics, Ivan reports. Robot cats, dogs and birds: Brian writes about how New York state is distributing thousands of robot animals to the elderly amid an "epidemic of loneliness." Apple bringing AI to the Vision Pro: Apple plans go beyond the previously announced Apple Intelligence launches on the iPhone, iPad and Mac. According to Bloomberg's Mark Gurman, the company is also working to bring these features to its Vision Pro mixed-reality headsets. Text-generating models like OpenAI's GPT-4o have become table stakes in tech. Rare are the apps that don't make use of them these days, for tasks that range from completing emails to writing code. But despite the models' popularity, how these models "understand" and generate human-sounding text isn't settled science. In an effort to peel back the layers, researchers at Northeastern University looked at tokenization, or the process of breaking down text into units called tokens that the models can more easily work with. Today's text-generating models process text as a series of tokens drawn from a set "token vocabulary," where a token might correspond to a single word ("fish") or a piece of a larger word ("sal" and "mon" in "salmon"). The vocabulary of tokens available to a model is typically determined before training, based on the characteristics of the data used to train it. But the researchers found evidence that models also develop an implicit vocabulary that maps groups of tokens -- for instance, multi-token words like "northeastern" and the phrase "break a leg" -- to semantically meaningful "units." On the back of this evidence, the researchers developed a technique to "probe" any open model's implicit vocabulary. From Meta's Llama 2, they extracted phrases like "Lancaster," "World Cup players" and "Royal Navy," as well as more obscure terms like "Bundesliga players." The work hasn't been peer-reviewed, but the researchers believe it could be a first step toward understanding how lexical representations form in models -- and serve as a useful tool for uncovering what a given model "knows." A Meta research team has trained several models to create 3D assets (i.e., 3D shapes with textures) from text descriptions, fit for use in projects like apps and video games. While there's plenty of shape-generating models out there, Meta claims its are "state-of-the-art" and support physically based rending, which lets developers "relight" objects to give the appearance of one or more lighting sources. The researchers combined two models, AssetGen and TextureGen, inspired by Meta's Emu image generator into a single pipeline called 3DGen to generate shapes. AssetGen converts text prompts (e.g., "a t-rex wearing a green wool sweater") into a 3D mesh, while TextureGen ups the "quality" of the mesh and adds a texture to yield the final shape. The 3DGen, which can also be used to retexture existing shapes, takes about 50 seconds from start to finish to generate one new shape. "By combining [these models'] strengths, 3DGen achieves very-high-quality 3D object synthesis from textual prompts in less than a minute," the researchers wrote in a technical paper. "When assessed by professional 3D artists, the output of 3DGen is preferred a majority of time compared to industry alternatives, particularly for complex prompts." Meta appears poised to incorporate tools like 3DGen into its metaverse game development efforts. According to a job listing, the company is seeking to research and prototype VR, AR and mixed-reality games created with the help of generative AI tech -- including, presumably, custom shape generators. Apple could get an observer seat on OpenAI's board as a result of the two firms' partnership announced last month. Bloomberg reports that Phil Schiller, Apple's executive in charge of leading the App Store and Apple events, will join OpenAI's board of directors as its second observer after Microsoft's Dee Templeton. Should the move come to pass, it'll be a remarkable show of power on Apple's part, which plans to integrate OpenAI's AI-powered chatbot platform ChatGPT with many of its devices this year as part of a broader suite of AI features. Apple won't be paying OpenAI for the ChatGPT integration, reportedly having made the argument that the PR exposure is as valuable as -- or more valuable than -- cash. In fact, OpenAI might end up paying Apple; Apple is said to be mulling over a deal wherein it'd get a cut of revenue from any premium ChatGPT features OpenAI brings to Apple platforms. So, as my colleague Devin Coldewey pointed out, that puts OpenAI's close collaborator and major investor Microsoft in the awkward position of effectively subsidizing Apple's ChatGPT integration -- with little to show for it. What Apple wants, it gets, apparently -- even if that means contentiousness its partners have to smooth over.

A former OpenAI researcher sees a clear path to AGI this decade. Bill Gates disagrees

Hello and welcome to Eye on AI. And a happy early July Fourth to my U.S. readers. This week, I'm going to talk about two starkly different views of AI progress. One view holds we're on the brink of achieving AGI -- or artificial general intelligence. That's the idea of a single AI model that can perform all the cognitive tasks a human can as well or better than a person. AGI has been artificial intelligence's Holy Grail since the field was founded in the mid-20th century. People in this "AGI is nigh" camp think this milestone will likely be achieved within the next two to five years. Some of them believe that once AGI is achieved we will then rapidly progress to artificial superintelligence, or ASI, a single AI system that's smarter than all of humanity. A 164-page analysis of this AGI is nigh and superintelligence-ain't-far-behind argument was published last month by Leopold Aschenbrenner, entitled "Situational Awareness." Aschenbrenner is a former researcher on OpenAI's Superalignment team who was fired for allegedly "leaking information," although he says he was fired after raising concerns to OpenAI's board about the company's lax security and vetting practices. He's since reemerged as the founder of a new venture capital fund focused on AGI-related investments. It would be easy to dismiss "Situational Awareness" as simply marketing for Aschenbrenner's fund. But let's examine Aschenbrenner's argument on its merits. In his treatise, he extrapolates recent AI progress in a more or less linear fashion on a logarithmic scale -- up and to the right. He argues every year we see what he calls "effective compute," a term that includes both the growth in the size of AI models and innovations that squeeze more power out of a model of a given size, resulting in about a 3x increase in capability (the term Aschenbrenner actually uses for the gain is "half an order of magnitude," or half an OOM for short). Over time, the increases compound. Within two years, you've got a 10x increase in "effective compute." Within four years, 100x, and so on. He adds to this what he calls "unhobbling" -- a catchall term for different methods to get AI software to do better on tasks at which the underlying "base" large language model does poorly. In this "unhobbling" category, Aschenbrenner lumps human feedback that trains an AI model to be more helpful and telling it to use external tools like a calculator. Combining the "effective compute" OOMs and the "unhobbling" OOMs, Aschenbrenner forecasts at least a five OOM increase in AI capabilities by 2027 and quite possibly more depending on how well we do on "unhobbling." Five OOMs is a 10,000x increase in capability, which he assumes will take us to AGI and beyond. He titled the section of his treatise where he explains this "It's this decade, or bust." Which brings me to the other camp -- which might be called the "or bust" camp. Among its members is Gary Marcus, the AI expert who has been a perpetual skeptic that deep learning alone will achieve AGI. (Deep learning is the kind of AI based on large, multi-layer neural networks, which is what all the progress in AI since at least 2010 is based on.) Marcus is particularly skeptical of LLMs, which he thinks are unreliable, plagiarism machines that are polluting our information ecosystem with low-quality, inaccurate content and are ill-suited for any high-stakes, real-world task. Also in this camp is deep learning pioneer and Meta chief AI scientist Yann LeCun, who still believes deep learning of some kind will get us to AGI, but thinks LLMs are a dead end. To these critics of the AGI-is-nigh camp, Aschenbrenner's "unhobbling" is simply wishful thinking. They are convinced that the problems today's LLMs have with reliability, accurate corroboration, truthfulness, plagiarism, and staying within guardrails are all inherent to the underlying architecture of the models. They won't be solved with either scale or some clever methodological trick that doesn't change the underlying architecture. In other words, LLMs can't be unhobbled. All of the methods Aschenbrenner lumps under that rubric are just kludges that aren't robust, reliable, or efficient. On the fringes of this "or bust" camp is Bill Gates, who said last week that he thought current approaches to building bigger and bigger LLMs could carry on for "two more turns of the crank." But he added that we would run out of data to feed these unfathomably large LLMs before we achieve AGI. Instead, what's really needed, Gates said, is "metacognition," or the ability of an AI system to reason about its own thought processes and learning. Marcus quickly jumped on social media and his blog to trumpet his agreement with Gates' views on metacognition. He also asked if scaling LLMs won't get us to AGI, why waste vast amounts of money, electricity, time, and human brain power on "two more turns of the crank" on LLMs? The obvious answer is that there's now billions upon billions of dollars riding on LLMs -- and that investment won't pay off if LLMs don't work better than they do today. LLMs may not get us to AGI, but they are useful-ish for many business tasks. What those two turns of the crank are about is erasing the "ish." At the same time, no one actually knows how to imbue an AI system with metacognition, so it's not like there are some clear alternatives into which to pour gobs of money. A huge number of businesses have now committed to AI, but are befuddled by how to get current LLM-based systems to do things that produce a good return on investment. Many of the best use cases big companies talk about -- better customer service, code generation, and taking notes in meetings -- are nice incremental wins, but not strategic game changers in any sense. Two turns of the crank might help close this ROI gap. I think that's particularly true if we worry a bit less about whether we achieve AGI this decade -- or even ever. You can think of AGI as a new kind of Turing Test -- AI will be intelligent when it can do everything well enough that it's impossible to tell if we're interacting with a human or a computer. And the problem with the Turing Test is that it frames AI as a contest between people and computers. If we think about AI as a complement to human labor and intelligence, rather than as a replacement for it, then a somewhat more reliable LLM might well be worth a turn of the crank. AI scientists remain fixated on the lofty goal of AGI and superintelligence. For the rest of us, we just want software that works, and makes our businesses and lives more productive. We want AI factotums, not human facsimiles. With that, here's more AI news. (And a reminder, we won't be publishing a newsletter on July 4, so you'll next hear from the Eye on AI crew on Tuesday, July 9.) Jeremy Kahn jeremy.kahn@fortune.com @jeremyakahn Before we get to the news...If you want to learn more about where AI is taking us, and how we can harness the potential of this powerful technology while avoiding its substantial risks, please check out my forthcoming book, Mastering AI: A Survival Guide to Our Superpowered Future. It's out next week from Simon & Schuster and you can preorder your copy here. If you are in the U.K., the book will be out Aug. 1 and you can preorder here. And if you want to gain 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. Amazon hires team from AI agent startup Adept. Amazon hired more than half the employees working at AI startup Adept, including its CEO and cofounder David Luan. Adept had been one of a clutch of startups trying to use LLMs to build a successful AI agent that could use business software to automate tasks such as building and analyzing spreadsheets and creating sophisticated slide decks. The deal was structured similarly to Microsoft's non-acquisition earlier this year of AI startup Inflection, with most of the staff being hired but the Big Tech company making a large, lump sum payment to the remaining startup to license its technology (and make its investors, which had pumped at least $400 million into Adept to date, whole). The deals have been seen as a possible way to escape antitrust issues, although antitrust authorities are still probing the Microsoft-Inflection arrangement. Amazon says the Adept team will contribute to its own "AGI Autonomy" effort, which is thought to be working on agent-like AI for Amazon to both sell through its AWS cloud service and to use to power a new, more agent-like Alexa personal assistant and perhaps other AI agents too. You can read more from CNBC here. In Los Angeles, an AI chatbot for students falls flat. The city had paid an AI startup called AllHere $6 million to develop an AI chatbot called Ed that could serve as a personal tutor and educational assistant for the 500,000 students in the LA public school system. But months after winning the contract, AllHere's chief executive and founder left the company and it furloughed most of its staff, the New York Times reported. AllHere delivered a version of Ed that was tested with 14-year-olds but was then taken offline for further refinement that might not happen now that AllHere has all but ceased operating. The school district has said it still hopes to have the chatbot widely available in September. Meanwhile, it has been left with a website AllHere built that simply collates information from other ed tech apps the district uses and doesn't have the interactivity of a chatbot. OpenAI's ChatGPT caught making up links to investigative stories from media partners. That's according to tests by journalism think tank and research group Nieman Lab, which wanted to see if OpenAI's chatbot could correctly cite and link out to investigative stories from a number of publications that have signed deals to license their content to OpenAI. These include the Associated Press, the Wall Street Journal, the Financial Times, The Times (U.K.), Le Monde, El Pais, the Atlantic, the Verge, Vox, and Politico. Nieman Lab found the chatbot could not reliably link out to landmark stories from these publications. OpenAI told the journalism think tank that it has not yet rolled out a citation feature for ChatGPT. Runway puts its latest text-to-video model Gen 3 into wide release. The sophisticated model, which can produce hyperrealistic and cinematic 10-second-long video clips from text prompts, is now available for anyone to use for free, the company said. Gen 3 is designed to compete with other text-to-video models, such as OpenAI's Sora, which is still being tested with select users and has not been made generally available, and Kling, a model from Chinese company Kuaishou. China is to develop 50 new AI standards by 2026. That's according to a story from Reuters, which cites China's industry ministry. The ministry says it will promulgate more than 50 national and industrial standards for AI deployment. Missed connections. Tuhin Chakrabarty, who recently received his PhD in computer science from Columbia University, is fast developing a reputation for coming up with some clever tests of large language models' true abilities in the domain on which they ought to perform best: language. In the past, he has tested LLMs' abilities to both write short stories and serve as copilots for human writers, and in both cases found LLMs lacking. Now he and a group of fellow researchers from Columbia and Barnard College are back with another paper that points to a surprising weakness in LLMs' language skills. They looked at whether LLMs could solve the New York Times's Connections game, in which a player is given a jumbled grid of four groups of four words that are in some way related. The player has to figure out how to regroup the words into their four categories based on the connections between them. Often the groupings are tricky, based on slang usages, alternative definitions, and homophones. They then compared how the LLMs performed with the scores of both novice and expert human players. It turns out that even the best performing LLM -- which was OpenAI's GPT-4o -- could only completely solve 8% of the Connections puzzles, significantly worse than both novice and expert human players. You can read the paper on the non-peer-reviewed research repository arxiv.org here. Can anyone beat Nvidia in AI? Analysts say it's the wrong question -- by Sharon Goldman Hollywood tycoon Ari Emanuel blasts OpenAI's Sam Altman after Elon Musk scared him about the future: 'You're the dog' to an AI master -- by Christiaan Hetzner Exclusive: Leonardo DiCaprio-backed AI startup Qloo clinches $20 million investment from Bluestone Equity Partners -- by Luisa Beltran These are the nine AI startups that VCs wish founders would pitch them -- by Allie Garfinkle July 15-17: Fortune Brainstorm Tech in Park City, Utah (register here) July 21-27: International Conference on Machine Learning (ICML), Vienna, Austria Should we use AI to evaluate other students, employees, and managers? There's an interesting Wall Street Journal article that looks at how teachers are increasingly embracing a range of AI tools designed to help them grade essays in subjects such as English and history. The teachers who are using the tool find it saves them time. And the teachers are supposed to have the final say, looking over AI grading software's assessment and deciding whether they concur. But of course, that's not necessarily how teachers will use these systems. The temptation will be to largely defer to the AI-generated grades. And that's a problem, partly because when the Journal tested different AI grading software on the same paper -- which had received a 97% from the original human teacher who graded it -- it received a wide range of grades, between 100% and just 62%. Some teachers the Wall Street Journal interviewed said they noticed that the AI grader was far tougher on students than they would tend to be. Others said the AI system seemed to call out both minor failings and major ones equally, which might shake students' confidence. "They're sixth-graders, I don't want to make them cry," one said, explaining why she was reluctant to use the AI grading software. Others pointed out that the AI could only judge writing in the paper itself -- not assess the work in light of the student's overall progression. It looked only at the end result and not at the process or effort a student might have made to produce that piece of work. As a result, some veteran teachers found it morally repugnant that some of their fellow educators would allow an AI system to stand in for their own professional judgment and, critically, their own empathy. I think this last point is key. As AI becomes more capable and more ubiquitous, there will be a lot of areas where it will be tempting to deploy AI when we really shouldn't. And one of those areas is, I think, cases in which we make high-consequence judgments about another human being. That would include legal proceedings, of course. But it would also include grades that impact a student's further educational prospects. And it would include employee performance reviews -- another area where I've heard some managers have begun turning to AI. In many of these areas, proponents of AI software have argued that the software's impartiality, the fact that it is never tired or having a bad day, means it should replace sometimes fallible human judgment. But AI systems can be fallible and inconsistent too. What's more, I think what we want in these cases is not actually impartiality. What we want is a fair application of human empathy. We want there to be the opportunity to appeal to the emotions of our evaluators, for them to be able to consider the whole of our circumstances, and to reflect on those circumstances based on their own life experiences. AI has no life experience and can never offer true empathy. The result may be impartial. But it may not be just.

OpenAI's reported decision to give Apple a board seat is very revealing

After last year's brief, chaotic ouster of CEO Sam Altman, OpenAI reformed its board and key partner Microsoft got a seat as a non-voting observer. Microsoft had been blindsided by the Altman episode, which had been set in motion by the old board, and this was a way of ensuring it wouldn't get any further nasty surprises from the company into which it had invested $13 billion. But now, according to Bloomberg, another non-voting observer is coming: Apple, as represented by App Store chief Phil Schiller. Apple hasn't invested anything into OpenAI, but it is integrating OpenAI's models into its devices (for free) and Schiller's role is reportedly part of that agreement. It's not hard to see how the situation may become awkward, as Microsoft and Apple are longstanding rivals that will soon both be using OpenAI's models to power AI assistants on their competing products. Even by the excessively cozy standards of the U.S. AI sector, this may be too close for comfort. There will have to be a lot of recusing going on during certain board discussions, and that might not be enough to make this work. But even without considering the potential for future disputes too deeply, there's quite a lot to digest about the immediate implications of letting Apple observe OpenAI's board meetings. The first and most obvious point is that both parties in the Microsoft-OpenAI love-in are increasingly hedging their bets. It's been clear for a while that Microsoft is doing this -- unsurprisingly, given last year's turmoil. It absorbed key staff from the AI startup Inflection in March and is now reportedly developing a GPT-rivaling AI model of its own, with Inflection cofounder Mustafa Suleyman leading the effort. Recently, there have also been hints of OpenAI asserting its own independence in the relationship. The Information reported last week that OpenAI is now making more revenue from directly selling access to its models than Microsoft is making doing the same thing. Building ChatGPT into MacBooks and iPhones is also a big deal, but giving Apple an observer role would confirm that OpenAI and Microsoft are not as tightly intertwined as once seemed the case. Conversely, the reported development also suggests that Apple may not be hedging its bets as much as it appeared to be doing just a few weeks ago. It didn't bring Altman on stage for the announcement of the Apple-OpenAI partnership -- he was in the audience -- and reporting since then has suggested that Apple may yet also incorporate Google and Anthropic's AI models into its platforms, as alternatives to OpenAI's ChatGPT. But unless it also plans to take observer roles on each of those companies' boards, Apple seems set to give OpenAI and its technology an elevated status for a long while yet. This may also say something about Apple's progress in developing its own top-tier AI models, which should obviate the need for any such partnerships -- when they appear, and if they prove competitive. I asked all three companies for comment, but none has been forthcoming. More news below -- and see you on Friday, as Data Sheet will be taking tomorrow off for the U.S. holiday. Tesla shipments. Tesla announced its quarterly sales numbers yesterday and, while deliveries were down 5% year-on-year to 444,000 vehicles, they were a little better than what Wall Street expected (440,000 or fewer) and also represented a 15% increase from Q1. As the Washington Post reports, the news lifted Tesla's share price by over 10%. EU vs Temu and Shein. Much like in the U.S., there is official concern in the EU about the flood of cheap products coming from Chinese e-commerce platforms such as Temu and Shein. Now, according to the Financial Times, Europe is considering closing off the loophole that allows such platforms to ship their goods at such low cost: the "de minimis" rule that says low-value goods (the EU's threshold is €150 or $161) can be imported without the need to pay duties. Figma's AI boo-boo. The user-interface design service Figma has disabled its AI-powered tool for designing apps after it emerged that asking it to make a weather app repeatedly produced an apparent rip-off of Apple's Weather app. As The Verge reports, Figma denies heavily training its AI on existing apps; it says it used OpenAI's GPT-4o and Amazon's Titan Image Generator G1 models. Either way, this can be chalked up as yet another controversy involving an AI seeming to regurgitate training material that someone else spent time and effort creating in the first place. -- The increase in Google's carbon emissions between 2019 and 2023, according to the company's latest environmental report, which pinned the blame on "increases in data center energy consumption and supply chain emissions." Google is trying to halve its emissions from the 2019 base year by 2030, but says energy-hungry AI creates "significant uncertainty" around achieving the target. Meta banned from mining data to train its AI models in Brazil, by the Associated Press Elon Musk got his mega-pay package, but Marc Benioff might not be so lucky, by Chloe Berger A former OpenAI researcher sees a clear path to AGI this decade. Bill Gates disagrees, by Jeremy Kahn Steve Ballmer, who was once Bill Gates' assistant, is now richer than his onetime mentor, by Marco Quiroz-Gutierrez Revolut billionaire Nik Storonsky embraces AI by launching 'truly systematic' $200 million VC firm, by Bloomberg Why Universal Hydrogen folded -- and why the company's mission will live on, by Allie Garfinkle Floppy-free Japan. Japanese Digital Minister Taro Kono has proclaimed victory in his war against the use of floppy disks in the government's systems. Floppy disks were once the most common form of portable data storage, but that hasn't been the case for over two decades now, and for most of the world they live on only as the thing that "save" icons are trying to depict. According to Reuters, Kono's ministry was set up during the COVID-19 pandemic, when Japan's reliance on archaic office technology became particularly troublesome. Kono, who took the reins in 2022 after a failed bid to become the country's leader, would also like to wean the civil service off fax machines.

The Prompt: The Crypto Miners Winning The AI Gold Rush

AI is now helping humans correct AI output. OpenAI has trained a new GPT-4-based model called CriticGPT to catch mistakes in code produced by ChatGPT. As generative AI models advance and their errors become more subtle, models like CriticGPT could help humans train the systems more efficiently by spotting more problems. Like other AI systems, CriticGPT is not fully accurate-but humans that use the tool to train AI outperform others by 60%. Now, let's get into the headlines. Popular AI chatbots like Microsoft Copilot and OpenAI's ChatGPT regurgitated disinformation about the presidential debate between President Joe Biden and former President Donald Trump on Thursday, NBC News reported. Answers produced by the chatbots repeated false claims from a conservative writer's post that there would be a delay in CNN's broadcast of the debate which would be used to edit parts of the debate's footage before it reached the public. CNN denied the claims. AI search startup Perplexity is increasingly citing as sources AI-generated blogs and social media posts, which contain inaccurate and contradictory information, Forbes has learned. The startup is drawing data from AI-generated materials on a wide variety of topics encompassing travel, politics, sports and medical information. In some instances, the answers generated by Perplexity's search engine also reflected the inconsistencies prevalent in the AI-generated sources, a study by AI content detection platform GPTZero found. Perplexity Chief Business Office Dmitri Shevelenko said in an email statement to Forbes that its system is "not flawless" and that it continuously refines its processes to identify high quality sources. Meanwhile, Amazon is investigating Perplexity to assess whether the startup violated its terms of service by scraping websites even after developers had attempted to prevent them from doing so, according to Wired. The investigation comes after Wired found a web crawler that was affiliated with the company had disregarded a common web standard that lets entities indicate which websites should not be scraped. Perplexity spokesperson Sara Platnick said PerplexityBot, the company's web crawler, had respected the web standard and had not violated terms of service. Plus, Quora's chatbot Poe is letting users download files of paywalled news articles, Wired reported. In an effort to beef up its artificial general intelligence team, Amazon has hired the cofounders and some employees of Adept, a startup that's building AI agents that perform various computer-related tasks. As part of the deal, Amazon will also license the two-year-old startup's technologies, AI models and datasets, the company said. Adept has raised over $400 million in funding from backers like General Catalyst and Greylock and is valued at over $1 billion. The deal mirrors a similar partnership struck by Microsoft, which in March hired most of Inflection's employees as well as the startup's CEO Mustafa Suleyman to head its consumer AI business. VC firm Benchmark is raising $425 million for its eleventh fund to back early stage AI startups, Forbes has learned. The firm's five partners -- Peter Fenton, Eric Vishria and Chetan Puttagunta, Sarah Tavel and Victor Lazarte -- plan to invest in AI companies within their area of expertise like consumer tech and cloud computing. The investment firm has already backed a number of AI startups including AI agent startup Sierra and video generation company HeyGen. Energy has become the hot commodity in the artificial intelligence world. Take cloud computing provider CoreWeave, which earlier this month inked a $3.5 billion deal with Core Scientific. The agreement involves CoreWeave paying $290 million annually over 12 years to host AI-related computing hardware at the Austin-based bitcoin miner's data centers. CoreWeave will also cover all the related capital expenditures. The deal was so good that Core Scientific's stock doubled to $10 in early June, leading some observers to view the company as a new "picks and shovels" play for AI. On June 26, CoreWeave announced a second contract for additional infrastructure, this one projected to bring Core Scientific $1.2 billion in revenue in the coming years. Core Scientific emerged from bankruptcy in January and is one of the largest bitcoin miners in North America. The soaring demand for heavy-duty computer capacity is driven by the energy needed for AI applications such as ChatGPT -- its queries require 10 times the electricity of traditional Google searches. That's an advantage for companies like Core Scientific that have access to cheap power in states such as Texas and North Dakota. Having sufficient power is vital when you consider that building and connecting new data centers to the grid can take as long as six years, according to the Lawrence Berkeley National Laboratory research center. "The demand is insatiable," says Adam Sullivan, Core Scientific's CEO. "If we just execute on what is within our current contracted power today, we'd be a top 10 data center company in the United States." Driven by the AI boom, data centers' energy demand could surge to 9% of U.S. power generation by 2030, according to the Electric Power Research Institute, which is more than double the current usage. A shift to AI operations for those with available infrastructure and energy capacity offers potentially compelling benefits. By replacing the volatility of bitcoin with more stable revenue from AI computing, miners can benefit from predictable budgets funded by established customers. This also helps miners boost income to be able to afford the high capital investment necessary to stay competitive with new mining equipment, concluded analysts at Morgan Stanley in an April report. Read the full article in Forbes. NBC Universal will use an AI-generated clone of iconic American sportscaster Al Michaels' voice to offer daily recaps of the Paris 2024 Olympics, the company said. The synthetic voice, trained on Michael's past NBC broadcasts, will be used to offer personalized coverage of the games via the app. Viewers can pick the sports and topics they are interested in receiving highlights of and the AI-generated voice can render 7 million different versions of recaps, pulling from 5,000 hours of live coverage of the event. 976 AI-generated news and information websites have been identified by misinformation tracking site NewsGuard. 54 AI-generated content farms are being aggregated by Google News, the report found. 43 Of those AI-generated news sites earn revenue from programmatic advertising. This semiconductor tech company is using light-based chips to help meet AI's growing energy demand for data centers: Toys 'R' Us released an advertisement made almost entirely with OpenAI's text-to-video AI tool, Sora, eliciting mixed reviews from viewers. (Sora created 80% of the advertisement and post production teams edited the video to add finishing touches.) The ad depicts the company's late founder, Charles Lazarus, as a young boy who dreams of a toy store. People were quick to point out that the boy's appearance changes throughout the video and objects keep melting into one another. On social media, people characterized the ad as "lame," "hollow" and "like a weird dream."

From AGI to ROI: The 6 AI debates shaping enterprise strategy in 2024

Don't miss OpenAI, Chevron, Nvidia, Kaiser Permanente, and Capital One leaders only at VentureBeat Transform 2024. Gain essential insights about GenAI and expand your network at this exclusive three day event. Learn More As I've been organizing VB Transform, our event next week focused on enterprise generative AI, I've noticed a stark shift in the scores of conversations I'm having with tech leaders. A year ago, it was all about how to embrace the magic of OpenAI's GPT-4 throughout the company. Now their focus is on practical implementation and ROI. It's as if the entire industry has hit reality mode. As we enter the second half of 2024, the artificial intelligence landscape is undergoing a seismic shift. The initial excitement following OpenAI's release of ChatGPT -- which became the fastest product in history to attract 100 million users -- has begun to wane. We're moving from an era of near unbridled hype to one of reality, where enterprises are grappling with how to implement AI technologies in real products. OpenAI CEO Sam Altman's pronouncements of "magic intelligence in the sky" sparked a frenzy among Silicon Valley developers, many of whom came to believe we were on the cusp of achieving human-level machine intelligence across all domains, known as artificial general intelligence (AGI). However, as 2024 progresses, a more nuanced narrative is emerging. Enterprises grounded in the practicalities of implementing AI in real-world applications are taking a more measured approach. The realization is setting in that while large language models (LLMs) like GPT-4o are incredibly powerful, generative AI overall has not lived up to Silicon Valley's lofty expectations. LLM performance has plateaued, facing persistent challenges with factual accuracy. Legal and ethical concerns abound, and infrastructure and business use cases have proved more challenging than anticipated. We're clearly not on a direct path to AGI as some had hoped. Even more modest promises, like autonomous AI agents, face plenty of limitations. And conservative technologies meant to "ground" AI with real data and accuracy, like RAG (retrieval-augmented generation), still have massive hurdles. Basically, LLMs still hallucinate like crazy. Instead, companies are focusing on how to leverage the impressive basic capabilities of LLMs already available. This shift from hype to reality is underscored by six critical debates that are shaping the AI landscape. These debates represent fault lines between the zealous believers in imminent superintelligence and those advocating for a more pragmatic approach to AI adoption. For enterprise leaders, understanding these debates is crucial. There are significant stakes for companies looking to exploit this powerful technology, even if it's not the godlike force its most ardent proponents claim. Don't read this wrong. Most enterprise leaders still believe that the technology has already produced profound benefits. During our recent AI Impact Tour, where meetings and events were held with Fortune 500 companies across the country, leaders openly discussed their efforts to embrace AI's promise. But these six debates will be central to discussions at our upcoming VB Transform event, scheduled for July 9-11 in the heart of San Francisco's SOMA district. We've curated the event based on extensive conversations with executives from the largest players in AI. The speaker lineup includes representatives from industry giants like OpenAI, Anthropic, Nvidia, Microsoft, Google, and Amazon, as well as AI leaders from Fortune 500 companies such as Kaiser Permanente, Walmart, and Bank of America. The live debates and discussions at Transform promise to shed light on these critical issues, offering in-person attendees a unique opportunity to engage with leaders at the forefront of enterprise AI implementation. The race to develop the most advanced LLM has been a defining feature of the AI landscape since OpenAI's GPT-3 burst onto the scene. But as we enter the second half of 2024, a question looms large: Is the LLM race over? The answer appears to be yes, at least for now. This matters because the differences between leading LLMs have become increasingly imperceptible, meaning enterprise companies can now select based on price, efficiency, and specific use-case fit rather than chasing the "best" model. In 2023, we witnessed a dramatic race unfold. OpenAI surged ahead with the release of GPT-4 in March, showcasing significant improvements in reasoning, multi-modal capabilities, and multilingual proficiency. Pundits assumed that performance would continue to scale as more data was fed into these models. For a while, it looked like they were right. But in 2024, the pace has slowed considerably. Despite vague promises from Altman suggesting more delights were coming, the company's COO Mira Murati admitted in mid-June that OpenAI doesn't have anything more in its labs than what is already public. Now, we're seeing clear signs of plateauing. OpenAI appears to have hit a wall, and its rival Anthropic has caught up, launching Claude 3.5 Sonnet, which outperforms GPT-4 on many measures. What's notable is that Claude wasn't able to leap far ahead; it's only incrementally better. More tellingly, Sonnet is based on one of Anthropic's smaller models, not its larger Opus model - suggesting that massive amounts of data training weren't necessarily leading to improvements, but that efficiencies and fine-tuning of smaller models were the key. Princeton computer science professor Arvind Narayanan wrote last week that the popular view that model scaling is on a path toward AGI "rests on a series of myths and misconceptions," and that there's virtually no chance that this scaling alone will lead to AGI. For enterprise leaders, this plateauing has significant implications. It means they should be leveraging the best individual LLMs for their specific purposes -- and there are now hundreds of these LLMs available. There's no particular "magical unicorn" LLM that will rule them all. As they consider their LLM choices, enterprises should consider open LLMs, like those based on Meta's Llama or IBM's Granite, which offer more control and allow for easier fine-tuning to specific use cases. At VB Transform, we'll dive deeper into these dynamics with key speakers including Olivier Godement, Head of Product API at OpenAI; Jared Kaplan, Chief Scientist and Co-Founder of Anthropic; Colette Stallbaumer, Copilot GM at Microsoft; David Cox, VP of AI Models at IBM; and Yasmeen Ahmad, Managing Director at Google Cloud. 2. The AGI hype cycle: peak or trough? As the pace of LLM breakthroughs slows, a larger question emerges: Have we reached the peak of inflated expectations in the AGI hype cycle? Our answer: Yes. This matters because companies should focus on leveraging existing AI capabilities for real-world applications, rather than chasing the promise of AGI. ChatGPT's release unleashed a wave of excitement about the possibilities of AI. Its human-like interactions, powered by massive amounts of training data, gave the illusion of true intelligence. This breakthrough catapulted Altman to guru status in the tech world. Altman embraced this role, making grandiose pronouncements about the future of AI. In November 2023, upon releasing GPT-4 Turbo, he claimed it would look "quaint" compared to what they were developing. He referred to AGI as possible in the next few years. These statements sparked massive enthusiasm among what we might call the spellbound zealots of Silicon Valley. However, the spell began to wear off. Altman's ejection from OpenAI's board in late 2023 (albeit temporary) was the first crack in the armor. As we entered 2024, his professions that AGI was close began to seem less convincing -- he tempered his predictions, emphasizing the need for further breakthroughs. In February, Altman said AGI might require up to $7 trillion of investment. Competitors narrowed the gap with OpenAI's leading LLM, and the steady improvements many had predicted failed to materialize. The cost of feeding more data to these models has increased, while their frequent logical errors and hallucinations persist. This has led experts like Yann LeCun, chief scientist at Meta, and others to argue that LLMs are a massive distraction and an "off-ramp" from true AGI. LeCun contends that while LLMs are impressive in their ability to process and generate human-like text, they lack the fundamental understanding and reasoning capabilities that would be required for AGI. That's not to say the hype has completely dissipated. The AI fever continues in some Silicon Valley circles, exemplified by the recent passionate four-hour video from Leopold Aschenbrenner, a former OpenAI employee, arguing that AGI could arrive within three years. But many seasoned observers, including Princeton's Narayanan, point to serious flaws in such arguments. It's this more grounded perspective that most enterprise companies should adopt. In conversations with enterprise AI leaders -- from companies like Honeywell, Kaiser Permanente, Chevron and Verizon -- I've consistently heard that the reality of AI implementation is much more complex and nuanced than the hype would suggest. While leaders are still enthusiastic about its potential, it's important not to get carried away with the idea that AI is improving so quickly that the next generation of the technology will solve the problems of the existing generation, says Steve Jones, EVP of CapGemini, a company that helps companies adopt AI. You've got to put in the controls now to harness it well: "Whether it's 20% or 50% of decisions will be made by AI in the next five years. It doesn't matter," he said in an interview with VentureBeat. The point is that your career success is based on the success of that algorithm, he says, and your organization is depending on you to understand how it works, and ensuring that it works well. "There's all the nonsense around AGI that's going on," he said referring to the continued hype among Silicon Valley developers who aren't really focused on enterprise deployments. But AI is "more of an organizational change than a technological one," he said, adding that companies need to harness and control the real, basic advancements LLMs already provide. Large companies are letting model providers do the heavy lifting of training, while they focus on fine-tuning models for their specific purposes. This more pragmatic approach is echoed by leaders across the finance, health and retail sectors we've been tracking. For instance, at JPMorgan Chase, Citi, Wells Fargo, and other banks I've talked with, the focus is on using AI to enhance specific banking functions, leading to practical applications in fraud detection, risk management and customer service. In healthcare, Dr. Ashley Beecy, medical director of AI operations at the NewYork-Presbyterian hospital system, provides another example of how big visions are being anchored instead by practical applications of AI. While she envisions an AI that knows everything about a patient, she says the hospital is starting with more practical applications like reducing the administrative burden on doctors by recording and transcribing patient visits. Beecy notes that much of the technical capability for the more ambitious version of AI is in place, but it's a matter of adjusting internal workflows and processes to allow this to happen, or what she called "change management." This will take a lot of work and testing, she acknowledged, and also require the sharing of ideas by national health organizations, since it will require larger structural change beyond her own hospital. At VB Transform, we'll explore this tension between AGI hype and practical reality with speakers from across the industry spectrum, providing attendees with a clear-eyed view of where AI capabilities truly stand and how they can be leveraged effectively in the enterprise. Speakers like Jared Kaplan, Chief Scientist at Anthropic, will discuss the current state of AI capabilities and the challenges ahead. We'll also hear from enterprise leaders who are successfully navigating this post-hype landscape, including Nhung Ho from Intuit and Bill Braun, CIO of Chevron. 3. The GPU bottleneck: infrastructure realities Is there a GPU bottleneck hurting the scaling of GenAI? Our answer: Yes, but it's more nuanced than headlines suggest. Why it matters: Enterprise companies need to strategically plan their AI infrastructure investments, balancing immediate needs with long-term scalability. The surge in AI development has led to an unprecedented demand for specialized hardware, particularly GPUs (Graphics Processing Units), that help run AI applications. Nvidia, the leading GPU manufacturer, has seen its market value skyrocket above $3 trillion, becoming the world's most valuable companies. This demand has created a supply crunch, driving up costs and extending wait times for this critical AI infrastructure. However, the bottleneck isn't uniform across all AI applications. While training large models requires immense computational power, many enterprise use cases focus on inference - running pre-trained models to generate outputs. For these applications, the hardware requirements can be less demanding. Jonathan Ross, CEO of Groq, a company developing innovative AI chips, argues that inference can be run efficiently on non-GPU hardware. Groq's language processing units (LPUs) promise significant performance gains for certain AI tasks. Other startups are also entering this space, challenging Nvidia's dominance and potentially alleviating the GPU bottleneck. Despite these developments, the overall trend points towards increasing computational demands. AI labs and hyperscale cloud companies that are training advanced models and want to stay leaders are building massive data centers, with some joining what I call the "500K GPU club." This arms race is spurring interest in alternative technologies like quantum computing, photonics, and even synthetic DNA for data storage to support AI scaling. However, most enterprise companies don't find themselves as constrained by GPU availability. Most will just use Azure, AWS and Google's GCP clouds, letting those big players sweat the costs of the GPU buildout. Take Intuit, one of the first companies to seriously embrace generative AI last year. The company's VP of AI, Nhung Ho, told me last week that the company doesn't need the latest GPUs for its work. "There are a lot of older GPUs that work just fine," Ho said. "We're using six or seven-year-old technology... it works beautifully." This suggests that for many enterprise applications, creative solutions and efficient architectures can mitigate the hardware bottleneck. At VB Transform, we'll delve deeper into these infrastructure challenges. Speakers like Groq's Jonathan Ross, Nvidia's Nik Spirin, IBM's director of Quantum Algorithms, Jamie Garcia, and HPE's Chief Architect Kirk Bresniker will discuss the evolving AI hardware landscape. We'll also hear from cloud providers like AWS, who are working on software optimizations to maximize existing hardware capabilities. 4. Content rights and LLM training: legal landmines ahead Is all content on the web free for training LLMs? Our answer: No, and this presents significant legal and ethical challenges. Why it matters: Enterprise companies need to be aware of potential copyright and privacy issues when deploying AI models, as the legal landscape is rapidly evolving. The data used to train LLMs has become a contentious issue, with major implications for AI developers and enterprise users alike. The New York Times and the Center for Investigative Reporting have filed suits against OpenAI, alleging unauthorized use of its content for training, which is just the tip of the iceberg. This legal battle highlights a crucial question: Do AI companies have the right to scrape and use online content for training without explicit permission or compensation? The answer is unclear, and legal experts suggest it could take up to a decade for this issue to be fully resolved in the courts. While many AI companies offer indemnification for enterprises using their services, this doesn't completely shield businesses from potential legal risks. The situation is further complicated by emerging AI-powered search engines and summarization tools. Perplexity AI, for instance, has faced criticism for summarizing paywalled articles, leading to a complaint from Forbes alleging copyright infringement. As the founder of VentureBeat, I have a stake in this debate. Our business model, like many publishers, relies on page views and advertising. If AI models can freely summarize our content without driving traffic to our site, it threatens our ability to monetize our work. This isn't just a concern for media companies, but any content creator. Any enterprise using AI models trained on web data could potentially face legal challenges. Businesses must understand the provenance of the data used to train the AI models they deploy. This is also key for finance and banking companies, which face big regulations around privacy and the usage of personal information. Some companies are taking proactive steps to address these concerns. On the training side, OpenAI is racing to strike deals with publishers and other companies. Apple has reportedly struck deals with news publishers to use their content for AI training. This could set a precedent for how AI companies and content creators collaborate in the future. At VB Transform, we'll explore these legal complexities in depth. Aravind Srinivas, CEO of Perplexity AI, will share insights on navigating these challenges. We'll also hear from enterprise leaders on how they're approaching these issues in their AI strategies. 5. Gen AI applications: transforming edges, not cores Are gen AI applications disrupting the core offerings of most enterprise companies? Our answer: No, not yet. Why this is important: While AI is transformative, its impact is currently more pronounced in enhancing existing processes rather than revolutionizing core business models. The narrative surrounding AI often suggests an imminent, wholesale disruption of enterprise operations. However, the reality on the ground tells a different story. Most companies are finding success by applying AI to peripheral functions rather than completely overhauling their core offerings. These applications are driving significant productivity gains and operational efficiencies. However, they're not yet leading to the massive revenue gains or business model shifts that some predicted. Executives at retail companies like Albertsons and AB InBev have told me they are eagerly looking for ways to impact their core, experimenting with "large application models" to predict customer purchasing patterns. In the pharmaceutical industry, there's hope that AI could accelerate drug discovery, though progress has been slower than many realize. Intuit provides an interesting case study here as well. Its business, based on tax and business code and terminology, is closer to the powerful language applications that LLMs provide, which explains why Intuit leaped ahead quickly, announcing its own Generative AI Operating System (GenOS) a year ago. It integrates AI assistants across products like TurboTax, QuickBooks, and Mailchimp. Still, its AI usage is focused on customer help, similar to what everyone else is using AI for. Apple's perspective is telling. They view AI as a feature, not a product - at least for now. This stance reflects the current state of AI in many enterprises: a powerful tool for enhancement rather than a standalone revolution. Caroline Arnold, an executive vice president of StateStreet, a major Boston-based bank, exemplifies this sentiment that generative AI is about productivity gains, but not a core revenue driver. At our Boston event in March, she highlighted AI's potential: "What gen AI allows you to do is to interact in a very natural way with large amounts of data, on the fly, and build scenarios... in a way that would take you much more time in a traditional way." While the bank's new LLM-infused chatbot quickly outperformed the existing helpdesk, it wasn't without challenges. The chatbot occasionally offered "weird answers," requiring fine-tuning. Four months later, State Street has yet to release its apps publicly, underscoring the complexities of enterprise generative AI adoption even at the edges. At VB Transform, we'll explore this nuanced reality with speakers like Nhung Ho, VP of AI at Intuit, and Bill Braun, CIO of Chevron, Daniel Yang, VP of AI for Kaiser Permanente, Desirée Gosby, VP of Walmart, and Christian Mitchell, EVP of Northwestern. They'll share insights on how they're integrating AI into their operations and where they see the most significant impacts. Why does this matter? AI agents represent a potential leap forward in automation and decision-making, but their current capabilities are often overstated. The concept of AI agents - autonomous systems that can perform tasks or make decisions with minimal human intervention - has captured the imagination of many in the tech world. Some, like former OpenAI employee Leopold Aschenbrenner, envision a not-to-distant future where hundreds of millions of AGI-smart AI agents run various aspects of our world. This, in turn, would squeeze a decade of algorithmic progress into a year or less: "We would rapidly go from human-level to vastly superhuman AI systems," he argues. However, most people I've talked with believe this is a pipe dream. The current state of AI agents is, in fact, far more modest than Silicon Valley enthusiasts even assumed they would be just a year ago, when excitement exploded around Auto-GPT, an agent framework that would supposedly allow you to do all kinds of things, including starting your own company. While there are promising use cases in areas like customer service and marketing automation, fully autonomous AI agents are still in their infancy, and face many challenges of staying on track with their jobs. Other emerging applications of AI agents include: These agents often use a lead LLM to orchestrate the process, with sub-agents handling specific tasks like web searches or payments. However, they're far from the general-purpose, fully autonomous systems some envision. Intuit's approach to AI agents is instructive. Nhung Ho revealed that while Intuit has built out infrastructure to support agentic frameworks, it has paused investments in that area. Intuit is waiting for the technology to mature before fully integrating it into their products. This cautious approach reflects the broader industry sentiment. While AI agents show promise, they're not yet reliable or versatile enough for widespread enterprise adoption in critical roles. At VB Transform, we'll explore the current state and future potential of AI agents. Speakers like Itamar Friedman, CEO of Codium AI, which is developing an autonomous coding agent, and Jerry Liu, CEO of LlamaIndex, will share their insights on this emerging technology. Conclusion: Navigating the AI landscape in 2024 and ebyond As we've explored the six critical AI debates shaping enterprise strategy in 2024, a clear theme emerges: the shift from hype to practical implementation. The key takeaways for enterprise leaders: The real AI revolution isn't happening in research labs pursuing AGI, but in offices worldwide where AI is being integrated into everyday operations. As Steve Jones of Capgemini said, "AI is more of an organizational change than a technological change." As we head toward VB Transform and into the second half of the year, remember that the most valuable AI implementation might not make headlines. It might be the one that saves your customer service team a few hours each day or helps your developers catch bugs more quickly. The question is no longer "Will AI change everything?" but "How can we harness AI to do what we do, better?" That's what will separate the AI leaders from the laggards in the years to come. And that's the conversion I believe will dominate at VB Transform.

Here's Why Analysts Say Machine Buying Will Be The Megatrend Of AI

Tech firms are scrambling to develop 'AI agents' that can perform complex, multi-step tasks on our behalf. While the current wave of enthusiasm for AI was ignited by large language models that can show and tell us things, there is firm consensus across the industry that the next and potentially much larger wave will be led by bots that can increasingly do things too. A lot of these tasks will inevitably involve some form of purchasing. So what happens when we let AI go out and do the shopping for us? Or negotiate our corporate contracts? At scale, the implications are vast, say researchers. AI isn't swayed by clever marketing slogans. It doesn't feel pressured by hard sales tactics. It's not impressed by influencer endorsements. It's not dazzled by pretty packaging. It's not bamboozled by gimmicks (and nor will it miss the crucial small print). It can't be wined and dined. And it doesn't care for tickets to whichever big game is coming up. But AI can make timely and highly informed purchases with unprecedented efficiency. It ignores low value features, filters out irrelevant information, and makes more rational, quicker purchases based on the vast amounts of data it can collect and consume about both the market and the needs of the person or company it serves. As a result, AI is increasingly being deployed by both consumers and companies for the purpose of machine buying. It already exists in the form of dedicated hardware like smart speakers, connected products capable of communicating their own needs, and - increasingly - buyer bots that can operate across any internet-connected device. These machine buyers are creating significant new value and making purchases that their human owners and operators previously wouldn't even often have time to research. As this trend continues, some analysts believe that machine buying will be as transformative to global business as e-commerce and will unleash a wave of economic growth comparable to the rise of emerging markets in the 1990s. Don Scheibenreif and Mark Raskino are senior analysts at Gartner and have conducted extensive research into the evolving impact of technology on purchasing, including through interviews with a wide range of business leaders. They see a two decade transition to full scale machine buying, with about 20% of business revenue in 2030 already coming from what they call 'custobots'. Just this week, for example, leading telcos added their voices to the consensus that machine buying is the future. Rather than just fending off malicious bots with captcha forms, businesses have begun thinking about how to also welcome bots as some of their most valuable customers, while also sending out their own buyer bots to conduct autonomous negotiations with their suppliers. This is leading to a radical rethink of how all businesses function, whether their customers are high street consumers or Fortune 500 companies. Scheibenreif and Raskino have compiled their analysis in When Machines Become Customers, which they describe as a field guide and survival manual for "this profound change to business in the 21st century". It's highly recommended reading for all C-suite leaders and procurement professionals. "It's a now thing" Until now, it's sellers who have been able to take the lead and phase in new automation technology - although not always smoothly for customers. We've all experienced a clueless chatbot, a QR code menu that doesn't load, or an "unexpected item in the bagging area". Nevertheless, the trend is clear. Automation keeps generating new value from greater efficiencies, while the technologies behind it keep improving, and we all keep accepting these new innovations as the new normal. In many parts of the world, people don't even bat an eyelid at the sight of autonomous delivery drones carrying groceries up and down the sidewalk. However, there's no reason to assume this trend would remain purely in the control of sellers. AI is already flipping the initiative, enabling buyers to introduce more automation to create efficiencies from their end too. Scheibenreif and Raskino point out that machines are already capable of handling each part of the buying process. They can weigh up reviews, receive messages from sellers, request more information, negotiate for the best deals, make payments, request support when needed, and share feedback. Gartner Vice President and analyst Michelle DeClue says "machines as customers is not just an 'in the future' type of thing. It's a now thing." When Scheibenreif and Raskino began exploring the subject of machine buying more deeply, they found examples of it in various forms emerging and fast evolving across every type of purchasing - all the way up to Fortune 500 companies that send out bots to conduct autonomous negotiations with their suppliers, including over multi-million dollar deals. By studying its ongoing evolution and impact, they concluded that machine buying will be the most significant and disruptive megatrend of AI over the next few decades, just as e-commerce was the most significant for the internet. The evolution of machine buying Scheibenreif and Raskino see buyer bots evolving in three stages. Prior to these, there's a zero 'announcer' stage that serves as the primordial soup for the emergence of machine buying. This is when technology simply tells us things that help prompt and inform purchasing, like a printer that says it's running low on ink or a car that tells you it's time for a maintenance inspection. This creates routine work for human owners who are therefore receptive to the idea that the machine could help take some of the hassle out of the process. Meanwhile, the manufacturers are certainly more than happy to help guide further spending. In the case of printers, for example, selling the ink cartridges later is actually a key part of the business model. This, therefore, leads to the first 'bound' stage of machine buying where machines evolve from merely announcing information to acting upon it, although still within the confines of a preset choice. Many people will have experienced this for the first time when their printer ordered its own ink, but this approach is now used by a wide variety of consumer products - from smart cars to connected toothbrushes. In the defense industry, this has been happening even earlier and on a far more complex scale. BAE Systems, whose products include multibillion dollar warships for the British Royal Navy, has been using machine learning starting with its Type 45 destroyers, which can diagnose their own maintenance needs and communicate them to its shore-based supply chain, drastically reducing the inefficiency of routine maintenance. So far, these examples are still led by sellers. However, consumers don't want their machines to be kept in walled gardens from where they can only order replenishments from their original manufacturers. Anyone currently developing AI agents for machine buying should be mindful of the hard lessons learned by Yahoo in earlier days of the internet. After obtaining a strong market position as a web directory, it developed a business model based largely on leading its users towards its own products. But Google soon leapfrogged Yahoo and achieved astronomical growth because it understood that being the leading web portal meant prioritizing the best results for users from across the web. This same kind of market pressure leads to the next phase of machine buying. Buyers will keep favoring the development of buyer bots that can operate at their most freely and make the best purchasing choices from across the widest possible markets. So bound machine buying inevitably becomes adaptable machine buying, in which the best purchases can be made among competing choices. Many people already have a home assistant in the form of a smart speaker, such as Amazon's Alexa, which can order a wide variety of products on a simple voice command and with ever increasing sophistication to help determine which products are most suitable among competing choices available. As the technology continues to evolve, crunch more data, and become ever more intuitive for users, next comes the autonomous stage of machine buying. This is where things get really interesting, according to Scheibenreif and Raskino. In this stage, buyer bots can mimic a full range of human buying behaviors, but with superhuman capabilities. Purchases are then increasingly made based on the inferred needs of the person or company that the AI is serving. "It is inevitable," argue Scheibenreif and Raskino, "that machines, now endowed with ever increasing levels of intelligence, will do more of our work for us, including our work as customers." AI Won't Take Your Job, But It Might Take Your Customers Elsewhere Some may resist this future out of concern that we might be losing something human on a significant scale. But does outsourcing the purchasing of laundry detergent to a machine really make you less human? Most buying is tedious work and not a good use of human time, nor something that humans are even particularly good at. The inefficiency of the way we currently shop can also be deeply unfair on poorer and otherwise less advantaged members of society who pay more for almost everything, from groceries to energy bills. "You may enjoy shopping some of the time," write Scheibenreif and Raskino. "Maybe you enjoy looking for clothes, or going to the bookstore, or buying a special gift for a friend or partner. But you probably don't enjoy shopping for toilet tissues, tires, or life insurance. And you almost certainly don't like making business purchases." Even in the absence of AI, most purchasing doesn't actually involve any meaningful human connections. Instead, it saps away time that could be used for exactly that. At large enterprises, the vast majority of supplier contracts are not being negotiated between humans in any significant way. Most suppliers are given standardized, 'cookie cutter terms' that get rolled over, even though both sides could find better terms and generate more value together if they simply had more time to negotiate with each other. It's just not economical for humans to properly research and spend time negotiating all the potential deals around them that could bring them value. That's why, where buyer bots have already been deployed by large enterprises for autonomous negotiations with suppliers, they haven't been replacing human work so much as they've been extending human capabilities to engage with suppliers who were previously unengaged. We don't worry that human connections are being lost when we step into an elevator that no longer has an attendant or when we board a plane that now also has an autopilot. Buyer bots, like modern elevators and modern aircraft, can help us make more human connections as an outcome of the value they deliver. Scheibenreif and Raskino say the rise of machine buying will be a 20 year change wave that will always require considerable human work managing these bots in order to ensure they create maximum value. Every context in which machine customers operate will be humanly conceived and enacted. So professional human buyers will have much to gain as they manage this transition. As with the internet, machine buying will very likely be a net positive for jobs, both directly and also through its wider value creation across the economy. But, as machine buying takes away significant drudgery and opens new markets, capturing their business will require radically rethinking every business function. Don't worry about AI taking your job, say Scheibenreif and Raskino, but do worry about AI taking your customers elsewhere. The jobs most at risk in an age of highly efficient machine buying are those that are currently being sustained by inefficient or under-informed human purchasing decisions, as well as those that depend on persuading humans to make purchasing decisions. If you are trying to get people to buy insurance that is unnecessary or a timeshare scheme with hidden fees, then it's fair to say you should already start reconsidering your long-term career prospects. Meanwhile, roles in marketing and sales will become mostly technical as companies increasingly focus on reverse engineering how to attract buyer bots, just as e-commerce led to the rise of optimization specialists focused on reverse engineering search engine results. You may not be able to take an AI out for dinner, but it is certainly hungry for data that you can feed it. It will be fascinating to see how this more efficient and sophisticated approach to purchasing on a wide scale can help incentivize wider benefits for society and the planet - such as by calculating how ethically sourced different products are and letting buyers choose based on those factors. While AI won't be dazzled by pretty packaging, it could extract the data on the carbon footprint of that packaging (along with the rest of the product) in order to judge which of competing products come out top by environmental standards. This is far from hypothetical. Fortune 500 companies already use their buyer bots to collect all kinds of data about their suppliers within the purchasing process in order to help inform their drive towards corporate social responsibility goals like diversity, equality, and inclusion. Ahead Of Schedule Since Scheibenreif and Raskino first published their book last year, we have more data and a better understanding of how their analysis is tracking against real world progress. If their conclusions are wrong, then it might just be in underestimating the scale and positive impact of machine buying. For a start, the pace of AI development has since exceeded expectations. In addition, the authors argued in their book that buyer bots acting on behalf of large enterprises will be tough negotiators for suppliers due to their ability to consume vast market data and reach out to the widest range of alternative suppliers. However, AI will always search for the most optimal way to achieve its objectives, which means playing games its own way. In negotiation, AI has recognized that this is not a zero-sum game and so has not been mimicking hardball negotiators. Instead, the more that AI has been used for autonomous negotiations between companies, the more that AI has learned to achieve the best outcomes through a collaborative approach that seeks to maximize value for both sides. This is smart not just for getting the most value out of any one deal, but also to ensure the long term health and motivation of the supply chain. In many cases, machine buying involves bots that are just going out to existing suppliers in the midst of their contracts and offering business growth opportunities as that helps both sides unlock more value. The more that people and companies use buyer bots, the more ways they discover those bots can create new value. Large enterprises originally measured only the hard savings, but there's now an increased focus on the way that machine buying can increase business velocity through better processes and governance, as well as stronger and more flexible relationships with suppliers. On an individual level, our purchasing power has a large influence on our quality of life and opportunities. In business, few things define a company more deeply than what it buys. While the global market will continue to change and there will almost certainly be more black swan events on the way, the people and companies able to react fast and extend their spending most efficiently will dominate the world of tomorrow. The age of machine buying isn't just inevitable. It's already begun.

We Need to Control AI Agents Now

In 2010 -- well before the rise of ChatGPT and Claude and all the other sprightly, conversational AI models -- an army of bots briefly wiped out $1 trillion of value across the NASDAQ and other stock exchanges. Lengthy investigations were undertaken to figure out what had happened and why -- and how to prevent it from happening again. The Securities and Exchange Commission's report on the matter blamed high-frequency-trading algorithms unexpectedly engaging in a mindless "hot potato" buying and selling of contracts back and forth to one another. A "flash crash," as the incident was called, may seem quaint relative to what lies ahead. That's because, even amid all the AI hype, a looming part of the AI revolution is under-examined: "agents." Agents are AIs that act independently on behalf of humans. As the 2010 flash crash showed, automated bots have been in use for years. But large language models can now translate plain-language goals, expressed by anyone, into concrete instructions that are interpretable and executable by a computer -- not just in a narrow, specialized realm such as securities trading, but across the digital and physical worlds at large. Such agents are hard to understand, evaluate, or counter, and once set loose, they could operate indefinitely. For all of today's concern about AI safety, including potentially existential risks, there's been no particular general alarm or corresponding regulation around these emerging AI agents. There have been thought experiments about an AI given (or setting for itself) an arbitrary and seemingly harmless goal, such as to manufacture as many paper clips as possible, only to cause disaster when it diverts all of humanity's resources toward that goal. But well short of having to confront a speculative monomaniacal superintelligence, we must attend to more pressing if prosaic problems, caused by decidedly nonspeculative contemporary agents. These can mess up, either through the malice of those who get them going, or accidentally, monkey's-paw style, when commissioned with a few ill-chosen words. For example, Air Canada recently experienced the latter when it set up a chatbot for customer assistance with a prompt to be helpful, along with access to the Air Canada website for use in answering customer questions. The bot helpfully explained a policy on bereavement fares in a way far more generous than the airline's actual policy. Air Canada tried to repudiate the bot's promises, and failed: A tribunal held that the customer was owed compensation. Read: This is what it looks like when AI eats the world Today's agents add up to more than a typical chatbot, with three distinct qualities. First, they can be given a high-level, even vague goal and independently take steps to bring it about, through research or work of their own. The idea is simple but powerful. For example, a year ago, an enterprising techie developed an AI that could order a pizza for him. He relied on software tools developed by companies such as OpenAI to create a "top-level AI" that could charter and command other AIs. That top-level AI was provided a goal -- order a pepperoni pizza by voice from a given phone number -- and then it went on to create its own task list and develop different versions of itself to perform those tasks, including prioritizing different steps in the list and producing a version of itself that was able to use a text-to-voice converter to make the phone call. Thus the AI was able to find and call a local pizzeria and place the order. That demonstrates a second quality of agents beyond planning to meet a goal: They can interact with the world at large, using different software tools at will, as you might when opening Excel or placing a DoorDash order while also browsing the web. With the invitation and blessing of companies such as OpenAI, generative-AI models can take in information from the outside world and, in turn, affect it. As OpenAI says, you can "connect GPTs to databases, plug them into emails, or make them your shopping assistant. For example, you could integrate a travel listings database, connect a user's email inbox, or facilitate e-commerce orders." Agents could also accept and spend money. This routinization of AI that doesn't simply talk with us, but also acts out in the world, is a crossing of the blood-brain barrier between digital and analog, bits and atoms. That should give us pause. A non-AI example jumps to mind as a nefarious road map for what may lie ahead. Last year, a man left a bag conspicuously containing wires and a lockbox outside Harvard Yard. Harvard police then received a call with a disguised voice warning that it was one of three bombs on campus, and that they'd all go off soon unless the university transferred money to a hard-to-trace cryptocurrency address. The bag was determined to be harmless. The threat was a hoax. When police identified and arrested the man who left the bag, it turned out that he had answered a Craigslist ad offering money for him to assemble and bring those items to campus. The person behind that ad -- and the threatening calls to Harvard -- was never found. The man who placed the wires pleaded guilty only to hiding out and deleting some potentially incriminating text messages and was sentenced to probation, after the authorities credited that he was not the originator of the plot. He didn't know that he'd joined a conspiracy to commit extortion. Read: Welcome to a world without endings This particular event may not have involved AI, but it's easy to imagine that an AI agent could soon be used to goad a person into following each of the steps in the Harvard extortion case, with a minimum of prompting and guidance. More worrying, such threats can easily scale far beyond what a single malicious person could manage alone; imagine whoever was behind the Harvard plot being able to enact it in hundreds or thousands of towns, all at once. The act doesn't have to be as dramatic as a bomb threat. It could just be something like keeping an eye out for a particular person joining social media or job sites and to immediately and tirelessly post replies and reviews disparaging them. This lays bare the third quality of AI agents: They can operate indefinitely, allowing human operators to "set it and forget it." Agents might be hand-coded, or powered by companies who offer services the way that cemeteries offer perpetual care for graves, or that banks offer to steward someone's money for decades at a time. Or the agents might even run on anonymous computing resources distributed among thousands of computers whose owners are, by design, ignorant of what's running -- while being paid for their computing power. The problem here is that the AI may continue to operate well beyond any initial usefulness. There's simply no way to know what moldering agents might stick around as circumstances change. With no framework for how to identify what they are, who set them up, and how and under what authority to turn them off, agents may end up like space junk: satellites lobbed into orbit and then forgotten. There is the potential for not only one-off collisions with active satellites, but also a chain reaction of collisions: The fragments of one collision create further collisions, and so on, creating a possibly impassable gauntlet of shrapnel blocking future spacecraft launches. Read: The big AI risk not enough people are seeing If agents take off, they may end up operating in a world quite different from the one that first wound them up -- after all, it'll be a world with a lot of agents in it. They could start to interact with one another in unanticipated ways, just as they did in the 2010 flash crash. In that case, the bots had been created by humans but simply acted in strange ways during unanticipated circumstances. Here, agents set to translate vague goals might also choose the wrong means to achieve them: A student who asks a bot to "help me cope with this boring class" might unwittingly generate a phoned-in bomb threat as the AI attempts to spice things up. This is an example of a larger phenomenon known as reward hacking, where AI models and systems can respond to certain incentives or optimize for certain goals while lacking crucial context, capturing the letter but not the spirit of the goal. Even without collisions, imagine a fleet of pro-Vladimir Putin agents playing a long game by joining hobbyist forums, earnestly discussing those hobbies, and then waiting for a seemingly organic, opportune moment to work in favored political talking points. Or an agent might be commissioned to set up, advertise, and deliver on an offered bounty for someone's private information, whenever and wherever it might appear. An agent can deliver years later on an impulsive grudge -- revenge is said to be a dish best served cold, and here it could be cryogenically frozen. Much of this account remains speculative. Agents have not experienced a public boom yet, and by their very nature it's hard to know how they'll be used, or what protections the companies that help offer them will implement. Agentics, like much of the rest of modern technology, may have two phases: too early to tell, and too late to do anything about it. In these circumstances, we should look for low-cost interventions that are comparatively easy to agree on and that won't be burdensome. Yale Law School's Ian Ayres and Jack Balkin are among the legal scholars beginning to wrestle with how we might best categorize AI agents and consider their behavior. That would have been helpful in the Air Canada case around a bot's inaccurate advice to a customer, where the tribunal hearing the claim was skeptical of what it took to be the airline's argument that "the chatbot is a separate legal entity that is responsible for its own actions." And it's particularly important to evaluate agent-driven acts whose character depends on assessing the actor's intentions. Suppose the agent waiting to pounce on a victim's social-media posts doesn't just disparage the person, but threatens them. Ayres and Balkin point out that the Supreme Court recently held that criminalizing true threats requires that the person making the threats subjectively understand that they're inspiring fear. Some different legal approach will be required to respond up and down the AI supply chain when unthinking agents are making threats. Technical interventions can help with whatever legal distinctions emerge. Last year, OpenAI researchers published a thoughtful paper chronicling some agentic hazards. There they broached the possibility that servers running AI bots should have to be identified, and others have made efforts to describe how that might work. Read: It's the end of the web as we know it But we might also look to refining existing internet standards to help manage this situation. Data are already distributed online through "packets," which are labeled with network addresses of senders and receivers. These labels can typically be read by anyone along the packets' route, even if the information itself is encrypted. There ought to be a new, special blank on a packet's digital form to indicate that a packet has been generated by a bot or an agent, and perhaps a place to indicate something about when it was created and by whom -- just like a license plate can be used to track down a car's owner without revealing their identity to bystanders. To allow such labels within Internet Protocol would give software designers and users a chance to choose to use them, and it would allow the companies behind, say, the DoorDash and Domino's apps to decide whether they want to treat an order for 20 pizzas from a human differently from one placed by a bot. Although any such system could be circumvented, regulators could help encourage adoption. For example, designers and providers of agents could be offered a cap on damages for the harm their agents cause if they decide to label their agents' online activities. Internet routing offers a further lesson. There is no master map of the internet because it was designed for anyone to join it, not by going through a central switchboard, but by connecting to anyone already online. The resulting network is one that relies on routers -- way stations -- that can communicate with one another about what they see as near and what they see as far. Thus can a packet be passed along, router to router, until it reaches its destination. That does, however, leave open the prospect that a packet could end up in its own form of eternal orbit, being passed among routers forever, through mistake or bad intention. That's why most packets have a "time to live," a number that helps show how many times they've hopped from one router to another. The counter might start at, say, 64, and then go down by one for each router the packet passes. It dies at zero, even if it hasn't reached its destination. Read: What to do about the junkification of the internet Agents, too, could and should have a standardized way of winding down: so many actions, or so much time, or so much impact, as befits their original purpose. Perhaps agents designed to last forever or have a big impact could be given more scrutiny and review -- or be required to have a license plate -- while more modest ones don't, the way bicycles and scooters don't need license plates even as cars do, and tractor trailers need even more paperwork. These interventions focus less on what AI models are innately capable of in the lab, and more on what makes agentic AI different: They act in the real world, even as their behavior is represented on the network. It is too easy for the blinding pace of modern tech to make us think that we must choose between free markets and heavy-handed regulation -- innovation versus stagnation. That's not true. The right kind of standard-setting and regulatory touch can make new tech safe enough for general adoption -- including by allowing market players to be more discerning about how they interact with one another and with their customers. "Too early to tell" is, in this context, a good time to take stock, and to maintain our agency in a deep sense. We need to stay in the driver's seat rather than be escorted by an invisible chauffeur acting on its own inscrutable and evolving motivations, or on those of a human distant in time and space.

Can AI outperform a wealth manager at picking investments?

When technology entrepreneur Edward Morris participated in the $5bn initial public offering of chip designer Arm last September, he adopted an entirely artificial intelligence-based strategy that would result in one of his most profitable investments to date. Morris -- who runs the consultancy Enigmatica, specialising in AI and prompt engineering (the creation of chatbot inputs that return the most desirable answers) -- says he conducted all the necessary due diligence on his investment using the popular AI-powered virtual assistant ChatGPT and made a 30 per cent return. He typically closes out at 10 per cent. In the past, Morris might have asked a human financial adviser to help with his investment activities. But, he considers these services "incredibly expensive" and, seeing first-hand the advancement of AI over the past few years, he was keen to try a new approach. He has no regrets. Morris claims that the chatbot -- while, essentially, only a text generator -- can improve his understanding of complex wealth management and finance topics, helps him find worthy investments such as the 2023 ARM listing and identifies discrepancies in his bank statements, just like a human financial adviser would. Additionally, Morris has linked AI tools to his WhatsApp and Telegram accounts so that he is alerted to investment opportunities via text message. Morris adds: "ChatGPT has given me a financial adviser in my pocket at all times that I can talk to and get advice from." Reflecting on his ARM investment, Morris says that the way he uses AI in his investments is "unbelievably simple". He says his first step is to find an investable stock. Then, he undertakes due diligence by firing questions to ChatGPT about the company's history, current activities, financials and any negative press. Morris says the AI-powered chatbot will then summarise this information and provide a rating on how well a stock might perform, helping people "make educated investments" without having to shell out large sums of money on a professional wealth management firm or expert. "Due diligence is something that used to take days upon days to do within wealth management and financial firms. That's not the case any more with AI," he explains. "Ninety nine per cent of the investment game is knowing if something is a good investment and ChatGPT seems to be absolutely incredible at creating that information and communicating it in different ways." While AI is not yet a proven tool for individual investors, Morris believes that wealth managers can also benefit from the technology. He says it allows wealth managers to "run their ideas past an extra set of eyes" and complete "time-consuming" tasks such as client risk-profiling questionnaires. They can also use it for helping clients get their estates in order, assessing the potential impact of economic policies and finding sector-specific investment opportunities, he claims. With these varying use-cases in mind, he urges wealth managers to upskill in AI and prompt engineering to get the most out of the technology in their day-to-day roles. "If you're a wealth manager, I'd say learn how to use ChatGPT properly and effectively. Don't just play with it for a bit and leave it. Give it time," advises Morris. "It can (and does) save people weeks', and sometimes months', worth of time." Sensing the looming AI revolution and its impact on the financial services sector, many of the biggest wealth management groups are already investing in this technology. For instance, Morgan Stanley has developed and rolled out an AI assistant designed to streamline the day-to-day tasks of its global wealth managers. Powered by OpenAI's large language model technology, the AI @ Morgan Stanley Assistant allows the bank's financial advisers to find relevant information from an internal database of more than 100,000 documents. One such financial adviser is Patrick Biggs, who explains that the chatbot enables him to "efficiently source and retrieve internal information" and summarise corporate processes so that he can spend more time with clients. "Before this technology, I'd have to wade through PDFs and documents of research to find what we needed, which was especially difficult because procedures can evolve, and the markets change every day," he says. Sal Cucchiara, chief information officer and head of wealth management technology at Morgan Stanley, says the success of this technology depends on several factors. "One, quality of the data used is critical," he stresses. "Two, [you need to] engage with the end user early in the process in addition to educating and partnering with teams across the organisation," he says. "Lastly, take a control-forward approach to the rollout and work hand-in-hand with legal, risk and compliance partners through every step." As AI continues to automate many aspects of wealth and investment management, a growing concern of wealth managers is whether it will one day take their jobs. But Mohamed Keraine, global head of digital, wealth and retail banking at Standard Chartered, does not think industry workers should fear the rise of AI. He views the technology as an "opportunity to complement human attributes" rather than "replace them". In particular, he expects AI to help wealth managers form stronger relationships with their clients by delivering enhanced, seamless virtual interactions and improving access to wealth management services. For instance, like many other banks, StanChart offers a 24/7 customer service chatbot in addition to a chat and collaboration tool called myRM. The latter allows users to chat with their relationship managers, transfer documents and files securely, and more. "[AI] will uncover a lot of opportunities in the way we offer advisory solutions and enable quicker and more accurate access to market insights and trends," he says. "[AI] will also drive a more proactive understanding of customer needs and an unprecedented ability to offer personal, instant and differentiating wealth solutions." John Mileham, CTO at online financial adviser Betterment, agrees that AI presents opportunities for both wealth managers and their clients. He explains that Betterment uses AI chatbots externally to answer customers' questions and requests "more quickly". And, internally, he says, AI is enabling the firm to automate manual processes ranging from the creation of meeting summaries and marketing copy to fixing software problems. Employees can then focus on "more strategic work". Zac Maufe, global head of regulated industries at Google Cloud, says wealth managers can use AI tools to analyse large volumes of client data -- such as their financial history and goals, tolerance to risk and demographics -- and use this information to develop more personalised investment plans and portfolios for customers. "Through continuous monitoring and real-time adjustments, AI can ensure clients stay on track towards their financial goals while advisers gain deeper insights to foster stronger relationships and offer relevant products and services," he says. Other AI use cases for wealth managers include automated trade execution, the automation of repetitive work, fraud detection, portfolio optimisation and real-time market insights, he adds. But, regardless of all these advancements, Maufe says wealth managers will still need to strike a balance between AI usage and "the human touch". He says: "Leveraging AI to enhance, not replace, the human element of wealth management is important since clients still value personalised advice and trust built through relationships." Although AI is improving the efficiencies of wealth managers and allowing people to access financial advice 24/7, this technology is not without its challenges when actually applied to financial advice or decision making. A major concern is whether AI may provide bad investment advice that results in users losing large sums of money. For example, Bloomberg reported in 2019 that Hong Kong-based entrepreneur Samathur Li Kin-kan lost $20mn when he used a robo investor service. Such problems could now be exacerbated by so-called AI hallucinations, in which chatbots generate false or entirely fictitious results. Mileham explains that these hallucinations, as well as other biases, can stem from the chatbots' underlying training data sets. "Investors should be very careful to evaluate the source of the financial advice they are relying on," he warns. "Generative AI is trained on massive data sets that go beyond good investing advice. It could draw incorrect inferences from inputs, and it might not guide you towards optimal strategies." Neil Sahota, co-author of Own the AI Revolution and an AI adviser at the UN, warns that AI systems often provide poor investment advice due to "limited personalisation" and a "lack of human empathy". He explains that AI wealth managers are powered by standardised algorithms that "may not fully account for the nuances of individual financial situations", such as "specific tax implications" and "unique financial goals". Sahota adds that these platforms often "lack the human touch essential to building trust and providing emotional support during volatile market conditions" and that human wealth advisers are best equipped to "offer reassurance and personalised advice". Robo investors are also susceptible to cyber attacks and technical issues, which can lead to data leaks and investment disruption, he warns. Because AI systems are typically trained on legacy data, Sahota suggests that they may also struggle to make decisions during "unprecedented market conditions". He says: "AI algorithms excel in stable environments but may struggle to adapt quickly to sudden economic shifts." Adam Rodriguez -- director of product at autonomous car technology company Waymo and an Arta Finance customer -- agrees that AI wealth managers may sometimes respond to an investment scenario in "an unexpected way" because this isn't reflected in their training data. However, he suggests that "established and proven" AI investors are better equipped to deal with this issue and, consequently, advises people only to "invest with reputable firms who have designed the system and built in the proper safeguards". Concerns aside, it seems robo investors have a bright future. In particular, the rise of agentic AI systems could allow robo investors to mimic the proven and winning strategies of businessman and investor Warren Buffett, suggests Nell Watson, an AI expert and author of Taming the Machine: Ethically Harness the Power of AI. She argues that agentic AI systems -- which use their own autonomy to set and meet complex goals with little human input -- would be able to "uncover key insights and patterns" by analysing large volumes of financial reports, market trends, news pieces and other reading materials at "incredible speeds". "Just as Buffett dedicates five to six hours daily to reading 500 pages, these AI systems can continuously ingest and analyse data 24/7, giving them an even more comprehensive and up-to-date knowledge base," she argues. Aside from content consumption, Watson believes that the technology could also one day be capable of deciphering the variables responsible for "market dynamics" and "company performance". "Using sophisticated machine learning algorithms and predictive modelling techniques, these systems can identify subtle correlations, curious anomalies, and forecast future trends with a high degree of accuracy," she says. "This allows them to spot undervalued 'diamonds in the rough' with strong fundamentals and growth potential -- the very essence of Buffett's value investing approach." However, acknowledging common concerns, she says this will depend on high-quality training data and underlying algorithms in addition to "robust" risk management frameworks and human oversight. Watson's belief in this technology is unwavering, though. She concludes: "The potential is immense -- just as Buffett has used his reading habit to build an unparalleled investment record, agentic AI's independent data processing capabilities could turn a small family office into the next Berkshire Hathaway, revolutionising the world of finance. The rise of the AI-powered 'super-investor' may be closer than we think."

How evolving AI regulations impact cybersecurity

While their business and tech colleagues are busy experimenting and developing new applications, cybersecurity leaders are looking for ways to anticipate and counter new, AI-driven threats. It's always been clear that AI impacts cybersecurity, but it's a two-way street. Where AI is increasingly being used to predict and mitigate attacks, these applications are themselves vulnerable. The same automation, scale, and speed everyone's excited about are also available to cybercriminals and threat actors. Although far from mainstream yet, malicious use of AI has been growing. From generative adversarial networks to massive botnets and automated DDoS attacks, the potential is there for a new breed of cyberattack that can adapt and learn to evade detection and mitigation. In this environment, how can we defend AI systems from attack? What forms will offensive AI take? What will the threat actors' AI models look like? Can we pentest AI -- when should we start and why? As businesses and governments expand their AI pipelines, how will we protect the massive volumes of data they depend on? It's questions like these that have seen both the US government and the European Union placing cybersecurity front and center as each seeks to develop guidance, rules, and regulations to identify and mitigate a new risk landscape. Not for the first time, there's a marked difference in approach, but that's not to say there isn't overlap. Let's take a brief look at what's involved, before moving on to consider what it all means for cybersecurity leaders and CISOs. US AI regulatory approach - an overview Executive Order aside, the United States' de-centralized approach to AI regulation is underlined by states like California developing their own legal guidelines. As the home of Silicon Valley, California's decisions are likely to heavily influence how tech companies develop and implement AI, all the way to the data sets used to train applications. While this will absolutely influence everyone involved in developing new technologies and applications, from a purely CISO or cybersecurity leader perspective, it's important to note that, while the US landscape emphasizes innovation and self-regulation, the overarching approach is risk-based. The United States' regulatory landscape emphasizes innovation while also addressing potential risks associated with AI technologies. Regulations focus on promoting responsible AI development and deployment, with an emphasis on industry self-regulation and voluntary compliance. For CISOs and other cybersecurity leaders, it's important to note that the Executive Order instructs the National Institute of Standards and Technology (NIST) to develop standards for red team testing of AI systems. There's also a call for "the most powerful AI systems" to be obliged to undergo penetration testing and share the results with government. The EU's AI Act - an overview The European Union's more precautionary approach bakes cybersecurity and data privacy in from the get-go, with mandated standards and enforcement mechanisms. Like other EU laws, the AI Act is principle-based: The onus is on organizations to prove compliance through documentation supporting their practices. For CISOs and other cybersecurity leaders, Article 9.1 has garnered a lot of attention. It states that High-risk AI systems shall be designed and developed following the principle of security by design and by default. In light of their intended purpose, they should achieve an appropriate level of accuracy, robustness, safety, and cybersecurity, and perform consistently in those respects throughout their life cycle. Compliance with these requirements shall include implementation of state-of-the-art measures, according to the specific market segment or scope of application. At the most fundamental level, Article 9.1 means that cybersecurity leaders at critical infrastructure and other high-risk organizations will need to conduct AI risk assessments and adhere to cybersecurity standards. Article 15 of the Act covers cybersecurity measures that could be taken to protect, mitigate, and control attacks, including ones that attempt to manipulate training data sets ("data poisoning") or models. For CISOs, cybersecurity leaders, and AI developers alike, this means that anyone building a high-risk system will have to take cybersecurity implications into account from day one. EU AI Act vs. US AI regulatory approach - key differences Feature EU AI Act US approach Overall philosophy Precautionary, risk-based Market-driven, innovation-focused Regulations Specific rules for 'high-risk' AI, including cybersecurity aspects Broad principles, sectoral guidelines, focus on self-regulation Data privacy GDPR applies, strict user rights and transparency No comprehensive federal law, patchwork of state regulations Cybersecurity standards Mandatory technical standards for high-risk AI Voluntary best practices, industry standards encouraged Enforcement Fines, bans, and other sanctions for non-compliance Agency investigations, potential trade restrictions Transparency Explainability requirements for high-risk AI Limited requirements, focus on consumer protection Accountability Clear liability framework for harm caused by AI Unclear liability, often falls on users or developers What AI regulations mean for CISOs and other cybersecurity leaders Despite the contrasting approaches, both the EU and US advocate for a risk-based approach. And, as we've seen with GDPR, there is plenty of scope for alignment as we edge towards collaboration and consensus on global standards. From a cybersecurity leader's perspective, it's clear that regulations and standards for AI are in the early levels of maturity and will almost certainly evolve as we learn more about the technologies and applications. As both the US and EU regulatory approaches underline, cybersecurity and governance regulations are far more mature, not least because the cybersecurity community has already put considerable resources, expertise, and effort into building awareness and knowledge. The overlap and interdependency between AI and cybersecurity have meant that cybersecurity leaders have been more keenly aware of emerging consequences. After all, many have been using AI and machine learning for malware detection and mitigation, malicious IP blocking, and threat classification. For now, CISOs will be tasked with developing comprehensive AI strategies to ensure privacy, security, and compliance across the business, including steps such as: Identifying the use cases where AI delivers the most benefit. Identifying the resources needed to implement AI successfully. Establishing a governance framework for managing and securing customer/sensitive data and ensuring compliance with regulations in every country where your organization does business. Clear evaluation and assessment of the impact of AI implementations across the business, including customers. Keeping pace with the AI threat landscape As AI regulations continue to evolve, the only real certainty for now is that both the US and EU will hold pivotal positions in setting the standards. The fast pace of change means we're certain to see changes to the regulations, principles, and guidelines. Whether its autonomous weapons or self-driving vehicles, cybersecurity will play a central role in how these challenges are addressed. Both the pace and complexity make it likely that we'll evolve away from country-specific rules, towards a more global consensus around key challenges and threats. Looking at the US-EU work to date, there is already clear common ground to work from. GDPR (General Data Protection Regulation) showed how the EU's approach ultimately had a significant influence on laws in other jurisdictions. Alignment of some kind seems inevitable, not least because of the gravity of the challenge. As with GDPR, it's more a question of time and collaboration. Again, GDPR proves a useful case history. In that case, cybersecurity was elevated from technical provision to requirement. Security will be an integral requirement in AI applications. In situations where developers or businesses can be held accountable for their products, it is vital that cybersecurity leaders stay up to speed on the architectures and technologies being used in their organizations. Over the coming months, we'll see how EU and US regulations impact organizations that are building AI applications and products, and how the emerging AI threat landscape evolves. Ram Movva is the chairman and chief executive officer of Securin Inc. Aviral Verma leads the Research and Threat Intelligence team at Securin. -- Generative AI Insights provides a venue for technology leaders -- including vendors and other outside contributors -- to explore and discuss the challenges and opportunities of generative artificial intelligence. The selection is wide-ranging, from technology deep dives to case studies to expert opinion, but also subjective, based on our judgment of which topics and treatments will best serve InfoWorld's technically sophisticated audience. InfoWorld does not accept marketing collateral for publication and reserves the right to edit all contributed content. Contact doug_dineley@foundryco.com. Next read this: Why companies are leaving the cloud 5 easy ways to run an LLM locally Coding with AI: Tips and best practices from developers Meet Zig: The modern alternative to C What is generative AI? Artificial intelligence that creates The best open source software of 2023

News outlets are accusing Perplexity of plagiarism and unethical web scraping | TechCrunch

Ambiguity around copyright laws and AI web crawlers complicate matters In the age of generative AI, when chatbots can provide detailed answers to questions based on content pulled from the internet, the line between fair use and plagiarism, and between routine web scraping and unethical summarization, is a thin one. Perplexity AI is a startup that combines a search engine with a large language model that generates answers with detailed responses, rather than just links. Unlike OpenAI's ChatGPT and Anthropic's Claude, Perplexity doesn't train its own foundational AI models, instead using open or commercially available ones to take the information it gathers from the internet and translate that into answers. But a series of accusations in June suggests the startup's approach borders on being unethical. Forbes called out Perplexity for allegedly plagiarizing one of its news articles in the startup's beta Perplexity Pages feature. And Wired has accused Perplexity of illicitly scraping its website, along with other sites. Perplexity, which as of April was working to raise $250 million at a near-$3 billion valuation, maintains that it has done nothing wrong. The Nvidia- and Jeff Bezos-backed company says that it has honored publishers' requests to not scrape content and that it is operating within the bounds of fair use copyright laws. The situation is complicated. At its heart are nuances surrounding two concepts. The first is the Robots Exclusion Protocol, a standard used by websites to indicate that they don't want their content accessed or used by web crawlers. The second is fair use in copyright law, which sets up the legal framework for allowing the use of copyrighted material without permission or payment in certain circumstances. Wired's June 19 story claims that Perplexity has ignored the Robots Exclusion Protocol to surreptitiously scrape areas of websites that publishers do not want bots to access. Wired reported that it observed a machine tied to Perplexity doing this on its own news site, as well as across other publications under its parent company, Condé Nast. The report noted that developer Robb Knight conducted a similar experiment and came to the same conclusion. Both Wired reporters and Knight tested their suspicions by asking Perplexity to summarize a series of URLs and then watching on the server side as an IP address associated with Perplexity visited those sites. Perplexity then "summarized" the text from those URLs -- though in the case of one dummy website with limited content that Wired created for this purpose, it returned text from the page verbatim. This is where the nuances of the Robots Exclusion Protocol come into play. Web scraping is technically when automated pieces of software known as crawlers scour the web to index and collect information from websites. Search engines like Google do this so that web pages can be included in search results. Other companies and researchers use crawlers to gather data from the internet for market analysis, academic research and, as we've come to learn, training machine learning models. Web scrapers in compliance with this protocol will first look for the "robots.txt" file in a site's source code to see what is permitted and what is not -- today, what is not permitted is usually scraping a publisher's site to build massive training datasets for AI. Search engines and AI companies, including Perplexity, have stated that they comply with the protocol, but they aren't legally obligated to do so. Perplexity's head of business, Dmitry Shevelenko, told TechCrunch that summarizing a URL isn't the same thing as crawling. "Crawling is when you're just going around sucking up information and adding it to your index," Shevelenko said. He noted that Perplexity's IP might show up as a visitor to a website that is "otherwise kind of prohibited from robots.txt" only when a user puts a URL into their query, which "doesn't meet the definition of crawling." "We're just responding to a direct and specific user request to go to that URL," Shevelenko said. In other words, if a user manually provides a URL to an AI, Perplexity says its AI isn't acting as a web crawler but rather a tool to assist the user in retrieving and processing information they requested. But to Wired and many other publishers, that's a distinction without a difference because visiting a URL and pulling the information from it to summarize the text sure looks a whole lot like scraping if it's done thousands of times a day. (Wired also reported that Amazon Web Services, one of Perplexity's cloud service providers, is investigating the startup for ignoring robots.txt protocol to scrape web pages that users cited in their prompt. AWS told TechCrunch that Wired's report is inaccurate and that it told the outlet it was processing their media inquiry like it does any other report alleging abuse of the service.) Wired and Forbes have also accused Perplexity of plagiarism. Ironically, Wired says Perplexity plagiarized the very article that called out the startup for surreptitiously scraping its web content. Wired reporters said the Perplexity chatbot "produced a six-paragraph, 287-word text closely summarizing the conclusions of the story and the evidence used to reach them." One sentence exactly reproduces a sentence from the original story; Wired says this constitutes plagiarism. The Poynter Institute's guidelines say it might be plagiarism if the author (or AI) used seven consecutive words from the original source work. Forbes also accused Perplexity of plagiarism. The news site published an investigative report in early June about how Google CEO Eric Schmidt's new venture is recruiting heavily and testing AI-powered drones with military applications. The next day, Forbes editor John Paczkowski posted on X saying that Perplexity had republished the scoop as part of its beta feature, Perplexity Pages. Perplexity Pages, which is only available to certain Perplexity subscribers for now, is a new tool that promises to help users turn research into "visually stunning, comprehensive content," according to Perplexity. Examples of such content on the site come from the startup's employees, and include articles like "A beginner's guide to drumming," or "Steve Jobs: visionary CEO." "It rips off most of our reporting," Paczkowski wrote. "It cites us, and a few that reblogged us, as sources in the most easily ignored way possible." Forbes reported that many of the posts that were curated by the Perplexity team are "strikingly similar to original stories from multiple publications, including Forbes, CNBC and Bloomberg." Forbes said the posts gathered tens of thousands of views and didn't mention any of the publications by name in the article text. Rather, Perplexity's articles included attributions in the form of "small, easy-to-miss logos that link out to them." Furthermore, Forbes said the post about Schmidt contains "nearly identical wording" to Forbes' scoop. The aggregation also included an image created by the Forbes design team that appeared to be slightly modified by Perplexity. Perplexity CEO Aravind Srinivas responded to Forbes at the time by saying the startup would cite sources more prominently in the future -- a solution that's not foolproof, as citations themselves face technical difficulties. ChatGPT and other models have hallucinated links, and since Perplexity uses OpenAI models, it is likely to be susceptible to such hallucinations. In fact, Wired reported that it observed Perplexity hallucinating entire stories. Other than noting Perplexity's "rough edges," Srinivas and the company have largely doubled down on Perplexity's right to use such content for summarizations. This is where the nuances of fair use come into play. Plagiarism, while frowned upon, is not technically illegal. According to the U.S. Copyright Office, it is legal to use limited portions of a work including quotes for purposes like commentary, criticism, news reporting and scholarly reports. AI companies like Perplexity posit that providing a summary of an article is within the bounds of fair use. "Nobody has a monopoly on facts," Shevelenko said. "Once facts are out in the open, they are for everyone to use." Shevelenko likened Perplexity's summaries to how journalists often use information from other news sources to bolster their own reporting. Mark McKenna, a professor of law at the UCLA Institute for Technology, Law & Policy, told TechCrunch the situation isn't an easy one to untangle. In a fair use case, courts would weigh whether the summary uses a lot of the expression of the original article, versus just the ideas. They might also examine whether reading the summary might be a substitute for reading the article. "There are no bright lines," McKenna said. "So [Perplexity] saying factually what an article says or what it reports would be using non-copyrightable aspects of the work. That would be just facts and ideas. But the more that the summary includes actual expression and text, the more that starts to look like reproduction, rather than just a summary." Unfortunately for publishers, unless Perplexity is using full expressions (and apparently, in some cases, it is), its summaries might not be considered a violation of fair use. AI companies like OpenAI have signed media deals with a range of news publishers to access their current and archival content on which to train their algorithms. In return, OpenAI promises to surface news articles from those publishers in response to user queries in ChatGPT. (But even that has some kinks that need to be worked out, as Nieman Lab reported last week.) Perplexity has held off from announcing its own slew of media deals, perhaps waiting for the accusations against it to blow over. But the company is "full speed ahead" on a series of advertising revenue-sharing deals with publishers. The idea is that Perplexity will start including ads alongside query responses, and publishers that have content cited in any answer will get a slice of the corresponding ad revenue. Shevelenko said Perplexity is also working to allow publishers access to its technology so they can build Q&A experiences and power things like related questions natively inside their sites and products. But is this just a fig leaf for systemic IP theft? Perplexity isn't the only chatbot that threatens to summarize content so completely that readers fail to see the need to click out to the original source material. And if AI scrapers like this continue to take publishers' work and repurpose it for their own businesses, publishers will have a harder time earning ad dollars. That means eventually, there will be less content to scrape. When there's no more content left to scrape, generative AI systems will then pivot to training on synthetic data, which could lead to a hellish feedback loop of potentially biased and inaccurate content.