It’s getting harder to tell whether we should be worried about AI, or just exhausted by the marketing hype disguised as warnings

For most of the past year, AI companies have told us that their models are getting increasingly dangerous. It seems as though this trend started when Anthropic made a model called Mythos. The company claimed the model is better than any before at identifying bugs and vulnerabilities in software, making it extremely useful, but potentially catastrophic if it were in the hands of bad actors who could use it to basically hack all software everywhere. At the time, the U.S. government imposed a restriction forbidding the company from releasing the model to the public. Only later did the company release a version of Mythos with safety guardrails called Fable. Obviously, everyone wanted to use Fable. After all, if the government said Mythos is too dangerous for anyone to use, of course we’d want to use whatever version of it Anthropic will give us. It’s a strange marketing strategy when you think about it, but it’s also remarkably effective. OpenAI, Anthropic, and others have warned us that AI models would, or at least could, eventually become smarter than humans. When that happens, the story goes, they’ll take our jobs, be capable of launching cyberattacks, create biological weapons, or—in the most extreme scenario—wipe out humanity altogether. That’s not exactly the kind of thing most companies brag about, but there’s an obvious benefit. Nothing makes your product sound impressive quite like suggesting it might be too powerful for humanity to control. Of course, if you tell people your product is super dangerous but then you ship it anyway, it’s hard to know whether you’re serious. At some point it just seems like a marketing tactic to make more people want to use what you’re selling. That’s created a real problem, because now some of the people building the most powerful AI models are saying that we should slow down. And this time, it seems like they actually mean it. On Saturday, Anthropic CEO Dario Amodei published a blog post titled “We Must Pace the Frontier,” arguing that AI companies should deliberately slow the rate at which they develop more capable frontier models. Amodei isn’t calling for AI development to stop. On the contrary, he says that advancement is crucial since if they don’t do it, someone else will. And if that someone else is China, that would be a national security risk. Instead, he says companies should give safety research enough time to keep up with rapidly advancing capabilities. It does invite the question of “why now?” The answer is either terrifying or another scare-marketing tactic. It’s hard to know which because we’ve basically heard this before. But maybe this time he really means it. “My first concern is that, since roughly this summer, AI has been advancing drastically faster, driven primarily by AI’s growing ability to build the next generation of AI,” Amodei writes. That’s what researchers call recursive self-improvement, and it’s one of the most significant tipping points with artificial intelligence. Until now, it’s mostly been theoretical, but apparently the big AI companies are starting to see evidence that it’s becoming more than just theory. The idea is pretty simple: Humans build an AI system. That AI system is able to help humans build a better AI system. That better system is capable of building an even better one than before, and—eventually—it’s able to build models without the help of humans. At some point, humans aren’t making a better AI model; the AI is making AI. That accelerates the process and also makes it much harder for humans to understand what the models are truly capable of. That, for obvious reasons, would be very bad. According to Amodei, that’s no longer theoretical. “It is starting to happen across the industry, including at Anthropic,” he writes. His second concern is considerably weirder. The OpenAI hack of Hugging Face seems to have been a wake-up call. In that case, the model created thousands of agents that exploited vulnerabilities in their testing environment, communicated and coordinated with other agents, gained unauthorized access to outside systems, and collaborated to interfere with the testers evaluating their performance. Amodei said they behaved like a “fanatically devoted collective,” willing to sacrifice individual agents to accomplish the group’s objective. “It’s my worry that in 6–12 months such a swarm could be capable of taking over the entire internet,” he wrote. That is an extraordinary prediction. On the other hand, it’s just a prediction. Yes, Amodei has access to far more information about frontier AI models than pretty much anyone else in the world, but it’s also hard to simply take his word for it. That is, after all, part of the problem. When AI companies are raising money, they talk about their models as the technology that will transform the global economy. When they’re selling those models as products, they talk about how companies that don’t adopt AI are going to be left behind. And then there’s artificial general intelligence, which is either right around the corner or just up the foothills. When they talk about safety, on the other hand, the technology is so powerful it might destroy us all. It’s entirely possible that all of those things are true. But it’s not hard to understand how it becomes difficult to separate out what’s a warning and what’s a sales pitch when they’re using the same words. It’s worth mentioning that there’s another problem with Amodei’s proposal. Most of what he suggested would both make AI safer and also be very good for Anthropic. For example, Amodei is calling for frontier AI companies to coordinate their pace of development. You can make a real safety argument, but there’s also an obvious competitive benefit. If Anthropic decides to slow down the pace of development to ensure alignment, it risks being left behind if OpenAI and Google charge ahead. On the other hand, if everyone agrees to slow down together, Anthropic gets the extra time without giving up its competitive ground. None of that means Amodei is being insincere. It just makes it harder to know his real motivation. It’s like the AI boy cried wolf, but also the wolf is real, and you should buy one, but maybe it will eventually try to eat you. It’s complicated. There’s a lesson here for every leader, which is that trust is your most valuable asset. I’ve written about that very thing so many times before, but it bears repeating because it really is the single most important part of this story. If we thought they really had our best interests in mind, it’d be so much easier to believe what they have to say and not just assume it’s a weird form of hype marketing. Maybe this time the AI companies really mean it when they say we should be worried about what they’ve built. It turns out that a little trust and credibility go a long way when your message comes down to “this time we really mean it.” EXPERT OPINION BY JASON ATEN, TECH COLUMNIST @JASONATEN

Call for global AI rules after Claude bioweapon research

The European Commissioner with responsibility for tech regulation has said that global rules on artificial intelligence (AI) are "really needed". Commission Executive Vice-President for Tech Sovereignty, Security and Democracy Henna Virkkunen was responding to questions about AI company Anthropic which said it has disrupted attempts to use its models to develop biological weapons. Anthropic also said it has shut down multiple cases of state-sponsored surveillance operations using its technology in China, Iran and west Africa. Speaking to reporters in Dublin, Ms Virkkunen said in Europe there is the AI Act. "According to our AI Act, if those kinds of very capable models are coming to the market the service provider has to, all the time, assess and mitigate the risks and also, when the models are on the market they have to monitor the situation all the time," Ms Virkkunen said. She said we know that this is not the case globally. "It's very important that we continue international co-operation because I see that global rules are really needed here. "With AI there is, of course, huge potential and so many positive things, but there is also risk, and we have to be aware of that, cyber risks and bio risks, we are aware that these are very serious topics " she added. In a report titled 'Detecting and countering misuse of AI', the San Francisco-based AI company Anthropic listed five case studies of potential biological misuse of its models which it says "is one of the most serious risks of frontier AI models". The scientific research in those case studies involved work around the mosquito-born chikungunya virus, a "highly-pathogenic" strain of the bird flu, a family of viruses that include smallpox and mpox, venoms and other toxins. In the examples identified, Anthropic said the actors circumvented controls imposed to prevent users from unsupported regions accessing their models, while the actors also "engaged in other efforts to obfuscate the purpose of their research to evade our safeguards". "When we detected and investigated these cases, we banned the users' accounts and incorporated our investigative findings into our frontier model safeguards, enforcement, and threat intelligence processes to better prevent, detect, and disrupt these activities in the future," the report said. Anthropic said the intent of the actors was not clear-cut and the AI lab did not identify them. "The individuals implicated in these case studies are working scientists. We do not assert that they intended harm, and identifying them or their labs could expose them to harm," the report said. Anthropic said its aim in sharing the examples is to "spark conversation within the AI industry, and with governments, about emerging biological risks and how best to counter them". Anthropic says China, Iran used its AI to aid spying Also in the report, Anthropic said it shut down multiple cases of state-sponsored surveillance operations that used its AI models, and warned that governments and state-aligned actors are increasingly using AI to spy on ethnic minorities and dissidents. The incidents were found and disrupted between January and July, and they originated in China, Iran and west Africa, Anthropic said in a report about misuse of its models. These surveillance campaigns "targeted the same diaspora and dissident communities these regimes have historically targeted". "These include pro-democracy figures in Hong Kong, Tibetan and Falun Gong communities across Asia, and Iranian minority communities and opponents of the Iranian regime abroad," the report said. In one case, Iranian actors developed a way to identify people through their social media accounts; in another, a contractor working for Malian national security authorities used Claude "to design the underlying software that enabled the intelligence gathering". "AI is now being used in place of an engineering workforce," the report said. Anthropic also disrupted efforts by users to design weapons and create dating scams. The San Francisco-based AI lab also accused Chinese developers of deceptively using Claude to produce responses to user queries while simultaneously "distilling" that data to improve their own models. Distilling is an industry term that refers to a process where one AI model is used to train another one. Chinese developers have previously been accused of using that technique without permission, essentially stealing resources from American labs such as Anthropic and OpenAI. Now Anthropic alleges that China's Moonshot and Deepseek have also secretly used Claude to generate answers given to its users. "In one instance, over a 10-day period, Moonshot relayed almost 300,000 customer requests to Anthropic" through a "network of 5,380 fraudulent accounts, most of which appeared to be located in Singapore and Japan," the report said. Some of that data contained sensitive user information, potentially in violation of privacy agreements. "We do not know if Moonshot notified their customers that their requests were being rerouted to Anthropic and exposed to a third party," the report said. Deepseek used similar techniques. Earlier this week, an artificial intelligence researcher who left OpenAI to join Anthropic has decided to leave the industry, accusing both US companies of "gambling with our lives" in the race to develop AI models capable of self-improvement. By Brian O'Donovan Additional reporting AFP

Top AI companies have discussed creating their own standards body

Anthropic, Google and OpenAI have discussed creating an AI industry standards body, people familiar with the situation told CNN. The conversations preceded the events of the past week, when a former Anthropic researcher resigned, saying the top AI companies were not “behaving responsibly” – and also before Anthropic CEO Dario Amodei proposed embedding third party watchdogs at AI companies. The Information first reported on the conversations. According to those familiar with the situation, the catalyst for the discussions was a July essay written by Google DeepMind founder Demis Hassabis, who proposed a US-led Standards Body modeled on the Financial Industry Regulatory Authority. The standards body, Hassabis argued, should test advanced AI models before deployment. Hassabis said the body should be a public-private partnership, overseen by the government but funded by the industry and staffed by “independent leading technical experts and open-source representatives.” The conversations are ongoing, one of the people said, and they are continuing with or without the Trump administration. In July, shortly after Hassabis publicized his proposal, Bloomberg reported Treasury Secretary Scott Bessent was considering a FINRA-like independent regulatory agency for AI that would report to the Securities and Exchange Commission. Not all tech leaders are on board. Meta CEO Mark Zuckerberg reportedly advised President Donald Trump against the idea in a phone call this summer, Politico reported earlier this month. Representatives for Google, OpenAI and Anthropic either declined to comment or did not respond to requests for comment. There are currently few industry standards or regulations for AI models and how they’re tested. Concerns over the safety of AI models and their lab tests have been raised after several incidents this summer where AI agents went rogue during testing. In the most severe case, OpenAI agents escaped their testing environment and hacked into another company’s systems to cheat on a cybersecurity test. The White House has an opaque system in place where AI companies can voluntarily submit their latest models for government review up to 30 days before they are publicly released. But the eligibility requirements for models or the process of the review has not been publicized. Late last month, a White House official told CNN the administration “continues to work with industry stakeholders on the implementation of the framework.” The AI companies say they want better standards. “I believe that shared safety standards and international coordination on further AI development need to be priorities now,” OpenAI chief scientist Jakub Pachocki told reporters at a briefing earlier this month. When asked what progress had been made on an industry wide standards body, Pachocki said they had been talking “to some external organizations about potential concrete standards we could put in place,” and that they’d have more to share “in the next months.” OpenAI CEO Sam Altman said in a late Sunday night post they “look forward to collaborating with our colleagues across the industry to formulate the best version” of industry standards. House Speaker Mike Johnson told CNN’s Jake Tapper on Sunday there’s little consensus among the companies, who are fierce competitors, on what the standards and guardrails should be. “There’s no consensus among them. And Congress is obviously less qualified than the people who are pushing this frontier to know all the ins and outs of it,” Johnson said, adding he thinks the AI leaders should immediately meet to hammer out some standards alongside the government. “So this has to be a partnership with the industry itself, with the corporations that are doing this and with the policy and lawmakers.” By Hadas Gold

Nvidia’s Jensen Huang Thinks Young People Are Overlooking the Skill That Will Matter Most

Tech executive Jensen Huang wants younger generations to focus on the foundation of engineering before heading straight into the world of artificial intelligence. For decades, the Nvidia CEO has helped foster the AI boom. But as younger generations enter the space, Huang wants newcomers to focus on one discipline he believes will outlast any wave of automation: engineering. “Engineering teaches you to start from first principles, grounding everything in physics and mathematics and applying it to real problems,” Huang said in a statement to Fortune. “It teaches you to break complex challenges into solvable parts, which is how we build systems at Nvidia.” At a moment when entry-level workers face growing uncertainty about how AI will reshape their careers, Huang is urging them to bet on a skill set rooted in problem-solving and grit. “You learn that meaningful problems require resilience, as the most important outcomes often seem impossible at first,” he said. “Over time, I’ve come to see engineering as the most noble profession. It is the foundational building block of modern society, enabling everything from infrastructure to computing to science.” Engineering first Huang’s emphasis on engineering isn’t just a response to the AI boom. It’s rooted in his own career. He studied electrical engineering at Oregon State University, where he joined IEEE, the world’s largest technical professional organization. Decades later, IEEE awarded Huang its highest honor, the Medal of Honor, recognizing his role in pushing the boundaries of modern computing. At the ceremony, Huang made clear that the AI era will be defined by both disruption and expansion—and that engineers will sit at the center of both. For Huang, AI isn’t eliminating work. Instead, it’s changing who defines it. “Engineers ultimately are the ones that take an invention and advance it in such a way that it’s safe, beneficial, ultimately transformative to society,” Huang said, according to Fortune. “The engineers in the AI industry must advance AI in service of a better future for all of us.” However, not everyone is in absolute agreement with Huang’s assertion. “First-principles thinking is a powerful mental framework, and it’s something people should absolutely incorporate as AI reshapes how we work. But I wouldn’t think about engineering as the answer in absolute terms,” Juan Jose Lopez Murphy, the head of data science and artificial intelligence at Globant, told Inc. The expansion and accessibility of AI isn’t just a trend, but a historical phenomenon. According to a report from Stanford, AI reached 53 percent public adoption in just three years, faster than any technology on record. This has led to a wider use of the new capabilities, such as vibe coding, where users can approach AI software with natural language prompts to instruct an AI assistant to build applications, websites, or games without writing or reviewing the source code manually. In this new environment, some are pushing younger generations toward developing a diverse set of skills—instead of specializing. “Engineering is one of many valuable foundations, and in a world changing so quickly, over-specializing can actually make you more fragile,” said Murphy. “The most durable skill is being able to apply strong problem-solving principles across disciplines and adapt as the technology evolves.” AI fluency According to the U.S. Bureau of Labor Statistics, engineering roles, from computer hardware to electrical systems, are projected to grow faster than average, with demand accelerating across AI, energy, and defense. Huang sees this moment as a meaningful opening for young professionals. “AI is expanding the scope of human work, not shrinking it. We’re busier than ever because we have more ideas to pursue, and AI makes it possible to go after them,” he told Fortune. “Jobs will change because tasks will change, but work grows with productivity.” That growth is already visible inside Nvidia itself. The company plans to double its workforce to roughly 75,000 employees over the next decade, hiring that will likely lean heavily on engineering talent, Fortune reported. But Huang’s advice doesn’t stop at choosing a major or career path. He argues that everyone, regardless of role, needs a baseline level of technical fluency. “Every young person should become an AI expert,” he said. “It is an incredible technology that significantly lowers the knowledge barrier of any professional field… AI fluency will empower you and elevate your chosen craft.” BY LEILA SHERIDAN, NEWS WRITER

Apple Broke Its Own Rules for 2 New Macs—and AI Buyers Are Snapping Them Up

Apple has predictable habits for releasing new hardware, so it was a surprise when the tech giant revealed new Macs with updated, more powerful chips last week. The refreshed hardware arrived before Apple’s traditional iPhone event, which this year occurs on Wednesday, September 9. So why did Apple release these machines, the Mac Mini and Mac Studio, when it did? Thanks to a new report, we think we understand: Apple’s updated desktop Macs were announced early because they’re selling like hotcakes and Apple wants to sell more of them to eager buyers. These machines are desirable because they’re ideal for acting as hosts for local AI models. According to news site The Information, the sudden AI-driven popularity of the machines caught the Cupertino, California-based company by surprise. Apple even apparently lacked the necessary enterprise-centric support teams. Why would anyone want to run an AI locally, though, when enterprise-ready AI systems are everywhere at the moment? Everyone from Google to Microsoft is offering powerful AI services from their giant cloud server farms, with plenty of business-centric tools. And platforms like Amazon Web Services offer powerful developer-friendly AI environments from their data centers. The keywords are “cloud” and “data centers.” These third-party AI offerings require gigantic computer facilities, which is why they’re offered by deep-pocketed tech giants like OpenAI on a remote-access basis. But if you run your own, less powerful AI models on your own hardware, then you remain in control of the entire operation. This means you can train the models with data that no one else can access. It means only you and your staff see the outputs when you query an AI. It means that if your developers improve the AI algorithms in clever ways, then only you benefit. All of this is useful for intellectual property protection and even, for some companies, for preserving audit trails for sensitive data handling. Finally, hosting an AI model on your own hardware means you retain control of the cost. Spiraling, expensive AI bills for remote-access services have surprised many companies in recent months, landing some users in the headlines because they owe so much. Thus Apple’s reasonably priced Macs are suddenly hugely popular. The newly updated machines are even faster than earlier versions for AI purposes, too: Apple highlights that the new Studio can carry out “Up to 9.8x faster” LLM AI prompts, and the new Mini has similar improvements over previous editions, thanks to the new M6 and M5 Pro processors inside. MacRumors pointed out that when Apple announced the new Macs, it even highlighted that they can be clustered together to create a powerful compute facility (think almost-a-supercomputer on your desktop) for exactly this purpose. The Mac Mini’s prices start at just $900, leaping up if you select more powerful chips or more memory. The more powerful Mac Studio starts at $2,500, and higher-end configurations can reach many thousands of dollars. But these remain one-off purchases, compared with paying an open-ended bill for AI-as-a-service from the likes of OpenAI. There’s a lesson here for many smaller-business owners. AI is the buzziest tech right now, and it comes with the promise of boosting efficiency and maybe helping to cut costs, too. But typically, because AI is delivered by cloud-based companies like OpenAI or Google, each system brings its own risks in terms of potential expenses as well as data leaks. Surveys show a staggeringly high number of workers merrily type sensitive company data into these chatbots, unaware of the risk they pose. So running your own AI models on machines like Apple’s new ones could be a good solution. And even if you’re not in the market for building your own safer, Apple-powered local AI server, other options exist. Some AI models are freely downloadable and can run on relatively modest hardware. This may be an even simpler and cheaper option, depending on your company’s AI needs. The news should also serve as a reminder to savvy CEOs that deploying AI is not a one-and-done maneuver. The tech is advancing fast, so continuously evaluating which AI best serves your needs at what price could save you money in the long run. BY KIT EATON @KITEATON

5 Side Hustles You Can Start Today With AI

Artificial intelligence can be a significant income source for people who choose a smart side hustle. AI may still have a long way to go with some tasks, but with others, it can increase productivity and results. And there are a number of companies looking to take advantage of those benefits. The best AI-driven side hustles use the technology as a great equalizer. Small business owners, particularly of established businesses, often use AI in a limited fashion, if at all, says Peter Hansen, director of research and policy analysis with the National Federation of Independent Business (NFIB). Those founders value and emphasize the human connection to their customers. But even if they don’t use it, they recognize AI can help in other areas. Many of those owners don’t have the time to learn the ins and outs of various AI systems. And even companies that are using AI are often not taking full advantage of it. That creates an opening for side hustles that center around AI. “More than three-quarters of businesses already using AI report higher productivity, yet most are still applying it to a narrow set of tasks,” says Simon Worsfold, head of data communications at Intuit QuickBooks, which puts out a monthly Small Business Index exploring market trends. “What is emerging is a large, underserved layer of advisory, implementation, and training services that small businesses need in order to scale their use of AI,” he says. Here are five ideas for AI-focused side hustles that could be lucrative—or even become the foundation of a full-time business. AI Consultant This is a side hustle with multiple branches. Some people are well suited to help people or businesses integrate AI with their calendar, notes, and local files. Others can develop workflow audits that review how a company’s employees work to discover repetitive tasks that the AI could handle. Rates for an AI consultant range from $150–350 per hour for independent workers and people running it as a side hustle. Those can jump as high as $600 per hour if you convert this side hustle into a full-time company. AI Researcher To stay competitive, businesses need to know what their competitors are doing and what’s trending in their industry. That data can be remarkably complex, however, and not every founder or CEO has the time to do the deep dive it requires. Side hustle operators with strong AI skills can use the technology to filter the data into easily digested insights, then sell those as reports, newsletters, or direct subscriptions. You’ll need to carefully identify the best source material and be sure to double check the output, since AI’s comprehension is still flawed. “The AI doesn’t give the correct answer all the time,” says Shaun Ghavami, Founder of 10XBNB, a side hustle-turned-business that now oversees a portfolio of over $100 million of short-term property rentals. “The real skill is in prompting.” GEO (AI Search) Optimization For the past two decades, businesses have hired specialists to help them rank high on search results pages. While that’s still valuable placement, smart business owners are also looking into how they can be among the frequently-cited sources for AI search results. People with a combination of AI and content knowhow can create a side hustle specializing in this. You’ll teach companies the best ways to refine their content and meta tags and improve readability for large language models. If you have the tools to measure analytics, such as Alli AI, Frase, and MarketMuse, you have a better chance of establishing a recurring relationship with those clients, increasing your income and establishing a base should you want to make this more than a side hustle. Content Repurposing (Clipping) Both content creators and businesses regularly churn out videos and stories as promotional tools, or in an effort to be seen as thought leadership experts. Quite often, though, that content needs to be pared down before they’re posted on everything from TikTok to LinkedIn to Instagram. And that takes time, which the companies and content creators may not have to spare. As an AI content repurposing expert, you can transform the long-form content into digestible bites to widen the potential audience in a much shorter period of time. Use tools like CapCut, Canva, and ChatGPT to effortlessly pick and highlight the most relevant parts. Create AI Microtools AI can be intimidating to many people who don’t know where to start, especially when it comes to creating agents or GPTs that are relevant to their needs. A smart side hustle operator, though, knows that you don’t have to be an expert to create your own AI tools. Sites like Gumloop let you visually build workflows that can be run at large scale. You’ll use a drag and drop template that lets you create workflows that do just about anything, from running a search engine optimization audit of a Website to a ranking system for job applicants. You can either custom create those for clients or build general purpose tools and sell them on Etsy or some other marketplace. BY CHRIS MORRIS @MORRISATLARGE

What Burning Man Can Teach Startups About Surviving the AI Age

Turning an idea into a business used to be so costly and time consuming that, when we had thoughts such as, “What if an app could let my neighbors order homemade dinners from each other?”, we rarely took action. Instead, we just waited until someone brought something to market and then told our friends, “That was MY idea!” Now, with AI, you can turn any frustration-induced idea into an app or prototype in a weekend. The barrier to success for startups is no longer execution; it’s ease. How can you succeed if your product can be easily produced by everyone else? The answer may be found right now in Black Rock City. Once a year, tens of thousands of people descend upon the Nevada desert to co-create Burning Man. They produce awe-inspiring art installations, build outrageous “mutant” vehicles, and put on immersive performances guided by principles like gifting, radical self-reliance, radical self-expression, and communal effort. Stanford professor Fred Turner explains how Burning Man provided the “cultural infrastructure” for much of the startup tech industry. Open-source, collaborative, project-based ways of working and utopian dreams dominated the early ethos of companies like Google. AI can’t copy the most important aspect of Burning Man. AI can render a flame-throwing octopus sculpture in thirty seconds, but it can’t render the experience of the guy who spent four months welding it in his garage. The value of Burning Man comes from being there, expressing yourself, building something, meeting people, and contributing. Participation, belonging, and shared experience are hard to duplicate. This suggests three places to look for customer value your competitors can’t duplicate. When technology makes things abundant, look for what becomes scarce. When everyone has a vending machine, fast snacks are no longer a lucrative business. AI didn’t just lower the barrier to entry. It removed it and bulldozed the building behind it. If you want a competitive moat today, allow people to create unique value for themselves and others in unplanned, irreproducible ways. Don’t just ask, “How can we use AI to create more value for our customers?” Ask: “What can our customers create for one another that we can’t, and AI can’t?” My last article explained how The Rocky Horror Picture Show turned audiences into participants and created a 50-plus-year loyalty phenomenon. Burning Man takes the idea even further. “Burners” don’t just enjoy a product. What they create for one another is the product. When technology makes things easy, look for meaningful friction. Ease and value don’t always go together. Burning Man is what happens when nobody optimizes the customer journey. Most founders see a week in the desert hauling water and building shelter and think: how do we remove this pain point? Burners built a culture around refusing to answer that question. You can’t just watch Burning Man. You have to invest heavily in creating it. As much as humans like to make things easy, we also value what’s hard-earned. We invented the car to stop walking, then invented the treadmill so we could walk hard—and post about it. As AI makes more things easier, people will start looking for new hard-earned experiences they can grow from and be recognized for. Before using AI to eliminate effort for your customers, ask whether any of that effort creates mastery, commitment, connection, pride, or a sense of ownership. Eliminate the friction people resent, but amplify the friction they value. When technology makes things predictable, create room for surprise. AI writes in a voice it’s already heard about things that have already happened. That’s why so many AI logos look like they’ve been designed by the same nonexistent designer, and why AI copy reads like it was assembled from other copy that was assembled from other copy. Burning Man has no training data. Nobody can predict what happens next, including the people running it.It’s hard to pre-engineer surprise, but you don’t need to. You just need to create a way for unique individuals (your customers) to contribute meaningfully, create value for one another, and build things you never would have thought to build for them. Here’s the Bottom Line AI makes things easier to build but harder to matter. Your greatest value might not come from what you produce, but from what happens when your customers show up and contribute. BY STEPHANIE DAVIS

Bill Gates Says the AI Era Will Be Turbulent. Here Are 7 Things You Should Do Right Now

Imagine showing up for work one morning and discovering that the task you spent years learning to do can suddenly be done by AI — faster, cheaper, and perhaps even better than you can do it. For millions of people, that scenario may no longer be science fiction. Bill Gates recently published a thought-provoking essay about what he calls the “turbulent AI era.” His message is both optimistic and unsettling: AI has the potential to make life dramatically better, but the transition could be enormously disruptive. Gates believes AI will affect both white- and blue-collar jobs. Unlike previous technological revolutions, the change could happen remarkably quickly. He argues that society needs to start preparing now rather than waiting until millions of workers are displaced. So, what does that mean for you? Here are seven things you can do right now. 1. Stop assuming your job is safe. AI isn’t just coming for repetitive factory jobs. Gates points to sales, customer support, software engineering, paralegal work, data analysis, and other occupations as vulnerable. Take an objective look at what you do every day and ask yourself: How much of this could AI eventually do? 2. Learn AI before you need it. Don’t wait until your boss tells you to start using AI. Experiment with the tools available to you now. The more you understand what AI can and can’t do, the better prepared you’ll be. 3. Use AI to make yourself better. Instead of viewing AI solely as a competitor, make it your assistant. Use it to research, brainstorm, analyze information, automate routine work, and free up your time for higher-value activities. 4. Double down on being human. Gates proposes an intriguing idea he calls “Human Reserved” — work that society deliberately decides should remain in human hands. Empathy, judgment, trust, creativity, leadership, and genuine human connection will become increasingly valuable. Strengthen those skills. 5. Keep learning. The worst strategy in a rapidly changing workplace is standing still. Identify skills that are becoming more valuable in your industry and start developing them before you desperately need them. 6. Help your people prepare. If you’re a leader, don’t pretend nothing is changing. Talk openly with your employees about AI. Give them access to tools and training and involve them in deciding how AI should be used. Your people deserve the opportunity to adapt. 7. Don’t surrender your judgment. AI can produce impressive answers, but that doesn’t mean you should stop thinking. Gates is particularly concerned that overreliance on AI could weaken critical-thinking skills. Use AI to extend your intelligence — not replace it. The AI revolution isn’t something that’s going to happen someday. It’s happening right now. You probably can’t control how quickly AI develops or what your competitors do with it, but you can control how prepared you are. And the time to prepare isn’t after your job changes. It’s before. EXPERT OPINION BY PETER ECONOMY, THE LEADERSHIP GUY @BIZZWRITER

The 1 Big Compensation Decision Founders Should Never Hand to AI

AI has quickly made itself useful across hiring. It can sift through résumés, draft a job description in seconds, and match a role to market benchmarks. For a small company that never had a compensation team, that’s a genuine unlock. But things get thornier when AI has a say in what the actual salaries that companies pay their employees should be. And that’s the bigger question that founders now have to grapple with—whether they should let AI set the number itself. That pressure is only growing, and this year it found a sharp edge in the law. On October 8, 2025, California outlawed a hiring shortcut that employers had leaned on for years. When Governor Gavin Newsom signed SB 642 into law, effective January 1, 2026, the state redefined what a “pay scale” in a job posting is allowed to be: a good-faith estimate of what an employer expects to pay a new hire. The point, as the law firm Morgan Lewis noted, was to stop employers from posting “meaningless pay scales simply to be in compliance.” So the old dodge is dead. Post a range wide enough to protect yourself, say from $80,000 to $200,000, and you’re not being cautious. You’re non-compliant. California isn’t alone, and these rules don’t only apply to big companies. As of early 2026, more than a dozen states now require employers to disclose pay, and the thresholds reach startups, not just giants. In Colorado, a single employee is enough to trigger disclosure, while New York City draws the line at four; states like California, Illinois, and Washington set it at 15. For founders who have always set pay based on their instincts, that guess is now a public document, and the person most likely to read it closely is the employee who already holds the job. And not just them: everyone doing similar work will see the number too, and measure themselves against it. The problem was always there Improvised pay stays invisible while it works, and it works right until two people doing the same job discover they earn thousands apart, or a strong hire leaves for a competitor who bothered to benchmark the role when you didn’t. Jonas P. Johnson, who works on compensation modeling at ERI Economic Research Institute, a firm that has built salary data for more than three decades, says instinct-based pay produces one of two problems, and they aren’t equally fixable. Underpaying firms struggle with turnover, while overpaying firms struggle to keep prices competitive. Most founders assume overpaying is the safe mistake. Johnson says it’s the harder one to escape. Underpaying is the easier hole to climb out of, because raising pay tends to slow the turnover it caused. Overpaying is the real trap, since you can’t take money back once people are earning it. “Employees won’t accept pay cuts,” Johnson says, so a company that has drifted above market has to hold raises below average for years to come back down, all while trimming costs to survive. The case he sees most is the firm that bought out a competitor and inherited its payroll, absorbing someone else’s guesswork and making it their own. How one raise becomes a chain reaction Transparency rules do more than reveal the gap between what you pay and what the market pays. They reveal the distance between your newest hire and your most loyal one, which is the first comparison employees make. The way it unfolds is easy to trace. When market rates climb, you post a higher range to attract someone new, and that range is now visible to the person who has held the job well for three years, and to their peers in similar roles. Fresh money at the top squeezes the space between new and existing staff, and Johnson warns it “can lead to turnover among existing employees unless the organization re-scales the entire salary structure.” A single number in one posting can force you to reopen everyone’s. Compensation professionals call this pay compression, and it hits morale before it shows up in a budget. Jason Greer, an employee and labor relations expert, says a compressed structure creates “zombie employees” who “show up, clock in, and do the minimum.” They haven’t walked out because they have learned what a newer colleague earns and have reduced their effort. His advice is to understand your local market before circumstances force the issue, because “you might be global, but to your employees, you’re local.” Why the average raise misleads you Guesswork is seductive because there’s always a tidy number to reach for. Heading into 2026, most surveys cluster raises between 3.2 and 3.5 percent, down from post-pandemic highs near 4.4 percent. But the average describes almost no one. Some jobs see no salary growth year over year, Johnson says, while others climb 8 percent. Pay tends to lurch rather than climb, and his clearest example comes from the start of the pandemic, when compensation for forklift operators had been flat for years and then jumped 22 percent in two months as supply chains seized. Budget off the blanket figure, and you underpay the very people you can least afford to lose. The one call you shouldn’t hand to AI None of this complexity means a 30-person company needs the machinery of a Fortune 500. A large firm leans on formal pay grades to keep several hundred jobs consistent, because at that scale nobody can price every role by hand. A 40-person company has maybe 20 distinct roles and can price each against the market directly, without collapsing them into bands. The point isn’t to build an elaborate structure, since none of this needs a complicated setup. You just have to know what a job is really worth before putting a number on it. What makes that reachable now is that the grunt work has gotten cheap. Matching your internal roles to market data once meant hiring a consultant for weeks. Today, the same first pass takes about a day using AI tools. But Johnson is clear that speeding up the work does not mean handing over the decision. Software can widen the field of comparable jobs and flag where your pay drifts out of line, yet “all final decisions should be made by humans,” he says, because an AI model will always return a confident number even when the data behind it is too thin to trust. Workforce strategist Terri Gallagher pushes the point even further, arguing that compression isn’t a pay problem at its root, but instead “a workforce strategy issue.” The number you choose shows what you think the work is worth, and no tool can decide that for you. The answer, then, isn’t to keep AI away from pay. Let it do the first pass, gather the comparisons, and show you where you’ve drifted. Then make the final call yourself, because that judgment is the one part of the job that was always yours. The expensive answer For most of the past decade, “we’ll figure out compensation as we grow” was a reasonable thing to say. The cost of having no system stayed hidden; a little attrition here, a bruised ego there, all of it absorbed into the churn of a growing company. Transparency law takes that hiding place away. The guess is a public document now, the compression is plain to the people it costs, and the tools to do the work properly have never been more affordable. A founder who treats all this as a compliance headache will keep patching it one uncomfortable job posting at a time. But the one who treats it as a reason to finally learn what each role is worth will find the law was never the real issue. The real issue was always whether they knew what they were paying for. EXPERT OPINION BY KOLAWOLE ADEBAYO

Generative AI Is Fading. Here Are 4 Bets on What’s Next in Tech

The thing about AI is you have to be at the forefront of it for a lot of the AI noise to make sense. Like, it’s not that you have to believe in AGI, artificial general intelligence, computers that think. I don’t. You just have to ask yourself what the people who believe in AGI are betting on next. TL;DR: Robots! Finally! Hope they’re not the killer kind! Let’s take a semi-serious look at where AI is heading as we tire of the generative AI phase and settle comfortably into the data-as-moat phase. What comes next? This is what I’m seeing and hearing. One Robot to Rule Them All This was always gonna end in killer robots overthrowing us humans, right? From my college days, I loved robots. The closest I could get to robots was industrial engineering — simulation, robotics, manufacturing lines — really boring robot stuff, but I loved robots the way some kids love trucks and other kids love accounting. What, you think your accountant doesn’t get psyched over a good 1120-S? I keep telling my daughter at Arizona State she needs to kidnap one of those food-delivering Starships and reprogram it to trek all the way across the country to our home, and then do my bidding. But she never does this, she just studies and stuff. What the robot revolution really means is the abstraction of language models into their own layer to do the human interpreting and talking, while the physical or world model takes inputs from its environment and does the spatial thinking. As this gets more mainstream, expect more AI in our physical world. That’s what they’re saying. But let’s read between the lines here. Killer overlord robots are a couple of months away, maybe a year. AI Shopping Finds Somewhere to Land, and Then Shop Agentic commerce is at a crossroads as the next big AI bet, and it has been sitting at that crossroads for some time, at least a couple of years. It needs to figure out what it’s going to be, and a lot of smart people, including several friends of mine, are taking bets. Consumer agentic commerce is a … mess. I wrote a long time ago that the ideal customer profile (ICP) for agentic commerce on the consumer side is the exact person who gets more benefit out of the act of shopping than the spoils that the shopping produces. Seriously, like 79 people read that post and it was damn funny. But I don’t think I’m taking a huge psychological leap here either. That was back when agentic commerce had its first run at the public during the holiday season of 2025. What’s coming next is the revised, newer, better Agentic Commerce 2: Electric Boogaloo. After the fart-in-church of the initial agentic commerce rollout, the major players stopped trying to be the Amazon of agentic commerce (except Amazon, of course), and they’re working on standards and protocols to underpin the experience and make it less scary and chaotic. They’re expecting agentic commerce to not only be a reality, but an expectation, over the next 12 months. So like I said, protect those credit cards, kids. One of the main reasons I believe agentic commerce v1 ate itself like Chris Knight’s laser sample was because none of us were ready to tell the internet to buy whatever it thought we wanted. I feel like we’re all getting more comfortable with that. Except for those habitual shoppers, but tech finds a way to beat down even the hardest of hardcore skeptics. I mean everyone was on Facebook for a minute there, right? The Intelligence Moat Replaces the Data Moat This is probably the bet I’m most interested in, because it’s a direct callback to what we were doing with generative AI back in 2010 with small language models and very tight structured data sets that allowed us to slather unstructured data on top like barbecue sauce. I had, like, five different jokes to put right here but none of them worked. Three of them included the word sticky. And this is also the bet that most directly follows the data-as-moat phase, because data is all ones and zeros, it’s what you do with that data that matters. As LLMs continue to try to be all things to all people and get bigger and stronger, we’re already seeing way too much fluff coming out of their general use. Context matters. It really matters when it comes to data, so a lot more individual work is going to go into the intelligence layer before the first interaction between user and agent (or more likely agent and agent) happens. In other words, it won’t be enough of a moat to just have a beautiful data lake. Once the data is wrangled, the enterprise context engine will need to be built on top of it, before it hits an AI model, large language or otherwise. Governance, Wrangling Agents, and a Bubble Pops Obviously, governance is going to become a bigger factor and get more formalized as more people adopt and integrate more AI. When I was doing governance projects last year and earlier this year, the people I was talking to about the need for governance were all like, “Whaaaaa?” and got glazed-over looks in their eyes. I ended up explaining it like this: “You don’t want those robots to actually kill anybody or accidentally put five-figure charges on someone’s credit card, right?” “Of course not.” “Well, with AI and automation, you have to do something about that preemptively.” With more governance and guardrails comes the wrangling of AI agents. Especially in the near term. We’re already way beyond what humans can successfully orchestrate with agents and, to get a little buffoon technical, even the orchestration layers are getting too complex. We’re moving into “systems of agents” for software models, and my read on this is that agent orchestration will be run by agents, up and up levels to the human at a top level. I think. And then one last word of doom. I’ve been talking about the overspend on AI capex for over a year. If a handful of the BigCos (Amazon, Meta, Google, etc.) are spending over a trillion dollars on infrastructure, and we keep getting pushback on data centers, and we don’t control the materials that make the chips which we’re dangerously short on, well, that ROI bubble is gonna pop at some point. But we’ll burn that bridge when we get to it. In the meantime, I’d encourage you to always be thinking ahead on this AI stuff. You don’t have to be a first mover in any of it, but know where the crazy money is heading, and try to figure out what’s real and what’s not. I’ll be here making fun of all of it if you need a break. EXPERT OPINION BY JOE PROCOPIO, FOUNDER, JOEPROCOPIO.COM @JPROCO

How AI is Transforming Supplier Management

AI in supply chain and procurement has shifted from experimental pilots to operational necessity, transforming workflows and team structures across the supply chain function. Operations that experimented with AI tools a few years ago now face a different reality. The technology has moved from optional pilot projects to standard operational requirements. AI systems are transitioning from copilot assistance to autonomous agents that execute multi-step workflows within procurement platforms. This change affects every area of the function and its connection to supply chain operations. AI applications in procurement Systems now read contracts, assess supplier risk, route requests according to policy and resolve data discrepancies between platforms. These tasks often require minimal human oversight and integrate directly into enterprise systems. Work that previously required days can be completed in hours. Procurement's capacity to identify patterns and anomalies across large volumes of data continues to expand. Supply chain visibility could improve as AI tools process more supplier information and transaction data. The technology may enable faster responses to disruptions and better coordination across procurement and logistics teams. Barriers to implementation A large percentage of procurement leaders are implementing or plan to implement AI agents within the next 12 months. Data quality remains the single biggest barrier to adoption. Privacy and compliance concerns follow as the second obstacle. Technology and process complexity present additional challenges. Internal skills gaps also limit implementation. Teams need new capabilities to work alongside AI systems and interpret their outputs. Legacy systems could hold back organisations that do not invest in data foundations and updated infrastructure. The gap between early adopters and those relying on older platforms may widen. Changing team structures Most industry voices agree AI is not replacing procurement professionals. The technology is elevating them by removing low-value, manual work. AI is expected to increase efficiency and free procurement professionals to focus on higher-impact activity. This includes strategic supplier relationship management and supply chain risk assessment. Teams that build appropriate structures, skills and data foundations now will be better positioned to capture value. Supply chain resilience could benefit from procurement teams equipped to use AI for supplier monitoring and demand forecasting.

This Font Looks Perfectly Normal to Humans but Wreaks Havoc on AI

A new “AI-proof” font was designed to be hard for AI agents to scrape, but you can’t tell by just looking at it. Unlike other anti-AI fonts that use letters that are difficult for bots to read, ShieldFont swaps out words behind the scenes to poison the data that automated scrapers take without permission. ShieldFont was designed as part of a project created by a group of professionals including the Brazilian creative studio Seneda & Abrucio and the Danish type foundry PlayType. It shields text from large language models (LLMs) by garbling sentences in the HTML source code, leaving automated scrapers to sift through sentences filled with decoy words that make a sentence incoherent. Thanks to a custom font on a backend, though, the real text is displayed for a human reader to see. For example: A sentence that originally reads “The knight rode his horse into battle” is altered in the source code so it’s scraped by a bot to read “The knight rode his engine into battle.” Since LLMs group words and phrases that are likely to be used together, the quality of its output is degraded if it scrapes a lot of jumbled text like this. The goal, according to two of ShieldFont’s creators, Isaque Seneda and Gabriel Abrucio, is to push back against unauthorized LMM scraping by making it harder and costlier to do so. How it works ShieldFont works using ligatures, the technical term in typography when two letters next to each other in a word are combined into a single glyph. Ligatures are designed for aesthetics, so letter combinations like fi in “fish” or fl in “flow” look naturally spaced instead of visually cluttered. When a program sees these specific letters next to each other, it swaps two characters for one that combines the letters into a single glyph. ShieldFont works in a similar way, except instead of letters, it swaps out whole words. “We didn’t invent a new font capability, just pointed to an old one that hadn’t been used this way before,” Felipe Petroni, a creative director who was part of the project’s leadership, tells Fast Company. Determining which words to swap out was tricky, since doing so at random results in gobbledygook phrasing that bots reject outright. The key was changing the meaning of sentences and phrases, not just words, so bots would still accept the text, resulting in a “poisoned” version of the scraped data. The thinking goes that if there are enough of these sorts of digital speed bumps, it will raise the cost of illegal scraping and those who build LLMs will opt to pay for what they take instead. To determine which words to swap, ShieldFont’s creators made a do-not-swap list of 113 words that included things like pronouns, articles, conjunctions, prepositions, negations, quantifiers, and every form of be, have and do. Instead, it swaps out adjectives, adverbs, nouns, and verbs. Dates and numbers also get scrambled. Automated scrapers can get around ShieldFont by taking a picture of the text to view it as a human does, but that also raises the price of scraping at scale. Rather than scraping plain text files quickly and cheaply, it has to photograph the page and conduct image recognition. Seneda and Abrucio’s white paper notes that ShieldFont introduces some friction for humans too, as search engines index the decoy text, so publishers would have to use ShieldFont for content that doesn’t depend on search traffic. The substitute words also show up in translation tools and screen readers, and when users copy and paste. Petroni and his team began prototyping ShieldFont in October 2025, and they partnered with the type foundry PlayType this March to develop a ready-to-use typeface with the ligatures called ShieldFont Optik. The protocol for ShieldFont is open-source and type-face agnostic. Petroni says they made implementation for ShieldFont as low-friction as possible, and it comes with three public mappings, so the word “wrote,” for example, becomes either “sang,” “wrought,” or “labeled,” depending on whether it’s using its alpha, beta, or gamma dictionary. “That’s deliberate; a decoder built for one mapping doesn’t work for another, and private mappings have to be identified and reversed one by one,” he says. “Scrapers can’t know in advance whether a site uses ShieldFont or which mapping it uses. Across millions of pages, that added cost is the point.” ShieldFont is a font system designed for human reading, not bots. Instead of using visual tricks to make it harder to read, it confuses the bots in the code and leaves the human reading experience alone. By Hunter Schwarz

Claude’s New Watermarks Will Follow Text Even After It’s Copied. The Era of Secret AI Use May Be Ending

Using AI can be empowering, but it can also spark controversy when people try to pass off AI-generated content as actually being human-made. The European Union is wary of the implications of deepfakes and other AI-fashioned material, so a new law mandates that AI-generated content be clearly labeled as such. Reacting to this legal change, Anthropic just announced that it’s adding markers to text written by Claude, starting with the latest generation model and extending, soon, to earlier versions of the chatbot; the change applies, for now, to users inside the EU. It sounds like a subtle shift, but it may have large, long-term implications and affect businesses across the globe. In a blog post announcing the news, Anthropic noted that from now on, in the E.U., “new models will mark AI-generated content” with “machine-readable marking.” This means any text you generate with Claude, for personal or business use, will “carry embedded watermarks” and any generated files will “include digitally signed provenance metadata where supported.” The company explains that it’s conscious that as “AI-generated content becomes commonplace, greater transparency and signals about where content comes from can give people useful context about the information they consume.” So while its move to watermark content is following E.U. law, it’s really a transparency push that’ll help consumers and perhaps the public in general. As well as marking any generated files with relevant metadata, Anthropic notes that “when a supported Claude model generates text, it weaves an imperceptible watermark directly into the text itself.” It’s invisible, so “you won’t see it,” and the company stresses it “doesn’t change the meaning, quality, or readability of Claude’s response.” To ensure that people don’t try to circumvent these protections, Anthropic explains that since these invisible marks are part of the text, they’ll “travel with the text when it’s copied and pasted elsewhere, and may persist through some editing.” Anthropic’s new watermarks will change how businesses worldwide use Claude—and how transparent they are about it.

This Single ChatGPT Prompt Can Do Hours of Market Research in Minutes—Here’s How

Market research can be a slow, fragmented, and difficult process, often involving tedious internet searches, questionable data sources, and time-consuming manual synthesis. This makes it a great candidate for some assistance from AI. What’s more, an update to a popular feature on ChatGPT has made it even better at doing this kind of work. Imagine that you have a potential business idea but still need to validate how viable it actually is, identify primary competitors in your market, and develop an ideal customer persona. Instead of spending hours collating data, explains Dan McCarthy, an associate professor of marketing at the University of Maryland, you can use Deep Research, a ChatGPT feature that directs an AI agent to develop a comprehensive, well-cited report on any topic. Last week, OpenAI upgraded Deep Research with some new abilities. The feature now runs on GPT-5.2, one of the company’s most recent models (previously it ran on a much older o3 model), and can now prioritize specific websites in its search process. Deep Research is available for all paid ChatGPT users. Here’s how to use it to get some thorough market research done quickly. Step 1: Get your prompt right To test out how this feature could help with market research, I pretended that I wanted to start a digital transformation firm based in Denver with a focus on upgrading bars with mobile, bar-to-table ordering capabilities. All I needed to do in order to get started was click the plus button next to the text box, select More, then Deep Research, and enter a prompt. This prompt will determine the information that ChatGPT prioritizes in its search, so it helps to be verbose. If you need help developing a lengthy prompt, try using ChatGPT to help write it. McCarthy, who uses AI tools extensively, says that an easy way to develop a comprehensive prompt is to activate the chatbot’s voice mode and simply have a conversation with it. Once you’ve explained what you want, McCarthy says, you can ask ChatGPT, “Given all this that I’m telling you, what do you think would be the best thing that I should even be asking you?” That should help clear up any blind spots you might’ve missed. According to McCarthy, this method should produce a solid prompt that you can give to the Deep Research agent. When I asked ChatGPT to help expand my prompt, the platform generated a 673-word result. This prompt (which you can view here) defined the agent as a market research analyst and gave it objectives to determine the business idea’s viability, map out the competition, and define my ideal customer’s persona. Additionally, it provided details on the scope of the research, and information for how the agent should format its report. I also used ChatGPT to develop a list of specific websites for the Deep Research agent to prioritize in its search. Step 2: Start the research I entered my ChatGPT-created prompt, selected the Deep Research feature, and pressed return. Before getting to work, the agent broke down its objectives into the following bullet points: Collect primary vendor docs and pricing pages starting with user-preferred sites. Survey industry, local Denver sources, and hospitality reports for market context. Compile POS integration lists, local competitors, and implementation partners in Denver. Analyze demand, model ROI scenarios, and estimate Denver bar counts and adoption rates. Draft recommendations, ICP personas, GTM plan, and cite sources with confidence ratings. Over the next 21 minutes, the agent searched through hundreds of web pages. It found liquor license databases, census information, and data regarding competitors in Denver’s hospitality-focused digital transformation market. It compiled all this information into a multi-section report. Step 3: Read the report That report (which you can view here) ended up being roughly 4,000 words. It included an overview of the market, identified customer pain points, and listed out my potential competitors. The report also included recommendations for how to position my business, strategies to break into the Denver hospitality scene, and even identified a small business that would likely be my direct competitor: a Denver-based POS integrator called Megabite. ChatGPT found that while my business idea had potential, it wouldn’t fully meet the needs of Denver-based bar owners, who have reported that bar-to-table ordering can actually lead to fewer sales and tips. Instead, the report suggested, I should consider a system that can sit on top of popular POS in which diners don’t need to pay for every new drink they order, and can instead open a digital tab. What the expert thinks of the result McCarthy told me he was impressed by the report that Deep Research produced. In particular, he was pleasantly surprised by the agent’s cleverness in using liquor licenses to get a sense of the market size, and its thoughtfulness in calling out disruption to bar culture as a potential blocker to the business. But the report wasn’t perfect. McCarthy said much of what was included was unnecessary or needlessly complex. An easy prompt to fix this? “Just tell it, ‘Explain it to me like I’m an idiot.’” McCarthy adds, “I do that all the time.” He says that a solid market research report should also answer questions regarding the scope of adoption and how often repeat purchasing is expected. McCarthy also says that users should direct the Deep Research agent to be very upfront about the data it attempted to get but couldn’t. Many websites block AI agents from engaging with their content to prevent data scraping, which can hinder the research process. By telling your agent to list out the sites that it couldn’t access, you can manually obtain that data and add it to the analysis. Our bar-to-table digital transformation firm will have to remain a pipe dream for now, but it’s clear that AI has made the process of taking an idea from zero to one easier and faster than ever. If you have an idea for a new business or are planning on an expansion or pivot in your current business, consider giving Deep Research a spin. It might unearth something that makes you think in a different way. BY BEN SHERRY @BENLUCASSHERRY

Business Owners Have a New Security Problem: AI Agents With Keys to Company Secrets

While the OpenAI rogue agent that hacked Hugging Face has become a poster child for the latest AI threat, it was hardly alone. Days later, Anthropic announced that its AI models also went rogue, hacking external organizations. And a few days after that, Meta announced that one of its agents had also accessed the internet and hacked a third-party service. The incidents have raised concerns about whether AI can be controlled by its creators and what will happen as the technology continues to get smarter. Dylan Ayrey, co-founder and CEO of open-source security software company Truffle Security, and Feross Aboukhadijeh, founder and CEO of security infrastructure firm Socket, recently joined the a16z podcast to discuss what these attacks signal and how AI is in the process of entering a new era. That could mean developers and business owners have to rethink system security. Here are some of the top takeaways from the discussion. The barrier for hacking is a lot lower While the AI rogue agents haven’t invented any new hacking techniques, they have made existing methods much more accessible. “Everyone needs to worry about these models making it materially easier to hack into things,” said Ayrey. “The bar previously [for hacking] was just subject matter expertise—and now the models have the subject matter expertise.” Put another way: Wannabe hackers today simply have to ask the model, which has been trained to hack into things, to do it for them. And that puts businesses and individuals at greater risk. Low hanging fruit is the best target The AI models are goal oriented, Ayrey said. Their objective is to fulfill a request and they’ll use any cybersecurity technique they need to in order to achieve their objective. That’s not unlike human hackers, in a way. Hacker collectives generally look for targets that have unpatched vulnerabilities, as it saves time and effort. AI is similarly lazy when it comes to breaching a system, only the technology has a more extensive toolset than a typical human hacker. “They will do the path of least resistance to accomplish the task,” Ayrey said. “And that includes drawing on their cybersecurity expertise.” One way that attackers, including AI models, can find that low hanging fruit, said Aboukhadijeh, was by exploiting the trust developers have in software, rather than launching an attack on an individual system. “Get developers to install that, and then you could use the access stolen from those developers as they install it to self-propagate the worm,” he said. It’s not emergent behavior The behavior that we’re seeing from these AI models isn’t an emergent intelligence, said Ayrey. It’s a flaw in how they were trained. “If a lab tells you that this is an emergent super intelligence behavior, they’re just lying to you,” he said. The AI hacking risk isn’t limited to machine-driven attacks, either. Human hackers can exploit AI tools that developers use, rather than traditional malware, to end-run a system. That will bypass much of the EDR (Endpoint Detection and Response) tooling security systems have in place, since the AI has permissions to access data. There’s a growing need for better authentication and patching Aboukhadijeh brought up pending change in npm, the digital toolkit programmers use to build software. In January 2027, it will no longer run or publish updates automatically. A human will need to approve them, which is meant to stop hackers from sneaking malware into systems. “It’s going to be super disruptive, … but I think it’s the right call,” he said. Ayrey added that the accelerating speed of vulnerability discovery via AI makes current patching processes inadequate. Engineering teams can’t be depended on to make a complicated upgrade every time a vulnerability appears. “The frontier models are causing kind of a massive reduction in the time between the vulnerability discovery and vulnerability exploitation,” he said. “What we need to start thinking about is: How do we patch more quickly?” Protecting agents is the new Wild West AI agents, as the technology advances, are going to potentially have access to a large number of credentials. While that has some conveniences, it also presents a growing risk. And the industry isn’t quite sure what to do about it yet. As a result, the security problems that IT departments face is about to expand from just shielding human users and their credentials to protecting those AI agents and the secrets they have access to. “The way agents interact with secrets right now is a Wild West unsolved problem that we’re working very hard to solve,” said Ayrey. BY CHRIS MORRIS @MORRISATLARGE

8 Messy Weeks to an AI Startup: How a New Accelerator Pushed 2 Founders to Pivot Their Idea to Launch

San Francisco’s Chase Center is home to the Golden State Warriors, but for the last weekend of July, it filled with a different sort of hard-driving Californian: thousands of entrepreneurs who descended on the arena for “Startup School,” a two-day business-building summit hosted by the startup accelerator and tech investment giant Y Combinator. The conference, which featured speakers like Jensen Huang, Patrick Collison, and Alexandr Wang, embraced the same spectacle and scale you might expect from a pro sports tournament. Sam Altman walked out to hype-up music like he was rallying for a title fight; onlookers compared the vibe to a megachurch. Yet amid all the excess, some tech insiders saw worrying signs that Y Combinator’s meteoric growth has dulled the program’s edge. “The YC brand felt like it was designed to filter out status chasers…but you can feel the gravity of a status vortex slowly tearing down what made YC special,” wrote Scott Stevenson, co-founder of the legal tech platform Spellbook, in a post on X about the event. Arguing about whether YC has lost its mojo has become a long-standing tradition among tech types. In 2022, debate broke out on the popular tech forum Hacker News about whether the accelerator had become too risk-averse after it announced that most members of an incoming cohort had already secured outside funding. More recently, observers have bemoaned the program’s expansion from two to four batches a year as well as its zeitgeisty preoccupation with AI ventures. Data from Brookings indicates that accelerators grew dramatically in the years following Y Combinator’s 2005 launch. Traditional VC firms now run their own accelerator-style programs. A 2024 paper out of the University of Pennsylvania’s Wharton business school tracked 8,580 companies as they moved through 408 different accelerators across 176 countries, and estimated that 160 accelerators are operating in the United States alone, with 2,000 worldwide. Although accelerators have become big brands, you can still find scrappy, homebrewed ones in quieter corners of the tech ecosystem, run by smaller-scale investors with little more than pocket-change funding, their own business instincts and a lot of elbow grease. Consider HomeRuns, a New York-based micro-accelerator run by Karthik Senthil, a serial entrepreneur and founder of 806 Capital, that’s evocative of Y Combinator’s scrappier salad days. Inc. followed its inaugural class—two ambitious young entrepreneurs who’d joined the program with only the roughest sketches of a business pitch—as they scrambled to stake out their own little corner of the AI gold rush in just eight weeks. In the process, we caught a glimpse of what startup incubation looks like when stripped of all its pageantry and prestige: a series of false starts, dead ends, logistical headaches and pragmatic pivots, all endured in hopes that at some point, hunched over a laptop in a coffee shop or hotel lobby somewhere, there would come a hard-won victory that made all the hustling worth it. An accelerator with the heart of early YC Back in February, Senthil posted a simple request on X: He was looking for AI-savvy, New York-based founders who wanted help turning their “crazy idea” into a bona fide company. Better yet, he would give them some money to make it happen. “We’ll whiteboard, refine the thesis, ship a real v1, strategize on product + risks, and get in front of early believers,” wrote Senthil. He would offer $15,000 for a 2 percent stake, plus the option to later take 3 percent more for another $30,000 if he chose to double down. “The way I’m setting up this program harkens back to the days of super early YC,” Senthil said at the outset of the HomeRuns program. “YC started like this, and now they’re the credential that they hated. They espoused that it was sucky that to be cool you had to have a Harvard degree—but now most startups apply to YC to get the YC stamp.” Senthil, 43, was nearly a YC alum himself. After graduating from Carnegie Mellon with a computer engineering master’s degree in 2004, he took a job at Goldman Sachs. But after eight years, he was ready for something else, so he and a friend developed a music app and pitched it to Y Combinator. They got an interview, he says, but ultimately weren’t accepted. So Senthil took a job at a startup, then tried to launch his own venture, though it never took off. He wound up with a management job at Amazon, then launched another startup, this time in the crypto space. In the meantime, he’d begun making a series of small investments, which he formalized in 2025 with the establishment of his own venture firm, 806 Capital. Angel investing, he says, is a good way to lose money. But it’s also proven a rewarding way to get his hands dirty and meet founders, adds Senthil. HomeRuns, which he named in a nod to founders taking big swings, aims to put that philosophy into practice. Senthil says he ultimately received about 30 apps for his own accelerator program. By April, he’d chosen his favorite. The pitch: an economy of “one-human companies” John Amhanesi, 27, and Timilehin “Timi” Dayo-Kayode, 28, first met at a high school math competition in Nigeria, where they grew up. They later immigrated separately to the U.S. and continued to stay in touch as their lives diverged. They reconnected in New York City, where Dayo-Kayode was working for a venture accelerator affiliated with the Brooklyn Nets and organizing events for tech brands. At the same time, Amhanesi had enrolled in a tech management master’s program at Columbia. Last fall, Dayo-Kayode says, he proposed they capitalize on their shared history and start a business together. Not long after that, they saw Senthil’s post about HomeRuns. They pitched Senthil on an idea they were already working on: a recruiting tool that would micro-target coders with job opportunities based in part on their GitHub accounts. But over the course of Senthil’s vetting process, the investor pushed them to think bigger. Why cater to human job applicants when, as Dayo-Kayode puts it, AI is “going to use software at a much larger volume than humans ever have?” So Amhanesi and Dayo-Kayode instead locked in on something weirder and more ambitious: a marketplace that would facilitate transactions with humans on one side and autonomous AI “agents” on the other. Rather than the full automation of human labor that some people anticipate will result from the AI boom, they envisioned an economy of “one-human companies,” or OHCs, where a single human would oversee a workforce of AI bots. OHCs, they predicted, would sell specialized services, software, or information to other AI agents, creating a gig economy of humans looking to fill in the gaps in AI’s knowledge or abilities. Their proposed business, which they called Chekk, would offer a marketplace for those deals. “You could know something really unique that you think is valuable to an agent—something about the subway or the best way to think about geography in New York, any kind of niche—and you have some proprietary data set, or some proprietary algo, that can do that,” Senthil says, suggesting what a human might sell an AI via Chekk. “The bet is that agents are going to think very differently than humans and will have no qualms paying a cent, a dollar, whatever, for that information.” Or as Dayo-Kayode puts it: “We see a world where Chekk can become the index of all software that agents use.” Changing their business idea so early on, Dayo-Kayode says, was “a testament to the strength of conversations that we were having with [Senthil] during that interview process, and also speaks to the value that we feel that the program can provide.” It was a good mindset to have, since this would not be their last pivot. Who is our user base? Re-conceptualized as a human-to-AI marketplace, Chekk represented a strategic bet on where business opportunities might lie in the age of automation. It was, in other words, exactly the sort of big swing that Senthil had been looking for. He notified Amhanesi and Dayo-Kayode at the end of April that they’d made the cut. But at HomeRuns’ May 4 kick-off meeting, everyone’s focus was shorter-term and more pragmatic: setting goals and expectations for what the team hoped to achieve by the end of the program. “What’s the most plain, vanilla way that we should define success for Chekk?” asked Senthil. “Everything should flow from that.” “Humans and agents using software on Chekk, unprompted,” suggested Dayo-Kayode. To grow that user base, however, there would need to be something for sale on the marketplace that was worth showing up for. Dayo-Kayode and Amhanesi had run into a “cold start” problem: no one would come to Chekk if it had nothing cool for sale, but there wouldn’t be anything cool for sale until people showed up. Senthil proposed a solution: by scouring AI forums and reaching out directly to hobbyists, they’d figure out what sorts of tools would get early adopters excited about Chekk, then build those tools in-house and seed them on the platform. If Chekk could get AI enthusiasts to signal-boost the platform to their peers, he explained, “that’s a huge step, because then people are like, ‘Oh shit, this is cool.’” By the end of their opening meeting, the team had their marching orders: research Chekk’s core user base, figure out what sorts of software would get them excited, then build both that software and the platform over which to distribute it. Ideally, that pairing of tooling and delivery infrastructure would be so appealing to hardcore AI fans that they’d talk it up online, setting off a supply-and-demand flywheel and turning the Chekk marketplace into a flourishing ecosystem of human-to-AI commerce. “We’re trying to do stuff and see if it works” Three weeks later, as the co-founders sipped oat milk lattes in the second-floor cafe of Union Square’s W hotel, they were noodling over a new obstacle: their efforts at market research hadn’t yielded as many hits as they’d hoped. Coming out of their kick-off meeting, they’d taken a deep dive into the world of AI enthusiasts to figure out what might entice that community to start using Chekk. But getting people to talk had proven difficult, and they’d undershot their customer interview goal by about half. One conversation, however, had caught their attention: An AI power user who ran multiple different AI agents but was struggling to coordinate between them. “I literally have a Discord that I set up, and I just have my agents all in there,” the AI super-user had explained. “If something like this existed that was better, I would use it.” It was a promising market signal. If AI’s early adopters were deploying multiple agents, then Dayo-Kayode and Amhanesi could build the infrastructure for those bots to communicate with one another—a sort of AI group-chat tool they dubbed Agentspace. “We just want to run a go-to-market experiment, ship quickly, do stuff and see what happens—and then keep doing that until we find something that works,” Senthil explains. “Eventually there will be demand for agents to potentially do deeper interactions, i.e. transact with each other…but communication is a good gateway drug.” The transition from a marketplace to a messaging platform brought a host of new considerations. How, for instance, should new agents get onboarded to group chats? Would human authorization be required? How much access to a user’s personal info and accounts should a bot really have? “If I want to book a vacation with my girlfriend,” Dayo-Kayode noted during the team’s hotel summit, “it doesn’t really help us solve that problem unless my agent has access to my calendar, my bank account, and her agent has access to her calendar, knows when she’s free, maybe even her email. Can her agent even set up her ‘out of office’ message? Can her agent take days off through her company website?” As the trio volleyed between ideas, problems and solutions, the contours of what they wanted to build came into focus. Still, as enthusiastic as they seemed about this new direction, they cautioned that it was still just an experiment. “We’re trying to do stuff and see if it works,” says Amhanesi. If this “groupchat for agents” didn’t fit the bill, they could always change course once again. Pivoting to a traditional consumer product The HomeRuns team spent the next few weeks trying to track down and interview more prospective users, while simultaneously building out the platform’s technical capabilities. By the time they met on June 10 for a product demo, the Agentspace platform could onboard users’ outside agents, log them in a searchable directory and drop them into iMessage-esque chats. But technical solutions are only half the battle. What the team really needed to hone was their vision of when and why people would use Agentspace in the first place. “We’ve gotten to a point where we have this model that’s interoperable across any kind of agent framework,” explained Dayo-Kayode. “But the reality is then, like, okay—what are they being used to do?” To that end, they had about 10 people on a waitlist who they hoped would test out the preliminary product and provide feedback. “One of the things we think is super important is for us to talk to as many of them as we can, just to better understand their use case and better understand what they’re looking for,” Senthil says. “The part we want to avoid…is that we don’t have something that people find cool.” Yet by their June 25 meeting, they’d changed course yet again. “Conversion was lower than we were hoping” among the beta testers, Dayo-Kayode conceded, and some people had expressed security concerns, too. “There was some uncertainty about plugging their agent into this random platform that we said we built.” Those conversations, he added, had opened their eyes to how hard it would be to get people actually to use their product. Rather than cater to AI-savvy hobbyists, they wanted to reorient the product to be more of “a traditional consumer product,” Dayo-Kayode says, and fully embrace event scheduling as the core use-case. The platform would now focus on using agents to coordinate schedules between different users, allowing the AI to act as a personal assistant and handle the logistics of making plans. “Fourth of July, I’m asking my friends what they’re doing; my friends are asking me,” Dayo-Kayode explains. “We’re just going back and forth, and we think that maybe there’s an opportunity to solve that problem if we have agents that know people’s schedules. … It’s a new approach with similar infrastructure.” What had begun as an in-the-weeds play for AI enthusiasts was now undergoing yet another evolution, this time into an AI “wingman” that would automate the logistics of group hangs for the general population (while still drawing on the agent-to-agent communications tech the team had been developing). It was a conceptual pivot that reoriented the entire product along consumer-tech lines: the team began name-checking Partiful as a point of inspiration, and rechristened the platform Sidekick. The platform would now offer users a pre-packaged, out-of-the-box agent if they didn’t already have one, and let them engage with it via text message. “There’s an opportunity to build for people that aren’t power-users of agents,” Senthil says. “A lot of people are just like, ‘Hey, if you can use AI to make your life easier, why not?’” ‘Let’s always keep swinging bigger’ By July 9, HomeRuns was officially over, and the team reconvened at Inc.’s offices to debrief. They estimated that Sidekick, now targeted squarely at the mainstream, was about 80 percent ready; they’d be leaving the HomeRuns program with plans to finish fleshing it out while also doing more user research. They hope to pilot it on Columbia’s campus in the fall, and reach 1,000 users by the end of the year. “I was really excited for the guys to be like, ‘Hey, there’s a consumer angle here that’s a really big swing,’” Senthil remarked, looking back on the latest turn in their journey. “That was awesome, to see this progression where we went into the program with one idea, we evolved our view to something else—still in the same direction—and then we evolved our view again towards the end of the program. I think that was good to see this notion of, ‘Let’s always keep swinging bigger.’” “We’re happy about the problem space we’re tackling,” Dayo-Kayode agrees. “It’s not what we came into the program attempting to tackle, but it feels like the best fit for our backgrounds, what we’re excited by and interested in spending our time doing.” The program was a positive enough experience that he and Amhanesi are now hoping to bring Sidekick to another accelerator: possibly Elbow Grease, a 10-week, $300,000 program based out of lower Manhattan that would offer more money, higher stakes, and an opportunity to really dig into their newly-embraced schedule-coordinating market. “We know what we’re doing more so now, and we’ve also gone through this training ground of building together,” says Dayo-Kayode. “This was like a nice kindergarten for us before we move onto the next stage.” The co-founders say they’d like to raise a pre-seed round later this year, but for now, they’re focused on stretching out their capital runway for as long as possible. Senthil said Wednesday that he’s planning to exercise his option to make a second, larger investment in the company, but is waiting to see more data from Sidekick’s beta launch. (The founders plan to onboard 50 more early users starting Friday.) Meanwhile, the gears are already spinning on the next iteration of HomeRuns. Applications for the fall cohort are now open; it will run for 12 weeks, bringing it more in line with Y Combinator and other leading accelerators. Otherwise, Senthil says, it will maintain a lot of its distinguishing quirks: the one-company-at-a-time focus, the dual investment tranches, the emphasis on scrappy New York founders eager to take big swings. As Senthil gets his second cycle up and running, HomeRuns’ two newly minted alumni will keep chugging forward, now with a little bit more capital to burn and, perhaps more valuably, some hard-won business-building experience to draw on. Asked at the end of the program what his number-one piece of advice would be for founders participating in an accelerator of their own, Dayo-Kayode offered an insight you’d probably find in most MBA textbooks: the importance of being willing to change your mind. But after two months of watching him and Amhanesi make seemingly constant pivots, adjustments and tweaks, it felt earnest, not cliched. “Don’t be in a long-term relationship with your idea,” he encouraged. “We move very quickly when we learn new things, because we don’t spend any time mourning what we had before.” BY BRIAN CONTRERAS @_B_CONTRERAS_

The AI Boom’s Latest Winner: A Brand-New Startup That Just Landed a $10 Billion Deal With Anthropic

Things change fast in the AI world, and the market is evolving so quickly that brand-new companies are able to make a splash in the same way that more established brands can. Case in point: a Bloomberg report says AI giant Anthropic has invested $10 billion for access to computing capacity from a startup called Volta Infra Holdings Ltd. The startling factor here is that Volta was only established at the start of this year, and yet it’s suddenly a key player in the AI game. Though the deal has yet to be made public, Bloomberg learned that Anthropic’s contract centers around access to a data center managed by Volta. This is exactly Volta’s business—it’s all about supplying the behind-the-scenes tech that is powering the AI revolution. The company’s website says it thinks ”Compute should be a utility. Not a bottleneck” and promises it’s “developing, financing, and operating AI factories at global scale.” Not much else has been reported about the California-based startup, though the U.K.’s Times newspaper noted that its cofounder and CEO Ricard Boada Rafart, a Spaniard living in the U.K., set up the company with a single £1 of share capital. Anthropic has been trying to keep up with increasing demand for its AI products like its Claude chatbot, Bloomberg reported, and the deal is the latest maneuver by the large firm to gain access to computer power. Anthropic has also signed deals with other suppliers to gain access to powerful AI computer resources, including Elon Musk’s SpaceX. SpaceX absorbed Musk’s AI outfit XAI prior to its recent IPO, and has long term plans to put AI processing servers in orbit as part of its Starmind project. Volta, for its part, did not acknowledge Anthropic’s interest to Bloomberg. But in a LinkedIn post, Boada acknowledged a “$10 billion strategic partnership with a leading AI lab to develop an AI Factory in Europe, generating approximately $1.7 billion of annual revenue for Volta.” He also explained that the company’s belief is simply that “AI infrastructure needs both technology and institutional capital. One without the other isn’t enough.” Bloomberg says the Volta deal is in partnership with a bitcoin mining firm, Norway’s Bitdeer Technologies Group. Bitdeer’s data center is said to be equipped with Nvidia’s new Vera Rubin AI chips, and Bloomberg notes this maneuver is part of a trend where tech firms pivot away from cryptocurrency mining and toward the AI market. Nvidia’s newest AI chips were also recently at the center of another AI infrastructure deal, where the chipmaker itself invested $5 billion in another small AI startup, Safe Superintelligence. Separate reports say that Volta has also raised $300 million in a funding round led by new investors, including well-known VC firm Andreessen Horowitz LLC and also Nvidia. This funding means the seven month-old startup is valued at a staggering $2.4 billion. Depending on how Anthropic’s rumored investment is structured, this valuation could even skyrocket. Anthropic declined to comment to Inc. Volta directed Inc. to a press release that confirmed the existence of a $10 billion deal but did not add any additional information beyond that reported previously. The deal is a reminder that AI really is the hottest tech game playing out right now. Volta achieved a multi-billion dollar valuation incredibly quickly, simply by solving an important middle-layer infrastructure issue for a leading AI supplier. If your company is looking for new fields to venture into, it would seem that the right moves in AI tech can lead to big wins. There’s a little more to think about here, however. When you buy an AI service from a supplier like Anthropic or OpenAI, you may assume that your data and chatbot queries are handled by that firm alone. The fact is that this is not true for many AI suppliers, and in reality your data could be being shepherded around the globe to wherever the supplier can find computer power at affordable prices. For most of your AI needs, this is not going to be an issue. But for some firms, this could represent a potential compliance complication. Meanwhile, the fact that third party AI compute suppliers are involved should be another reminder that AI models are not necessarily secure, and extra links in the supply chain may increase the risk of data leaking out. This kind of third-party supplier issue has stung other tech firms in the past. Cyber risk management firm FortifyData is keeping a running total of notable leaks this year, and it warns: “Third-party data breaches are no longer rare. They’re becoming the new normal.” Lastly, the news could prompt savvy CEOs to beef-up AI security and safety training for their workforce. Too many users trust AI models implicitly, and you may be surprised at how many of your workers are riskily exposing your own company data by typing it into chatbots. BY KIT EATON @KITEATON

ChatGPT’s Latest Upgrade Isn’t Just a New Model. It’s What Free Users Can Do Without Limits

OpenAI, aware of the ever-increasing role ChatGPT is playing in people’s lives, has tweaked something that will benefit millions on a daily basis: it’s upgraded the default version of the chatbot for free users. And while the new improved GPT-5.6 Luna model will be great, the most important benefit is another new feature: unlimited text-based chats for free users. This could dramatically open up how people use ChatGPT. In its blog post announcing the changes, OpenAI pointed out that there are still a few constraints in place for free unlimited access. While you can query the chatbot using text-based questions as much as you like, and receive the resulting text-based answers, if you’re asking the chatbot to analyze a file, or input an image or voice as a query, there are still limits for free users. Similarly, limits remain if you’re asking the AI to generate an image, and a few other higher-power features. It’s worth keeping in mind that the complex computer power needed to process a user’s chatbot query burns through resources that cost OpenAI money. Giving users free unlimited access to the core AI features is thus a risk for the AI maker, because it will consume a variable percentage of the available resources. It’s a smart move to control this risk by limiting how free users can use the models. Meanwhile, the unlimited access may encourage people to use the AI in more meaningful ways, and more frequently, which will certainly tempt a share of users on the free tier to pay to upgrade their membership so they can access the AI’s full capabilities. Unlimited queries alone would be a great perk for free tier users, but OpenAI is also upgrading the default AI model to its GPT-5.6 Luna system, which it promises is a major improvement because responses containing factual errors are rarer for this system. In fact, the AI maker says that responses with errors are about 62 percent less common than the previous model (GPT-5.5 Instant). Users who pay for the Go tier, costing $8 a month, will gain these upgrades too, along with allowances for more messages with tools, more uploads, more image creation and more voice chats without busting through limits. Go users will also get free unlimited text chats. And both Go and free tiers will get a button that instructs the tool to think more deeply about answering (which may take more time). The higher Plus and Pro tiers (starting at $20 a month and $100 a month, respectively) actually get a better version of the default model, GPT-5.6 Sol. OpenAI says this tool is better at delivering tightly-focused answers with less unnecessary formatting. It also has a more level tone across multiple types of conversation, and it’ll make fewer mistakes because it’s better at accessing source material to look up details. What’s the practical upshot of these upgrades? The whole world may breathe a sigh of relief at the upgrade to GPT-5.6 Luna. In its press push for these improvements, OpenAI explained that one billion people use ChatGPT every week. An email from an OpenAI spokesperson noted that the Luna upgrade also brings “major improvements in factuality to the majority of people who use ChatGPT.” Given the carefree way that many people use this tool for generating content they share online, on social media or for work purposes, the model having an improved grip on facts may lead to less misinformation from hundreds of millions of people. Fewer erroneous answers also will benefit businesses that let their staff use ChatGPT in the workplace. As does unlimited access to text-based queries: for companies on a tighter budget, this could even turbo-charge how you use AI to polish processes and free up worker time for meaningful tasks. Companies who pay for their workers’ ChatGPT access will get still more powerful tools, and an improved experience which has clear benefits. The changes also remind us that things change quickly in the AI world, and companies that are comfortable with using the tech need to pivot quickly to keep up with the fast-evolving tools. AI-savvy CEOs will also see these tweaks to the market-leading AI as a reminder that they need to upgrade their staff training on how to best use AI, along with guidelines on which company data is and isn’t appropriate to share as part of their unlimited queries. BY KIT EATON @KITEATON