In “The Optimist: Sam Altman, OpenAI, and the Race to Invent the Future,” Wall Street Journal reporter Keach Hagey examines our AI-obsessed moment through one of its key figures — Sam Altman, co-founder and CEO of OpenAI. Hagey begins with Altman’s Midwest childhood, then takes readers through his career at startup Loopt, accelerator Y Combinator, and now at OpenAI. She also sheds new light on the dramatic few days when Altman was fired, then quickly reinstated, as OpenAI’s CEO. Looking back at what OpenAI employees now call “the Blip,” Hagey said the failed attempt to oust Altman revealed that OpenAI’s complex structure — with a for-profit company controlled by a nonprofit board — is “not stable.” And with OpenAI largely backing down from plans to let the for-profit side take control, Hagey predicted that this “fundamentally unstable arrangement” will “continue to give investors pause.” Does that mean OpenAI could struggle to raise the funds it needs to keep going? Hagey replied that it could “absolutely” be an issue. “My research into Sam suggests that he might well be up to that challenge,” she said. “But success is not guaranteed.” In addition, Hagey’s biography (also available as an audiobook on Spotify) examines Altman’s politics, which she described as “pretty traditionally progressive” — making it a bit surprising that he’s struck massive infrastructure deals with the backing of the Trump administration. “But this is one area where, in some ways, I feel like Sam Altman has been born for this moment, because he is a deal maker and Trump is a deal maker,” Hagey said. “Trump respects nothing so much as a big deal with a big price tag on it, and that is what Sam Altman is really great at.” Techcrunch event Save now through June 4 for TechCrunch Sessions: AI Save $300 on your ticket to TC Sessions: AI—and get 50% off a second. Hear from leaders at OpenAI, Anthropic, Khosla Ventures, and more during a full day of expert insights, hands-on workshops, and high-impact networking. These low-rate deals disappear when the doors open on June 5. Exhibit at TechCrunch Sessions: AI Secure your spot at TC Sessions: AI and show 1,200+ decision-makers what you’ve built — without the big spend. Available through May 9 or while tables last. Berkeley, CA | June 5 REGISTER NOW In an interview with TechCrunch, Hagey also discussed Altman’s response to the book, his trustworthiness, and the AI “hype universe.” This interview has been edited for length and clarity.  You open the book by acknowledging some of the reservations that Sam Altman had about the project —  this idea that we tend to focus too much on individuals rather than organizations or broad movements, and also that it’s way too early to assess the impact of OpenAI. Did you share those concerns? Well, I don’t really share them, because this was a biography. This project was to look at a person, not an organization. And I also think that Sam Altman has set himself up in a way where it does matter what kind of moral choices he has made and what his moral formation has been, because the broad project of AI is really a moral project. That is the basis of OpenAI’s existence. So I think these are fair questions to ask about a person, not just an organization. As far as whether it’s too soon, I mean, sure, it’s definitely [early to] assess the entire impact of AI. But it’s been an extraordinary story for OpenAI — just so far, it’s already changed the stock market, it has changed the entire narrative of business. I’m a business journalist. We do nothing but talk about AI, all day long, every day. So in that way, I don’t think it’s too early. And despite those reservations, Altman did cooperate with you. Can you say more about what your relationship with him was like during the process of researching the book? Well, he was definitely not happy when he was informed about the book’s existence. And there was a long period of negotiation, frankly. In the beginning, I figured I was going to write this book without his help — what we call, in the business, a write-around profile. I’ve done plenty of those over my career, and I figured this would just be one more. Over time, as I made more and more calls, he opened up a little bit. And [eventually,] he was generous to sit down with me several times for long interviews and share his thoughts with me. Has he responded to the finished book at all? No. He did tweet about the project, about his decision to participate with it, but he was very clear that he was never going to read it. It’s the same way that I don’t like to watch my TV appearances or podcasts that I’m on. In the book, he’s described as this emblematic Silicon Valley figure. What do you think are the key characteristics that make him representative of the Valley and the tech industry? In the beginning, I think it was that he was young. The Valley really glorifies youth, and he was 19 years old when he started his first startup. You see him going into these meetings with people twice his age, doing deals with telecom operators for his first startup, and no one could get over that this kid was so smart. The other is that he is a once-in-a-generation fundraising talent, and that’s really about being a storyteller. I don’t think it’s an accident that you have essentially a salesman and a fundraiser at the top of the most important AI company today, That ties into one of the questions that runs through the book — this question about Altman’s trustworthiness. Can you say more about the concerns people seem to have about that? To what extent is he a trustworthy figure?  Well, he’s a salesman, so he’s really excellent at getting in a room and convincing people that he can see the future and that he has something in common with them. He gets people to share his vision, which is a rare talent. There are people who’ve watched that happen a bunch of times, who think, “Okay, what he says does not always map to reality,” and have, over time, lost trust in him. This happened both at his first startup and very famously at OpenAI, as well as at Y Combinator. So it is a pattern, but I think it’s a typical critique of people who have the salesman skill set. So it’s not necessarily that he’s particularly untrustworthy, but it’s part-and-parcel of being a salesman leading these important companies. I mean, there also are management issues that are detailed in the book, where he is not great at dealing with conflict, so he’ll basically tell people what they want to hear. That causes a lot of sturm-und-drang in the management ranks, and it’s a pattern. Something like that happened at Loopt, where the executives asked the board to replace him as CEO. And you saw it happen at OpenAI as well. You’ve touched on Altman’s firing, which was also covered in a book excerpt that was published in the Wall Street Journal. One of the striking things to me, looking back at it, was just how complicated everything was — all the different factions within the company, all the people who seemed pro-Altman one day and then anti-Altman the next. When you pull back from the details, what do you think is the bigger significance of that incident? The very big picture is that the nonprofit governance structure is not stable. You can’t really take investment from the likes of Microsoft and a bunch of other investors and then give them absolutely no say whatsoever in the governance of the company. That’s what they have tried to do, but I think what we saw in that firing is how power actually works in the world. When you have stakeholders, even if there’s a piece of paper that says they have no rights, they still have power. And when it became clear that everyone in the company was going to go to Microsoft if they didn’t reinstate Sam Altman, they reinstated Sam Altman. In the book, you take the story up to maybe the end of 2024. There have been all these developments since then, which you’ve continued to report on, including this announcement that actually, they’re not fully converting to a for-profit. How do you think that’s going to affect OpenAI going forward?  It’s going to make it harder for them to raise money, because they basically had to do an about-face. I know that the new structure going forward of the public benefit corporation is not exactly the same as the current structure of the for-profit — it is a little bit more investor friendly, it does clarify some of those things. But overall, what you have is a nonprofit board that controls a for-profit company, and that fundamentally unstable arrangement is what led to the so-called Blip. And I think you would continue to give investors pause, going forward, if they are going to have so little control over their investment. Obviously, OpenAI is still such a capital intensive business. If they have challenges raising more money, is that an existential question for the company? It absolutely could be. My research into Sam suggests that he might well be up to that challenge. But success is not guaranteed. Like you said, there’s a dual perspective in the book that’s partly about who Sam is, and partly about what that says about where AI is going from here. How did that research into his particular story shape the way you now look at these broader debates about AI and society? I went down a rabbit hole in the beginning of the book, [looking] into Sam’s father, Jerry Altman, in part because I thought it was striking how he’d been written out of basically every other thing that had ever been written about Sam Altman. What I found in this research was a very idealistic man who was, from youth, very interested in these public-private partnerships and the power of the government to set policy. He ended up having an impact on the way that affordable housing is still financed to this day. And when I traced Sam’s development, I saw that he has long believed that the government should really be the one that is funding and guiding AI research. In the early days of OpenAI, they went and tried to get the government to invest, as he’s publicly said, and it didn’t work out. But he looks back to these great mid-20th century labs like Xerox PARC and Bell Labs, which are private, but there was a ton of government money running through and supporting that ecosystem. And he says, “That’s the right way to do it.” Now I am watching daily as it seems like the United States is summoning the forces of state capitalism to get behind Sam Altman’s project to build these data centers, both in the United States and now there was just one last week announced in Abu Dhabi. This is a vision he has had for a very, very long time. My sense of the vision, as he presented it earlier, was one where, on the one hand, the government is funding these things and building this infrastructure, and on the other hand, the government is also regulating and guiding AI development for safety purposes. And it now seems like the path being pursued is one where they’re backing away from the safety side and doubling down on the government investment side. Absolutely. Isn’t it fascinating?  You talk about Sam as a political figure, as someone who’s had political ambitions at different times, but also somebody who has what are in many ways traditionally liberal political views while being friends with folks like — at least early on — Elon Musk and Peter Thiel. And he’s done a very good job of navigating the Trump administration. What do you think his politics are right now? I’m not sure his actual politics have changed, they are pretty traditionally progressive politics. Not completely — he’s been critical about things like cancel culture, but in general, he thinks the government is there to take tax revenue and solve problems. His success in the Trump administration has been fascinating because he has been able to find their one area of overlap, which is the desire to build a lot of data centers, and just double down on that and not talk about any other stuff. But this is one area where, in some ways, I feel like Sam Altman has been born for this moment, because he is a deal maker and Trump is a deal maker. Trump respects nothing so much as a big deal with a big price tag on it, and that is what Sam Altman is really great at. You open and close the book not just with Sam’s father, but with his family as a whole. What else is worth highlighting in terms of how his upbringing and family shapes who he is now? Well, you see both the idealism from his father and also the incredible ambition from his mother, who was a doctor, and had four kids and worked as a dermatologist. I think both of these things work together to shape him. They also had a more troubled marriage than I realized going into the book. So I do think that there’s some anxiety there that Sam himself is very upfront about, that he was a pretty anxious person for much of his life, until he did some meditation and had some experiences. And there’s his current family — he just had a baby and got married not too long ago. As a young gay man, growing up in the Midwest, he had to overcome some challenges, and I think those challenges both forged him in high school as a brave person who could stand up and take on a room as a public speaker, but also shaped his optimistic view of the world. Because, on that issue, I paint the scene of his wedding: That’s an unimaginable thing from the early ‘90s, or from the ‘80s when he was born. He’s watched society develop and progress in very tangible ways, and I do think that that has helped solidify his faith in progress. Something that I’ve found writing about AI is that the different visions being presented by people in the field can be so diametrically opposed. You have these wildly utopian visions, but also these warnings that AI could end the world. It gets so hyperbolic that it feels like people are not living in the same reality. Was that a challenge for you in writing the book? Well, I see those two visions — which feel very far apart — actually being part of the same vision, which is that AI is super important, and it’s going to completely transform everything. No one ever talks about the true opposite of that, which is, “Maybe this is going to be a cool enterprise tool, another way to waste time on the internet, and not quite change everything as much as everyone thinks.” So I see the doomers and the boomers feeding off each other and being part of the same sort of hype universe. As a journalist and as a biographer, you don’t necessarily come down on one side or the other — but actually, can you say where you come down on that? Well, I will say that I find myself using it a lot more recently, because it’s gotten a lot better. In the early stages, when I was researching the book, I was definitely a lot more skeptical of its transformative economic power. I’m less skeptical now, because I just use it a lot more.

In April, Paul Graham, the founder of the tech startup accelerator Y Combinator, sent a tweet in response to former YC president and current OpenAI CEO Sam Altman. Altman had just bid a public goodbye to GPT-4 on X, and Graham had a follow-up question.  “If you had [GPT-4’s model weights] etched on a piece of metal in the most compressed form,” Graham wrote, referring to the values that determine the model’s behavior, “how big would the piece of metal have to be? This is a mostly serious question. These models are history, and by default digital data evaporates.”  There is no question that OpenAI pulled off something historic with its release of ChatGPT 3.5 in 2022. It set in motion an AI arms race that has already changed the world in a number of ways and seems poised to have an even greater long-term effect than the short-term disruptions to things like education and employment that we are already beginning to see. How that turns out for humanity is something we are still reckoning with and may be for quite some time. But a pair of recent books both attempt to get their arms around it with accounts of what two leading technology journalists saw at the OpenAI revolution.  In Empire of AI: Dreams and Nightmares in Sam Altman’s OpenAI, Karen Hao tells the story of the company’s rise to power and its far-reaching impact all over the world. Meanwhile, The Optimist: Sam Altman, OpenAI, and the Race to Invent the Future, by the Wall Street Journal’s Keach Hagey, homes in more on Altman’s personal life, from his childhood through the present day, in order to tell the story of OpenAI. Both paint complex pictures and show Altman in particular as a brilliantly effective yet deeply flawed creature of Silicon Valley—someone capable of always getting what he wants, but often by manipulating others.  Hao, who was formerly a reporter with MIT Technology Review, began reporting on OpenAI while at this publication and remains an occasional contributor. One chapter of her book grew directly out of that reporting. And in fact, as Hao says in the acknowledgments of Empire of AI, some of her reporting for MIT Technology Review, a series on AI colonialism, “laid the groundwork for the thesis and, ultimately, the title of this book.” So you can take this as a kind of disclaimer that we are predisposed to look favorably on Hao’s work.  With that said, Empire of AI is a powerful work, bristling not only with great reporting but also with big ideas. This comes across in service to two main themes.  The first is simple: It is the story of ambition overriding ethics. The history of OpenAI as Hao tells it (and as Hagey does too) is very much a tale of a company that was founded on the idealistic desire to create a safety-focused artificial general intelligence but instead became more interested in winning. This is a story we’ve seen many times before in Big Tech. See Theranos, which was going to make diagnostics easier, or Uber, which was founded to break the cartel of “Big Taxi.” But the closest analogue might be Google, which went from “Don’t be evil” to (at least in the eyes of the courts) illegal monopolist. For that matter, consider how Google went from holding off on releasing its language model as a consumer product out of an abundance of caution to rushing a chatbot out the door to catch up with and beat OpenAI. In Silicon Valley, no matter what one’s original intent, it always comes back to winning.   The second theme is more complex and forms the book’s thesis about what Hao calls AI colonialism. The idea is that the large AI companies act like traditional empires, siphoning wealth from the bottom rungs of society in the forms of labor, creative works, raw materials, and the like to fuel their ambition and enrich those at the top of the ladder. “I’ve found only one metaphor that encapsulates the nature of what these AI power players are: empires,” she writes. “During the long era of European colonialism, empires seized and extracted resources that were not their own and exploited the labor of the people they subjugated to mine, cultivate, and refine those resources for the empires’ enrichment.” She goes on to chronicle her own growing disillusionment with the industry. “With increasing clarity,” she writes, “I realized that the very revolution promising to bring a better future was instead, for people on the margins of society, reviving the darkest remnants of the past.”  To document this, Hao steps away from her desk and goes out into the world to see the effects of this empire as it sprawls across the planet. She travels to Colombia to meet with data labelers tasked with teaching AI what various images show, one of whom she describes sprinting back to her apartment for the chance to make a few dollars. She documents how workers in Kenya who performed data-labeling content moderation for OpenAI came away traumatized by seeing so much disturbing material. In Chile she documents how the industry extracts precious resources—water, power, copper, lithium—to build out data centers.  She lands on the ways people are pushing back against the empire of AI across the world. Hao draws lessons from New Zealand, where Maori people are attempting to save their language using a small language model of their own making. Trained on volunteers’ voice recordings and running on just two graphics processing units, or GPUs, rather than the thousands employed by the likes of OpenAI, it’s meant to benefit the community, not exploit it.  Hao writes that she is not against AI. Rather: “What I reject is the dangerous notion that broad benefit from AI can only be derived from—indeed will ever emerge from—a vision of the technology that requires the complete capitulation of our privacy, our agency, and our worth, including the value of our labor and art, toward an ultimately imperial centralization project … [The New Zealand model] shows us another way. It imagines how AI could be exactly the opposite. Models can be small and task-specific, their training data contained and knowable, ridding the incentives for widespread exploitative and psychologically harmful labor practices and the all-consuming extractivism of producing and running massive supercomputers.”  Hagey’s book is more squarely focused on Altman’s ambition, which she traces back to his childhood. Yet interestingly, she also  zeroes in on the OpenAI CEO’s attempt to create an empire. Indeed, “Altman’s departure from YC had not slowed his civilization-building ambitions,” Hagey writes. She goes on to chronicle how Altman, who had previously mulled a run for governor of California, set up experiments with income distribution via Tools for Humanity, the parent company of Worldcoin. She quotes Altman saying of it, “I thought it would be interesting to see … just how far technology could accomplish some of the goals that used to be done by nation-states.”  Overall, The Optimist is the more straightforward business biography of the two. Hagey has packed it full with scoops and insights and behind-the-scenes intrigue. It is immensely readable as a result, especially in the second half, when OpenAI really takes over the story. Hagey also seems to have been given far more access to Altman and his inner circles, personal and professional, than Hao did, and that allows for a fuller telling of the CEO’s story in places. For example, both writers cover the tragic story of Altman’s sister Annie, her estrangement from the family, and her accusations in particular about suffering sexual abuse at the hands of Sam (something he and the rest of the Altman family vehemently deny). Hagey’s telling provides a more nuanced picture of the situation, with more insight into family dynamics.  Hagey concludes by describing Altman’s reckoning with his role in the long arc of human history and what it will mean to create a “superintelligence.” His place in that sweep is something that clearly has consumed the CEO’s thoughts. When Paul Graham asked about preserving GPT-4, for example, Altman had a response at the ready. He replied that the company had already considered this, and that the sheet of metal would need to be 100 meters square.

Scientific research across fields like chemistry, biology, and artificial intelligence has long relied on human experts to explore knowledge, generate ideas, design experiments, and refine results. Yet, as problems grow more complex and data-intensive, discovery slows. While AI tools, such as language models and robotics, can handle specific tasks, like literature searches or code analysis, they rarely encompass the entire research cycle. Bridging the gap between idea generation and experimental validation remains a key challenge. For AI to autonomously advance science, it must propose hypotheses, design and execute experiments, analyze outcomes, and refine approaches in an iterative loop. Without this integration, AI risks producing disconnected ideas that depend on human supervision for validation. Before the introduction of a unified system, researchers relied on separate tools for each stage of the process. Large language models could help find relevant scientific papers, but they didn’t directly feed into experiment design or result analysis. Robotics can assist in automating physical experiments, and coding libraries like PyTorch can help build models; however, these tools operate independently of each other. There was no single system capable of handling the entire process, from forming ideas to verifying them through experiments. This led to bottlenecks, where researchers had to connect the dots manually, slowing progress and leaving room for errors or missed opportunities. The need for an integrated system that could handle the entire research cycle became clear. Researchers from the NovelSeek Team at the Shanghai Artificial Intelligence Laboratory developed NovelSeek, an AI system designed to run the entire scientific discovery process autonomously. NovelSeek comprises four main modules that work in tandem: a system that generates and refines research ideas, a feedback loop where human experts can interact with and refine these ideas, a method for translating ideas into code and experiment plans, and a process for conducting multiple rounds of experiments. What makes NovelSeek stand out is its versatility; it works across 12 scientific research tasks, including predicting chemical reaction yields, understanding molecular dynamics, forecasting time-series data, and handling functions like 2D semantic segmentation and 3D object classification. The team designed NovelSeek to minimize human involvement, expedite discoveries, and deliver consistent, high-quality results. The system behind NovelSeek involves multiple specialized agents, each focused on a specific part of the research workflow. The “Survey Agent” helps the system understand the problem by searching scientific papers and identifying relevant information based on keywords and task definitions. It adapts its search strategy by first doing a broad survey of papers, then going deeper by analyzing full-text documents for detailed insights. This ensures that the system captures both general trends and specific technical knowledge. The “Code Review Agent” examines existing codebases, whether user-uploaded or sourced from public repositories like GitHub, to understand how current methods work and identify areas for improvement. It checks how code is structured, looks for errors, and creates summaries that help the system build on past work. The “Idea Innovation Agent” generates creative research ideas, pushing the system to explore different approaches and refine them by comparing them to related studies and previous results. The system even includes a “Planning and Execution Agent” that turns ideas into detailed experiments, handles errors during the testing process, and ensures smooth execution of multi-step research plans. NovelSeek delivered impressive results across various tasks. In chemical reaction yield prediction, NovelSeek improved performance from a baseline of 24.2% (with a variation of ±4.2) to 34.8% (with a much smaller variation of ±1.1) in just 12 hours, progress that human researchers typically need months to achieve. In enhancer activity prediction, a key task in biology, NovelSeek raised the Pearson correlation coefficient from 0.65 to 0.79 within 4 hours. For 2D semantic segmentation, a task used in computer vision, precision improved from 78.8% to 81.0% in just 30 hours. These performance boosts, achieved in a fraction of the time typically needed, highlight the system’s efficiency. NovelSeek also successfully managed large, complex codebases with multiple files, demonstrating its ability to handle research tasks at a project level, not just in small, isolated tests. The team has made the code open-source, allowing others to use, test, and contribute to its improvement. Several Key Takeaways from the Research on NovelSeek include: NovelSeek supports 12 research tasks, including chemical reaction prediction, molecular dynamics, and 3D object classification. Reaction yield prediction accuracy improved from 24.2% to 34.8% in 12 hours. Enhancer activity prediction performance increased from 0.65 to 0.79 in 4 hours. 2D semantic segmentation precision improved from 78.8% to 81.0% in 30 hours. NovelSeek includes agents for literature search, code analysis, idea generation, and experiment execution. The system is open-source, enabling reproducibility and collaboration across scientific fields. In conclusion, NovelSeek demonstrates how combining AI tools into a single system can accelerate scientific discovery and reduce its dependence on human effort. It ties together the key steps, generating ideas, turning them into methods, and testing them through experiments, into one streamlined process. What once took researchers months or years can now be done in days or even hours. By linking every stage of research into a continuous loop, NovelSeek helps teams move from rough ideas to real-world results more quickly. This system highlights the power of AI not just to assist, but to drive scientific research in a way that could reshape how discoveries are made across many fields. Check out the Paper and GitHub Page . All credit for this research goes to the researchers of this project. Also, feel free to follow us on Twitter and don’t forget to join our 95k+ ML SubReddit and Subscribe to our Newsletter. NikhilNikhil is an intern consultant at Marktechpost. He is pursuing an integrated dual degree in Materials at the Indian Institute of Technology, Kharagpur. Nikhil is an AI/ML enthusiast who is always researching applications in fields like biomaterials and biomedical science. With a strong background in Material Science, he is exploring new advancements and creating opportunities to contribute.Nikhilhttps://www.marktechpost.com/author/nikhil0980/This AI Paper Introduces ARM and Ada-GRPO: Adaptive Reasoning Models for Efficient and Scalable Problem-SolvingNikhilhttps://www.marktechpost.com/author/nikhil0980/This AI Paper Introduces WEB-SHEPHERD: A Process Reward Model for Web Agents with 40K Dataset and 10× Cost EfficiencyNikhilhttps://www.marktechpost.com/author/nikhil0980/This AI Paper Introduces MMaDA: A Unified Multimodal Diffusion Model for Textual Reasoning, Visual Understanding, and Image GenerationNikhilhttps://www.marktechpost.com/author/nikhil0980/This AI Paper Introduces Differentiable MCMC Layers: A New AI Framework for Learning with Inexact Combinatorial Solvers in Neural Networks

Posted on : May 28, 2025 By Tech World Times Business  Rate this post Starting a new business is exciting. But money is often the first big hurdle. Without funds, it’s hard to build products, hire staff, or run ads. That’s where Startup Funding becomes important. It helps entrepreneurs turn ideas into real businesses. In this article, we share the top 10 startup funding sources for new entrepreneurs in 2025. 1. Personal Savings Most founders start with their own money. It’s simple and quick. You don’t owe anyone else. Using savings shows you believe in your idea. But only use what you can afford to lose. Avoid risking your rent or emergency funds. 2. Friends and Family This is a common early funding source. People who know and trust you may help. They might give you money or offer loans. But always keep it professional. Write down terms and make repayment plans clear. It avoids confusion and protects relationships. 3. Angel Investors Angel investors are wealthy individuals who support startups. They often invest in early stages of business. They bring both money and advice. They usually take equity in return. Search platforms like AngelList or attend startup events to meet them. 4. Venture Capital (VC) VC firms invest big money in fast-growing startups. They look for high returns and innovation. VC funding is best for tech or scalable startups. They take equity and sometimes want a say. It’s competitive, so prepare strong pitch decks. 5. Business Incubators and Accelerators These programs help startups grow fast. They offer funding, training, and mentorship. Some popular examples are Y Combinator and Techstars. They often end with demo days to attract investors. You may give up a small equity share. 6. Crowdfunding Crowdfunding is when many people fund your idea online. Sites like Kickstarter and Indiegogo make this easy. You offer early access or rewards instead of equity. It works well for products people can see. Be sure to promote your campaign heavily. 7. Bank Loans Banks offer loans for startups and small businesses. They usually require a credit check and a business plan. Interest rates vary. Some banks need collateral. If approved, loans give quick access to funds. This option doesn’t dilute your ownership. 8. Government Grants Many governments support small businesses. They offer grants for innovation, research, or job creation. Grants don’t require repayment or equity. But they involve paperwork and clear guidelines. Search local government or small business websites. 9. Business Competitions Pitch competitions can offer funding and exposure. Entrepreneurs present their ideas to a panel of judges. Winners receive cash prizes or investment offers. Even if you don’t win, you gain feedback. Search for startup contests in your city or industry. 10. Corporate Venture Funds Big companies often invest in small startups. They want access to new ideas and technologies. These corporate funds work like VCs. But they may also offer partnerships or clients. Look for companies related to your industry. Tips for Choosing the Right Source Choosing the right Startup Funding source is important. Think about your business stage and goals. Do you need quick cash or long-term help? Can you give up equity or not? Always read the terms and plan your pitch. How to Get Ready for Funding Before applying for funding, get prepared. Make a clear business plan and pitch deck. Know your numbers: costs, sales, profits, and growth plans. Show why your idea is different or better. Practice your pitch until you feel confident. Pros and Cons of Funding Options Each Startup Funding source has pros and cons. Here’s a quick comparison for easy understanding: SourceProsConsSavingsFull controlRisk of personal lossFriends/FamilyEasy accessCan hurt relationshipsAngelsSmart moneyGive up equityVCsLarge fundsHigh pressure to growIncubatorsSupportiveEquity shareCrowdfundingNo equity neededNeeds marketingBank LoansKeep full ownershipInterest, credit checksGrantsFree moneySlow and competitiveCompetitionsWin fundingNo guaranteeCorporate FundsBig supportMight want exclusivity Choose what fits your situation best. Final Thoughts Finding Startup Funding takes effort, but it’s possible. Start small, and build trust with each step. Many successful companies began with tiny investments. You don’t need millions to get started. Pick the right funding source and stay focused. With patience and planning, your startup can grow. Keep learning, keep networking, and never stop pitching. Your big break might be one meeting away. Tech World TimesTech World Times (TWT), a global collective focusing on the latest tech news and trends in blockchain, Fintech, Development & Testing, AI and Startups. If you are looking for the guest post then contact at techworldtimes@gmail.com

Reasoning tasks are a fundamental aspect of artificial intelligence, encompassing areas like commonsense understanding, mathematical problem-solving, and symbolic reasoning. These tasks often involve multiple steps of logical inference, which large language models (LLMs) attempt to mimic through structured approaches such as chain-of-thought (CoT) prompting. However, as LLMs grow in size and complexity, they tend to produce longer outputs across all tasks, regardless of difficulty, leading to significant inefficiencies. The field has been striving to balance the depth of reasoning with computational cost while also ensuring that models can adapt their reasoning strategies to meet the unique needs of each problem. A key issue with current reasoning models is the inability to tailor the reasoning process to different task complexities. Most models, including well-known ones like OpenAI’s o1 and DeepSeek-R1, apply a uniform strategy—typically relying on Long CoT across all tasks. This causes the “overthinking” problem, where models generate unnecessarily verbose explanations for simpler tasks. Not only does this waste resources, but it also degrades accuracy, as excessive reasoning can introduce irrelevant information. Approaches such as prompt-guided generation or token budget estimation have attempted to mitigate this issue. Still, these methods are limited by their dependence on predefined assumptions, which are not always reliable for diverse tasks. Attempts to address these issues include methods like GRPO (Group Relative Policy Optimization), length-penalty mechanisms, and rule-based prompt controls. While GRPO enables models to learn different reasoning strategies by rewarding correct answers, it leads to a “format collapse,” where models increasingly rely on Long CoT, crowding out more efficient formats, such as Short CoT or Direct Answer. Length-penalty techniques, such as those applied in methods like THINKPRUNE, control output length during training or inference, but often at the cost of reduced accuracy, especially in complex problem-solving tasks. These solutions struggle to achieve a consistent trade-off between reasoning effectiveness and efficiency, highlighting the need for an adaptive approach. A team of researchers from Fudan University and Ohio State University introduced the Adaptive Reasoning Model (ARM), which dynamically adjusts reasoning formats based on task difficulty. ARM supports four distinct reasoning styles: Direct Answer for simple tasks, Short CoT for concise reasoning, Code for structured problem-solving, and Long CoT for deep multi-step reasoning. It operates in an Adaptive Mode by default, automatically selecting the appropriate format, and also provides Instruction-Guided and Consensus-Guided Modes for explicit control or aggregation across formats. The key innovation lies in its training process, which utilizes Ada-GRPO, an extension of GRPO that introduces a format diversity reward mechanism. This prevents the dominance of Long CoT and ensures that ARM continues to explore and use simpler reasoning formats when appropriate. The ARM methodology is built on a two-stage framework. First, the model undergoes Supervised Fine-Tuning (SFT) with 10.8K questions, each annotated across four reasoning formats, sourced from datasets like AQuA-Rat and generated with tools such as GPT-4o and DeepSeek-R1. This stage teaches the model the structure of each reasoning format but does not instill adaptiveness. The second stage applies Ada-GRPO, where the model receives scaled rewards for using less frequent formats, such as Direct Answer or Short CoT. A decaying factor ensures that this reward gradually shifts back to accuracy as training progresses, preventing long-term bias toward inefficient exploration. This structure enables ARM to avoid format collapse and dynamically match reasoning strategies to task difficulty, achieving a balance of efficiency and performance. ARM demonstrated impressive results across various benchmarks, including commonsense, mathematical, and symbolic reasoning tasks. It reduced token usage by an average of 30%, with reductions as high as 70% for simpler tasks, compared to models relying solely on Long CoT. ARM achieved a 2x training speedup over GRPO-based models, accelerating model development without sacrificing accuracy. For example, ARM-7B achieved 75.9% accuracy on the challenging AIME’25 task while using 32.5% fewer tokens. ARM-14B achieved 85.6% accuracy on OpenBookQA and 86.4% accuracy on the MATH dataset, with a token usage reduction of over 30% compared to Qwen2.5SFT+GRPO models. These numbers demonstrate ARM’s ability to maintain competitive performance while delivering significant efficiency gains. Overall, the Adaptive Reasoning Model addresses the persistent inefficiency of reasoning models by enabling the adaptive selection of reasoning formats based on task difficulty. The introduction of Ada-GRPO and the multi-format training framework ensures that models no longer waste resources on overthinking. Instead, ARM provides a flexible and practical solution for balancing accuracy and computational cost in reasoning tasks, making it a promising approach for scalable and efficient large language models. Check out the Paper, Models on Hugging Face and Project Page. All credit for this research goes to the researchers of this project. Also, feel free to follow us on Twitter and don’t forget to join our 95k+ ML SubReddit and Subscribe to our Newsletter. NikhilNikhil is an intern consultant at Marktechpost. He is pursuing an integrated dual degree in Materials at the Indian Institute of Technology, Kharagpur. Nikhil is an AI/ML enthusiast who is always researching applications in fields like biomaterials and biomedical science. With a strong background in Material Science, he is exploring new advancements and creating opportunities to contribute.Nikhilhttps://www.marktechpost.com/author/nikhil0980/This AI Paper Introduces WEB-SHEPHERD: A Process Reward Model for Web Agents with 40K Dataset and 10× Cost EfficiencyNikhilhttps://www.marktechpost.com/author/nikhil0980/This AI Paper Introduces MMaDA: A Unified Multimodal Diffusion Model for Textual Reasoning, Visual Understanding, and Image GenerationNikhilhttps://www.marktechpost.com/author/nikhil0980/This AI Paper Introduces Differentiable MCMC Layers: A New AI Framework for Learning with Inexact Combinatorial Solvers in Neural NetworksNikhilhttps://www.marktechpost.com/author/nikhil0980/This AI Paper Introduces GRIT: A Method for Teaching MLLMs to Reason with Images by Interleaving Text and Visual Grounding

Perplexity this week released Perplexity Labs, a new tool for Pro users that can craft reports, spreadsheets, dashboards, and visual representations, to meet users’ increased demand for AI productivity tools with greater autonomy and ever more sophisticated capabilities. The platform, a rival to Anthropic Claude, OpenAI’s ChatGPT, and Google Gemini, can even work on its own for 10 minutes (or more) as it reasons through complicated assignments. “Labs underscores a broader shift toward multi-agent AI systems that plan, execute, and refine full workflows,” said Thomas Randall, research lead for AI at Info-Tech Research Group. Designed to handle more complex assignments Perplexity launched Perplexity Search, its proprietary search engine, in December 2022, just after ChatGPT dropped, and earlier this year released Deep Research (now to be rebranded as Research), which scours the web, reads papers, reasons through materials, and creates comprehensive reports for users. The company says that Perplexity Labs is like “having a team” that can bring projects from ideation to reality. The platform creates reports, spreadsheets, dashboards and simple web apps. It can perform at least 10 minutes of self-supervised work, uses web browsing, writes and executes code to handle tasks like organizing data or applying formulas, and can create charts and images. “In some respects, this is a continuation of Perplexity’s original capabilities as an AI-driven search engine that provides deeper answers,” said Hyoun Park, CEO and chief analyst at Amalgam Insights. Indeed, Perplexity explained that Labs was designed to handle more complex assignments than Deep Research. “While Deep Research remains the fastest way to obtain comprehensive answers to in-depth questions —  typically within 3 or 4 minutes — Labs is designed to invest more time (10  minutes or longer) and leverage additional tools, such as advanced file generation and mini-app creation,” Perplexity wrote in a blog post. “This expanded capability empowers you to develop a broader array of deliverables for your projects.” With its longer research workflow, Perplexity Labs can generate spreadsheets, visual representations, and high-quality reports, the company said. It iteratively searches through hundreds of sources, reasons about that data, and refines its approach as it gets deeper into a project, similar to the way in which a human researcher might approach a new area of study. To create interactive dashboards without the need for coding expertise or external development tools like Ploty and Dash, users just describe what they’re visualizing in natural language, and Labs will generate it in real-time. Dashboards could, for instance, visualize business finances or other complex datasets, incorporating clickable elements to allow non-technical users to quickly act on insights. In one example from the blog, Perplexity prompted Labs from the position of a leader at a tech consulting firm looking to create a potential customer list. It specified that it wanted to partner with US B2B companies in seed, series A, or series B stages, and asked Labs to list 20 relevant companies and include key details including contact information. Labs compiled a comprehensive dataset of potential customers, organizing them by stage (A, B, or seed) and identified their core focus, intended customers, and funding to date. The platform cited links from Forbes, YCombinator, and Exploding Topics that it had used as sources. When further prompted, it crafted introductory emails to the CEOs of the series A startups. To simplify workflows, Labs arranges generated files in a dedicated tab for easy access, supports integration with other tools such as Google Sheets, and allows users to pull out and format citations to bring credibility to its research. Finished materials can be exported as PDFs or documents, or converted into a shareable Perplexity Page. Pro subscribers ($20 a month) can now work with Labs on Web, iOS, and Android; Mac and Windows apps are coming soon. A good fit for enterprise users? This new capability joins an increasingly competitive space, as users look for AI productivity tools that are ever-more performant and can handle more and more tasks autonomously. Park pointed out that Perplexity Labs is a response to tools and models such as OpenAI o1-pro (launched in March), Claude Opus 4 (released in May), and Google’s recent Flow and Firebase announcements. “There is a massive Hunger Games in the AI world right now,” said Park. “Every major vendor is ferociously trying to one-up each other in providing more functionality, either in a native model or with an agency of AI agents designed to work together and create digital assets such as documents, apps, and videos.” However, Perplexity Labs does provide differentiation from other providers in the market, Info-Tech’s Randall noted. In particular, Perplexity is betting that users will prefer a “low-cost, open, tool-agnostic sandbox” for web crawling, code execution, and the creation of finished artifacts including mini web apps. “These capabilities cannot yet be found in other enterprise platforms, such as Microsoft or Google offerings,” said Randall. But enterprises should approach Perplexity Labs with a governance-first mindset, he emphasized. Assets live in Perplexity’s cloud and, for now, lack the private data grounding and compliance controls that CIOs expect, and that they find in tools such as Microsoft Copilot or Google Gemini. From an enterprise perspective, Park noted, the biggest challenge is that every asset-creating model and agent is “still opaque” when it comes to understanding the assumptions, training, and reliability of assumptions used to create a document or app. He compared it to the way the iPhone bypassed BlackBerry and Windows through “sheer consumer delight.” “At some point, AI vendors seeking serious business usage will need to provide more transparency and governance tools to the business world, just as mobile device management and mobile security solutions eventually came to the iPhone,” said Park. Otherwise, businesses may be compelled to build their own clunky but secure versions of Perplexity Labs, “which are guaranteed to be less accurate and useful just [based on] the history of business apps trying to imitate viral consumer apps,” he said.

Crypto criminals can’t hide The single largest cryptocurrency heist in history took place one day in late February, when hackers exploited system vulnerabilities in Bybit, a Dubai-based crypto exchange, siphoning off a whopping $1.5 billion in digital assets within minutes. Bybit’s security team immediately launched an investigation that would eventually involve the FBI and several blockchain intelligence companies. Among those involved from the beginning were the experts at TRM Labs, a San Francisco-based company of around 300 that analyzes the blockchain networks which power cryptocurrency transactions to investigate—and prevent—fraud and financial crimes. “Literally from the first minutes, we were involved,”  says Ari Redbord, the company’s global head of policy, “working with Bybit and law enforcement partners like the FBI to track and trace funds.” The attack was soon attributed to a North Korean state-sponsored hacker organization commonly known as Lazarus Group. Lazarus has been blamed for a series of high-profile cybercrimes in recent years, including the 2014 hack on Sony Pictures Entertainment, the 2016 digital heist from the Bangladeshi central bank and, more recently, billions of dollars in digital currency thefts. TRM was among the first to attribute the Bybit attack after detecting an overlap between the blockchain resources used here and those used in Lazarus’s previous thefts. Since then, the company has harnessed its expertise in tracking crypto to keep law enforcement abreast of where the stolen funds are headed, following them from blockchain to blockchain and through clever concealment mechanisms. “We were very much built for an investigation like this,” Redbord says. Today, TRM’s investigators probe cryptocurrency thefts, ransomware attacks, and phishing scams. They help investigate other crimes that involve digital currencies, from child pornography to drug trafficking. The company’s free, public platform Chainabuse, launched in 2022, helps people report fraud, hacking, blackmail, and other crypto-related crimes. Clients in the cryptocurrency and finance industries harness the company’s software and data about blockchain transactions to identify funds associated with criminal activity and to flag suspicious transactions. Law enforcement agencies around the world enlist TRM’s tools—and sometimes even the company’s own investigators. Demand for such investigators is growing. TRM—which stands for Token Relationship Management—has raised about $150 million in total funding to date, from notable backers that include the venture arms of PayPal, American Express, and Citi, as well as Goldman Sachs. The investment bank led TRM’s most recent, late-stage funding round, which closed in January for an undisclosed amount, according to the research firm PitchBook. Meanwhile, the crypto ecosystem is likely to experience positive growth throughout 2025, according to a recent analysis by PitchBook. So too will crypto crimes: Illicit operations took $40 billion worth of crypto last year, according to Chainalysis, another blockchain security company—far more than the roughly $10 billion in venture capital funding that flowed into the above-board crypto sector in the same span, and more even than crypto’s 2022 VC funding peak of $29.8 billion. Roles like TRM’s will become more urgent if the government continues to abdicate its regulatory duties. Last month, the Trump administration shuttered a Justice Department unit that targeted crypto-related crimes. Yet crypto sits at the nexus of so many of the president’s domestic interests—fentanyl, counterterrorism, border security, and fraud. For TRM and rivals like Chainalysis and Elliptic, all of which have already won millions of dollars in federal contracts, the future is bright. From NFTs to crypto fraud One paradox of Bitcoin, Ethereum, and other cryptocurrency systems is that while they’re widely thought to provide anonymity, with users exchanging funds based not on real names and physical addresses, but on so-called digital addresses—unique and lengthy strings of alphanumeric characters that serve as a given account’s sole identifier—the records of those transactions are still public. A common ledger logs every payment, tying each transaction to those that came before, all the way back to the tokens’ minting. And once information becomes known about one transaction and the people or organizations behind the addresses involved, it becomes possible to trace those funds back and forth through time and from address to address. That allows clever observers to follow the money and deduce where funds came from, who other counterparties may be, and which transactions likely involved some of the same parties, like how investigators might piece together who used an anonymous burner phone based on the numbers they called. It’s a limitation to anonymity that Bitcoin’s pseudonymous creator Satoshi Nakamoto alluded to in the groundbreaking paper describing cryptocurrency’s underpinnings. And it’s one that computer scientist Sarah Meiklejohn and colleagues at the University of California San Diego showed to be a reality in a widely cited 2013 paper that demonstrated concretely how Bitcoins could be grouped by likely common owner—and how those owners could sometimes be identified from a database of known addresses. And that database, Meiklejohn and colleagues showed, could be assembled by a determined researcher simply doing ordinary business on the blockchain and recording the addresses used by the various vendors, exchanges, and other parties they transact with. While not the first company to run with Meiklejohn’s ideas on tracking the transfer of cryptocurrencies—rival Chainalysis, for one, launched in 2014—TRM offered the first-ever platform compatible with the Ethereum blockchain, widely used both for its own currency and assets like non-fungible tokens, or NFTs. At the time, “all of these blockchain intelligence companies had built their entire data architecture on the Bitcoin blockchain,” Redbord says, “because Bitcoin was entirely synonymous with cryptocurrency, and vice versa.” TRM began in 2018 as CEO Esteban Castaño and CTO Rahul Raina’s effort to capitalize on NFTs’ trendiness. After demoing an easy-to-use analytics tool they’d built to help understand NFT market movement to a friend with his own blockchain-based startup, Castaño and Raina decided to pivot. Their creation could be its own product with wide appeal—the same blockchains which track NFTs also manage cryptocurrencies—Castaño says that while “nobody had ever gotten excited about any of the other NFT applications we were building,” this was different. Describing their friend and his employees’ reactions, he says, “it was the first time they’d seen on-chain activity visualized in a way they could understand.” Talking to potential customers soon revealed a critical use case beyond basic customer analytics: understanding the flow of funds on the blockchain to avoid unwittingly participating in money laundering. A now-pivoted TRM publicly launched in 2019 with a tool it planned to sell to blockchain businesses looking to comply with anti-money-laundering regulations. But a more proactive use case soon arose that suggested even bigger opportunities. A friend reached out to say he’d fallen victim to a cryptocurrency hack and wanted to know if TRM could help find the missing money. With the company’s tool, “we could see in clear daylight where the money was,” Castaño says. “So we got in touch with the Secret Service, we got in touch with the FBI, and that was the initial pull into that market.” By the time TRM Labs emerged from Y Combinator, in 2019, fighting and preventing fraud and other crime had become its primary focus. ‘They’re threat hunters’ Many TRM senior leaders and investigators honed their expertise over years in law enforcement, working at police agencies across the world. Redbord, the global policy head, served for more than a decade as a U.S. federal prosecutor and spent two years working on money laundering and national security at the Treasury Department before joining the company. Chris Janczewski, head of global investigations, previously served as a special agent at IRS Criminal Investigations, where he was instrumental in recovering cryptocurrency stolen in the infamous 2016 hack on the Bitfinex exchange; in the time between theft and recovery, the digital coins’ value had ballooned to $3.6 billion, making it the largest federal government seizure in history. The laptop Janczewski used in the investigation is now in the Smithsonian’s permanent collection. “They’re threat hunters,” Redbord says of TRM’s investigators. “Our terror financing expert is out there communicating on password-protected Telegram channels with mujahideen, who will send him a crypto address. He’ll take that address and label it terror financing, and then we use AI and machine learning to build on that attribution.” With investigators around the globe, the company is able to track illicit funds around the clock. “Things like Bybit, you can’t have just one investigator doing that,” says TRM senior investigator Jonno Newman. Being based in Australia, in a time zone close to that of North Korea, made it easy for Newman to help out in the early days of the still-ongoing Bybit investigation. It also helped that he had previously led TRM’s investigation into an earlier hack attributed to North Korea, in 2023, where more than $100 million in cryptocurrency was reported stolen from thousands of blockchain addresses on the digital coin storage tool Atomic Wallet. Then, Newman says, the hackers began obfuscating the stolen funds’ origins and ultimate destination, shuffling their plunder between different virtual addresses and cryptocurrencies. They relied on so-called mixers, which hold and combine coins from multiple sources before disbursing them to new addresses, and cross-chain bridges, which let users convert funds from one cryptocurrency to another. Hackers would later use a similar playbook in moving the Bybit funds. As a result of TRM’s automated fund tracker across bridges, a service it has offered since 2022—an industry first, CEO Castaño says—investigators were able to closely monitor where the Atomic Wallet funds headed, tipping off law enforcement as needed about opportunities to freeze or seize them. “It was early mornings and late nights trying to keep up with the laundering process.” says Newman of the investigation. The former head of South Australia Police’s cybercrime training and prevention unit and author of a recent children’s book about the crypto world, he says “it becomes this almost cat-and-mouse game about where they are going to go next.” TRM’s products at least make the game playable. “When you’re following the money, it used to be that you would reach a dead end when the money went to a different blockchain,” Castaño says. “But with TRM, tracing across blockchains is seamless.” Cautious optimism for blockchain security Not everyone believes TRM’s tech can fully deliver on its promise, at least from a legal perspective. J.W. Verret, an associate professor at George Mason University’s Antonin Scalia Law School who has testified as an expert witness in crypto-related matters, cautions that most testimony based on blockchain forensics tools should be viewed as potentially fallible, “They are useful for developing leads at the start of an investigation,” he says, but can be overly relied on like “the long history of junk forensic science—handwriting analysis, bitemark analysis, stuff that’s all kind of later proven to be unreliable.” For its part, Verret says, TRM Labs offers tools that are less prone than some of its competitors to false positives because the company is more careful about how it establishes associations between blockchain addresses and criminal activity. Meanwhile, last September, TRM announced the creation of the T3 Financial Crime Unit, a partnership with the organizations behind the Tron blockchain and Tether stablecoins to combat the use of those technologies for money laundering. By January, TRM said the partnership had helped freeze more than $100 million in USDT—Tether’s stablecoin pegged in value to the U.S. dollar—found to be tied to criminal activity. That figure has since more than doubled, with the total now including nearly $9 million linked to the massive Bybit heist. “In the seven months since launch, T3 has worked with law enforcement to freeze over $200 million linked to illicit activity ranging from terror financing to money laundering to fraud,” Castaño says. “And when you think about how much crime is financially motivated, adding a $200 million expense to criminals’ balance sheet is a huge win for deterring crime.” But even as TRM jockeys for pole position in a competitive industry, cybercriminals continue to develop new methods of stealing and hiding funds through complex blockchain machinations, often by taking advantage of crypto efficiency gains that make it easier to move more money faster. That will only continue as criminals deploy AI to automate scams and potentially even money laundering—and investigators use new AI and machine learning techniques, along with ever-growing blockchain datasets, to track them more efficiently and coordinate with law enforcement to stop them and seize their funds. And since blockchain ledgers last forever, crypto criminals are risking more than perhaps they realize, according to Castaño. “You’re betting not only that TRM and law enforcement won’t be able to identify your illicit activity today, but that we won’t be able to do it in the future,” he says. “Because the record is permanent.” And that’s the most powerful advantage investigators possess.

Web navigation focuses on teaching machines how to interact with websites to perform tasks such as searching for information, shopping, or booking services. Building a capable web navigation agent is a complex task because it requires understanding the structure of websites, interpreting user goals, and making a series of decisions across multiple steps. These tasks are further complicated by the need for agents to adapt in dynamic web environments, where content can change frequently and where multimodal information, such as text and images, must be understood together. A key problem in web navigation is the absence of reliable and detailed reward models that can guide agents in real-time. Existing methods primarily rely on multimodal large language models (MLLMs) like GPT-4o and GPT-4o-mini as evaluators, which are expensive, slow, and often inaccurate, especially when handling long sequences of actions in multi-step tasks. These models use prompting-based evaluation or binary success/failure feedback but fail to provide step-level guidance, often leading to errors such as repeated actions or missing critical steps like clicking specific buttons or filling form fields. This limitation reduces the practicality of deploying web agents in real-world scenarios, where efficiency, accuracy, and cost-effectiveness are crucial. The research team from Yonsei University and Carnegie Mellon University introduced WEB-SHEPHERD, a process reward model specifically designed for web navigation tasks. WEB-SHEPHERD is the first model to evaluate web navigation agents at the step level, using structured checklists to guide assessments. The researchers also developed the WEBPRM COLLECTION, a dataset of 40,000 step-level annotated web navigation tasks, and the WEBREWARDBENCH benchmark for evaluating PRMs. These resources were designed to enable WEB-SHEPHERD to provide detailed feedback by breaking down complex tasks into smaller, measurable subgoals. WEB-SHEPHERD works by generating a checklist for each task based on the user’s instruction, such as “Search for product” or “Click on product page,” and evaluates the agent’s progress against these subgoals. The model uses next-token prediction to generate feedback and assigns rewards based on checklist completion. This process enables WEB-SHEPHERD to assess the correctness of each step with fine-grained judgment. The model estimates the reward for each step by combining the probabilities of “Yes,” “No,” and “In Progress” tokens and averages these across the checklist. This detailed scoring system enables agents to receive targeted feedback on their progress, enhancing their ability to navigate complex websites. The researchers demonstrated that WEB-SHEPHERD significantly outperforms existing models. On the WEBREWARDBENCH benchmark, WEB-SHEPHERD achieved a Mean Reciprocal Rank (MRR) score of 87.6% and a trajectory accuracy of 55% in the text-only setting, compared to GPT-4o-mini’s 47.5% MRR and 0% trajectory accuracy without checklists. When tested in WebArena-lite using GPT-4o-mini as the policy model, WEB-SHEPHERD achieved a 34.55% success rate, which is 10.9 points higher than using GPT-4o-mini as the evaluator, while also being ten times more cost-efficient. In ablation studies, the researchers observed that WEB-SHEPHERD’s performance dropped significantly when checklists or feedback were removed, proving their importance for accurate reward assignments. They also showed that multimodal input, surprisingly, did not always improve performance and sometimes introduced noise. This research highlights the critical role of detailed process-level rewards in building reliable web agents. The team’s work addresses the core challenge of web navigation—evaluating complex, multi-step actions—and offers a solution that is both scalable and cost-effective. With WEB-SHEPHERD, agents can now receive accurate feedback during navigation, enabling them to make better decisions and complete tasks more effectively. Check out the Paper and GitHub Page. All credit for this research goes to the researchers of this project. Also, feel free to follow us on Twitter and don’t forget to join our 95k+ ML SubReddit and Subscribe to our Newsletter. NikhilNikhil is an intern consultant at Marktechpost. He is pursuing an integrated dual degree in Materials at the Indian Institute of Technology, Kharagpur. Nikhil is an AI/ML enthusiast who is always researching applications in fields like biomaterials and biomedical science. With a strong background in Material Science, he is exploring new advancements and creating opportunities to contribute.Nikhilhttps://www.marktechpost.com/author/nikhil0980/This AI Paper Introduces MMaDA: A Unified Multimodal Diffusion Model for Textual Reasoning, Visual Understanding, and Image GenerationNikhilhttps://www.marktechpost.com/author/nikhil0980/This AI Paper Introduces Differentiable MCMC Layers: A New AI Framework for Learning with Inexact Combinatorial Solvers in Neural NetworksNikhilhttps://www.marktechpost.com/author/nikhil0980/This AI Paper Introduces GRIT: A Method for Teaching MLLMs to Reason with Images by Interleaving Text and Visual GroundingNikhilhttps://www.marktechpost.com/author/nikhil0980/This AI Paper Introduces Group Think: A Token-Level Multi-Agent Reasoning Paradigm for Faster and Collaborative LLM Inference

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More We’re seeing AI evolve fast. It’s no longer just about building a single, super-smart model. The real power, and the exciting frontier, lies in getting multiple specialized AI agents to work together. Think of them as a team of expert colleagues, each with their own skills — one analyzes data, another interacts with customers, a third manages logistics, and so on. Getting this team to collaborate seamlessly, as envisioned by various industry discussions and enabled by modern platforms, is where the magic happens. But let’s be real: Coordinating a bunch of independent, sometimes quirky, AI agents is hard. It’s not just building cool individual agents; it’s the messy middle bit — the orchestration — that can make or break the system. When you have agents that are relying on each other, acting asynchronously and potentially failing independently, you’re not just building software; you’re conducting a complex orchestra. This is where solid architectural blueprints come in. We need patterns designed for reliability and scale right from the start. The knotty problem of agent collaboration Why is orchestrating multi-agent systems such a challenge? Well, for starters: They’re independent: Unlike functions being called in a program, agents often have their own internal loops, goals and states. They don’t just wait patiently for instructions. Communication gets complicated: It’s not just Agent A talking to Agent B. Agent A might broadcast info Agent C and D care about, while Agent B is waiting for a signal from E before telling F something. They need to have a shared brain (state): How do they all agree on the “truth” of what’s happening? If Agent A updates a record, how does Agent B know about it reliably and quickly? Stale or conflicting information is a killer. Failure is inevitable: An agent crashes. A message gets lost. An external service call times out. When one part of the system falls over, you don’t want the whole thing grinding to a halt or, worse, doing the wrong thing. Consistency can be difficult: How do you ensure that a complex, multi-step process involving several agents actually reaches a valid final state? This isn’t easy when operations are distributed and asynchronous. Simply put, the combinatorial complexity explodes as you add more agents and interactions. Without a solid plan, debugging becomes a nightmare, and the system feels fragile. Picking your orchestration playbook How you decide agents coordinate their work is perhaps the most fundamental architectural choice. Here are a few frameworks: The conductor (hierarchical): This is like a traditional symphony orchestra. You have a main orchestrator (the conductor) that dictates the flow, tells specific agents (musicians) when to perform their piece, and brings it all together. This allows for: Clear workflows, execution that is easy to trace, straightforward control; it is simpler for smaller or less dynamic systems. Watch out for: The conductor can become a bottleneck or a single point of failure. This scenario is less flexible if you need agents to react dynamically or work without constant oversight. The jazz ensemble (federated/decentralized): Here, agents coordinate more directly with each other based on shared signals or rules, much like musicians in a jazz band improvising based on cues from each other and a common theme. There might be shared resources or event streams, but no central boss micro-managing every note. This allows for: Resilience (if one musician stops, the others can often continue), scalability, adaptability to changing conditions, more emergent behaviors. What to consider: It can be harder to understand the overall flow, debugging is tricky (“Why did that agent do that then?”) and ensuring global consistency requires careful design. Many real-world multi-agent systems (MAS) end up being a hybrid — perhaps a high-level orchestrator sets the stage; then groups of agents within that structure coordinate decentrally. For agents to collaborate effectively, they often need a shared view of the world, or at least the parts relevant to their task. This could be the current status of a customer order, a shared knowledge base of product information or the collective progress towards a goal. Keeping this “collective brain” consistent and accessible across distributed agents is tough. Architectural patterns we lean on: The central library (centralized knowledge base): A single, authoritative place (like a database or a dedicated knowledge service) where all shared information lives. Agents check books out (read) and return them (write). Pro: Single source of truth, easier to enforce consistency. Con: Can get hammered with requests, potentially slowing things down or becoming a choke point. Must be seriously robust and scalable. Distributed notes (distributed cache): Agents keep local copies of frequently needed info for speed, backed by the central library. Pro: Faster reads. Con: How do you know if your copy is up-to-date? Cache invalidation and consistency become significant architectural puzzles. Shouting updates (message passing): Instead of agents constantly asking the library, the library (or other agents) shouts out “Hey, this piece of info changed!” via messages. Agents listen for updates they care about and update their own notes. Pro: Agents are decoupled, which is good for event-driven patterns. Con: Ensuring everyone gets the message and handles it correctly adds complexity. What if a message is lost? The right choice depends on how critical up-to-the-second consistency is, versus how much performance you need. Building for when stuff goes wrong (error handling and recovery) It’s not if an agent fails, it’s when. Your architecture needs to anticipate this. Think about: Watchdogs (supervision): This means having components whose job it is to simply watch other agents. If an agent goes quiet or starts acting weird, the watchdog can try restarting it or alerting the system. Try again, but be smart (retries and idempotency): If an agent’s action fails, it should often just try again. But, this only works if the action is idempotent. That means doing it five times has the exact same result as doing it once (like setting a value, not incrementing it). If actions aren’t idempotent, retries can cause chaos. Cleaning up messes (compensation): If Agent A did something successfully, but Agent B (a later step in the process) failed, you might need to “undo” Agent A’s work. Patterns like Sagas help coordinate these multi-step, compensable workflows. Knowing where you were (workflow state): Keeping a persistent log of the overall process helps. If the system goes down mid-workflow, it can pick up from the last known good step rather than starting over. Building firewalls (circuit breakers and bulkheads): These patterns prevent a failure in one agent or service from overloading or crashing others, containing the damage. Making sure the job gets done right (consistent task execution) Even with individual agent reliability, you need confidence that the entire collaborative task finishes correctly. Consider: Atomic-ish operations: While true ACID transactions are hard with distributed agents, you can design workflows to behave as close to atomically as possible using patterns like Sagas. The unchanging logbook (event sourcing): Record every significant action and state change as an event in an immutable log. This gives you a perfect history, makes state reconstruction easy, and is great for auditing and debugging. Agreeing on reality (consensus): For critical decisions, you might need agents to agree before proceeding. This can involve simple voting mechanisms or more complex distributed consensus algorithms if trust or coordination is particularly challenging. Checking the work (validation): Build steps into your workflow to validate the output or state after an agent completes its task. If something looks wrong, trigger a reconciliation or correction process. The best architecture needs the right foundation. The post office (message queues/brokers like Kafka or RabbitMQ): This is absolutely essential for decoupling agents. They send messages to the queue; agents interested in those messages pick them up. This enables asynchronous communication, handles traffic spikes and is key for resilient distributed systems. The shared filing cabinet (knowledge stores/databases): This is where your shared state lives. Choose the right type (relational, NoSQL, graph) based on your data structure and access patterns. This must be performant and highly available. The X-ray machine (observability platforms): Logs, metrics, tracing – you need these. Debugging distributed systems is notoriously hard. Being able to see exactly what every agent was doing, when and how they were interacting is non-negotiable. The directory (agent registry): How do agents find each other or discover the services they need? A central registry helps manage this complexity. The playground (containerization and orchestration like Kubernetes): This is how you actually deploy, manage and scale all those individual agent instances reliably. How do agents chat? (Communication protocol choices) The way agents talk impacts everything from performance to how tightly coupled they are. Your standard phone call (REST/HTTP): This is simple, works everywhere and good for basic request/response. But it can feel a bit chatty and can be less efficient for high volume or complex data structures. The structured conference call (gRPC): This uses efficient data formats, supports different call types including streaming and is type-safe. It is great for performance but requires defining service contracts. The bulletin board (message queues — protocols like AMQP, MQTT): Agents post messages to topics; other agents subscribe to topics they care about. This is asynchronous, highly scalable and completely decouples senders from receivers. Direct line (RPC — less common): Agents call functions directly on other agents. This is fast, but creates very tight coupling — agent need to know exactly who they’re calling and where they are. Choose the protocol that fits the interaction pattern. Is it a direct request? A broadcast event? A stream of data? Putting it all together Building reliable, scalable multi-agent systems isn’t about finding a magic bullet; it’s about making smart architectural choices based on your specific needs. Will you lean more hierarchical for control or federated for resilience? How will you manage that crucial shared state? What’s your plan for when (not if) an agent goes down? What infrastructure pieces are non-negotiable? It’s complex, yes, but by focusing on these architectural blueprints — orchestrating interactions, managing shared knowledge, planning for failure, ensuring consistency and building on a solid infrastructure foundation — you can tame the complexity and build the robust, intelligent systems that will drive the next wave of enterprise AI. Nikhil Gupta is the AI product management leader/staff product manager at Atlassian. Daily insights on business use cases with VB Daily If you want to impress your boss, VB Daily has you covered. We give you the inside scoop on what companies are doing with generative AI, from regulatory shifts to practical deployments, so you can share insights for maximum ROI. Read our Privacy Policy Thanks for subscribing. Check out more VB newsletters here. An error occured.

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More We’re seeing AI evolve fast. It’s no longer just about building a single, super-smart model. The real power, and the exciting frontier, lies in getting multiple specialized AI agents to work together. Think of them as a team of expert colleagues, each with their own skills — one analyzes data, another interacts with customers, a third manages logistics, and so on. Getting this team to collaborate seamlessly, as envisioned by various industry discussions and enabled by modern platforms, is where the magic happens. But let’s be real: Coordinating a bunch of independent, sometimes quirky, AI agents is hard. It’s not just building cool individual agents; it’s the messy middle bit — the orchestration — that can make or break the system. When you have agents that are relying on each other, acting asynchronously and potentially failing independently, you’re not just building software; you’re conducting a complex orchestra. This is where solid architectural blueprints come in. We need patterns designed for reliability and scale right from the start. The knotty problem of agent collaboration Why is orchestrating multi-agent systems such a challenge? Well, for starters: They’re independent: Unlike functions being called in a program, agents often have their own internal loops, goals and states. They don’t just wait patiently for instructions. Communication gets complicated: It’s not just Agent A talking to Agent B. Agent A might broadcast info Agent C and D care about, while Agent B is waiting for a signal from E before telling F something. They need to have a shared brain (state): How do they all agree on the “truth” of what’s happening? If Agent A updates a record, how does Agent B know about it reliably and quickly? Stale or conflicting information is a killer. Failure is inevitable: An agent crashes. A message gets lost. An external service call times out. When one part of the system falls over, you don’t want the whole thing grinding to a halt or, worse, doing the wrong thing. Consistency can be difficult: How do you ensure that a complex, multi-step process involving several agents actually reaches a valid final state? This isn’t easy when operations are distributed and asynchronous. Simply put, the combinatorial complexity explodes as you add more agents and interactions. Without a solid plan, debugging becomes a nightmare, and the system feels fragile. Picking your orchestration playbook How you decide agents coordinate their work is perhaps the most fundamental architectural choice. Here are a few frameworks: The conductor (hierarchical): This is like a traditional symphony orchestra. You have a main orchestrator (the conductor) that dictates the flow, tells specific agents (musicians) when to perform their piece, and brings it all together. This allows for: Clear workflows, execution that is easy to trace, straightforward control; it is simpler for smaller or less dynamic systems. Watch out for: The conductor can become a bottleneck or a single point of failure. This scenario is less flexible if you need agents to react dynamically or work without constant oversight. The jazz ensemble (federated/decentralized): Here, agents coordinate more directly with each other based on shared signals or rules, much like musicians in a jazz band improvising based on cues from each other and a common theme. There might be shared resources or event streams, but no central boss micro-managing every note. This allows for: Resilience (if one musician stops, the others can often continue), scalability, adaptability to changing conditions, more emergent behaviors. What to consider: It can be harder to understand the overall flow, debugging is tricky (“Why did that agent do that then?”) and ensuring global consistency requires careful design. Many real-world multi-agent systems (MAS) end up being a hybrid — perhaps a high-level orchestrator sets the stage; then groups of agents within that structure coordinate decentrally. For agents to collaborate effectively, they often need a shared view of the world, or at least the parts relevant to their task. This could be the current status of a customer order, a shared knowledge base of product information or the collective progress towards a goal. Keeping this “collective brain” consistent and accessible across distributed agents is tough. Architectural patterns we lean on: The central library (centralized knowledge base): A single, authoritative place (like a database or a dedicated knowledge service) where all shared information lives. Agents check books out (read) and return them (write). Pro: Single source of truth, easier to enforce consistency. Con: Can get hammered with requests, potentially slowing things down or becoming a choke point. Must be seriously robust and scalable. Distributed notes (distributed cache): Agents keep local copies of frequently needed info for speed, backed by the central library. Pro: Faster reads. Con: How do you know if your copy is up-to-date? Cache invalidation and consistency become significant architectural puzzles. Shouting updates (message passing): Instead of agents constantly asking the library, the library (or other agents) shouts out “Hey, this piece of info changed!” via messages. Agents listen for updates they care about and update their own notes. Pro: Agents are decoupled, which is good for event-driven patterns. Con: Ensuring everyone gets the message and handles it correctly adds complexity. What if a message is lost? The right choice depends on how critical up-to-the-second consistency is, versus how much performance you need. Building for when stuff goes wrong (error handling and recovery) It’s not if an agent fails, it’s when. Your architecture needs to anticipate this. Think about: Watchdogs (supervision): This means having components whose job it is to simply watch other agents. If an agent goes quiet or starts acting weird, the watchdog can try restarting it or alerting the system. Try again, but be smart (retries and idempotency): If an agent’s action fails, it should often just try again. But, this only works if the action is idempotent. That means doing it five times has the exact same result as doing it once (like setting a value, not incrementing it). If actions aren’t idempotent, retries can cause chaos. Cleaning up messes (compensation): If Agent A did something successfully, but Agent B (a later step in the process) failed, you might need to “undo” Agent A’s work. Patterns like Sagas help coordinate these multi-step, compensable workflows. Knowing where you were (workflow state): Keeping a persistent log of the overall process helps. If the system goes down mid-workflow, it can pick up from the last known good step rather than starting over. Building firewalls (circuit breakers and bulkheads): These patterns prevent a failure in one agent or service from overloading or crashing others, containing the damage. Making sure the job gets done right (consistent task execution) Even with individual agent reliability, you need confidence that the entire collaborative task finishes correctly. Consider: Atomic-ish operations: While true ACID transactions are hard with distributed agents, you can design workflows to behave as close to atomically as possible using patterns like Sagas. The unchanging logbook (event sourcing): Record every significant action and state change as an event in an immutable log. This gives you a perfect history, makes state reconstruction easy, and is great for auditing and debugging. Agreeing on reality (consensus): For critical decisions, you might need agents to agree before proceeding. This can involve simple voting mechanisms or more complex distributed consensus algorithms if trust or coordination is particularly challenging. Checking the work (validation): Build steps into your workflow to validate the output or state after an agent completes its task. If something looks wrong, trigger a reconciliation or correction process. The best architecture needs the right foundation. The post office (message queues/brokers like Kafka or RabbitMQ): This is absolutely essential for decoupling agents. They send messages to the queue; agents interested in those messages pick them up. This enables asynchronous communication, handles traffic spikes and is key for resilient distributed systems. The shared filing cabinet (knowledge stores/databases): This is where your shared state lives. Choose the right type (relational, NoSQL, graph) based on your data structure and access patterns. This must be performant and highly available. The X-ray machine (observability platforms): Logs, metrics, tracing – you need these. Debugging distributed systems is notoriously hard. Being able to see exactly what every agent was doing, when and how they were interacting is non-negotiable. The directory (agent registry): How do agents find each other or discover the services they need? A central registry helps manage this complexity. The playground (containerization and orchestration like Kubernetes): This is how you actually deploy, manage and scale all those individual agent instances reliably. How do agents chat? (Communication protocol choices) The way agents talk impacts everything from performance to how tightly coupled they are. Your standard phone call (REST/HTTP): This is simple, works everywhere and good for basic request/response. But it can feel a bit chatty and can be less efficient for high volume or complex data structures. The structured conference call (gRPC): This uses efficient data formats, supports different call types including streaming and is type-safe. It is great for performance but requires defining service contracts. The bulletin board (message queues — protocols like AMQP, MQTT): Agents post messages to topics; other agents subscribe to topics they care about. This is asynchronous, highly scalable and completely decouples senders from receivers. Direct line (RPC — less common): Agents call functions directly on other agents. This is fast, but creates very tight coupling — agent need to know exactly who they’re calling and where they are. Choose the protocol that fits the interaction pattern. Is it a direct request? A broadcast event? A stream of data? Putting it all together Building reliable, scalable multi-agent systems isn’t about finding a magic bullet; it’s about making smart architectural choices based on your specific needs. Will you lean more hierarchical for control or federated for resilience? How will you manage that crucial shared state? What’s your plan for when (not if) an agent goes down? What infrastructure pieces are non-negotiable? It’s complex, yes, but by focusing on these architectural blueprints — orchestrating interactions, managing shared knowledge, planning for failure, ensuring consistency and building on a solid infrastructure foundation — you can tame the complexity and build the robust, intelligent systems that will drive the next wave of enterprise AI. Nikhil Gupta is the AI product management leader/staff product manager at Atlassian. Daily insights on business use cases with VB Daily If you want to impress your boss, VB Daily has you covered. We give you the inside scoop on what companies are doing with generative AI, from regulatory shifts to practical deployments, so you can share insights for maximum ROI. Read our Privacy Policy Thanks for subscribing. Check out more VB newsletters here. An error occured.