Enterprises with a Neural Spine

21 Jan 2026 15:45h - 16:30h

Session at a glance

Summary

This discussion, moderated by Ina Turpen Fried, focused on the transformation of businesses in the AI-native era and the competition between traditional companies and AI-first startups. The panelists, including Yutong Zhang, Bipul Sinha, Richard Socher, Ioannis Antonoglou, and Ravi Mhatre, explored how organizations are adapting to AI integration and the challenges they face.

A key theme was the emergence of “bring your own AI to work” practices, where employees use AI tools personally to enhance productivity, often paying out of their own pockets. The discussion highlighted how AI-native companies operate with extremely lean teams, sometimes having more AI agents than human employees, creating unprecedented operational leverage. However, traditional enterprises face significant barriers to AI adoption, particularly around compliance, governance, and risk management, with chief risk officers concerned about hallucinations and security vulnerabilities.

The panelists emphasized the importance of developing trust in AI systems through better technology, transparency, and citation capabilities. They noted that trust builds over time as AI systems demonstrate consistent accuracy, similar to how people gradually trusted computers. The conversation explored how AI is transforming user interfaces, with natural language potentially replacing traditional software interactions, making “software invisible” as users access capabilities through conversational agents.

Looking ahead, the panelists predicted several developments for the coming year: AI systems building better AI, the emergence of truly long-term AI agency, continuous learning capabilities in models, and the widespread application of reinforcement learning to business-specific workflows. The discussion concluded that the biggest companies in 20 years will likely be those starting today as AI-native organizations.

Keypoints

Major Discussion Points:

– AI-Native vs. Legacy Companies: The competitive landscape between established companies with existing workflows and new startups built from scratch with AI-first approaches, including the advantages and challenges each face

– Trust, Compliance, and Enterprise AI Adoption: The critical barriers preventing large organizations from moving AI pilots to production, particularly around governance, risk management, and the need for “trust-based technologies” to enable faster adoption

– Agent Orchestration and Multi-Agent Systems: The emerging need for sophisticated systems to coordinate multiple AI agents working together, and the lack of mature science around optimizing these complex human-AI collaborative workflows

– The Future of Software and User Interfaces: The prediction that traditional software will become “invisible” as natural language interfaces powered by AI agents replace complex GUI interactions and manual processes

– Long-Term AI Agency and Autonomous Systems: The evolution toward AI systems capable of working independently over extended periods (like weekly cycles) and the potential for AI to eventually build better AI systems than humans can

Overall Purpose:

This appears to be a panel discussion at Davos focused on exploring how AI is transforming business operations, the competitive dynamics between AI-native startups and established companies, and predictions for the future of AI adoption in enterprise settings.

Overall Tone:

The discussion maintains an optimistic and forward-looking tone throughout, with participants sharing insights as industry experts and investors. The conversation is collaborative rather than confrontational, with panelists building on each other’s points. The tone becomes increasingly speculative and visionary toward the end as participants make predictions about radical changes they expect within the next year.

Speakers

– Ina Turpen Fried: Moderator/Host (mentions doing a daily newsletter for Axios)

– Bipul Sinha: Business executive focused on AI compliance and governance infrastructure for enterprises

– Richard Socher: CEO/Founder of U.com, AI company executive with expertise in search engines and LLMs

– Ravi Mhatre: Investor, focuses on AI startup investments and enterprise AI adoption

– Yutong Zhang: AI company executive, works on consumer AI applications with tens of millions of users globally, has experience with Chinese AI market

– Ioannis Antonoglou: AI company founder/executive building frontier-level reasoning models, focuses on open source AI development

Additional speakers:

None – all speakers mentioned in the transcript are included in the provided speakers names list.

Full session report

Discussion Report: AI-Native Business Transformation at Davos

Executive Summary

This panel discussion, moderated by Ina Turpen Fried from Axios, examined how businesses are adapting to AI-native approaches. The conversation featured Yutong Zhang, Bipul Sinha, Richard Socher (CEO of U.com), Ioannis Antonoglou, and Ravi Mhatre. The discussion explored the differences between traditional enterprises and AI-first companies, barriers to AI adoption, and predictions for AI’s future development.

Key themes included the operational advantages of AI-native companies, compliance challenges preventing enterprises from scaling AI pilots to production, and the transformation of software interfaces through AI agents. The panelists concluded with specific predictions about AI capabilities expected within the coming year.

Key Discussion Points

AI-Native Companies and Operational Efficiency

Yutong Zhang highlighted the remarkable efficiency ratios achieved by some AI-native companies, noting examples of organizations with “less than 10” people but “hundreds of agents” working alongside them. Zhang’s company, which has “300 people” and builds both models and applications serving “tens of millions of users,” represents this new operational model.

The moderator framed this as competition between established companies with legacy processes and startups that can build AI-first from the ground up. The discussion suggested that companies beginning with AI-native approaches have significant advantages over those trying to retrofit existing workflows.

Enterprise AI Adoption Barriers

Bipul Sinha identified the primary challenge facing large organizations: “The biggest difficulty overall we see in our customer base is they have a lot of pilots and they love the outcome of those pilots, but go from pilot to production, their biggest worry is compliance risk.”

The panelists discussed various approaches to building trust in AI systems. Zhang emphasized technological transparency, suggesting that showing citations, references, and reasoning chains can help build confidence in AI decisions. Antonoglou noted that trust develops gradually as AI systems prove reliable over time, similar to how people learned to trust computers.

Richard Socher stressed the importance of systematic evaluation frameworks, arguing that successful AI adoption requires principled, scientific approaches to testing and validation rather than ad-hoc methods.

Software Evolution and User Interfaces

A significant portion of the discussion focused on how AI will change software interaction. Zhang described a vision where software becomes “invisible” as users interact through natural language with AI agents rather than traditional interfaces. This connects to the broader theme of AI agents handling complex workflows autonomously.

Ravi Mhatre noted that the dramatic decrease in AI inference costs (100-1000x reduction in the past year) is making intelligence-driven automation economically viable for many new use cases. He suggested this will enable AI to automate areas where traditional software was never built.

The conversation referenced Marc Andreessen’s famous prediction that “software will eat the world,” with the suggestion that now “AI will eat software.”

Consumer AI Adoption Patterns

Zhang observed interesting adoption patterns in consumer AI, noting that users are willing to pay for AI tools that enhance their workplace productivity, describing this as people bringing their own AI solutions to work environments.

The discussion touched on China’s AI landscape when the moderator asked about the interplay between startups and tech giants, though Zhang’s response focused more on general patterns of AI adoption and efficiency gains.

Future AI Capabilities – Panelist Predictions

The discussion concluded with specific predictions from each panelist about AI developments expected within the coming year:

Richard Socher made perhaps the most striking prediction: “That AI will build the best AI, not people,” suggesting a shift toward automated model development.

Yutong Zhang predicted the emergence of “really long agency,” describing AI systems that can work autonomously for extended periods like human employees.

Ravi Mhatre anticipated “continuous learning” capabilities, allowing AI models to adapt and update their knowledge through ongoing interaction with their environment.

Bipul Sinha focused on reinforcement learning (RL), predicting better understanding of how to adapt generic models to specific business workflows and processes.

Ioannis Antonoglou agreed with the long agency prediction and emphasized that multi-agent system dynamics will become increasingly important as AI agents proliferate.

Workforce and Human-AI Collaboration

The panelists generally agreed that the future involves human-AI collaboration rather than complete replacement. Antonoglou emphasized that humans will remain essential, working with AI to increase productivity. However, Zhang’s examples of companies with hundreds of agents supporting fewer than ten people suggest this collaboration may involve more extreme ratios than traditionally anticipated.

Antonoglou provided a framework for AI-first thinking: examining existing workflows and rethinking them from first principles, knowing that AI exists and will continue improving.

Key Takeaways

The discussion revealed several important insights about AI adoption in business contexts:

– AI-native companies are achieving unprecedented operational efficiency ratios through human-agent collaboration

– The primary barrier to enterprise AI adoption is compliance and governance concerns rather than technical limitations

– Trust in AI systems can be built through transparency mechanisms and systematic evaluation frameworks

– The dramatic reduction in AI inference costs is enabling new forms of automation

– Software interfaces may fundamentally change as natural language interaction with AI agents becomes prevalent

– The panelists expect significant advances in AI autonomy and self-improvement capabilities within the coming year

Conclusion

This discussion provided practical insights into current AI adoption challenges and future trajectories. The conversation moved beyond general AI trends to explore specific operational models, compliance barriers, and technical predictions. The panelists’ focus on systematic evaluation, trust-building, and human-AI collaboration offers actionable guidance for organizations navigating AI transformation, while their specific predictions about autonomous agents and AI-driven development suggest significant changes ahead in how businesses operate and software is created.

Session transcript

Ina Turpen Fried

The Path of Seclusion A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by

Bipul Sinha

A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by

Richard Socher

A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by

Ina Turpen Fried

A film by A film by

Ravi Mhatre

A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by A film by to

Ina Turpen Fried

ask you to represent an entire country, especially one as large as China. Help us understand how the interplay is going between these startups that have been born since the AI models and some of the older companies in China, some of the tech giants in China. What are you seeing?

Yutong Zhang

Yutong Zhang Yeah, I think also like for organizations, I think one approach is, I think like Ravi said, is like more top-down, the management have the AI native mindset. But the other perspective we see interesting is bring your own AI to work. So we actually have a consumer app where we have tens of millions of users from the globe.

We actually observed that we have the user research and interview with them, and a lot of people that are using AI at work, and they are willing to pay out of their own pocket. So I think which means AI can really enhance the productivity in a lot of the jobs that we are doing today. I think that would be very interesting.

But from the organization level, I think a couple of things may be interesting to see. I think one is the human-to-agent ratio, you know, the company right now, all the startup company are really, really lean and small. We have like 300 people, but we are like building models and building applications.

And we saw some company, if they are purely just building applications, they have people like less than 10, but they have, you know, hundreds of agents helping them to running a lot of things on an operational level.

So I think right now, AI giving all the companies a really high operation leverage.

Ina Turpen Fried

Yeah. I mean, I do think that that’s one of the discussions that I’ve heard start in Davos that I think is going to really continue throughout the year is this competition is no longer you and your existing competitors.

If you’re an existing startup or existing company, it’s really you against somebody who started from scratch with no workflow, with no process, with no legacy that says, well, I can do that. And obviously the incumbents don’t come to this with no assets. I don’t think that.

But can you talk and whoever wants to jump in first, but I’m curious, you know, obviously every company of more than one person has a workflow among people. How do you make that an asset? How do you use AI within the organization while still recognizing that it’s humans that have to absorb the chain?

And I’m curious if anyone maybe has examples from their own company of how you’re harnessing AI while recognizing the fact, I don’t think anyone’s company here fits in the category of more agents than people, or at least not a lack of people.

Bipul Sinha

The difficulty of like AI adoption is we heard about bring your own AI to work, but the particularly American companies, the compliance, the governance actually restricts how much data you can feed to AI that you want to bring to work.

So what we have done is we have really created a kind of like a compliance infrastructure saying here are the set of data and here are the set of certified models that you are allowed to interact with.

And then based on that, we are actually applying coding, legal, marketing, customer support. So every function is working within that category. The biggest difficulty overall we see in our customer base is they have a lot of pilots and they love the outcome of those pilots, but go from pilot to production, their biggest worry is compliance risk.

And then chief risk officers, CISOs are coming down and saying, how do I ensure that these agents are within their guardrails? Are they hallucinating? Are they compromised by nation state threat actors?

Ina Turpen Fried

And I think, won’t that be where some of the tension is between the older companies that may be slowed by their legacy, but already understand those things. And the one person startup that vibe coded this amazing demo, but suddenly wants to go into an industry where they have to meet compliance. So the flip side is, you know, what are some of the challenges for these AI native companies that might be, you know, eight people and 800 agents, but have to follow a set of rules and laws?

Richard Socher

I mean, we definitely have more agents than people inside U.com. I think one of the biggest challenges that we see for companies that want to embrace AI is that they often don’t have yet a benchmark or eval set mindset. And this is, I think, a big divide.

If your company can identify what are the inputs and what do you classify as a correct output, and you create a set, and it’s not just like, oh, I tried these three things and it worked, or I tried these four things and it didn’t work, so I don’t want to use that tool.

You got to make it into a principled, more scientific process. And if you have that benchmark, we love working with those folks, because then we win, because we have the best models and answers and accuracy and like citations and all that. But the vast majority of companies out there have not yet thought about benchmarking and creating sort of scientific ways to evaluate their AI models and their agents.

Ina Turpen Fried

And Ioannis, you obviously, you know, are an AI native company in the sense that you were born of and for this mission, but also have been around long enough that I’m guessing there’s a few people involved as well.

Talk about how you view both for your own company, but also for your customer base, this idea of how, you know, it’s going to be a mix of legacy companies that have humans and some of these new startups that started with very few people.

Ioannis Antonoglou

Yeah, maybe I’ll just like start by saying, you know, going back to the previous conversation of like what’s an AI first in my mind, and AI first is the idea of just like seeing what are the workflows that you have and what are the things that you’ve done, and then just like try to rethink them from like first principles, knowing that like AI exists and knowing that AI will just become better.

And of course, you know, the new workflow will have like humans in the loop. It’s not going to be just AI. And you need to just be really careful on how you design and architect the system so that you can maximally use your people and ensure that like they’re used like human labor is actually used where it’s necessary and is used where it’s like the most efficient.

So I don’t see a world, at least for like the next few years where, or actually the foreseeable future where like humans will be completely out of the picture. I think that like there’s going to be a combination of humans and AI is just working closely together to just increase productivity and ensure that like everything falls into place. So like now in terms of, I think like what’s happening with traditional industries is that they have a lot of the know how on like the things that they need to be solving and the problems that they need to be solving, especially like with respect to compliance, with respect to their customers and their industry, but they don’t really fully understand AI.

And this is where it’s important to actually like work with, you know, other companies like AI companies that have the deep understanding of how AI works and what it’s capable of to kind of like have this synergy of like, you know, bringing the industry experience and bringing the AI experience, working together to rethink and re-architect the workflow so that they’re optimized for the AI age.

Ina Turpen Fried

And Ravi, you’re obviously investing in companies. I’m curious to like how you look at, like, what do you want to hear from a startup CEO in terms of how much are they relying on people versus AI? Do you want to hear from-

Ravi Mhatre

If I can just go back to one thing that I mean, I really be curious as to Yannis’ view, but people alluded to it that today the business leaders for the large enterprises, these are not AI first companies, they do, even if they are trying to reimagine what access to unlimited intelligence could do to reinvent their businesses.

This need for trust and safety, it’s turned into a compliance mindset, which actually slows down the adoption and the usage of AI in their businesses. And what we need is really, I believe it’s a set of both technologies and then a mindset that allows for AI to be safe and trusted, because trust, if there were trust-based technologies embedded in AI, then I think these more traditional enterprises and business leaders could be in a lean-in mode where they say, as I scale this, I know that safety will be resident in it, and so I can go fast as opposed to thinking about as I introduce this new technology, how do I go slow or why should I be more conservative?

And I’m just curious, Yannis, I mean, I think you alluded to the fact that right now it’s evals and business leaders in traditional enterprises, they don’t have the wherewithal or the expertise in their companies to come up with sophisticated evals.

And so we think that one of the critical accelerants for adoption, at least in democratic AI-based societies, has got to be productization of trust-based technologies somewhere inside or alongside of the models, and Yannis is actually building frontier-level reasoning models that are open source, the companies based in the U.S., but how do you think about the trust-based technologies, again, as those kinds of systems are diffused into the world and into businesses that would allow them to be used kind of in a good way?

Ioannis Antonoglou

Yeah, that’s a really good question, actually, because this is, I guess, one of the main obstacles in the adoption of AI. How do you build the trust and how do you ensure that what you think the AI is doing is actually what it’s doing? And actually, it boils down to having better technology.

When JudgePT came out in 2022, I think there were all these hallucinations and everyone was really skeptical of anything that they would just read on JudgePT and they would just double-check it, and in most cases it was actually wrong.

But now, for the most part, when you just read something on your JudgePT, you trust it. You’re like, okay, probably it’s right. You’ve actually double-checked it a few times, and over time, it won your trust.

So you saw it being correct time and again. And over time you actually developed this trust with AI. And I think that AI is not, it’s something that we will be interacting a lot in different ways in the near future.

And we’ll just start developing this trust with AI systems in general. We’ll just start to trust them more. The same way that, even when computers came out, people didn’t trust them.

But now everyone trusts the computer, it will just do the things that you told them to do. So.

Ina Turpen Fried

I was just gonna ask you, Tong, is that something that is, is it approached differently in China? Or is it a fairly similar notion of trust and the need for trust in order for businesses to use AI?

Yutong Zhang

Yeah, I think actually trust can be enabled by technology. And the trust is also a perception of the users, right? It’s like, when we first launched the application, we are connected with search tools.

And also, we show all the citations, all the reference, where AI got the information from the web, from the files to the specific sentence. This type of layer of interaction actually help people to feel more trusted toward the technology. And also, I think a chain of thought is another thing that AI was like a black box before.

They give you a number from a complex calculation. You don’t know where the number is coming from. But right now, with all the reasoning steps, we can see the whole trajectory of how AI, using tools, gathering information together to synthetic the answer.

So I think definitely that is very important. AI cannot just providing the direct answer. It have to just unbox all the thinking process, all the tools, all the data sources.

So I think there are a lot of infrastructure to be built there to make people feel, OK, they can trust the technology with not only execution and also a lot of decisions on how to run the business. Or they can really let the LLM to decide a lot of the things. If people can define what good looks like, having all the expert opinion, being the rubrics, and then we can just really building a scalable decision system driven by LLM.

Ina Turpen Fried

And people, I’m curious how you’re, because others may have thoughts on this, and feel free to jump in after people. But you’re building a business to help other companies build for this AI world. And I’m curious, today, and I imagine this will shift, how much are you thinking of helping the larger established businesses?

And how much of your thought is shifting to, how do I make sure I’m a good partner for that generation of couple people startups? Where’s the energy today? And how much are you thinking about the latter?

Bipul Sinha

For larger corporation, the question is what we were talking about is the trust. And this is what we are focused on. We are helping them adopt agentic work.

What we have understood is that many of these organizations have documents which says, this is what you’re allowed to do. This is what you’re not allowed to do. But it is very hard for them to translate that into a set of rules that your agents and your models can perform, because those are probabilistic systems.

And this is a fixed set of rules. So what we are working with our customers is, can we take this set of rules, train a model that will be a judgment model on your agentic work? So you are now judging an LLM with another LLM that is based on your own set of rules, similar to what you were describing.

As a result, you don’t have to create rules. Let the neural drive the rules. So LLM can judge LLM.

And then you have better confidence in terms of the accuracy of these results. So these are the directions that our customers want to go. They don’t want to be in the business of translating their business rules into stopping agents.

They want to enable agents. But they want to enable agents in a way that they have a higher confidence in hallucination rates, third-party cyber attacks, surface area is bigger. So those are the issues that we are solving for.

Ina Turpen Fried

So it sounds like you still see your business as helping larger companies adopt AI. And I’m curious, maybe Richard, do you have thoughts on how that needs to shift over time so that even upstarts are thinking, how do I help these new generation? Because I imagine when we’re sitting here a year or two from now, somebody on this panel is going to be running a billion dollar or more valuation.

Although now that seems like nothing these days. That’s a pre-seed startup. But some substantial valuation of company with only a few people.

And those will be some of your customers. Richard, how do you think about that?

Richard Socher

I mean, in general, of course, everyone loves scale. And large customers can bring scale. But large customers also have just much, much longer procurement processes.

And working through those is hard as a small startup. So it’s a trade-off. The more smaller startups you have, the faster you can move.

And you can hope that some of those startups actually will scale and keep growing. And then when it comes to trust, it’s something that is really important for us, too. Actually, in 2022, we filed patents and shipped the first connection of a search engine with an LLM so that you actually have citations.

And I think citations have now obviously been everywhere. And they are a way to build that trust over time. And that’s, again, coming back to that delegation capability.

Part of delegating well is knowing when to trust and when to trust but verify. And then kind of build that over time. But yeah, overall, my hunch is the biggest companies in 20 years are those that are starting right now and are AI native.

Ina Turpen Fried

And Yutong, you mentioned early on that some of your customers are actually individuals bringing your chatbot in to help them do their jobs better, whether their employer is AI first or not. How do you think about who the customer is today for your chatbots, but also who the customer of the future is?

Yutong Zhang

Yeah. So I think a very important change is I’m not saying that software will disappear. But I think software will become invisible.

So I think right now it’s like if we want to create a Word document or we need to remember all the Excel formulas to using Excel or create a very beautiful, presentable PowerPoint slides, people are directly using the GUI, clicking the hundreds of buttons, remember all the formulas.

But I think in the future, definitely, human will just using natural language to access all the powers from all the tools and all the softwares through agent. I think AI can help us to get access to all the softwares. I think that’s the way that it will change how we work.

A lot of the companies where people have access to AI as a superpower to help with their work. Somebody said this week, and I apologize, I’m stealing someone else’s wisdom, but there was the famous Andreessen saying of software is eating the world and that AI is eating software. Is that your sense?

I don’t know. I just think that it more like a UI and UX transformations because we are relying on a lot of the clicks and the keyboards for the previous interactions. But I think the interaction going forward will be more natural that people, as long as they can describe their intent, they can access, they know how to use AI, and then AI can help them to access all the powers that offer by existing softwares.

And if there is no existing softwares, AI can use in their coding capabilities to write personalized tools or on the spot to help to deliver the end results. I think that’s definitely.

Ravi Mhatre

And I mean, I would agree 100% with you, Yutong Zhang. I mean, it partly is related to the fact that people don’t appreciate that it really is an exponential curve in terms of the improvement in AI or intelligence kind of capabilities. In the last year, on average, the cost per token for inference, on average, went down 100x.

In many cases of use cases, went down 1,000x. So when that’s happening, just this concept of intelligence being able to essentially make software disposable where it can be vibe-coded with robustness as in the Quad Coast case, and it then can be mostly production-ready, the ability for just someone to imagine any task that you want to automate via a piece of software and use intelligence to make that happen.

If we’re on this exponential curve of how powerfully intelligence can sort of show up for a given cost parameter, I think you will see that software becomes invisible and intelligence will be sort of the new kind of language in which people express what they want automation behind the scenes to do, and it will just be kind of a continuous flow of that.

Ina Turpen Fried

If we take that as the assumption, does anyone want to challenge the assumption that that’s where we’re headed?

Richard Socher

My hunch is there will be some software where you really want it to be just really well done, and it’s a very common use case, like Instagram or something, right? You could hack up an Instagram, but then you don’t have the network behind it, you don’t have the people contributing their content and stuff. But I agree that especially in enterprise, if you do it cleverly and you have all the complex kind of permission sets and reporting and so on, if you can automate that and then innovate on top of that, then it would be very powerful.

Ravi Mhatre

Yeah, I think this will expand the use cases in places where software doesn’t get built and you have more manual processes and workflows, that will likely go away. And again, the rate of change, because it’s more of an exponential improvement in the capabilities of intelligence, I think will surprise people. But I agree with you also that applications where you have many, many users and so fit and finish and certain things about robustness of the design capability matter, that will still for a longer period of time, maybe ultimately in the limit.

that also can be automated with some combination of engineering people who use AI to create leverage in their workflows but also have to do some of the design.

Ina Turpen Fried

So if we take that as the assumption of where we’re headed, what are we not considering in terms of how that comes? For example, I think one of the things we underestimated with AI is just how fast humans would be able to adapt it. If we assume as a starting point that we are going to be in this world where businesses are going to create these layers that just do what needs to be done without creating another piece of software, what might we be missing?

What do we also have to pay attention to before we run full speed ahead? What still needs to be done?

Bipul Sinha

In my mind, agentic orchestration is an area that is going to be very important. Because if you think about all the software that we have built in the last 15, 20 years to automate the enterprise for digital transformation, this new agentic orchestration layer will make use of those software.

It’s just that human beings will not be involved in the process. And over time, it will go deeper and deeper and deeper and make the workflow engine totally useless. So ultimately, you will have the data structure and then you have agentic orchestration.

So what is less understood or not yet been focused on is this agentic orchestration layer that takes an input from you as what do I want to accomplish? And then how do you create and orchestrate the agents? So in my mind, if you fast forward 10, 15 years, you will have data and agent orchestrator that can create any kinds of workflow that you need, whether you want to take a business order from a customer, or you want to run a campaign to your customer or prospects, or you want to support your customer.

It will dynamically generate that workflow for you, run that workflow, create the result, give you the outcome and say this is what has happened. But that’s a process that is an orchestration process.

Richard Socher

I would say maybe there are two things. One is right now, AI is we’re moving away from more and more from infrastructure that was initially built for humans, and now AI is kind of starting to use it more and more. We’ve seen this like GPUs, they’re initially for people playing games, and we train them.

Now we’re building new hardware that is only made for AI. We see this as you move up the stack, like a lot of AI models used to use Google as their search engine or SERP APIs and so on. Now they’re actually transitioning and we’re building an actual search engine just for LMs, for them to search for us and then summarize everything.

I think the other thing that we’re massively missing and not considering is what the next level might be. And the way I can imagine that from history is, imagine when the IBM Watson team just won Jeopardy. Everyone’s high-fiving, you’re like, good job to the team that worked on geography questions.

Good job to the team that worked on music questions and celebrity questions and so on. But by the way, the 800 of you are going to be one neural network because we can learn all of that now and replace a manual system with a learned system. Right now, the system of building models is actually done manually.

Really smart people that sometimes you have to pay $50, $100 million to are manually creating intuitions on what works when they create new models and then it takes them months to merge a new model and then you have GPD, you know, 5.3 or something.

That part itself can also be automated and that will be sort of recursive self-improving superintelligence. These are models that sort of get better on their own.

Ina Turpen Fried

Exactly. And Giannis, how about from you, what do you think some of the things that aren’t problems today because we aren’t AI native, but when we get AI native, things that become problems or challenges?

Ioannis Antonoglou

Yeah, actually, I’ll just go back to the agent orchestration and I think it’s not just like a problem of agent orchestration, it’s rather the problem of multi-agent. And when you have a system that, you know, it’s both humans and many agents interact with each other and what the dynamics of like these systems are, it’s not a well understood system. We don’t have like any science really behind that.

We don’t know how to optimize these models to operate better in multi-agent systems. You know, there is like some early research on that, but like it’s not mature enough. And I think that like this, as there’s a proliferation of like agents and proliferation of like systems with like many, many agents interact with each other, this will become more of a pressing problem.

Ina Turpen Fried

So for a last question for each of you, what’s something that sounds kind of radical today, but you think will be obvious by the end of the year? By the time we get to Davos next year, sounds like a great prediction today, but you think it’ll be common sense wisdom by the time we get to rejoin next year?

Richard Socher

That AI will build the best AI, not people.

Yutong Zhang

I think it’s a really long agency. I think before that we really have a useful, scalable AI system, the AI need to have really long agency like human. You know, I was working with my team, it’s like we have a weekly meeting and we set up the goal and we discuss roughly about a task and then they just go off and they can autonomously working on their own and everything, and then we can meet at the next week to see the results.

So I think the really long agency will probably happen.

Ina Turpen Fried

Any other thoughts?

Ravi Mhatre

I think we’ll start to see the first signs about the kind of frontier of intelligence capability, the ability for some form of more continuous learning, it may not be perfect, but the ability for models to be adaptive for how they interact with their environment and then kind of update their core knowledge set.

Bipul Sinha

For businesses, RL is not properly understood today. The challenge with a lot of businesses are there are no like generic employees, everybody does a specific task for the business and that is their business processes that is defined for that business. And what is not well understood is how is RL going to take that specific workflow for the business and do reinforced learning to make a generic model work for business.

That we’ll see next year as the big thing. Today it’s a little theoretical, but that’s what will accelerate the adoption of AI in the enterprise.

Ina Turpen Fried

Ioannis, you get the last word.

Ioannis Antonoglou

Yeah, I mean, I feel like I’ll agree with Yutong that really long agency is going to be like commonplace.

Ina Turpen Fried

Well, I hope we get the chance to discuss all that next year. I’m sure it will be a year of great change. Thank you so much, Ioannis, Bipol, Richard, Yutong and Ravi for an incredible discussion.

If you’re enjoying this, I do a daily newsletter for Axios. You can go to axios.com slash newsletters and you’ll be hearing from folks like this in your inbox every day. Thank you.

Thank you. Thank you, Bipol, Richard, Ioannis, Bipol, Richard, Ioannis, Bipol, Richard, Ioannis, Bipol, Richard, Ioannis, Bipol, Richard, Ioannis, Bipol, Richard, Ioannis, Bipol, Richard, Ioannis, Bipol, Richard, Ioannis, Bipol, Richard, Ioannis, Bipol, Richard, Ioannis, Bipol, Richard, Ioannis, Bipol, Richard, Ioannis.

Yutong Zhang

Speech speed

150 words per minute

Speech length

873 words

Speech time

347 seconds

Current interactions require remembering Excel formulas, clicking hundreds of buttons in PowerPoint, and navigating complex GUIs, while future interactions will be natural language-based

Major discussion point

Software Evolution and User Interface Changes

Topics

Digital business models | Digital standards

Agreed with

– Ravi Mhatre
– Richard Socher

Agreed on

AI will fundamentally transform software development and user interfaces

Disagreed with

– Ravi Mhatre
– Richard Socher

Disagreed on

Software replacement vs. software persistence

Long-term agency will become commonplace, with AI systems able to work autonomously for extended periods like human employees

Explanation

Zhang predicts that AI systems will develop the ability to work independently for long periods, similar to human employees who can take on tasks in weekly meetings and work autonomously until the next check-in. This represents a significant advancement in AI capability.

Evidence

She compares this to working with her team where they have weekly meetings, set goals, discuss tasks roughly, and team members work autonomously until the next meeting

Major discussion point

Future AI Capabilities and Predictions

Topics

Future of work | Digital business models

Agreed with

– Ioannis Antonoglou

Agreed on

Long-term autonomous AI agency will become commonplace

Ina Turpen Fried

Speech speed

116 words per minute

Speech length

1234 words

Speech time

634 seconds

Competition is no longer between existing competitors but between established companies and startups that begin with no legacy workflows or processes

Explanation

Fried argues that the competitive landscape has fundamentally shifted in the AI era. Traditional competition between similar companies is being replaced by competition between established organizations with legacy systems and new startups that can build from scratch without existing workflows or processes constraining their AI implementation.

Major discussion point

AI-Native Companies vs. Legacy Organizations

Topics

Digital business models | Future of work

Ioannis Antonoglou

Speech speed

184 words per minute

Speech length

720 words

Speech time

233 seconds

ChatGPT’s evolution from 2022 when it had many hallucinations and users were skeptical, to now when people generally trust its responses after seeing it be correct repeatedly

Major discussion point

Trust and Compliance in AI Adoption

Topics

Privacy and data protection | Digital business models

Agreed with

– Yutong Zhang

Agreed on

Trust in AI systems can be built through technological transparency and repeated positive experiences

Multi-agent system dynamics and optimization will become critical as agent proliferation increases

Explanation

Antonoglou identifies multi-agent systems as a key challenge, noting that the dynamics of systems with both humans and many agents interacting are not well understood. There’s insufficient scientific understanding of how to optimize models to operate better in these complex multi-agent environments.

Evidence

There is some early research on multi-agent systems, but it’s not mature enough for the proliferation of agents expected

Major discussion point

Future AI Capabilities and Predictions

Topics

Digital business models | Digital standards

Agreed with

– Yutong Zhang

Agreed on

Long-term autonomous AI agency will become commonplace

Richard Socher

Speech speed

169 words per minute

Speech length

801 words

Speech time

283 seconds

The biggest companies in 20 years will be those starting now as AI-native organizations

Explanation

Socher predicts that the most successful companies two decades from now will be those that are being founded today with AI-native approaches from the beginning. This suggests that starting fresh with AI capabilities provides a significant long-term competitive advantage over trying to retrofit existing organizations.

Major discussion point

AI-Native Companies vs. Legacy Organizations

Topics

Digital business models | Future of work

Companies need benchmark and evaluation set mindsets to create scientific processes for evaluating AI models rather than ad-hoc testing

Explanation

Socher argues that successful AI adoption requires companies to develop systematic approaches to testing AI performance. Instead of informal testing like ‘I tried these three things and it worked,’ companies need to identify correct inputs and outputs and create principled, scientific evaluation processes.

Evidence

His company U.com has more agents than people and wins when working with companies that have proper benchmarks because they have the best models, answers, accuracy, and citations

Major discussion point

Trust and Compliance in AI Adoption

Topics

Digital standards | Digital business models

Some software applications with strong network effects and complex user bases will still require traditional development approaches

Explanation

While agreeing that much software could become automated, Socher argues that certain applications like Instagram will still need traditional development because they require well-designed user experiences and depend on network effects and user-generated content that can’t be easily replicated through automated coding.

Evidence

Instagram as an example – you could hack up an Instagram clone, but you wouldn’t have the network of users contributing content

Major discussion point

Software Evolution and User Interface Changes

Topics

Digital business models | Network security

Agreed with

– Yutong Zhang
– Ravi Mhatre

Agreed on

AI will fundamentally transform software development and user interfaces

Disagreed with

– Yutong Zhang
– Ravi Mhatre

Disagreed on

Software replacement vs. software persistence

AI will build the best AI systems, not people, representing a shift toward automated model development

Explanation

Socher predicts that the manual process of building AI models will itself become automated. He compares this to how individual expert teams were replaced by neural networks, suggesting that the current system of highly paid experts manually creating models over months will be replaced by recursive self-improving AI systems.

Evidence

Historical analogy of IBM Watson’s Jeopardy team with 800 specialists being replaced by a single neural network that could learn all domains

Major discussion point

Future AI Capabilities and Predictions

Topics

Digital business models | Digital standards

Bipul Sinha

Speech speed

145 words per minute

Speech length

Digital business models | Future of work

Ravi Mhatre

Speech speed

116 words per minute

Speech length

773 words

Speech time

398 seconds

AI-first companies need trust-based technologies embedded in AI systems to enable faster adoption by traditional enterprises

Explanation

Mhatre argues that traditional enterprises are held back by a compliance mindset that slows AI adoption. What’s needed are trust-based technologies built into AI systems that would allow business leaders to scale AI confidently, knowing that safety is inherent rather than requiring conservative, slow approaches.

Major discussion point

Workforce Transformation and Human-AI Collaboration

Topics

Privacy and data protection | Digital business models

The cost of AI inference has decreased 100-1000x in the past year, making intelligence-driven automation economically viable for many use cases

Explanation

Mhatre highlights the exponential improvement in AI economics, with inference costs dropping dramatically. This cost reduction enables intelligence to make software disposable, where robust, production-ready software can be generated on-demand for any automation task someone can imagine.

Evidence

On average, cost per token for inference went down 100x in the last year, with many use cases seeing 1000x cost reductions

Major discussion point

Software Evolution and User Interface Changes

Topics

Digital business models | Economic

Agreed with

– Yutong Zhang
– Richard Socher

Agreed on

AI will fundamentally transform software development and user interfaces

AI will expand automation into areas where software was never built, replacing manual processes and workflows

Explanation

Mhatre argues that the exponential improvement in AI capabilities will primarily impact areas where traditional software development never occurred, automating manual processes and workflows that were previously too complex or niche for software solutions.

Major discussion point

Software Evolution and User Interface Changes

Topics

Future of work | Digital business models

Agreed with

– Yutong Zhang
– Richard Socher

Agreed on

AI will fundamentally transform software development and user interfaces

Disagreed with

– Yutong Zhang
– Richard Socher

Disagreed on

Software replacement vs. software persistence

Continuous learning capabilities will emerge, allowing models to adapt and update their knowledge through environmental interaction

Explanation

Mhatre predicts that we’ll see the first signs of more advanced AI capabilities, specifically the ability for models to continuously learn and adapt based on their interactions with their environment, updating their core knowledge set rather than remaining static after training.

Major discussion point

Future AI Capabilities and Predictions

Topics

Digital business models | Digital standards

Agreements

Agreement points

AI will fundamentally transform software development and user interfaces

Speakers

– Yutong Zhang
– Ravi Mhatre
– Richard Socher

Arguments

Software will become invisible as humans use natural language to access all software capabilities through AI agents rather than traditional GUI interactions

The cost of AI inference has decreased 100-1000x in the past year, making intelligence-driven automation economically viable for many use cases

AI will expand automation into areas where software was never built, replacing manual processes and workflows

Some software applications with strong network effects and complex user bases will still require traditional development approaches

Summary

All three speakers agree that AI will dramatically change how software is created and used, with natural language interfaces replacing traditional GUIs and intelligence-driven automation becoming economically viable, though some applications may still require traditional development

Topics

Digital business models | Future of work | Digital standards

Trust in AI systems can be built through technological transparency and repeated positive experiences

Speakers

– Yutong Zhang
– Ioannis Antonoglou

Arguments

Trust can be built through technology by showing citations, references, and chain-of-thought reasoning to make AI decision-making transparent

Trust develops over time as AI systems prove their reliability, similar to how people gradually trusted computers

Summary

Both speakers agree that trust in AI systems develops through technological solutions that provide transparency (citations, reasoning chains) and through repeated positive interactions over time

Topics

Privacy and data protection | Digital business models

Humans and AI will work together rather than AI completely replacing humans

Speakers

– Yutong Zhang
– Ioannis Antonoglou

Arguments

Individual employees are willing to pay out of their own pocket for AI tools that enhance their productivity at work

The future involves rethinking workflows from first principles to maximize human efficiency while using AI where most effective

Humans will remain essential in the foreseeable future, working closely with AI to increase productivity rather than being completely replaced

Summary

Both speakers envision a collaborative future where AI enhances human productivity rather than replacing humans entirely, with workflows redesigned to optimize the human-AI partnership

Topics

Future of work | Digital business models

Long-term autonomous AI agency will become commonplace

Speakers

– Yutong Zhang
– Ioannis Antonoglou

Arguments

Long-term agency will become commonplace, with AI systems able to work autonomously for extended periods like human employees

Multi-agent system dynamics and optimization will become critical as agent proliferation increases

Summary

Both speakers predict that AI systems will develop the ability to work independently for extended periods, though this will create new challenges in managing complex multi-agent systems

Topics

Future of work | Digital business models | Digital standards

Similar viewpoints

Both speakers identify compliance and trust as major barriers to enterprise AI adoption, with traditional companies being held back by risk concerns that could be addressed through better trust-based technologies and compliance frameworks

Speakers

– Bipul Sinha
– Ravi Mhatre

Arguments

American companies face significant compliance and governance restrictions that limit AI data usage, requiring certified models and compliance infrastructure

Companies struggle to move from successful AI pilots to production due to compliance risks and concerns about hallucination and security threats

AI-first companies need trust-based technologies embedded in AI systems to enable faster adoption by traditional enterprises

Topics

Privacy and data protection | Data governance | Cybersecurity

Both speakers predict that AI development itself will become automated, with systems capable of self-improvement and continuous learning rather than relying on manual human development processes

Speakers

– Richard Socher
– Ravi Mhatre

Arguments

AI will build the best AI systems, not people, representing a shift toward automated model development

Continuous learning capabilities will emerge, allowing models to adapt and update their knowledge through environmental interaction

Topics

Digital business models | Digital standards

Both speakers agree that the competitive landscape has fundamentally shifted, with AI-native startups having significant advantages over legacy organizations constrained by existing processes

Speakers

– Ina Turpen Fried
– Richard Socher

Arguments

Competition is no longer between existing competitors but between established companies and startups that begin with no legacy workflows or processes

The biggest companies in 20 years will be those starting now as AI-native organizations

Topics

Digital business models | Future of work

Unexpected consensus

AI systems will achieve human-like autonomous work capabilities

Speakers

– Yutong Zhang
– Ioannis Antonoglou
– Richard Socher

Arguments

Long-term agency will become commonplace, with AI systems able to work autonomously for extended periods like human employees

Multi-agent system dynamics and optimization will become critical as agent proliferation increases

AI will build the best AI systems, not people, representing a shift toward automated model development

Explanation

It’s unexpected that all speakers, despite representing different companies and perspectives, unanimously agree that AI will soon achieve human-like autonomous work capabilities. This represents a significant consensus on a very ambitious prediction about AI advancement

Topics

Future of work | Digital business models | Digital standards

Traditional software development paradigms will be fundamentally disrupted

Speakers

– Yutong Zhang
– Ravi Mhatre
– Richard Socher
– Bipul Sinha

Arguments

Software will become invisible as humans use natural language to access all software capabilities through AI agents rather than traditional GUI interactions

AI will expand automation into areas where software was never built, replacing manual processes and workflows

The cost of AI inference has decreased 100-1000x in the past year, making intelligence-driven automation economically viable for many use cases

Reinforcement learning will be better understood for adapting generic models to specific business workflows and processes

Explanation

The unanimous agreement that traditional software development will be fundamentally disrupted is unexpected given that these speakers represent companies that currently operate within existing software paradigms. Their consensus suggests this transformation may happen faster than typically anticipated

Topics

Digital business models | Digital standards | Future of work

Overall assessment

Summary

The speakers demonstrate remarkable consensus on several transformative predictions: AI will fundamentally change software development and user interfaces, trust can be built through technological transparency, humans and AI will collaborate rather than compete, and AI systems will achieve long-term autonomous capabilities. There’s also strong agreement that compliance and trust issues are major barriers to enterprise adoption, and that AI-native companies have significant competitive advantages over legacy organizations.

Consensus level

High level of consensus with significant implications – the speakers, despite representing different companies and roles, largely agree on the direction and timeline of AI transformation. This suggests these predictions may be more certain and imminent than typically assumed, particularly regarding the disruption of traditional software development and the emergence of autonomous AI agents. The consensus also indicates that current barriers (compliance, trust) are well-understood and likely to be addressed soon.

Differences

Different viewpoints

Software replacement vs. software persistence

Speakers

– Yutong Zhang
– Ravi Mhatre
– Richard Socher

Arguments

Software will become invisible as humans use natural language to access all software capabilities through AI agents rather than traditional GUI interactions

AI will expand automation into areas where software was never built, replacing manual processes and workflows

Some software applications with strong network effects and complex user bases will still require traditional development approaches

Summary

Zhang and Mhatre argue that software will become largely invisible or disposable through AI automation, while Socher maintains that certain applications with network effects and complex user bases will still require traditional development approaches.

Topics

Digital business models | Future of work | Network security

Unexpected differences

AI model development automation timeline

Speakers

– Richard Socher
– Bipul Sinha

Arguments

AI will build the best AI systems, not people, representing a shift toward automated model development

Reinforcement learning will be better understood for adapting generic models to specific business workflows and processes

Explanation

While both discuss future AI capabilities, Socher predicts complete automation of AI model development by the end of the year, while Sinha focuses on more incremental improvements in reinforcement learning for business applications. This represents different views on the pace and nature of AI advancement.

Topics

Digital business models | Digital standards

Overall assessment

Summary

The speakers show remarkable consensus on the transformative potential of AI but differ significantly on implementation approaches, timelines, and the extent of change. Main disagreements center on software evolution, trust-building mechanisms, and the pace of AI advancement.

Disagreement level

Low to moderate disagreement level. While speakers share similar visions of AI transformation, their different perspectives on implementation paths and timelines could lead to significantly different strategic decisions for businesses and policymakers. The disagreements are more about approach and pace rather than fundamental opposition to AI adoption.

Partial agreements

Similar viewpoints

Both speakers identify compliance and trust as major barriers to enterprise AI adoption, with traditional companies being held back by risk concerns that could be addressed through better trust-based technologies and compliance frameworks

Speakers

– Bipul Sinha
– Ravi Mhatre

Arguments

American companies face significant compliance and governance restrictions that limit AI data usage, requiring certified models and compliance infrastructure

Companies struggle to move from successful AI pilots to production due to compliance risks and concerns about hallucination and security threats

AI-first companies need trust-based technologies embedded in AI systems to enable faster adoption by traditional enterprises

Topics

Privacy and data protection | Data governance | Cybersecurity

Both speakers predict that AI development itself will become automated, with systems capable of self-improvement and continuous learning rather than relying on manual human development processes

Speakers

– Richard Socher
– Ravi Mhatre

Arguments

AI will build the best AI systems, not people, representing a shift toward automated model development

Continuous learning capabilities will emerge, allowing models to adapt and update their knowledge through environmental interaction

Topics

Digital business models | Digital standards

Both speakers agree that the competitive landscape has fundamentally shifted, with AI-native startups having significant advantages over legacy organizations constrained by existing processes

Speakers

– Ina Turpen Fried
– Richard Socher

Arguments

Competition is no longer between existing competitors but between established companies and startups that begin with no legacy workflows or processes

The biggest companies in 20 years will be those starting now as AI-native organizations

Topics

Digital business models | Future of work

Takeaways

Key takeaways

AI-native companies are operating with unprecedented efficiency ratios, with some having hundreds of agents supporting fewer than 10 people, fundamentally changing competitive dynamics

The primary barrier to AI adoption in enterprises is compliance and governance concerns rather than technical limitations, with companies struggling to move from successful pilots to production

Trust in AI systems is built through transparency mechanisms like citations, chain-of-thought reasoning, and showing data sources, similar to how trust in computers developed over time

Software interfaces will become invisible as natural language becomes the primary way humans interact with all digital tools and capabilities through AI agents

The cost of AI inference has dropped 100-1000x in the past year, making intelligence-driven automation economically viable for widespread use

Future workforce will involve human-AI collaboration rather than replacement, requiring companies to rethink workflows from first principles to optimize both human and AI capabilities

Multi-agent system dynamics and orchestration will become critical challenges as AI systems proliferate in business environments

Resolutions and action items

None identified – this was a panel discussion focused on sharing insights and predictions rather than making decisions or assigning tasks

Unresolved issues

How to effectively translate business compliance rules into actionable guidelines for probabilistic AI systems

The science and optimization methods for multi-agent systems where humans and AI agents interact dynamically

How traditional enterprises can develop the expertise needed to create sophisticated evaluation benchmarks for AI systems

The balance between helping large established businesses adopt AI versus supporting new AI-native startups as customers

How to manage the transition period where some software applications will remain traditional while others become AI-automated

The implications and timeline for recursive self-improving AI systems that can build better AI without human intervention

Suggested compromises

Using LLM-based judgment models to evaluate other LLMs based on company-specific rules rather than trying to translate business rules into fixed constraints

Implementing ‘bring your own AI to work’ policies within certified compliance frameworks that balance innovation with risk management

Creating partnerships between traditional companies (with industry expertise) and AI companies (with technical expertise) to combine domain knowledge with AI capabilities

This prediction about recursive self-improvement represents a fundamental shift in how AI development itself will evolve. It suggests we’re approaching a inflection point where the manual process of AI development becomes automated.

Impact

This comment served as a powerful conclusion that tied together earlier themes about exponential improvement and automation, suggesting that even the most sophisticated human work (AI research) will be transformed by AI itself.

Overall assessment

Explanation

This is fundamental to the future of business automation, as it involves creating systems that can understand high-level objectives and automatically create the necessary workflows to achieve them.

Disclaimer: This is not an official session record. DiploAI generates these resources from audiovisual recordings, and they are presented as-is, including potential errors. Due to logistical challenges, such as discrepancies in audio/video or transcripts, names may be misspelled. We strive for accuracy to the best of our ability.