March 18, 2025

ikayaniaamirshahzad@gmail.com

Is AI a Bubble or a Revolution? Human[X] Asks: Why Not Both?


Editor’s note: The author moderated a panel at Human[X].

LAS VEGAS — The big takeaways from the first-ever Human[X] conference: AI agents are everywhere, and need orchestration and governance. Models are improving rapidly — but trust is a work in progress.

And man, there’s a lot of money sloshing around in the AI space these days: the industry grew to global market size of $184 billion in 2024, up from about $134 billion the previous year, according to Statista figures.

The Human[X]conference, founded by Stefan Weitz and Jonathan Weiner, is dedicated to showcasing how organizations are using AI — and helping investors and business decision-makers learn more about this fast-growing industry.

Human[X] — launched last spring with an initial $6 million investment by VCs like Primary Venture Partners, Foundation Capital, FPV Ventures and Andreessen Horowitz — drew more than 6,500 people registered for the initial outing in Las Vegas.

In his welcome speech last Monday, Weitz told the crowd why he and Weiner started Human[X]: to focus on real-world implementation and its attendant benefits and challenges.

“The conversation is, frustratingly to me, binary,” Weitz said. “It’s either utopia or dystopia. It’s either a benevolent robot overlord or it’s going to be Skynet with a LinkedIn profile. There is no apparent middle ground between Utopia and Judgment Day.

“And reality, as we all know, is more complex. So that’s why we try to build Human[X], not to feed the hype, not to create more fear, but to engage in conversations about what’s actually happening, what’s worked, what hasn’t and what we do next.”

On Wednesday, the conference unveiled the results of a survey of more than 1,000 U.S. business leaders, conducted by the pollster HarrisX. Seventy-five percent of respondents said their organization has a dedicated AI strategy.

Most leaders surveyed said they are spending between 10% and 25% of their budgets on AI initiatives; 37% of survey participants said they expect their AI investments to grow significantly over the next three years,

Meanwhile, the field’s biggest players keep churning out new advancements. On Tuesday, OpenAI unveiled its new AI agent framework; on Wednesday, Google dropped Gemma 3, a collection of lightweight, open models.

In his opening address on Monday, Weitz addressed the question of whether the AI boom was a bubble or the opening shots of a coming revolution in the way we work and live.

While he acknowledged that “signs of a bubble are trending,” he noted that previous bubbles have nevertheless resulted in long-lasting societal change; for instance, in the early 1900s, roughly 2,000 companies made cars. While nearly all of those makers failed, cars still transformed society. So, too, did the digital startups of the 1990s, most of which folded.

The AI boom might be “a little frothy,” Weitz said, but it’s still a potential revolution.

However, he cautioned: “Hype is dictating decisions, and the race to not get left behind is pushing companies and governments to move fast, whether or not they actually understand what they’re building. So the problem isn’t just that AI is overhyped, it’s that the hype itself is making us irresponsible. It’s kind of like the Fyre Festival with better algorithms.”

The next Human[X], slated for April 7-9, 2026, will be held in San Francisco, Weitz announced Tuesday — an acknowledgment of how Bay Area-centric the industry and its funders are.

AI Agents and Governance

Because Human[X] attracted lots of investors and business leaders, it made for a noticeably more extroverted crowd than the usual at tech events. For example, tablemates at Human[X] actually asked each other questions at lunch and breakfast, rather than remaining buried in their phones.

Through those conversations, along with the sessions, some key themes emerged. Among them, Agentic AI’s current hotness is presenting governance challenges.

There’s a reason why agents are on the rise, and it’s got at least as much to do with profit as productivity, Yash Sheth, co-founder and COO of Galileo, a generative AI evaluation company, told The New Stack.

The interest in AI software “has been in an assistive manner so far,” Sheth said. “But RAG and chatbots generating documents and briefs still have the human in the loop. The true [return on investment] from AI will be only through automation, that you can automate massive work streams and transform your backend processes to be more efficient.

Yash Sheth, COO and co-founder of Galileo

Yash Sheth, of Galileo, said it’s no mystery why AI agents have taken off: “That’s the true ROI of AI.”

“You can have multiple businesses interact with each other and really perform complex actions on their own. That’s the true ROI of AI.”

In essence, AI agents infuse AI into robotic process automation, Sheth said, “to automate some of the hard-coded processes and make it more robust.”

He added, “What AI is bringing to the table is a generalization of rules in that automation process. So I think fundamentally, if you understand, why are people so crazy about agents, it’s because that’s going to accomplish tasks end to end. “

Governing all those AI agents is a big task — and an opportunity for vendors. On Monday, Boomi, which specializes in integration as a service, announced a beta trial of its AI Studio platform, which it plans to move into general availability in May.

Boomi’s customers have deployed more than 25,000 agents, Mani Gill, the company’s vice president of product, told The New Stack. “We got our customers thinking of agents and using agents, and naturally they’re like, Hey, how do I better understand what data these agents have access to?’

Mani Gill of Boomi

Mani Gill, of Boomi, said Ai agents will sprawl, just as apps, APIs and data have, in modern enterprise systems.

The company also saw a pattern emerging: Just as applications, data and APIs sprawled, so would AI agents. “So we started talking to our customers about, ‘Hey, as you’re thinking about your agentic journey, would this be a value to you to be able to manage across all of these agents? And the concept is very similar to API management where I’ve got all these APIs. How do I understand them across my landscape?”

He added, “We led our customers there a little bit, but it also is unfolding in front of them.”

Boomi AI Studio provides a platform for AI agent design, governance and orchestration. There are four components:

  • Agent Designer: Lets users create and deploy AI agents using Generative AI prompts, through no-code templates, using trusted data and security guardrails.
  • Agent Control Tower: The centerpiece of AI Studio, according to Gill, the control tower “provides governance, but also compliance and auditing,” with all that monitoring of both Boom and third-party AI agents in a central place.
  • Agent Garden: A space that allows users to interact with their AI agents using natural language. Design, testing, deployment and tool development are enabled in the Agent Garden. “They can learn, continuously learn and nurture that,” Gill said.
  • Agent Marketplace: “We’re working with our partners to use that design capability to create agents that then our customers can just use as templates.” The Agent Marketplace resides in Boomi Marketplace (formerly Boomi Discover).

Other players are also crowding into the AI governance space. Holistic AI, a Software as a Service company that offers end-to-end AI deployment management. Founded only five years ago, it’s been seeing 50 to 60% growth each year, fueled by customers like Unilever, Raj Patel, Holistic AI’s AI transformation lead, told The New Stack.

The Holistic AI platform, Patel said, includes “observability and evidence-based backing for your decisions as a business — whether you should deploy AI or not, and when you do deploy it, do you have the responsible AI guardrails, ethics, observability in place.”

The idea is to not only govern AI applications and agents but also determine if the applications and agents should be built in the first place.

“Data science teams cost hundreds of thousands of dollars, in order to build a team and spend six months testing and then deploying,” Patel said. “You want to know very early on if this is something that you want to explore and what are the mitigations that you need to put in place in order to make this happen.”

Governance, Patel said, is a gap waiting to be filled as more organizations take up generative AI.

“One of the key deficiencies in the market is they see governance as a checkbox exercise,” he said. “At the moment, it’s something that should be one and done. It’s really not like that anymore.

“If you want to be able to effectively deploy AI in your business, there is a continuum of checks that need to be done, and you need to have a system in place that supports an AI governance strategy that allows that.”

Making LLMs More Human

On Wednesday morning, Sean White, CEO of Inflection AI, a three-year-old company that specializes in training and tuning large language models for enterprises, spoke to the Human[X] audience about his company’s ongoing efforts to make LLM-based chatbots more conversational.

The company’s Pi.ai, a personal assistant chatbot used by 35 million people, began as a way to release its frontier models, a term used for cutting-edge models. White, formerly chief research and development director at the Mozilla Foundation, joined Inflection a year ago. “When I joined, a large part of the shift was taking all of that and seeing if we could then apply that to the enterprise.”

Sean White, CEO of Inflection AI.

Sean White, of Inflection AI, said a lot of languge models result in user experiences that “either start off where they are just book reports, or they actually are just not good to talk to.”

Pi.ai has been intentionally developed as an emotionally intelligent online assistant. It’s central to Inflection’s mission, White said.

“We really believe this is a new generation of user interfaces and experiences,” he told TNS. “It’s not just the computational system or the UX system — we don’t want to build a crappy user experience. A lot of these systems either start off where they are just book reports, or they actually are just not good to talk to.”

Inflection AI, he said, has put a lot of effort into making models less like “book reports” and more, well, human, fine-tuning for nuance and context.

“We have collected over 10 million of these examples of good conversation, of emotional intelligence. We have this very large dimensional space in which we kind of want the qualities of, is it keeping the conversation going? Is that utterance sarcastic?”

Alongside newer startups like Inflection AI at Human[X] were companies that began before the post-2022 explosion in demand for generative AI tools. Unbabel, started in 2013, doing machine translation. “We got a community of translators from all around the world that would post-edit this machine translation, Gil Coelho, head of product at Unbabel, told The New Stack.

And now, because machine translation has improved so substantially, Coelho said, “We have a generation of models, which we call power LLMs, and they’re state of the art right now for machine translation across most of the languages.”

Gil Coelho, of Unbabel.

Gil Coelho, of Unbabel, said his company wants developers to start building on top of Widn.ai, its recently open sourced translation technology.

In addition, Unbabel can perform quality estimation — using another AI model to predict the confidence the company has in a particular translation. ”So basically, one model does the translation, and then the other model will say, ‘Hey, I have high confidence,’ or ‘I have low confidence on the translation,’” Coelho said. “And if I have low confidence, I’m going to send this to a human” to check the translation.

A few months ago, Unbabel released open source versions of some of its proprietary, “secret sauce” translation technology, under the banner Widn.ai, and encourages developers to build on its components.

“That’s something that was a big shift in terms of our strategy,” said Coelho. “We just thought it made sense. We’ve been building these and we want to make it available to a lot more people, a lot more developers, a lot more builders, and not just keep it within the Unbabel platform.”

Moving Beyond ‘Prompt and Pray’

Ahead of Human[X], the eight-year-old company AI21 Labs unveiled Jamba 1.6, the latest iteration of its open LLM based on the hybrid transformer-Mamba-mixture of experts (MoE) architecture.

And in alignment with Human[X]’s emerging theme of AI orchestration, It introduced Maestro, an AI planning and orchestration system, on Tuesday.

The problem that Maestro is meant to address: To help overcome the issues of trust that hamper AI adoption in production at enterprises.

While consumer adoption of AI tools is rising, “in the enterprise, it’s a very different story,” Ori Goshen, AI21 co-founder and co-CEO, told The New Stack. “You see a lot of a lot of experimentation, a lot of these charismatic demos, very little workloads that actually go to production.”

While it could be that the enterprise market simply isn’t educated enough yet, Goshen said, “There’s a more fundamental issue here: to get these to work in mission-critical workflows, you have to build trust around those systems. They have to be robust.

“That’s kind of the basic piece. And we’ve been working with customers; we’re seeing their pain. It’s really painful to get something from a flashy demo to actual workflows that actually work in production.”

The current approach to building AI applications, Goshen said, is the underlying culprit. Typically a developer AI builder within the enterprise takes an agentic framework like a React, like Cloudchain, or CrewAI, or AutoGen or any of these, and then it uses a language model or a reasoning model, to figure out what the system is going to do.

“So it lets the language model basically plan and operationalize the workflow, which, again, works for demos but breaks in reality. We call this method ‘prompt and pray.’”

Ori Goshen, co-founder and co-CEO of AI21 Labs.

Ori Goshen of AI21 Labs: “It’s really early days. I think there are lots of questions of, how do you govern? How do you create more control?

Workarounds are possible, but not sustainable as the system grows and gets more complex. In their hard code, Goshen said, developers might “put clear checkpoints within, they call the LLM to get the dynamic part of the processing. That method indeed gives you more control, but it’s rigid and brittle and it’s hard to scale.”

Maestro is a model-agnostic system, Goshen said: “It learns the specific enterprise environment. So it learns the APIs, the tools, the data sources … it understands the environment, and then it’s training by doing offline simulation. And then, when a task is received by the system, it creates a structured plan that is explicit, so you can actually trace it”

The Maestro system is currently in private preview, Goshen said, with the expectation that it will roll out to general availability in Q2 of 2025.

As for the conference, Goshen cautioned against letting the hype get ahead of the reality engineering teams face.

“It’s really early days,” he said. “I think there are lots of questions of, how do you govern? How do you create more control? But I think the fundamental, the real fundamental part, is, how do we trust these systems?”


Group Created with Sketch.





Source link

Leave a Comment