One of the biggest talking points coming out of AI4 wasn’t a product announcement or benchmark. It was the keynote discussion between three of the most influential voices in AI: Andrew Ng, Geoffrey Hinton, and Fei-Fei Li. 

Their differing perspectives generated a lot of conversation online. They challenged each other on AI’s impact on jobs, regulation, education, and the future of technology itself.

What stood out to me wasn’t who was right. It was how their disagreements revealed the questions the AI industry is still trying to answer.

Humans Still Have the Context Advantage

As the conversation shifted from AI capabilities to education, reskilling, and the future of work, the moderator asked a practical question: 

How should we prepare people for a future where AI becomes part of nearly every job?

While Hinton warned that “routine intellectual labor is going to be done by AI,” Ng added that “humans for a long time will have a fundamental context advantage.”

At first glance, those ideas may seem contradictory. They’re not.

AI is becoming remarkably capable at routine cognitive work, but capability isn’t the same as context.

Ng explained that humans possess something today’s AI systems still lack: years of accumulated contextual knowledge. We understand our organizations, customers, business priorities, relationships, constraints, and the countless unwritten rules that shape good decisions. We know when an idea looks plausible but would never work in practice.

A model may know more facts than anyone in the room, but it still doesn’t know your business, and that distinction matters.

The future of enterprise AI won’t be determined solely by which model is the smartest. It will be determined by which systems deliver the right context at the right moment.

The Future of Work Is More Complicated Than “AI Takes Jobs”

The panel also pushed back on one of AI’s biggest headlines: that jobs will simply disappear.

The discussion wasn’t about whether AI will change work. Everyone agreed it will. The real debate is how.

Ng emphasized that jobs are changing, developers are broadening their scope, and tasks are being automated. He described how developers are already using AI to automate portions of their work and expand into broader responsibilities. Instead of specializing narrowly, they’re increasingly becoming full-stack builders capable of owning larger portions of a product. He has made this case publicly before, trading the “jobpocalypse” narrative for what he calls an “AI jobapalooza.”

Hinton’s view is that people who worry about jobs are worrying for good reason. He pointed to routine intellectual work, the kind done in call centers, as the clear near-term target. His analogy was manual labor after the mechanical digger arrived. The machine did the digging better, workers moved to other manual jobs, and offices and services absorbed them. His worry is that once AI handles routine intellectual labor, there is no obvious next place for those workers to go.

Li reframed the whole thing in layers. Every existing job is made of many tasks. Some tasks get automated, some stay untouched, and some get more of your time because the tedious parts sped up. On top of that sit three categories: jobs that transform, jobs that get created, and jobs that get genuinely displaced. Rather than asking whether AI replaces jobs, she encouraged the audience to think about how it changes the work inside those jobs.

All three perspectives can be true. Some tasks will disappear, while others will become dramatically faster, and entirely new responsibilities will emerge.

A developer may spend less time writing boilerplate code and more time designing systems. A marketer may spend less time producing first drafts and more time shaping strategy. A customer service representative may handle fewer routine requests and more complex customer conversations.

The organizations that benefit most won’t simply automate work. They’ll redesign work.

Productivity isn’t the Same as Prosperity

One of the strongest moments of the keynote came from Li: “Increased productivity does not translate to shared prosperity.”

AI is already making individuals and organizations more productive. Hinton even described a future where AI could drive “massive increases in productivity.”

But productivity is only one measure of success.

If AI allows one employee to complete five times as much work, what happens next?

Do organizations create new opportunities? Do employees take on broader responsibilities? Do customers receive better service? Or do companies simply reduce headcount?

Those are business decisions, not technological ones. The future of AI won’t be defined only by what the technology can do. It will also be shaped by how organizations choose to apply it. 

The point that stayed with me is that productivity and prosperity are not the same word. AI reliably raises productivity. Whether that turns into shared prosperity is a choice we make, not a result we are handed.

The Best AI Systems Increase Human Agency

The keynote repeatedly returned to education, but the lessons apply far beyond the classroom.

Li reminded the audience: “The core of learning is agency and motivation.”

Hinton offered an optimistic vision of AI tutors helping students learn more effectively through individualized instruction.

Despite approaching the topic from different perspectives, both speakers arrived at a similar conclusion: 

The goal isn’t to remove humans from the process. It’s to help people become more capable.

That’s true whether you’re teaching students, helping developers build software, enabling customer support teams, or assisting knowledge workers.

The best AI systems don’t replace human judgment. They amplify it.

Security Is More Nuanced Than “Open vs. Closed”

Another part of the keynote that caught my attention was the discussion around open source. It was one of the most nuanced parts of the keynote, and it was a good reminder that our industry often oversimplifies complex issues.

Hinton began by making an important distinction that is often lost in AI conversations: open source is not the same as open weights.

Open source means publishing the code so others can inspect and improve it. Open weights means releasing the trained model itself, making it easy for anyone to download, fine-tune, and adapt.

Hinton explained that he opposed releasing open weights for years because it dramatically lowered the cost of repurposing powerful models for harmful uses. But he also acknowledged that the landscape has changed. The capable open-weight models are already here. As he put it, the barrier he once worried about has largely disappeared.

Ng approached the question from a different angle. Rather than focusing on risk, he focused on competitiveness.

He argued that open models such as DeepSeek and Moonshot’s Kimi K3 are already gaining global adoption. That matters not only economically but strategically. AI models influence how people access information, solve problems, and increasingly interact with the digital world. In that sense, widespread adoption becomes a form of soft power.

His concern wasn’t that openness is inherently risky; it was that fear-driven policy could leave American open models unable to compete globally.

Then, Li argued that treating openness as a binary choice is itself the mistake.

Drawing on examples from nuclear physics and the Human Genome Project, she pointed out that society has long balanced openness and control. Scientific research can remain open while sensitive materials, infrastructure, or applications are governed differently.

Her message was simple: stop asking whether AI should be open or closed. Instead, ask what should be open, for whom, and at which layer of the technology stack.

That framing resonated with me because it’s the same kind of thinking we need across AI infrastructure.

Enterprise AI isn’t built from a single component. Models, retrieval systems, business logic, data, evaluation, and applications all have different requirements for openness, governance, and control.

My biggest takeaway here wasn’t that one speaker “won” this debate. It was that “open versus closed” is the wrong question.

The better question is: Open at which layer?

Context Is the Real Infrastructure Challenge

Listening to the discussion, I kept coming back to Ng’s observation about context. It also reinforces something I’ve seen repeatedly in enterprise AI:

Models often don’t fail because they can’t reason. They fail because they lack the context needed to reason well.

Ng described using AI to brainstorm ideas. Some suggestions were genuinely useful, while others were obviously wrong. The model generated all of them, but the human knew immediately which ones didn’t fit the situation.

That’s the difference between intelligence and understanding. 

For enterprise AI, the next challenge isn’t simply making models more intelligent. It’s giving them the context they need to make intelligent decisions.

A model can’t reliably use information that it can’t access. It can’t distinguish fresh information from stale information unless the system tells it. It can’t understand business rules, user permissions, customer history, or organizational priorities unless those signals are part of the retrieval process.

Without access to current data, business logic, organizational knowledge, and user-specific information, even the most capable model can produce an answer that is fluent, convincing (and completely wrong).

That makes context an infrastructure problem. The model may generate the response, but the system determines whether the model has the right information to generate the right response.

It’s why retrieval, ranking, freshness, personalization, and grounding matter so much.

This is exactly the problem platforms like Vespa.ai are designed to solve.

Rather than treating retrieval, ranking, inference, and serving as separate systems stitched together with multiple network hops, Vespa.ai brings them into a single query pipeline that runs computation where the data lives. AI applications can retrieve fresh information, combine lexical, semantic, and structured signals, apply business-specific ranking logic, personalize results, and perform machine-learned inference in real time. And all of this happens before the model generates a response.

As AI moves from answering questions to making decisions and taking actions, delivering the right context, not just more context, becomes the competitive advantage.

My Biggest Takeaway

The Ai4 keynote didn’t provide definitive answers about the future of AI. Instead, it showed why thoughtful disagreement is valuable.

AI may create enormous abundance, but it may also create significant disruption, and it will almost certainly transform how we work.

But the lesson I’ll remember most wasn’t about jobs, regulation, or model capabilities. It was Ng’s observation that humans retain a context advantage.

As foundation models become increasingly capable, competitive advantage won’t come from simply using AI. It will come from building systems that connect AI to the right information, the right business logic, the right user signals, and the right moment.

The model may be smart, but the system that understands your business will be the one that creates value.