I caught up recently with my friend and mentor Yacov Salomon, who has spent most of his career somewhere around the intersection of math, software, data science, and AI.
Yacov has a Ph.D. in mathematics, taught machine learning at UC Berkeley, spent much of his career building startups, and today is an EVP of Software Engineering at Salesforce leading the technology organization behind Service Cloud. His history with AI goes back to the late 1990s, when his high-school final project was a two-layer neural network for pathfinding around obstacles.
After the holiday weekend, the Jensen Huang and OpenAI story was suddenly everywhere. People were texting and emailing me about it, and it came up on every client call after the holiday weekend. Our board had also been encouraging me for a while to write more about AI, so I figured this was a great reason to call someone whose perspective I trust.
What I wanted from Yacov was not another prediction about which model wins. I wanted his perspective on a more basic question: if AI really is crossing into a new level of capability, what actually changes for the rest of us?
Jonny: Let's start with the obvious question. What is AGI?
Yacov: I don't think there is one exact mathematical definition that everyone agrees on.
There are Turing-style notions people use to try to draw the line. Others focus more on autonomy: whether the system can operate without a person directing each individual step. And some definitions include the social side of intelligence, like collaboration and forming relationships.
I think another useful way to look at it is as a progression. Traditional machine learning and data science were largely about pattern recognition. Early LLMs could still be understood that way too, only now the patterns were in language.
What has become interesting is the emergence of capabilities beyond that. These deep networks have an enormous number of degrees of freedom, and we are seeing models combine information and form genuinely new ideas in ways that are harder to describe as simple pattern recognition.
That is where the AGI conversation starts to become meaningful.
Jonny: So when Jensen says AGI has arrived, what does that actually mean to you? Do you care about the label?
Yacov: I don't see AGI as the end of the road at all. Astra is one point in an ongoing progression. There will be a GPT-7, a GPT-8, a GPT-9, and plenty of other models after those.
What feels different is that we've crossed an artificial line where the models are capable enough that it is reasonable to start classifying them differently. We can debate exactly where that line belongs, but I think the AGI declaration is more useful as a stamp of approval: the technology has reached a level where we should expect the world around it to change.
AGI is not the end of the road. It is a stamp of approval that the world is about to change.
Jonny: What do you think people get wrong when they try to imagine what that change will look like?
Yacov: I think we're still limited by what we already know how to imagine.
Take electricity. The first electric light replaced gaslight. Both were about light. Nobody saw electricity arrive and immediately thought about computers.
The internet was similar. Some of its earliest applications were basically mail moved online. We tend to take a fundamentally new technology and first use it to reproduce the old world.
AI may be at a similar stage. A lot of what we're doing today is taking work we already understand and adding AI to it. That can be incredibly useful, but it doesn't mean those are the applications we'll ultimately find most important.
Jonny: If you're running a large company today, how should that change the way you think about AI?
Yacov: One thing I would start thinking about is that agents are increasingly becoming users of systems.
If the workforce used to mean humans, maybe now you start thinking about a hybrid workforce: humans working alongside agents, and humans themselves increasingly augmented by personal agents.
Historically, business software was designed around a person on the other side of the interface. That assumption starts to change.
There are three interfaces I would look at:
- Agents and humans.
- Agents and systems.
- Agents and other agents.
Jonny: What does that look like in practice?
Yacov: Start with agents and humans. Inside a company, that means thinking about how agents participate in the places where people already work. Slack is a good example. If agents become part of the workforce, what should they know, when should they act, and when should they involve a person?
But the same question exists outside the company.
If you're a retail or consumer business, your customers are increasingly spending time researching and making decisions on new surfaces. That may be GPT, Claude, or whatever comes next. Businesses have to think about how they show up there in the same way they have historically thought about search, websites, and other channels.
Then there is the interface between agents and systems. If an agent is going to do useful work, it needs a way into the systems where the business actually runs. APIs become even more important. MCP is part of that same broader shift.
And security has to move with it. The same technology that gives companies more capable agents gives attackers a massive boost in finding and exploiting vulnerabilities. The more you make systems available to intelligent agents, the more seriously you have to think about access, permissions, and protecting your data.
Jonny: You also made an interesting point that infrastructure sometimes becomes more important, not less.
Yacov: When email arrived, people predicted that the postman would disappear. In one sense, they were right: we send far fewer letters than we used to.
But then e-commerce happened.
Suddenly FedEx, UPS, and the postal system became incredibly important because somebody was already regularly going to every home. Existing infrastructure became the conduit for something people had not originally imagined.
I think we should expect some of that with AI too. Existing systems and channels will not necessarily disappear. Their role may change.
Jonny: So what should companies actually be doing now?
Yacov: I would start with the basics.
- Think about how agents will interact with your people.
- Think about how agents will interact with your systems.
- Think about where your customers will increasingly discover and interact with your business.
- And think about the security and controls required to make all of that possible.
You do not need to predict the final form of AGI to start preparing for those changes.
My takeaway
The part of Yacov's perspective that stuck with me is that the AGI label itself is probably less important than what it tells us about the direction of travel.
For the companies we work with at Harmonyze, the immediate opportunity is not to add AI everywhere. It is to identify where a more capable system can actually understand the business, interact with the systems where the business runs, and help the people responsible for making decisions.
That is already a meaningful shift.
And if Yacov's lightbulb analogy is right, we should probably assume that many of the most important applications have not been invented yet.