Sam Altman: AI Will Become Infrastructure, Driving Societal Reconstruction | Vital Views

Oasis Capital has been emphasizing for the past two years that AI is not merely a technological shift, but a social revolution.

In a recent public conversation between Sam Altman and Vinod Khosla — partner at Khosla Ventures, an early investor in OpenAI — we see a similar assessment: behind the iterative advances in AI technology lies a social restructuring more rapid and more thorough than the Industrial Revolution.

When we talk about AI, we are not merely discussing the iteration of tools, but the leap of civilization.

Oasis Capital has compiled and shared excerpts from this conversation.

Enjoy.

Altman cuts straight to the point: between 2035 and 2050, the pace of technological evolution will defy imagination.

Fortune 500 turnover will happen faster than ever before — old giants dying from inertia, new enterprises rapidly scaling through AI, software generated on demand, established moats collapsing overnight.

The reordering of the commercial landscape will be the defining theme of this era.

Khosla: Imagine the world between 2035 and 2050. What do you think it will look like?

Altman: I do think that during this period, the speed of technological change will far exceed what we can currently comprehend. Making precise predictions is extremely difficult — will we have started building Dyson spheres? Will we have achieved nanobot applications? But I expect many breakthroughs of that magnitude will occur.

Perhaps the difference in day-to-day human experience won't be so dramatic — evolution moves slowly, and biology has shaped deep-seated drives over long stretches of history. Yet at the technological level, within the "technology stack," what individuals are capable of accomplishing will be fundamentally different from today.

Khosla: Taking this further, what happens to existing companies? I mentioned in a talk this morning that as we enter the 2030s, Fortune 500 companies may die off faster than at any point in history. Which survive and which get eliminated will depend largely on their choices and responses. But overall, the pace of elimination will accelerate significantly. Do you agree?

Altman: I'm not entirely sure. My intuition is that it will be somewhat faster, though I haven't thought deeply about it. Most of my time outside OpenAI has been spent on software companies, and I've always felt I more or less understood the "physics" of how software companies operate. If the future becomes one where any software you want can be generated instantly — you type a request into an AI chat interface and it immediately writes a great application — that means no need to buy a SaaS product; instead, you simply say "run" and the software appears. That shift may not be far off.

For Fortune 500 companies managing complex supply chains, though, physical-world change is always slower. But I'd bet most incumbent companies won't adapt quickly enough and will suffer major losses. My sense is that the growth speed of new companies and their rate of eating into existing markets is accelerating. It's like a massive exponential curve — new companies can grow faster and capture larger shares. OpenAI itself is an example of this rapid growth.

As AI doctors, AI accountants, and AI engineers become reality, the thing humans have been most proud of over centuries of evolution — intelligence — faces continual new challenges.

The definition of human professional roles begins to be questioned.

Altman acknowledges that AI will surpass humans in most knowledge-based positions, yet biological instincts mean people still crave companionship and motivation from "real others." Perhaps a mediocre human teacher still moves students more than a perfect AI tutor.

Care, motivation, and authentic emotion may become humanity's final moat.

Khosla: What jobs will still exist in that timeframe? Many entrepreneurs are already building AI doctors, AI therapists, AI oncologists — this is happening today. I suspect that whether these current companies succeed or future competitors do, by 2035 AI will have maturely covered at least 80% of knowledge work.

Altman: But there are many jobs that I think people won't want to hand over to AI, or that many people don't want AI to do. And entirely new roles will emerge that people prefer to have filled by humans.

I think we are biologically wired to care deeply about whether the other party is a person — this may be an ingrained instinct at the deepest biological level. For example, you might have an extremely capable AI teacher, but it may not motivate you the way a mediocre human teacher can. I fully believe this is possible. Simply knowing whether someone is a "real person" makes a profound difference.

Khosla: I may disagree — I think AI teachers could do more and understand you better.

Altman: I also think there will soon be AI investors better than you, no doubt. But personally, I still enjoy having dinner with you more. When you tell me "good job" or "you should do this," it motivates me more than a reminder from AI would.

Khosla: I completely agree, but undoubtedly my work, like most people's, is no safer than other professions — in practical terms, perhaps even less so.

Altman: In fact, I've been wondering: could we try building an AI venture capitalist? I've thought about this for a long time. At current model capabilities, it might already be possible — it could even be an interesting side project.

Khosla: That would be an interesting side project (laughs).

Though Sam, you just had a child. I do think people will have more time in the future — mothers won't have to return to work 16 weeks after giving birth; they'll have more freedom. We'll also have more time to care for elderly parents.

I agree on the importance of human relationships, but my view may differ in service areas like education and healthcare. What I'm really thinking is: what work is AI fundamentally incapable of? And in the near term, who is driving these changes?

Altman: To go deeper on the teacher question — objectively speaking, I probably learned more from reading Wikipedia than from any teacher. But when I look back at my actual learning journey, the most important moments were always tied to specific people: they connected with me, took interest in me, cared about me, understood me — you could feel this care. While AI can replicate this relationship to some degree, overall the process will be stranger, more complex, and more uneven than it appears on the surface.

AI may indeed be capable of almost every job, but ultimately we'll find that deep biological human instincts are extremely difficult to replace.

Khosla: I completely agree. In fact, I mentioned earlier today that biological instincts won't evolve away. Humans will still pursue status, still compete, attention economies will still have "thought leaders," we'll still care about children, parents, and family — we'll just have more time for these things.

The next trillion-dollar opportunity won't be another OpenAI.

New giants will be born in the blank spaces that emerge after AGI arrives, just as the transistor spawned the entire computer industry. Investors still chasing existing winners will only miss the rise of entirely new species.

From a capital perspective, the imperative is to embrace the "new possibilities" unlocked by AI.

Khosla: So if we look 18 months ahead to the end of 2026, how do you think AI capabilities will have changed compared to today? And compared to the two and a half years since ChatGPT emerged?

Altman: I'm not sure how to measure this change. If we're talking about "feel," I think going from 0 to the first version of ChatGPT was the biggest shock most people experienced: it simply didn't exist before, and suddenly it did. Though imperfect, the zero-to-one significance was enormous. Then we perhaps went from 1 to 10, which should theoretically be greater progress, but most people didn't feel it as intensely. Maybe the next 18 months will take us from 10 to 100. But I sense people have already accepted that AGI is coming; life goes on, people continue with their affairs.

At the macro level, the inputs haven't changed dramatically. We keep discovering better algorithms, so the results from Scaling Laws become increasingly steep; we keep building bigger computers, developing stronger chips, and connecting them; we keep finding more and higher-quality data. Next may come a new phase of combining these systems in novel ways to achieve some form of "continuous learning" — systems that run indefinitely and keep getting smarter.

But the through-line of recent years has really been: better algorithms, more compute, more data. I wish I could offer something more profound and unique, but that's what it is. Some algorithmic breakthroughs have indeed been remarkable — our progress on reasoning, and the original unsupervised learning idea, fall into this category. There are smaller breakthroughs too, but overall it's sustained accumulation at the intersection of research and industry, much like the transistor in history.

Khosla: Another question I'm thinking about: when will AI scientists begin to dominate most AI research?

Altman: I think this will be very gradual. Right now, OpenAI researchers using codecs might have 10% of PRs AI-generated, then gradually 20%, 30%, and eventually it starts testing new model architectures on its own.

But overall, researchers remain in the driver's seat — just more efficient. A researcher might say, "I still did 100% of the research, just with better tools." But if that researcher's output is now 2x, or even 10x what it was before, do you count that as AI doing 90% of the research, or 0%? AI isn't autonomously completing the entire research loop. I think this will be a hybrid acceleration process, difficult to measure with a single percentage.

What matters more is the pace of progress — the rate of advancement, the acceleration, is the ultimate metric. Every year going forward, we'll move faster in research because tools are better. Whether you say AI is assisting humans, humans are assisting AI, or AI is doing research on its own, the end result is: faster speed. Not just algorithmic progress, but the entire supply chain — if AI helps us build data centers faster, develop new chip designs faster (including some very cutting-edge ones), that counts too.

So as long as AI involvement makes overall progress faster, you can say it's doing research.

Khosla: To me, a key test is whether AI can propose hypotheses, test them itself, and revise them. If it can complete that loop, that would be a "virtuous madness" of development.

Altman: I have a slightly different perspective. What I care more about is whether research can get better faster. Whether the hypothesis comes from AI itself, or AI enables humans to propose hypotheses they otherwise couldn't, makes no difference to me.

Khosla: Yes, many LPs ask me: under this acceleration, won't existing leaders keep extending their advantage? If research truly enters an acceleration phase, then you and other top players can maintain leadership. This makes it harder for new entrants unless they find completely orthogonal approaches. As you know, we're trying some orthogonal paths ourselves. But does this mean OpenAI will gain greater valuation advantage? This is what many people wonder about.

Altman: If I were an LP, I would spend 0% of my time trying to invest in another AI research lab, and 100% of my time thinking about "what comes next." Generally, most global investment attention goes toward chasing the previous round's winners — and that almost never makes real money. The real returns almost always come from extremely early, unproven projects born from newly emerging possibilities, as long as you have unique insight into them.

So right now everyone wants to invest in the "next OpenAI," but I believe the next trillion-dollar company won't be another AGI research lab. It will more likely be something entirely new built using AGI as a novel technology. When OpenAI was starting out, most people wanted to invest in the "next Facebook" or some new cryptocurrency.

Khosla: Here's an interesting story: when we invested in OpenAI, it was the only time in 20 years I wrote an apology letter to all my LPs saying, "I know this looks strange, but we're doing it anyway."

Altman: I'll only tease you about that slightly, but it's fine. Anyway, I think this is an incredibly exciting period because the openness of new frontiers is unprecedented. Going back to when OpenAI was founded, breakthroughs then came more from research — the cryptocurrency market driving GPU availability, and some new developments. But the list of things truly worth doing was very short then.

Now, I think there are vastly more directions worth pursuing — that's what people should be investing in. Whether research returns ultimately flow to OpenAI or other institutions is no longer the most important question from a capital allocation standpoint. There will be near-free AGI in the future, and companies will create enormous value from it, but what truly matters is the new opportunities spawned by leveraging AGI.

Using the transistor analogy: back then, only a few companies were true "transistor companies," and most later disappeared, leaving only a handful of survivors. Today, transistors are in nearly every device around us, yet no one calls them "transistor companies." It became an underlying technology that enabled a whole new generation of companies, including OpenAI. I think we're at a similar moment now, incredibly exciting. The task of capital is to chase the future, not be trapped in the past.

Khosla: So going back to the ChatGPT launch — what surprised you most? And how did your thinking evolve based on user behavior feedback?

Altman: Let me rewind slightly. The traditional tech company path is to start as a product company, then add a research lab once the product matures. This worked in some cases, like Xerox PARC, and failed in others. But OpenAI is the only counterexample I know of: we started as a research lab, then "clumsily" added a company later. Four and a half years in, we realized that constrained by scaling laws and capital requirements, we had to build a large company, and that required products.

We had GPT-3 at the time, and I kept pushing the team to find a product direction. But while GPT-3 was cool, it wasn't enough to support a truly usable product. Paul Graham had given advice: no matter what, build an API — something unexpected will always happen. So we made GPT-3 into an API and let the world try it. The world found exactly one truly profitable application: copywriting. Some companies quickly grew to billion-dollar valuations, but beyond that there were almost no success stories.

However, we also had something called "Playground" where people could test prompts and get feedback. It became an unexpected "hit." Though model quality was poor then (not even GPT-3.5), a small number of users would chat with it all day. Clearly, conversation was what users actually wanted, so we decided to try making the model more suitable for dialogue, no longer relying on complex prompts, and launched a chat interface.

There was actually intense internal debate at the time: if it's just casual chatting without clear learning or task objectives, won't users lose interest? This discussion nearly made us abandon the idea, but we ultimately launched this chat interface version, allowing users to chat freely.

Retention during testing was terrible — most users churned quickly — but the few who stayed showed continuously growing usage frequency. This was an important insight: as long as a product has even slight retention, it may have potential, even if only 5%, it could still grow. Most products default to this curve eventually hitting zero, and we didn't intuitively realize this at the time.

Khosla: I don't know what your currently disclosed ChatGPT user count is, but what's your vision for ChatGPT's future?

Altman: OpenAI's vision is to build a small, refined set of products and a platform that can integrate with other services, becoming users' default "personal AGI": a system that knows you, connects to your materials, and operates according to your preferences.

If you want to use it through a chat interface, you can; if you want to experience it in new ways through social or entertainment products, you can; if you want to use agents to accomplish large amounts of work, you can. You could even "bring your intelligence with you" into any other service.

In the future, this will expand to new services, and I think there will be new computing form factors too, which we hope users will have. But broadly, people will build very important relationships with AI that help them be more efficient, better, happier, and live more smoothly — and we want OpenAI to be that presence.

Khosla: So it's not just an on-call intelligent agent?

Altman: In some sense it is, but the greater value lies in creating excellent experiences and integrating them in different ways.

Khosla: And to expand to the next billion users — is it the same use cases, or does it expand to different scenarios?

Altman: I don't want to go through the product roadmap item by item, but I think we're still in the "terminal" phase, and I personally quite like terminal interaction. However, as we build the next generation of computer interfaces (not necessarily changes in computer appearance, but a similar leap), AI will become more accessible, more powerful, and more easily accepted by the majority of people.

As free and accessible AI capabilities accelerate into billions of people's daily lives, education, healthcare, and scientific research are being reshaped.

But emerging challenges of compute scarcity, wealth distribution, energy demands, and regulatory frameworks — AI may bring the strongest "deflationary dividend" in history, making basic resources accessible to all.

How to ensure fair distribution and effective governance will become the key test for future order.

Khosla: Before discussing global-level AI impact, I want to ask my favorite question, and also leave time for audience questions: when do you think we'll see a 10-person company with $1 billion in revenue? Has such a company already emerged? Will it happen soon? Or is it impossible? Of course I have my own biases.

Altman: I'd bet such a company has either already been born or will emerge in the next few years.

Khosla: I completely agree. I've also been guessing such a company may already exist. Thinking about how value is created and what it takes to create value is itself quite mind-bending.

Altman: This is also why I find the "AI for Science" direction so interesting. It's entirely conceivable that a single new drug generates over $1 billion in revenue, with its discovery and clinical advancement requiring just one person plus compute from 50,000 GPUs.

Khosla: I just received an article this morning, haven't finished reading it yet — the headline says a treatment for macular degeneration was discovered purely through AI. You're exactly right, and this could happen in entertainment or many other fields too.

So let's talk about the global question: how do we ensure AI's benefits spread more broadly and more fairly? There's concern about "the strong getting stronger," because perhaps only a few hundred people globally truly understand the scale and speed of this massive transformation. Can you talk about how to drive more equitable distribution and access to AI outcomes, domestically and globally?

Altman: I don't want to seem like I'm deflecting this question, nor do I want to focus only on negatives — I think there's something genuinely important worth discussing here. ChatGPT is now probably the world's fifth-largest website, and if it stays on its current trajectory, it will likely become number one. In the future, billions of people will use AGI for free — everyone will get quality medical advice, everyone will receive good education, everyone can freely request any software be generated and use it directly. This is how technology works.

I think capital-driven approaches are very effective here. Of course, this doesn't mean we don't need corrections in some areas — I do think there are issues that require different handling. But broadly, technology benefits the world because it puts tools in people's hands, and when they're free or low-cost, people can create amazing results.

Many people feel that "the world isn't ready, people can't handle this, only a few understand" — there are papers on this. But I think people actually know quite well what they need, and their ability to learn new technologies is quite strong. This isn't theoretical speculation; it's happening at massive scale globally right now. I think this is very good, and shows that existing global systems and incentive mechanisms work in many ways.

Of course, AI will also bring some unprecedented situations. For example, compute may become extremely scarce in the future, requiring us to democratically decide which problems to prioritize and how to allocate massive compute. This may be different from the past. You can imagine capital scrambling for compute, driving prices sky-high and creating severe shortages — that would be a bad situation. But my first answer is: then produce vastly more compute. While I'm not 100% certain of this solution, I tend to believe AI should push the world toward greater equality rather than exacerbating division.

Khosla: Finally, to save time, I'll put three questions on the table at once: one, the role of government; two, my expectation that AI will cause severe deflationary economics in the 2030s; and three, the risk of "sentient AI."

You can choose to answer one, or all of them.

Altman: Hmm, I'll answer the deflation one. I think the future economy will be highly deflationary, but the massive new wealth generated has to go somewhere. I hope it makes water, food, healthcare, education, access to nature, time with family — all these things affordable for everyone.

As for those who want to play "status games," they can bid things up — auction a da Vinci painting to $1 trillion, or buy an entire galaxy as a show of status. These people can satisfy their vanity while also creating value in the process.

But there's a more interesting question here: if we want all goods and services to deflate dramatically and become extremely cheap, while humans remain ambitious, creative, and willing to work hard, what do we use to measure value and achievement? How does the excess wealth flow? This is itself quite an interesting design problem.

(This article is only a partial excerpt and compilation of the conversation. Please click "read original" to view the full video.)