With a reasonable token cost, one person can now accomplish what once required a hundred-person elite engineering team. This option was entirely nonexistent for startups in the past—but it’s a reality today.

Author: Houlang Evolution Planet

Source: Wall Street Journal

In a systems design class at Stanford, OpenAI CEO Sam Altman discussed a正在发生的变化: with just a reasonable token cost, one person can now accomplish what once required a hundred-person elite engineering team. This option was previously nonexistent for startups but has now become a reality. He also spoke about the magic of scale, how users discovered ChatGPT and Codex, how AI is becoming a new utility like electricity, and why betting against LLM scaling is currently the worst possible move.

Below is the transcript of Sam Altman's talk:

One person, with the right token costs, can accomplish what a hundred-person team can do.

Now, with a reasonable token cost, one person can accomplish what once required an exceptional hundred-person engineering team at a startup. This was previously impossible and not even an option for startups—it’s now a reality.

So, the ambition you can afford, the speed you can achieve, and the number of things you can push forward at once are all completely different. This isn’t incremental change—it’s a transformation of the very set of options available.

I don’t understand why this observation holds true, nor do I have a theory that fully satisfies me, so I feel a bit uneasy recommending it to others. But empirically, it has held up: every most interesting thing I’ve observed in my career has been related to emergent properties arising from scale.

When something has worked in an interesting way on a smaller scale, and you can scale it up to an unprecedented level, it’s usually a great idea—and most people don’t do nearly enough of it. This applies to AI model scaling laws, bringing more brilliant minds together to solve a problem, and also to the scale of companies and economies.

That’s what I truly learned at YC. At the time, everyone said YC was too big and should shrink, arguing that investing in only 10 companies per batch was optimal. Many very smart people said this. It was tempting because it would drastically reduce the workload. But much of YC’s magic comes from the network effects within each batch—a phenomenon that only emerges at scale. No one had previously invested in startups at this scale in the same way, so no one had discovered this yet.

II. Challenges of Scaling: Things will break in unpredictable ways

People don’t often scale up, partly because as things grow, they break faster and in increasingly unpredictable ways. If you truly want to scale something up, there will always be a few small issues. And there will always be very smart people telling you: Don’t be too ambitious, don’t go too big—let’s try something smaller.

So, scaling is fundamentally a system design problem. Taking our effort to scale AI models as an example, the questions we had to address included: Is it technically feasible at all? No one had ever attempted training across 10,000 or 100,000 GPUs—it required massive engineering resources; How do we solve the capital requirements? Where is the business model? How can you justify taking on such risk? Researchers also asked: If you have this much compute power, why bet everything on a single project? Why not distribute it across multiple projects?

It’s crucial to break down these challenges one by one and address each individually. This is a universal approach across nearly all scaled domains.

Three: The hardest for humans to rebuild—clear goals, clear plans, and clear decision-making methods.

When scaling, the hardest part to restructure is often not the technology, but the human element. What experience do you have with involving humans in large-scale organizational systems?

My answer is: clear goals, a clear plan, a clear explanation of how to get there, and how to make decisions along the way. This is very important.

Returning to our example of scaling deep learning: many people said it wouldn’t work, that there were these issues, and that we needed a more diversified portfolio. But once we said, “No, we’re betting on scaled deep learning. This is our direction. If we’re wrong, we’ll fail—but we’re doing it anyway. Here’s why: we believe in what the world will look like if we get there.” That clarity itself is incredibly powerful.

Moreover, humans have not evolved to be good at thinking about exponential growth. It is difficult for people to imagine that scaling laws will continue to grow exponentially, that income will grow exponentially, or that organizations can withstand exponential complexity. From my experience, truly getting people to reason from first principles about “why this will happen” takes a very long time.

Four, ChatGPT was not planned—it was discovered by users.

We built GPT-3 and needed to generate revenue because we wanted to scale up to computers costing billions or tens of billions of dollars. GPT-3 was a cool demo, but we couldn’t find a product to build around it. We tried a few things, but none worked. We knew the models would improve, but we also wanted to get the revenue engine running early. So we said: since we couldn’t think of a product, let’s just make it an API and hope others can build products on top of it.

In the summer of 2020, we released the GPT-3 API. Initially, it had almost no traction. About a month later, it unexpectedly went viral on Twitter. On that same day, several different developers shared cool things they had built with it, and others began trying it out.

But if you go back and use GPT-3 or 3.5, you’d be shocked at how poor the models were compared to the excitement they generated. People tried all sorts of things, and the only business that truly worked in a significant way was copywriting—and even that wasn’t very good or particularly exciting. At the time, we thought: we just have to wait for better models.

Although only copywriting has been successfully implemented, developers have already learned how to input a prompt and chat with it. We’ve seen many people unable to use the API for business purposes, yet they use their API keys purely for chatting. We realized: people clearly want a user-friendly chatbot.

We had GPT-4 at the time, but there was also a ready model in between called 3.5. We also discovered a new post-training method that helped the model better follow instructions and chat more naturally. We thought: the API might only be a $10 to $20 million run-rate business, but there was one thing people really loved. Following YC’s principle—see what users like, then build it—we built a chatbot around it.

We didn’t think it would be that successful when we launched it. It was originally more like a research demo meant to convince others to build chat products and purchase APIs. Then it went viral unexpectedly.

Another thing I learned from YC is that when something starts growing rapidly but isn’t yet good enough, you essentially have a product that’s guaranteed to explode. For five days, traffic spikes and drops, and everyone says, “It’s just a hype cycle”; then the next day, the peak is even higher before dropping again, and people still call it hype. By the fourth or fifth day, I knew: we understand how this will unfold—we have a potential killer product on our hands.

By day five, we gathered everyone and said: This is an emergency—but a good one. We have to build a company and a product at the same time. Then came two months of wild growth. We’d figure out the business model later—first, we had to charge something, just to keep our computer bills from burning us alive. This wasn’t a long-term solution, but back then, we just needed to get it running.

Five: Codex was originally the master plan: programming is the way the model controls the world.

Before ChatGPT, our plan was actually to fully focus on coding. We knew these models could write code and recognized it as a valuable domain. But our internal belief at the time was: programming is how models control computers, while robotics is how models control the physical world. If you build a sufficiently intelligent model that can write code, control robots, and operate robotic actuators, you can enable intelligence to act for you in the real world.

It took us some time to get there. Codex had already become solid earlier this year, but 5.5 was the real turning point, when people started doing incredible things with it.

The current pipeline is roughly: pre-training, mid-training, post-training, followed by RL and supervised feedback loops. I believe this is indeed the current pipeline, but I anticipate we’ll undergo a major rewrite. I don’t know when or exactly how it will happen, but the very notion of it being a “pipeline” feels a bit odd—it doesn’t seem optimal.

What is the optimal solution? I believe this is the problem AI should research and solve. We’ve set the goal: by this September, to use computing power equivalent to 500,000 A100s as AI research interns; by March 2028, to possess a complete, end-to-end, highly talented AI researcher capable of discovering entirely new architectures on its own.

So, with the current pipeline and architecture, we might cross that line: AI can produce incredible work.

Six: AI is becoming the new utility: Selling "intelligence" is hard to relate to

I believe we are creating a new utility. It’s not common. Electricity is a utility, the internet is a utility, water is too—there aren’t many of these things. So, we don’t have many ready-made analogies or experiences to teach us how to explain it to the world.

I recently studied what happened when electricity became a utility. Early electricity companies didn’t market “electricity,” because no one knew what it was or why they would want it. Electricity sounded frightening: it would enter your home, could kill you, and felt completely unlike anything from the previous world. Perhaps they tried selling or marketing electricity, but it didn’t work.

Later, what they started selling was the light of night. What you receive from us is not electricity, but the light of night. By the way, the same thing can be used for many other purposes. But people would ask: Why would I want that? They’d say: One day, it could even help you do your laundry. People would respond: No, that’s too far off—I can’t accept that.

I don’t know what “light of the night” corresponds to in AI. But even if we’re completely right—that intelligence will become a new utility that every company, every customer, and every government needs to connect to—you’ll have some kind of OpenAI token subscription plugged into everything, running continuously and performing amazing tasks for you. At least for now, I don’t think it’s accurate to compare it to “we’re selling intelligence,” because people don’t resonate with that.

If we want to become a new utility, we need to find a way to explain to the world what it truly means to have this smart pipeline that lets you do anything you want.

As a consumer, business, or individual, you will ultimately think in terms closer to tokens—or even one level above tokens. You won’t care much about where the hardware is, what specific chips are used, or what’s powering it—all of that will be abstracted away. What matters to you is: Can I use the system extensively? Is it affordable? Is it well-made?

Now, it’s like a token. As we enter a world where everyone has a continuously running agent serving them constantly and being useful, you might think even higher—just as when you pay your phone bill, you think, “I’m buying talk time and data, accessing the entire system,” not about the specific hardware in the cell tower or how it connects to the internet. The underlying hardware is abstracted away.

Seven: Betting that LLM scaling will fail is the most misleading assessment currently.

Regarding Yann LeCun’s view that “LLMs are a dead end”: First, in terms of achieving human-level intelligence, these models already far surpass humans in some aspects and fall far short in others. For example, they appear to be significantly worse than humans in tasks requiring very long-term planning and high-level judgment.

On the other hand, just yesterday, one of our models discovered and disproved a conjecture related to an Erdős problem—a question that brilliant minds had studied for a long time. Many intelligent scientists, I don’t know if LeCun is one of them, even recently said this would never happen. Yet the model did it. Now mathematicians are asking: Is mathematics coming to an end? What does this mean for our field?

So, LLMs clearly have the ability to discover new knowledge and perform certain intelligent tasks that humans cannot. They will continue to scale. The extent to which they can outperform humans in various tasks remains to be seen, but I suspect it will be substantial.

Moreover, as we previously discussed, the lack of belief in exponential growth has indeed been hindered by a generation of scientists who were overly certain that "scaling won't produce anything." Meanwhile, others simply looked at charts and said, "It still looks great—let’s keep going."

World models are clearly important, and things like robots need them. But betting on the scaling of LLMs at this stage would be a failure, and I find it highly misleading.

VIII. Recommendation for "One Person Frontier Lab": Focus on research into reasoning to make intelligence affordable and abundant.

If I were in this class and working on the "One-Person Frontier Lab" project, what would I do?

First, your project is excellent. Many very smart people are currently researching training ideas, and I promise we’ll soon have incredible models. But I believe we’re underinvesting in delivering large-scale, low-cost intelligence.

So perhaps I’ll look into the inference stack: how do we make this incredible intelligence cheap and abundant? I think this area is underinvested in. And ultimately, all leading labs will largely have to become inference companies.

It might be a bit late to change this project now, but better late than never. Do what you want to do.

Conclusion

This interview outlines Sam Altman’s key insights on scale, product, and the future of AI: one person with the right amount of tokens can accomplish the work of a hundred-person team; scale often produces emergent properties beyond consensus, but most people don’t push hard enough to unlock them; both ChatGPT and Codex were born from “see what users like, then build it”; AI is becoming a new utility, yet we haven’t yet found the “light at night” to explain it to the world; the current capability pipeline is not optimal and requires a major rewrite; and in the face of exponential growth, don’t let your identity distort your judgment of the truth.

As Altman said: "If you tie your identity to whether something specific succeeds or fails, you're too attached to your identity to let go and see the truth." This may be the most important reminder for each of us in this exponential era.