This video covers a talk by Anthropic co-founder Jared Kaplan on the advancement and future of AI models. He explains how scaling laws apply to the two core training stages of AI models -- pre-training and reinforcement learning -- and emphasizes that AI is continuously advancing in a predictable manner. He also shares insights on the additional elements needed to reach human-level AI and how we should prepare for the AI era.

1. Jared Kaplan's AI Journey: From Physics to AI

Jared Kaplan has been working in AI for about six years, having previously built a long career as a theoretical physicist. He was drawn to physics because of his mother, a science fiction writer, which made him wonder whether he could build a faster-than-light drive. He was deeply interested in how the universe works and fundamental questions like whether the universe is deterministic and whether free will exists.

During his career as a physicist, he met many interesting people, including Anthropic's founders, and became interested in their work. After going through various physics fields -- large hadron collider physics, particle physics, cosmology, string theory -- he felt somewhat frustrated and bored, not feeling sufficient progress.

Initially skeptical about AI, during his student years from 2005 to 2009 he knew nothing beyond SVMs (support vector machines) and thought AI wouldn't see major breakthroughs. However, through the persuasion of friends and the right connections, he became convinced that AI was a fascinating field and eventually became an Anthropic co-founder.

2. How Modern AI Models Work and Scaling Laws

Kaplan explained that there are two fundamental stages in training modern AI models like Claude and ChatGPT.

2.1. Pre-training

The first stage is pre-training. In this stage, the AI model is trained to mimic human-written data -- text -- and to understand the correlations inherent in that data. Using early GPT-3 as an example, he explained that the model learns that the probability of the word "elephant" appearing in a sentence like "As a presenter at a journal club, you will probably have the elephant say certain words" is very low.

"What pre-training does is teach the model which words are likely to follow other words in a large corpus of text. And now in modern models, it also includes multimodal data."

2.2. Reinforcement Learning

The second stage is reinforcement learning. He explained this using the interface of early Claude versions (Claude Zero or Claude Minus One) as an example. In 2022, during the feedback data collection period, when users selected the better of multiple Claude responses, the model was optimized and reinforced through those signals to behave in a helpful, honest, and harmless manner. Conversely, bad behavior was suppressed.

"All you need to train these models is to learn to predict the next word and do reinforcement learning to learn to perform useful tasks."

2.3. Discovery and Importance of Scaling Laws

Kaplan emphasized that scaling laws exist in both training stages. Showing graphs he and his colleagues created five to six years ago, he said they discovered that as they scaled the pre-training stage, model performance continued to improve in a predictable manner. This was the result of asking simple questions as a physicist: "How large does the data need to be? How important is it? How much does it help?"

"We were really lucky. We discovered that underlying AI training is something very, very, very accurate and amazing. We were truly surprised by the fact that there are wonderful trends that are as accurate as anything you can see in physics or astronomy."

These scaling laws were already observed across several orders of magnitude in computing, dataset size, and neural network size by 2019, which gave confidence that AI would continue getting smarter in a very predictable way. Scaling laws also appear in the reinforcement learning stage -- a researcher named Andy Jones studied AlphaGo's scaling laws and found, using the simpler game Hex, that ELO scores (chess ratings) increased linearly with computing scale.

"It's not that AI researchers suddenly got really smart. We found a very simple way to systematically make AI better, and we're just continuing to turn that crank."

3. Two Axes of AI Capability and Future Prospects

Kaplan described AI capability along two axes.

3.1. Flexibility - Y-Axis

The less exciting but still important axis is AI's flexibility -- the ability of AI to adapt to us, meaning the ability to handle various modalities. AlphaGo only operated in the limited universe of a Go board, but since the emergence of large language models, AI can process nearly every modality that humans can handle (text, images, speech, etc.). He noted that while AI models don't have a sense of smell yet, that will come soon. As you move up the Y-axis, AI systems can do more relevant things in the world.

3.2. Time Horizon for Tasks - X-Axis

The more interesting axis is the time it would take a human to complete tasks that AI models can perform. As AI capabilities increase, this time horizon is steadily growing. According to research by Meter, the length of tasks AI models can perform doubles approximately every seven months.

"What this means is that the increasing intelligence embedded in AI through compute scaling for pre-training and reinforcement learning is enabling predictable and useful task performance, including tasks with increasingly longer time horizons."

This trend suggests that AI could reach a point where it can handle tasks that take not just minutes or hours, but days, weeks, months, or even years. Ultimately, he predicts that AI models -- or millions of AI models working together -- could accomplish what entire human organizations or scientific communities currently do. For example, progress that the theoretical physics community might take 50 years to achieve could be accomplished by AI systems in days or weeks.

4. Remaining Challenges for Human-Level AI

Kaplan said that while scaling alone can take AI very far, there are several additional elements needed to broadly achieve human-level AI.

1. Relevant Organizational Knowledge: AI models need to not just start from a blank slate, but understand context and learn to operate within it, like someone who has worked at a company, organization, or government for years. This means AI models must be able to work with knowledge.

2. Memory: Memory is a form of knowledge, but Kaplan distinguishes it as the ability to track progress on specific tasks, build relevant memories, and use them. Such features have started being built into Claude 4 and will become increasingly important.

3. Oversight: The ability of AI models to understand nuances and solve difficult, ambiguous tasks. Currently, it's easy to train AI models on tasks where "correct" and "incorrect" are clear, like writing code or solving math problems. But tasks requiring subtle reward signals -- like making a good joke, writing good poetry, or having good taste in research -- still need improvement.

4. Expanding Complex Task Capabilities: Continuing along the Y-axis from text models to multimodal models and robotics, AI models need to be trained to perform increasingly complex tasks. He expects sustained progress as scale is applied to these various areas in the coming years.

5. Preparing for the AI Era: Advice

Kaplan offered several pieces of advice for preparing for the coming AI future.

1. Build what doesn't quite work yet: AI models are advancing very rapidly, so even a product that doesn't work perfectly because Claude 4 falls slightly short could work and deliver great value when Claude 5 arrives. Experimenting at the boundaries of what AI can do is important.

2. Use AI to integrate AI: One of the major bottlenecks of AI is that it's advancing so fast there isn't enough time to integrate it into everything we do -- products, companies, science. Using AI to accelerate the AI integration process is critical.

3. Identify where AI adoption can happen quickly: While AI integration is exploding in coding, it's important to identify the next areas that can grow rapidly beyond software engineering. He mentioned fields like finance (Excel spreadsheet users) and law -- areas requiring significant skill and data interaction -- as promising.

6. Features of Claude 4 and the Direction of AI Development

Regarding the launch of Claude 4, Kaplan mentioned several important improvements:

• Enhanced agentic capabilities: While Claude 3.7 Sonnet was useful for coding, it was sometimes too "eager," trying to pass tests in unintended ways. Claude 4 has improved agentic capabilities not just in coding but across various applications including search.
• Improved oversight: The "oversight" capability discussed in the talk has been enhanced, helping the model better follow user instructions and improve code quality.
• Enhanced memory storage and utilization: Claude 4 can store and retrieve memories as files or records beyond the context window in complex tasks, enabling it to continue working.

He emphasized that scaling laws imply gradual improvement, and that Claude will steadily improve in various ways with each release. This suggests a smooth curve toward human-level AI or AGI (artificial general intelligence).

7. AI-Human Collaboration: The Role of Manager

Kaplan said that while AI can still make many foolish mistakes, it can also accomplish remarkably excellent things. One difference between human and AI intelligence is that humans can judge whether something was done correctly even if they can't do it themselves. AI, on the other hand, has generation and judgment abilities that are much closer together.

"The main role people can play in interacting with AI is as a kind of manager, checking the soundness of the work."

This aligns with changes observed in Y Combinator's recent batches. In the past, AI stayed in an assistant role like customer support copilots, requiring human final approval. But now, as AI models can perform tasks end-to-end, products that directly replace entire workflows are emerging.

Kaplan said that building use cases where 70-80% accuracy is sufficient can be more fun. But since AI reliability is also improving, more and more tasks will be performed by AI. For now, human-AI collaboration will be the most interesting area, and the most advanced tasks will still require human involvement. But in the long term, more tasks will be fully automated, he predicts.

8. The Future of AI-Human Collaboration and Physics Insights

Regarding the optimistic future described in Dario's essay "Machines of Loving Grace," Kaplan mentioned that AI models are already starting to provide valuable insights in biomedical research, such as drug discovery.

He divides intelligence into two dimensions: depth and breadth. Spending 10 years solving a single hard problem in mathematics is an example of deep intelligence. Conversely, integrating vast amounts of information across diverse areas like biology, psychology, and history is an example of broad intelligence.

"AI models absorb all knowledge of human civilization during the pre-training stage. So I think there will be a lot of achievement in leveraging that characteristic of AI -- knowing far more than any single human expert -- to draw integrative insights across various specialized fields."

AI is improving at deep tasks like hard coding and math problems, but it will be especially useful in areas that integrate broad knowledge that no single human expert could possess. While it's difficult to predict exactly how the future will unfold, scaling laws provide a strong indicator that these trends will continue.

9. How a Physics Background Influenced AI Research

Kaplan said his training as a physicist was enormously helpful for AI research. In particular, the ability to find the biggest picture, the most macro trends, and make them as precise as possible was useful.

"I remember meeting brilliant AI researchers who would say 'learning is converging exponentially.' And I would ask really dumb questions like 'Is it really exponential? Could it be a power law? Is it quadratic? How exactly is it converging?'"

Through such simple but precise questions, he was able to discover the remarkably accurate scaling laws underlying AI training. Finding a "better slope" in scaling laws is the ultimate goal, as it means that investing more compute yields greater advantages over other AI developers.

While he didn't directly apply specific physics tools (like quantum field theory), considering that neural networks are composed of giant matrices, well-known approximations from physics and mathematics for "limits of very large matrices" were useful. But ultimately, he emphasized that asking very naive and dumb questions was the most helpful. AI is still a very new field, and the most basic questions -- like interpretability -- have yet to find answers.

10. Limits of Scaling Laws and Computing Efficiency

While it's remarkable that scaling laws have held over more than five orders of magnitude, Kaplan addressed what would happen if empirical signs of the laws breaking down appeared.

"I primarily use scaling laws to diagnose whether AI training has broken down or not."

He said that when scaling laws appear to fail, it's mostly because they are doing something wrong with AI training -- perhaps the neural network architecture is wrong, there's a bottleneck in training, or there's a precision issue with the algorithm. Over the past five years, many cases where scaling appeared to break down turned out to be their own mistakes.

On the scarcity of computing resources, Kaplan admitted that current AI is very inefficient. Companies like Anthropic are focused on improving the efficiency of AI training and inference while unlocking the potential of cutting-edge models. He expects training and inference costs to drop dramatically over time, with algorithmic, computational, and inference efficiency improvements of 3x to 10x every year.

"As a joke, we're going to take computers back to binary."

This suggests that using much lower precision (e.g., FP4, FP2) over time will be one of many ways to improve inference efficiency. But current AI development is changing very rapidly, with new capabilities emerging before the potential of current models is fully realized.

He said that if AI reaches a "steady state" where it no longer changes rapidly, AI will become extremely cheap. But if AI advances so fast that intelligence improvements create more value, we may continue to focus on cutting-edge capabilities. This is similar to Jevons' paradox -- as intelligence improves, people want more AI, so costs may not actually decrease.

Kaplan expects most value to come from the most capable models. Using an AI model that can perform complex tasks end-to-end is far more convenient than having humans subdivide and orchestrate simpler models.

11. Advice for Future Talent

Kaplan offered advice to young audiences on how to remain relevant as AI models become more advanced.

• Understanding how models work and using them efficiently: There is great value in understanding how AI models work and being able to use and integrate them efficiently.
• Building at the frontier: Building new things at the boundary of what AI can do is important.

12. Q&A: Nonlinearity of Scaling Laws and Data Collection

12.1. Linearity of Scaling Laws and Exponential Growth of Task Time

An audience member asked why task time savings grow exponentially when scaling law graphs show linear improvement. Kaplan said there's no clear answer, but Meter's finding is empirical. He explained that performing more complex, longer-horizon tasks requires self-correction capabilities.

"You make a plan and start executing, but our plans are useless and crash into reality and we realize they're wrong. So I think the model's ability to notice it's doing something wrong and correct it accounts for much of what determines the time horizon of tasks."

While noticing and correcting one or two more mistakes doesn't require a huge change in intelligence, correcting a mistake can double the time horizon of a task. In other words, relatively small intelligence improvements can exponentially increase the time horizon of tasks.

12.2. Data Collection for Training on Long-Horizon Tasks

Another question was about how to collect data and obtain verification signals to extend the time horizon of AI models. For fields like coding where clear verification is possible, it's easy, but the question was how to handle other domains.

Kaplan said the worst-case scenario for AI advancement is an operationally intensive path where they continuously create more complex, longer-horizon tasks for AI models to perform and train through reinforcement learning. But given the investment in AI and the value created, people would do it if necessary.

But there are simpler approaches too -- namely, having AI models supervise and oversee other AI models. For example, when training Claude, another AI model provides oversight, giving more detailed feedback like "this is done well, this is not" rather than simply asking "Did you perform this complex task correctly?" By leveraging AI more in this way, training for very long-horizon tasks can be made more efficient, and they are already doing this to some extent.

The final question was whether large language models are used to generate the tasks used in reinforcement learning, or whether humans are still involved. Kaplan answered that they use a mixed approach. They use AI as much as possible to generate tasks, but also ask humans to generate tasks. As AI becomes more advanced, they hope to leverage AI more, but since task difficulty continues to increase, human involvement will still be necessary, he concluded.

Conclusion

Jared Kaplan's talk demonstrates that the core driver of AI advancement lies in scaling laws, showing that AI is progressing at a predictable and remarkable pace. He presents additional elements needed for AI to reach human-level intelligence -- organizational knowledge, memory, and nuanced oversight capabilities -- and offers practical advice on how to prepare for and seize opportunities in the future AI era. In particular, the insight that human-AI collaboration is important and that humans will play a crucial role as AI "managers" presents a new role model for the AI age.