Next-token prediction builds broad world models, while reinforcement learning on chain of thought helps models solve harder problems.
2
Agents will move from completing tasks to co-creating with people through tools, long-running work, and interfaces that adapt to intent.
3
Trust in asynchronous agents depends on giving people ways to inspect, edit, verify, and correct model outputs.
Summary
Karina Nguyen describes two scaling approaches that have shaped recent AI products. Pre-training through next-token prediction gives models broad knowledge and an internal model of how sequences and parts of the world work. Post-training and reinforcement learning then make those capabilities more useful, while chain-of-thought training helps models spend more computation on difficult problems. Nguyen expects agents to combine reasoning, tools, long context, and creativity. She draws on work on Claude, ChatGPT Tasks, Canvas, and Claude in Slack to explain the product problems this creates. Familiar interfaces can make unfamiliar capabilities easier to use. Canvas shows how people and models can edit, research, code, and verify work together. Nguyen also predicts that people will create software, games, and interactive learning experiences through blank-canvas interfaces that adapt to their intent. The hard problem is trust, especially when agents work for hours without direct supervision.
Next-token prediction gives models a broad model of the world
Nguyen describes pre-training as a world-building machine. A model learns by predicting what comes next, and the token can be a word, string, pixel, or another type of data. Since sequences follow from earlier actions, prediction requires some understanding of how the world works. This produces broad multitask learning. Translation and factual knowledge can be easier because examples appear often in online material. Math, long-form creative writing, and coherent storytelling are harder because they require more computation, planning, and sustained consistency. Nguyen says creative writing remains an open research problem because it is difficult to measure quality and maintain a plot over a long text.
Post-training makes broad model capabilities useful in products
Nguyen places early products such as GitHub Copilot in the period when researchers were scaling pre-training. Copilot's code autocomplete benefited from training on billions of code tokens, then reinforcement learning from human and AI feedback made the raw capability more useful. Post-training teaches models to complete function bodies, understand docstrings, generate multiline completions, and predict or apply diffs. She says there is still much to explore in post-training models so they can reason through complex codebases.
Reinforcement learning on chain of thought lets models spend more time solving hard problems
The next research paradigm Nguyen describes is scaling reinforcement learning on chain of thought, as used in OpenAI's o1 model. During training, the model learns how to think through difficult tasks from useful feedback signals. For harder applications, such as medical problem solving, models need time to work through a problem and use tools in environments where they can check their outputs. Nguyen says chain-of-thought research still has unresolved questions about faithfulness, measuring whether reasoning reflects the real process, and whether a model can recognize a wrong direction and backtrack.
Longer reasoning changes how people should interact with models
When models take 15 seconds or much longer to answer, the interface has to change. Nguyen describes streaming summaries of a model's thoughts as one approach to communicating progress without making people wait without information. This creates its own design problems. The agent era she describes combines complex reasoning with browsing, search, computer use, long context, and work over a long horizon. She sees co-innovation as a later stage that adds creativity through human and AI collaboration.
Reasoning models can speed product research and create training data
The new scaling approaches change product development because highly reasoning models can help create faster evaluation cycles. Teams can distill their behavior into smaller models and use them to generate synthetic data, post-training sets, and reinforcement-learning environments. Nguyen gives multiplayer collaboration as an example. A team could simulate different users, condition generated data on those users, and train a model for a new type of collaborative task. Models can also learn new tools from a few examples through in-context learning, which lets developers test product ideas quickly.
Familiar product patterns make unfamiliar model capabilities easier to use
Nguyen says the 100K context feature worked well because file uploads are familiar to people who already work with documents. A giant chat containing the same context would have been less natural. ChatGPT Tasks uses another familiar pattern, scheduled reminders, but extends it into recurring activities such as continuing a story, searching for topics, or helping someone learn a language with interactive visual material. Her lesson is that product features should allow modular combinations as model capabilities grow.
Trust requires collaborative ways to inspect and correct long-running work
A central product problem is connecting real-time interaction with asynchronous completion. A person might ask an agent to research or write code for ten hours and return with a solution. Nguyen identifies trust as the bottleneck. People need ways to verify and edit model outputs, then provide feedback while the work is happening so the model can improve. Claude in Slack was an early attempt at a virtual teammate, using Slack's tools, image uploads, and multiplayer collaboration. Nguyen also points to recurring Slack channel summaries as an example of delegated work.
Canvas turns writing, coding, research, and analysis into shared work
Nguyen describes Canvas as a flexible interface for fine-grained collaboration. The model can act as a co-creator or co-editor, search while generating a report, and answer questions about whether an output is correct. The same interface could support multiple people or multiple agents, such as a critic and an editor. Canvas can also help with pair programming, search API documentation, and analyze uploaded CSV files in real time. Nguyen says its flexibility comes from training for collaborative writing and coding alongside tool use.
A blank canvas could become the software people need for their immediate intent
Nguyen expects software creation to become available to people who have never written code. She describes using Canvas on a plane to create a game and imagines people making tools, visualizations, and businesses for themselves. She predicts that people will click less on Internet links and access information through model-generated, personalized outputs. A request to learn about the solar system might produce an interactive visualization rather than plain text. In her view, an AI interface could begin as a blank canvas and change into an IDE for coding, or into writing and plotting tools for creating a novel.
"The interface to AI is a blank canvas that kind of self morphs into your intent."22:30
Who should watch
You are building an agent that needs to work with people over a long period rather than return one answer.
Your product exposes a new model capability and you need a familiar interaction pattern that people can understand.
You are deciding how users will review, edit, and correct work delegated to an agent.
