I’ve always been a nerd and an intensely curious person, but over the last year, I’ve been amazed to see how effectively I can learn with AI. Yes, lots of people are coming to these kinds of realizations in their own way. But this post is my way of condensing my experiences in the last several months into words.
Recently, the release of GPT-6 Astra ⤴ has got me really thinking about how human-agent interfaces can support a new kind of sustained inquiry. On a really difficult AI evaluation benchmark, ARC-AGI-3, Astra’s best reported score rose from 62.7% in the standard harness to 99.9% with a provider adapter that preserved reasoning state and compacted longer conversations.1
What we’re seeing with Astra is just the beginning of a new wave of human-AI collaboration, because it illustrates how much the harness around a highly capable model can affect its ability to serve the needs of the human. We don’t need the AI model to be perfect at reasoning － we just need more personal agents to help us reason about the world better. Hopefully, that translates into a world where more humans are able to think critically and act on their understanding.
As agents take on more of an investigative role in the early stages of learning, I keep coming back to this question: which parts can I hand over to AI, and which parts do I need to work through myself if I want to become a better critical thinker?
Does AI make us intellectually lazy?#
There’s a reasonable concern that AI makes us intellectually lazy. When ChatGPT can explain a concept or summarize a paper in seconds, it’s easy to skip the useful friction involved in thinking through a problem. I’ve definitely been guilty of this myself.
Getting a quick answer can be useful when the task is trivial. But when I want to learn something deeply, I need to work through why an explanation makes sense, where it might fail, and how I assess that failure to inform my next step. The ability to engage in a back-and-forth with an agent gives me a totally new way to practice such a reasoning loop.
When GPT-4 was released in 2023, OpenAI demonstrated the idea of a Socratic tutor2 that guided students with questions so they could work out an answer themselves. That approach feels all the more relevant today, with models like Sol and Astra in the mix. When I explain my thought process and ask the agent to challenge it, I begin to notice gaps in my understanding. Reconstructing the explanation in my own words makes those gaps harder to gloss over.
The agent can search and compare a large number of sources concurrently, while leaving room for me to make a prediction, question an assumption, or decide what evidence would change my mind. Those are the parts of the interaction that one needs to exercise, to become better at critical thinking. Just like going to the gym to build muscle, one needs to exercise their own reasoning skills, too.
From questions to actions#
Over the last few months, I’ve been researching how to better approach my diet, calisthenics, tea-drinking, and sauna + cold plunge routines in a more holistic way. Doing any one of these well is rewarding on its own, but learning about them together has made me curious about how they fit into my daily life, and my daily routines and overall well-being have significantly improved as a result.
With my agents, I begin with hypothetical questions, work through recent scientific literature, pose further questions through the agents’ assistance, and follow interesting reading material they cite, getting myself into research rabbit holes. Eventually, with enough cycles, an initially crude question about a topic (“How does sauna help with exercise adaptation?”) becomes practical: what can I do with this understanding?
Acting on an idea gives me something to bring back to the conversation with the agent. It also gives me a reason to distinguish what a blog post or paper supports, what I’m hypothesizing, and what I’ve actually noticed in the real world. Feeling better doesn’t, by itself, tell me which change helped or whether my explanation is right. That distinction is part of the reasoning I want sharpened by these conversations with the agent.
The approach I’ve been using is definitely working well so far: I’ve been feeling healthier, happier, and more mentally relaxed as these routines have become part of my life. The continuous stream of learning has also taken me somewhere I didn’t anticipate: into new conversations with other people.
Human connection keeps our learning porous#
In my earlier post on how a Taiwanese oolong changed the way I look at tea, I wrote about how curiosity about a particular tea led me to explore the processing, chemistry, and geography of tea. It not only opened up a fascinating rabbit hole about the vast world of tea, but it also led me to walk into tea shops and tea-focused events IRL where I meet people who enjoy drinking tea. The same holds for my time at the sauna, too: I’ve met all sorts of interesting people I would never have met otherwise.
What started off as a personal health-focused inquiry has become a way to meet new people and have new conversations. Meeting other people adds a totally different set of experiences, perspectives and interests to my thought process. Recently, this has been leading me to ideas I hadn’t yet considered. Some of those become hypothetical questions I explore with an agent later. Learning with AI has changed where I go and who I meet. Those encounters, in turn, change what I want to learn.
I picture these sorts of interactions as a cross section of a patch of earth that represents how much I know about a topic (see the figure below). The porous topsoil in the image is a passing conversation with another human, or hearing about someone else’s experience, which seeds a new question. The rabbit holes are where the AI agent and I can dive in deeper on those questions, where I examine its explanations and connect it to things I already know.
Although AI can introduce unfamiliar ideas too, I’ve repeatedly found that humans in the real world are a source of true randomness, allowing me to explore new ideas in a way that pure human-AI interactions cannot live up to (over time, the human and the agent influence each other too much).
What I’ve valued most is how the two modes of daily interaction (human-AI, and human-human) become part of a life I’m enjoying more in aggregate. With the kinds of AI interfaces we use today (voice, chat, etc.), it’s become quite seamless to explore a new interest, act on it, and find people in the real world to share it with.
Building toward better mental models#
Each time an observation or conversation helps me revise an explanation, I carry that understanding forward. I can ask a more precise question in the future, or recognize an assumption I previously missed. This is where I notice the compounding benefits of human-AI synergy.
Every person in tech likely knows about the “S-curve”. Recently, I’ve been picturing the “human-AI synergy S-curve” below. Rather than a predefined learning trajectory, the part in the middle with the upward slope represents the increasing returns from the combined human and AI feedback loop I described above. Right now, I find myself right in the middle of that slope. 🚀
I picture the plateau towards the right as an “ideal state”. There’s a point beyond which working with AI yields diminishing returns. Once I’ve worked through the main questions in a topic, another conversation with ChatGPT may refine my understanding without changing it very much. What once took effort has become familiar, and further progress may depend on encountering something my current explanation can’t account for.
This is why human connection in the real world is so important, and is very much part of that feedback loop. It’s what provides new experiences, the opportunity for a failed prediction, or a fresh perspective that in turn can give me a better question to pursue. Then, this can potentially kickstart another S-curve.
Learning how to learn#
The biggest change I’ve noticed over the last year is in my relationship to not knowing. A gap in my understanding no longer feels like a dead end, nor does it make me feel inadequate. Inquiry has become rewarding in itself, and I have more ways to pursue a question when it occurs to me. And most importantly, I have more ways to act on what I learn.
The combination of human and AI feedback loops has made me more comfortable with uncertainty, and more willing to explore it.
As agents become more able to carry out detailed investigations autonomously, we can spend more of our attention on what the findings mean, where we’re unconvinced, and what we could try next. This is where the user interface plays a crucial role. A useful interface should support that exchange wherever it begins, seamlessly moving thoughts between the human and the agent. Ultimately, it should also support the human in acting on what they learn, and in reflecting on what they learned from that action.
The arrival of Astra-class models will open up new questions on how humans will interact with them in the future. Learning new things has never been more fun, but hopefully, it will also make us better thinkers and more effective actors in the world!
Footnotes#
-
Greg Kamradt, OpenAI’s GPT-6 Astra on ARC-AGI-3 ⤴, September 3, 2026. Best reported semi-private scores use different reasoning settings: standard at max, provider adapter at high. The custom code tools were observed separately in PRO-LONG. ↩
-
OpenAI, GPT-4 ⤴, March 14, 2023, “Steerability: Socratic tutor.” ↩