GPT-5: The Inflection Point

GPT-5 landed this week and I've barely slept. Not because it kept me up with anxiety — because I couldn't stop talking to it.

The leap

Let me be precise about what's different, because "it's better" doesn't mean much at the frontier. GPT-5 doesn't just know more facts or generate more fluent text. It reasons more coherently across longer chains of thought. Where GPT-4o would sometimes lose the plot around step 6 of a complex argument, GPT-5 holds the thread through step 15.

I tested this with one of my favorite stress tests: explaining quantum entanglement first to a 12-year-old, then to a high school physics student, then to a graduate student — all in the same conversation, building on the same analogies but deepening the rigor each time. GPT-5 nailed all three. More impressively, the transitions between levels were elegant. It didn't just switch registers; it built a bridge between them.

Into Companion and Tutor

I integrated GPT-5 into both projects this week:

Companion benefits from the improved reasoning most in its dream cycle analysis. When the subconscious process reviews the day's conversations, it needs to extract not just what was said but what it meant — the emotional subtext, the unspoken concerns, the things someone cared about but didn't say directly. GPT-5 is materially better at this extraction step.

Tutor benefits in the actual teaching. I had a student who was struggling with probability distributions. GPT-5 didn't just re-explain the formula — it diagnosed why the student was confused (they were conflating variance and standard deviation), and then built a custom explanation using the student's interest in basketball to make the distinction concrete. That's not just reasoning about math. That's reasoning about a person learning math.

How far we've come

I was looking back at my first post on this blog — May 2024, when GPT-4o's vision capabilities felt like a revelation. I was feeding it photos of wildflowers on my hikes and it could identify them. That felt magical.

Fourteen months later, we have a model that can hold a nuanced, multi-turn conversation, reason through complex problems, adapt its explanations to individual learners, and do it all with a consistency that makes it genuinely useful in production. The distance we've traveled is staggering.

And yet — the fundamental excitement hasn't changed. It's still the feeling of this thing understands something. Each generation understands a little more.

What's next

I think the frontier is shifting. Raw model quality still matters — GPT-5 is proof of that — but increasingly, the interesting work is in how you use these models. Architecture around the model. Memory systems. Personality customization. Dream cycles. The model is the engine; the car is what we build around it.

And I've got cars to build.

Live curiously and give generously.