Solving this gap–making retrieval continuous and fluid the way it is for people—feels like the missing piece that would make these systems dramatically more useful.

For models, though, retrieval is a discrete event: pulling a handful of snippets into a limited context window and hoping they’re the right ones. If the net misses, you get the kind of silly, out-of-touch answers that make it seem like it forgot everything.

Humans don’t work this way. Our retrieval is messy but constant. When you’re talking to a friend, the context of your history together just lives in you. You’re not running a search in the background. You don’t have to consciously “retrieve” every fact. It’s fluid.

At any given moment the model has to decide which pieces of your history matter most (and more importantly, which don’t). That filtering is where responses break down, leaving you wondering if it remembers anything at all.

My guess is that this is a retrieval problem. The memories exist somewhere as vectors, but pulling the right ones into the flow of a live conversation is hard.

It clearly does know a lot about you. (If you’ve never tried, start a fresh chat and ask it to tell you everything it knows about you and your situation in life.)

ChatGPT’s “memory” is so frustrating sometimes. It’s like talking to a close friend who knows your whole life story but randomly forgets crucial details mid-conversation.

Shoutout to @adrian_horning_ . Took his web scraping course a while back and finally found a fun side project to work on where I can utilize some of the skills learned from him.

This isn’t some vibe coded side-project thrown together on a weekend either. It’s a big, well-known company. Moral of the story: Security isn’t just about login walls. If your frontend is still bleeding data under the hood, you don’t really have protection.

when I opened the developer tools, the app was leaking data. Like everything I was hoping to find and more. All just sitting there. 🤯

Last night I fell down a rabbit hole while poking around a real estate data site. On the surface, it looked locked down: certain pages forced you to log in, blocking everything with a pop-up. But... 👇🧵

Paradoxically, these claims feel more true and less true than ever before. Yes, you can go from zero to a working prototype faster than at any point in history. But without the right foundation, you’re building on sand.

On the surface it feels plausible, because technically, you can get something that looks like a CRM in minutes. But in practice, it rarely holds up under real-world use.

And it’s not just individuals. I’ve heard a different flavor of this story at a higher level: execs asking their dev teams, “Why can’t we just get AI to build us a CRM in a weekend?”

I’m not trying to throw shade here, I use these tools myself and love them. Just an observation of the gap between promise and reality.

The second you start talking about structure or code, the person looks at you like you’re speaking another language: “Why are we talking about that? AI is supposed to handle it.”

You really can one-shot an app in a few prompts, and people are. But in my coaching sessions, I’m increasingly seeing apps that look “almost done,” except for one broken button or workflow somewhere, with a hornet’s nest of brittle logic underneath.

And even before LLMs started eating everything, I’d often hear from beginners: “This is way more complicated than I thought.” AI is now the new version of that same challenge.

The “No Code” movement has always had a marketing problem. “Build anything without code” was never really true, at least not to the extent it was promised. 🧵

Google's Nano Banana is quite impressive. First time in a while I've been blown away by a vision model. I took this picture and gave it the following prompt: "Put the baby in an astronaut suit and have him sitting on the moon looking at the stars with wondering eyes."