Smarter Rats vs. Artificial Brains: Which Path Will Make Us Evolve?
Read on Substack · 2025-06-09
Full article
I got tired of ChatGPT. Then I realized why that matters for all of us. The choice between tools that amplify us and tools that replace us.
I've been using less and less ChatGPT in the last 3 weeks. So much so that I thought of not renewing the pro subscription anymore. Although the thought of not being able to access the memory on hitting the submit button after vomiting my thoughts and doubts on the chat window to get an answer that'll make me think if I should just ask myself the same question and realise that I'm okay without the GPT, made me pass. But back to the point, I used it less and less.
Does that mean I'm over my AI hoe phase? No. I'm still scrolling on YouTube / X recommendations about new tools in the market, an AI that'll listen you all day secretly, and AI that'll create shopify sections for you, a revolutionary AI tool that'll search the "ENTIRE" web for everything I have to ask. One of the above I started using and god it's so helpful. More on that later.
Why I Started Using ChatGPT Less
Why did I end up using less and less ChatGPT? Is it because I bought the pro version of Claude? Maybe. Maybe not.
Like every other power user, I started seeing through its responses. It made more and more sense to me how generic it is, and how there are certain instructions given to the chat app, are also, generic, but there's some sort of direction in them. Like 80% of the direction given by engineers, and the other 20% with my 'instructions' + whatever memory I've been able to build on the app.
But so what? It still has access to more than you could ever search on the app, you still need its help to solve your small and big doubts about how dating apps work (that's what I was interested in few weeks back, when I stopped using it).
The Memory Problem That Changed Everything
Have you ever noticed, how the app builds on your chats, your prompts, there's a small memory module that keeps getting updated, with every chat. Very early on, mine got full. I thought it's normal (and it is, no surprises there). But, what I realised later, that even Open AI's LLM stores everything but can't possibly access everything. Humanly and now, artificial humanly also, impossible.
So when my memory got full, or to put it more accurately, my memory built with a highly compressed version of my chats, got full, I had an option to remove few of those saved excerpts, to make space for the new ones. When I read those, things started making sense to me. Like how when in my mind, I asked a question on the chat, it would have gone to our chat about a relevant topic 45 days ago, picked up everything and added it to my current question, and thought about it, looked up relevant things, frame them perfectly, and give it back to me, in the way I must've mentioned in the prompt. For unlimited times. For $20/month. I realised how dumb it sounded. Even though when today all expectations are broken with the revolutionary product launch reels we're seeing on IG, it is still dumb of me to believe that.
Send this to your friends directly from here, would love to get more eyes on this.
The Real Question: What Do We (I) Want From AI?
Back to my point, do we want more contribution on how our AI agent is responding to us? Like a 60-40 maybe instead of 80-20? Is this thought and its solution as linear as just increasing the subscription cost from $20 to $40? Would it be that simple? The tech gods on my reading list say otherwise. And general logic says otherwise too.
Clearly using ChatGPT has made me dependent on my screen's text more than my intuition and what I'd like to call, in this context we're in, my own memory. So it wouldn't be linear. To be honest, I don't know what it would actually be. Although I know what helped me afterwards.
Why AI Shouldn't Be a Generalist Like Us
It’s very much visible that I'm not in the phase of "1 prompt of mine could have powered a bulb or two in my room, for an hour". But the general response to my questions, that becomes complex with every prompt, doesn't help. And being overly appreciative of how complex my question definitely doesn't help. Sorry GPT 4o update I know we had a moment or two where you almost fell in love with me for my questions.
I was talking about how a small change in my routine has a cascading affect over other habits of mine. Maybe I was wrong to do that with this tool
What helps? A couple of things. Humans are now being appreciated for being Jack of all trades. Generalist. Multitasker across multiple domains. I don't think we want the AI agents to do that too. I don't think it helps. If we're going to ✨co-exist✨, we'll have to be opposites. This is highly subjective. Not my opinion, but the level of how general a generalist can be and how specific a specialist can be. But for now, we'll stick to the can co-exist if we're opposites.
If the goal is to make AI assistants, the best one, for many tasks, it has to be freakishly good at what we're expecting it to be good at. That's where I ended up looking outside of Open AI (hello claude code). That's also when I realised being vanilla and being flavourless are 2 very different things. You'd think opting for flavourless would give you so much of freedom. But our minds aren't made to think. They're made to avoid thinking. As much as it can. (Book: Why Don't Students Like School, by Daniel T. Willingham) More on the open-source bit in the next post.
What else helps? Being direct. This helps for problems outside this post as well. Especially when the agent is being used as an agent. And not a 60 year old wise librarian who has read all the books in the room. It's kind of annoying when the agent tries to become your bestie. Giving unsolicited advice just because it "knows" you. Apologies for making this paragraph a rant. But being direct really helps. Claude gets that. 100 points to Claude for that. I do appreciate the assumptions that are taken while giving an answer. But they do more harm than good. A wrong assumption is worse than a follow up question. Especially when it's backed by a limited and summarised memory of your life.
Enough of the comparison between 2 tools.
The Core Problem: Lifelong Learning vs Reinforcement Learning
Do I expect the tools to get better? Yes, overtime they'll make this post look like a stupid rant, I'm hoping a year or two. But, do I expect the tools to become specialized generalists? Like us? I don't. Kind of don't want that to happen. Maybe I'm delusional. But as of today the facts say so. In better language, I believe it's called learning that's continuous. Lifelong Learning. Something that fits perfectly against Reinforcement Learning. Many AI Apps are being optimised for RL and not Lifelong Learning. They suck at lifelong learning. I'm sure they'll get better.
What is Lifelong Learning?
Simply put, it's all your lifelong learnings, and based on the experiences, you retain most of it. The meaning of retain here is way way bigger than the retain that's being used with AI context. It means that when you go to a medical school to learn about human body, you don't end up forgetting your childhood and your mom's favourite TV show.
What is Reinforcement Learning?
Agent's ability of learning to maximize rewards through trial and error. When you put a rat in a maze with cheese at the end, it tries different paths, gets rewarded when it finds the cheese, and gradually learns the optimal route. The rat's goal is simple: get cheese faster. Current AI is like a very smart rat that can only do maze-running, not actual rat-like general learning (like remembering an old rat friend).
You see how great AI is at this? Pattern matching at it’s peak.
What Makes Us Human (And I’m Grateful)
Us humans don't have to learn how to do lifelong learning. And now to the next point that is, our senses. In no way I mean that understanding and processing text is not enough. But, humans being able to see, hear, touch, taste, and smell. I'm in no position of knowledge to comment on how and when an AI startup will jump onto this, or when a tech revolution like this will come for startups to jump on it. I'm nobody to say that. Really. But, just makes me feel quite grateful. That's all this paragraph is for.
Billions will be spent, compute, energy and many other resources, for the computer to be able to see everything we see, and understand what they see the way we do, and then to put it to words what they understand after seeing, the way we do. What our brain does in whatever the unit below microseconds is.
The Central Question and My Bet
What question I'll be thinking about for next few days. Do we want smarter and smarter rats, or more artificial brains, imitating us? If the goal is to evolve and evolve, which one will help us do that?
One more realisation in the same direction is, that AI/LLMs are getting better at things like code generation, image generation, tutoring, and helping humans process text, faster. If I'm framing it correctly, it can pattern match BRILLIANTLY, and generate plausible responses. Retrieving and recombining information is where it's the best option. Out of the list mentioned above, one thing that's missed is Reasoning. That's a loaded word can't deny.
Over time, we've seen the tools to get better with code, images, text processing. With every model, the skill, and more importantly, the speed, is getting significantly better. Although the 'reasoning' is in its phase of diminishing returns. AI being able to push out efficient responses, very quickly, is different from the quality of reasoning it's doing. That's where it's struggling.
Will This Be Solved Next Year?
But wait, it has gotten better in last 24-36 months. Maybe next year this will be solved too?
I believe the only way it has become so intelligent and better at reasoning is because of more and more compute. Even if we keep money, land, water and compute aside (we can't), the approach of 'pattern matching' is not something that'll solve for human-like reasoning. Human-like. It won't be sustainable. It's not a buzzword here, it really won't be.
What This Means: AI as the Perfect Sidekick?
What it does solve for, is a great sidekick. For tasks that require 1) Pattern Matching and 2) Reinforcement Learning. There are quite a few. Many actionable prompts come under this umbrella. Write a code snippet that does X. Create an image that looks like Y. Z company did 45 Mn of revenue last year, how much did its competitors do in our neighboring countries?
Managing Expectations, Not Making Predictions
This choice between smarter rats and artificial brains matters for human development because it helps us manage our expectations better. I believe before planning, our expectations cloud our judgment. What might seem like the perfect button to double click on, may not be the button for all our problems.
I'm at a position where my words will not change the course of anything. It's just a brain dump. Whatever I thought of, it's here. But the thinking process matters. Instead of urgency about collective decisions, maybe we focus on individual expectation management. Being honest about what we're actually choosing between.
This is where my thinking has led me so far. The question remains open, but at least I know what I'm actually choosing between. If it reads like I'm leaning toward "smarter rats," maybe that's genuinely how I feel - but I'm not trying to convince anyone else. I wrote this because this is how I think about it right now.
Connected ideas
These links connect the published piece to working idea notes. The article text above retains its published wording.