Thoughts on AI

 

Andrej Karpathy, during his podcast interview with Dwarkesh Patel. October 17, 2025.

Hype.

In the midst of all the hype around how AI is going to take all our jobs and Elon Musk telling us to give up on the idea of humans producing code in 2027, it's refreshing to hear saner voices making more nuanced statements and predictions about AI and our future. 

Andrej (pronounced Andre) Karpathy is one such voice. I will talk more about him a bit later in this post.

While AI has produced some incredible results, there have also been plenty of failed predictions. Geoffrey Hinton (considered the godfather of AI) predicted in 2016 that AI will completely replace radiologists by 2021. Here we are in 2026 and radiologists are flourishing like never before!

We live in a world that is undergoing disruption by AI just as it was once turned topsy-turvy by the inventions of printing, electricity, personal computers, the internet, smartphones, and CRISPR.

Each time hype-mongers told us that everything is going to change. Each time many things changed and yet many other things stayed the same (we still put our pants on one leg at a time). Many hype cycles have come and gone. Take for example the failed predictions around virtual reality and 3D TV for home or blockchain as the new internet for everything. It's not that the technologies weren't real. Just that some folks were too eager to generate hype and cash in on the new craze.

For perspective, I took an AI course as part of my computer science degree curriculum in the late 80s. That was four decades ago! The subject of my project was "a rule-based AI inference engine," written in Prolog. The point being that AI is not new. It has been around a very long time. AI has gradually been getting better and the latest advances are indeed awesome! Instead of hand-coded rules-based inference (my college project), we're now doing data-driven, learned models. And yet, a lot of the predictions about the imminent AI takeover are likely overblown.

LLMs and the AI Landscape

LLMs.

AI is about a lot more than large language models (LLMs). But LLMs have found some broad applications and are a huge advancement over the rule-based systems and simpler statistical methods that were used previously for natural language processing (NLP). 

This jump to a somewhat broadly accessible application is driven by a combination of deep learning methods applied using neural networks, self-supervised learning-based pre-training (it's the P in GPT), and a transformer+attention-based architecture (it's the T in GPT). 

Wow, that's a lot a jargon. But the key point here is that the current LLMs are just text predictors -- and text is not everything. For example, there is presently no Copilot for creating PowerPoint slides with visuals arranged according to my PowerPoint templates and brand guidelines or styleguide. In other words, if you're thinking AI has reached its peak, that's not at all true. AI has a long way to go -- so buckle up and enjoy the ride.

Context.

There's an excellent book by Pedro Domingos from 2015 that introduces the reader to what the author refers to as the five tribes of machine learning algorithms. Pedro argued in his book that the master algorithm will be one that takes the best of each of the five learning approaches. 

So I asked ChatGPT how that prediction has panned out in the case of ChatGPT. (Note that LLMs, which arrived in 2017, are a kind of ML algorithm.) ChatGPT's assessment is that today's LLMs exhibit a: 

connectionist/neuroscience-inspired backpropagation and gradient-based optimization core, 

Bayesian/probabilistic objective, 

analogizer/support vector machine (SVM)/supervised learning-like behaviors,

symbolist/inverse deduction scaffolding, 

and a minimal evolutionary role.

So, ChatGPT (a type of LLM) does sort of leverage all five tribes to varying degrees. And the book is a valuable introduction to competing algorithms. But the thesis is considered flawed because Domingos pushed too aggressively for the idea that a single algorithm combining all five tribes will ultimately be the winning master algorithm. Maybe there is no single best learning method. The best method might depend on the use case. And, of course, Domingos might still turn out to be correct in the long term.

Education.

Karpathy recently released a 200-line Python MicroGPT that serves as an incredible educational tool to understand how the GPT algorithm works at its most fundamental algorithmic level. It's also a good way to verify which of Domingos' five tribes are most represented in the algorithm that is considered one of the winning algorithms of today. (Of course, that is not to say that the other tribes aren't going to end up winning in future versions of the master algorithm.)

Evolution.

If we define intelligence as the algorithm or the ability to process facts and knowledge as the facts, then evolution has designed the brain or the algorithm as a result of natural selection (aka survival of the fittest). 

Humans are born with a fully designed algorithm and no data. To that extent humans are similar to LLMs in that they learn as they encounter data (examples or what to emulate and what not emulate). However, LLMs haven't figured out the natural selection bit. Poorly performing LLMs don't die and go away. They keep trying. Or perhaps they stop getting funded and do get replaced?

However, when we say LLMs are not evolutionary we mean there's no cross-pollination and mutation happening between LLMs over generations that is finally resulting in a winning algorithm. (In this context, Sapolsky's discussion on neural Darwinism is also interesting in view of how the human brain implements selection as it matures.)

Limitations.

When Generative AI (GenAI), of which GPT is an example, first became a thing a few years back, one of my early thoughts was that this is just plagiarism or cheating or copying. And it is. GenAI can only create new images based on what it has seen before. So, if our use of GenAI ends up putting creative producing humans out of work, GenAI will run out of new art to plagiarise. And that will be the end of GenAI.

During a podcast interview with Dwarkesh Patel, Karpathy was asked if he tried AI-assisted methods to create nanochat, which he describes provocatively as "The best ChatGPT that $100 can buy." His response was that AI coding tools are very bad at writing original code, i.e. coding paradigms that have never been written before (e.g. nanochat). And that's because they can only emulate what they've seen before. They're not smart enough to create anything new.

My Take.

I initially embraced Claude, which I still like for non-technical topics such as history. I wasn't happy with the graphing capabilities in Excel and wanted something I could customize and control. So, I used the conversational approach to build a 4,000-line data visualizer that takes any data in CSV format and graphs it. I prefer the conversational approach to a more intrusive approach (e.g. repo-aware) wherein I feel that the developer loses control over the direction of the end product and often doesn't quite understand how the code was put together. 

Later I experimented with Gemini, Perplexity, Copilot, and others before settling on ChatGPT for technology related work. Claude's training data usually lacks the most recent 6-9 months. Compared with Claude, ChatGPT is more proactive about leveraging retrieval augmented generation (RAG) to bring it's knowledge up to date, which is critical when working on cutting edge technical topics.

Reinforcement Learning.

RL was famously credited for the 2017 DeepMind AlphaGo victory over the world champion. But Karpathy is very critical of RL. He’s pointing at a real, classic RL pain point: credit assignment with sparse, delayed reward. He describes RL for LLMs as "sucking supervision through a straw" and complains that if you only get a single reward at the end, that signal gets spread over the whole trajectory of actions taken to get to the end -- so you can end up reinforcing junk steps along with the good ones.

As an example, If I need to move to a point three steps behind me and I end up doing so by walking around the block (taking three right turns), vanilla RL would credit the entire trajectory of my process because I got to the correct result.

Entropy.

Model collapse is when a model trained on its own (or another model's) generated data starts to lose diversity and drift away from the true data distribution over generations of training -- eventually producing bland, repetitive, lower-quality outputs. Learning reaches a plateau. 

Another symptom is overfitting (memorizes training distribution, generalizes poorly). This is why talking to other people is considered healthy because it increases entropy (uncertainty, disorder).

Can facts be distracting?

In the same podcast, Dwarkesh asks a wonderful question about something we're all aware of. We remember very little of our childhood even though that's when our most important learning was happening. Does the brain purposely avoid collecting memories until the child's brain has matured to a point? 

Humans in general are bad at memorizing. This is why they try to find patterns rather than memorize facts. We know how to get to our favorite restaurant even though we may not remember the exact cross streets. Karpathy thinks LLMs should bring their cognitive core down from 1 trillion parameters to 1 billion parameters to become more effective. In other words, get closer to the human ratio of what to remember and what to lookup as needed.

Is AI revolutionary?

I've always been amazed when some of my clients object to using AI. They might say we don't allow AI for transcribing or summarizing meeting notes. That always makes me cringe. It's like saying I don't do online banking. I'm going to ask my employer to send me physical checks and I'm going to schedule an hour every Saturday morning to go to my bank branch and physically deposit those checks. Not realizing that those deposits are held in that same online system you're refusing to use.

If you're that particular about not using a certain technology, then you first need to understand that technology and how it's being used. In this example, the online-banking-averse person should then immediately convert his/her cash into gold (or similar) and put it in a safety deposit box.

What clients or people refusing to use AI fail to understand is that AI isn't revolutionary like the invention of electricity. It's on a continuum of various computer science advances that have been happening since at least 1948. An LLM is just another algorithm. So, where do you draw the line on what's AI and what's not AI? 

Unknown to many clients (and folks) we've actually been using AI for spam filtering, data anomaly detection, managing credit card fraud, optical character recognition (OCR), auto-correct, sentence completion, web search, and much more at least since 1950s. Except that no one referred to these technologies as AI and equally I don't recall anyone complaining about their use in personal life or corporate settings. So, I was heartened to hear Karpathy say that AI is really just one more advancement in CS; it's not anything radical like the invention of electricity.

GDP.

The big question that Dwarkesh asks repeatedly (too many times, I think) is will AI result in a 10x per-capita GDP growth in the developed world similar to the industrial revolution (which increased GDP from 0.2% to 2%, according to Dwarkesh). Karpathy's response (also repeatedly) was that the industrial revolution did not result in a discrete 10x jump (it happened gradually over a long period of time such that it was not attributable to any one thing) and AI will be no different.

Karpathy.

To me Andrej Karpathy is a cross between Richard Feynman and Salman Khan. He has a rare ability to simplify things to help you understand. Eureka Labs founded by him is the new Khan Academy. He feels that the way AI is going right now is disempowering humans and the fix is better education. I totally LOVE how he talks about education as the not so easy job of "building ramps to knowledge". He favors a one-on-one tutoring style that makes each student feel neither under-challenged nor over-challenged, which is uniquely possible to deliver with AI. It's a laudable goal and I wish him all the success.

Closing.

In closing, it's easy to see why software for financial services companies, governments, trading platforms, e-commerce, or any non-toy application where the cost of failure is high (e.g. leaking PII) is not going to be developed using AI for a very long time. 

As Karpathy says, sometimes the biggest value-add you can offer to a client is to advise them not to use AI for a specific use case. It's important to understand the current state of AI well enough to know when to use it and when not to, so that you don't overprescribe it in the to-a-hammer-everything-looks-like-a-nail fashion, thereby "calibrating yourself to what exists".

You might say that humans are perhaps even more likely to make mistakes than AI. That might eventually become true even if it isn't yet. But we can background check and interview humans. We can't know if or when an AI decides to go rogue. It's great to see all of the advances, but let's maintain some perspective while we get all giddy about it.