Model Takes

Why Learn About Neural Networks Today?

Taran Rai — Wed, 01 Apr 2026 19:13:23 GMT

Yesterday I gave my annual intro to neural networks lecture to a lovely cohort of MSc students, and this year prepping for it got me thinking about how much the AI/ML space has changed recently.

We can now vibe-code software, websites, and even fairly basic apps. We can use LLMs to act as copilots in development, and even run AI agents that handle operational tasks, usually with some form of a human-in-the-loop.

These models can do a lot, but I posed a question: why bother learning neural networks at a level that requires understanding basic concepts, even with some maths, at a toy problem level?

My answer is simple. It keeps you grounded.

There’s a lot of magical thinking in this space. People anthropomorphise their favourite LLMs or take LLM outputs at face value because they sound plausible, especially in domains where they don’t even have enough experience or expertise to challenge them!

I usually suggest to folks (closer to the field) that if you want a feel for what’s going on under the hood, then train a basic neural network on a small problem yourself. This will force you to face their limitations directly, rather than treating the entire thing like a magical black box.

You’ll see how much effort it takes to train a model from data preparation through to choosing appropriate evaluation metrics. You might end up with a good model, but the inconsistencies will be glaringly obvious.

Now imagine scaling this to the gargantuan models we have today. Even though LLMs aren’t directly comparable (their capabilities are indeed super impressive), the underlying fundamentals in training don’t just disappear. They don’t go away but become harder to see and reason about. Going through this process may give you a much clearer appreciation of that.

It’s also difficult to predict where this space is heading in the next 5 years. There are a lot of conflicting opinions, even among experts. Scaling models may only take you so far, and if that’s the case, then a different paradigm and direction may need to be undertaken.

And how does that happen? When we go back to the drawing board and revisit the basics.

Advanced Voice Mode??

Taran Rai — Sat, 17 Aug 2024 19:34:42 GMT

For the past few weeks, there have been a number of posts showcasing the latest capabilities of OpenAI's ChatGPT's Advanced Voice Mode. Right now, only a select few users have the privilege of accessing and playing around with the latest update1, and some of the demo’s are pretty impressive. Here is a snippet from a longer clip by @CrisGiardina on X that really stood out for me:

After watching this clip, I thought about how I could use this feature to enhance my own Punjabi language learning. My ability to speak Punjabi is poor - I can mostly understand what is being said to me, but I often struggle to reply back with coherent sentences. This has always been the case when speaking with my Nani who isn’t as familiar with English. I pulled out some Punjabi books from my youth2 but quickly found them to be cumbersome and static, so I'll probably switch to LLMs for a more practical learning experience. When this version of advanced voice mode becomes available for all subscribers, I’m going to make the most of it by practising Punjabi speaking exercises.

LLMs can make excellent teaching companions and have the the ability to create a bespoke personalised learning experience. Adding a realistic low-latency voice mode enhances the utility of these models even more. However, learning languages is just one use case. I’ve already used the current voice mode on ChatGPT as a knowledge companion when going for lengthy walks ~~or drives~~3. It’s kind of like having a limited version of J.A.R.V.I.S. from Iron Man in your pocket4.

As of August 2024, although, apparently, OpenAI plans to enable it for all paying customers at some point this Autumn.

Long story short, me and my bro used to attend Punjabi school when we were kids but were thrown out because we never took it seriously. There’s a letter in the Punjabi alphabet called ਙ “nganngaa” which kind of sounds like the Punjabi word for “naked” (“nangā”). This used to crack us up all the time (we were kids man). Nonetheless, it was the last straw for the Punjabi teacher.

Crossed out just in case someone has a problem with it 😉.

I say “limited” because these LLMs currently can’t perform many actions on your behalf - like calling out your Iron Man suit.

When writing with AI-assisted tools, please actually check your outputs

Taran Rai — Sun, 28 Apr 2024 16:43:26 GMT

The other week I came across a LinkedIn post exposing the obvious use of large language models (LLMs) like ChatGPT in academic publications clearly without any (or minimal) checking of outputs.

There are plenty of examples you can find online (just search LinkedIn and X). Not only is this embarrassing, but it also begs the question on how something like this could have passed peer review.

By now, most of us should know that outputs from LLMs have the potential to hallucinate or confabulate; that is, they can fabricate details or produce absolute garbage unrelated to the input prompt. However, in my humble opinion, it's acceptable to use AI tools like LLMs to assist in writing articles only when outputs are double-checked... But let’s not take the p*ss like the examples above…

Hello World! 🤖

Taran Rai — Mon, 29 May 2023 22:09:45 GMT

I have been thinking about Keeping Humans in the Loop for a while now. Interest and excitement around Artificial Intelligence (AI) has skyrocketed in the past year thanks to chatbot apps like ChatGPT and text-to-image generation apps like Midjourney. However, with this surging interest is the inevitable hype cycle and sometimes unrealistic demands made on this technology.

It’s been almost a year since ChatGPT’s inception and we have seen some pretty wild developments, including attempts at building AI agents (a system that can “think” or make decisions), the creation of several opensource Large Language Models (LLMs) that can give ChatGPT a run for its money (for certain tasks) and even text-to-video apps like RunwayML.

We have also seen a rise in backlash against the use of AI. Protests have broken out in Hollywood over AI potentially being used to write (or edit) scripts being used to take the digital likeness of an actor without informed consent and fair compensation. Significant voices such as Geoff Hinton and Yoshua Bengio who are both considered “Godfathers” of AI, have warned us about the existential threats of AI, even causing Hinton to leave his role at Google to speak more openly about “dangers” of the technology he helped to develop. Mustafa Suleyman (founder of Inflection AI and co-founder of DeepMind) has warned us about the “containment problem”, which describes how humanity has never really successfully contained a technology, where we may be engulfed in an AI tsunami rather than a wave, causing disruption like never seen before. There was also the proposed six-month pause letter that was circulated earlier this year, aimed at fostering a collaborative effort in slowing down AI developments worldwide (wishful thinking).

I have mentioned hype around AI, although, this time it feels different. There is something about this AI wave that cannot be dismissed as merely a trend. Maybe it’s the constant developments that are almost happening weekly, or perhaps it is because I too am slowly being influenced by the techno-optimism that surrounds me daily. Whatever your thoughts and opinions, we simply cannot turn back the clock. This technology at least in its current form is here to stay, and we (humans) have much work to do in determining where it fits in to give us the most benefit.

I view technology as a means to empower humanity and improve our lives. In particular, I see AI as a tool to augment our abilities, to essentially remove the repetitive nature of some undesirable tasks and to enhance our understanding of complicated topics. This view considers us humans as being a part of the loop. When it comes to AI technology, ensuring humans are in the loop not only complements our capabilities but also prevents the black-box nature of the current AI algorithms from blindly overtaking our decision-making.

A little about me: I obtained a PhD in AI in digital pathology in 2022, where my focus was on deep learning for necrosis and mitosis detection in canine sarcoma biopsy images. I have worked as a data scientist at the Veterinary Health Innovation Engine (vHive) and have collaborated on some exciting data science projects with animal health focused pharmaceuticals and massive conglomerates. I now currently lead applied AI initiatives at Surrey DataHub where we are focused on researching, developing and integrating AI technologies thoughtfully and responsibly, ensuring they drive real-world outcomes. I will use these experiences as a data scientist, applied AI researcher and fellow human being to share my thoughts and sometimes provide breakdown explanations for some of the complicated topics that are trending. To be honest, I will also write these posts as a sort of archive for myself. I have the tendency to jump between topics, read several articles and then just forget about them or hope they just appear in my head when the time is right. I can already tell you that this rarely happens. In fact, you should see the number of bookmarks I have saved on LinkedIn and Twitter for “future” use.

For now, let’s remain in the loop.

Taran

[Following a trend on Threads, where Midjourney generated this image based on my Threads bio fed as a text prompt: “applied ai and data science, phd in ai and digital pathology”.]

Subscribe now