The end of year is a great time for spending time with family, but also perfect for tinkering with new tech. And naturally, this means experimenting with AI tools. I'm writing this note to share my personal experience and a couple of takeaways - take it with a grain of salt.
1. Creating an application from scratch
By now most people would be experienced with using ChatGPT to answer simple questions; unless you are a coder, you probably wouldn't have experienced the magic of an almost completely AI-driven workflow.
At around September last year, news articles were abuzz with the idea that AI could be "agentic" - i.e. programmes could independently perform complex tasks and make decisions to achieve a certain objective.
We are not yet at the point where we would completely rely on a network of agents to run complex systems for us. But for developers, agents are already available to help with coding and application creation and can add massive value, right off the shelf. (Examples of these tools are Cursor, Cline, Lovable, Bolt.DIY and others - but the number of tools keeps growing every day.)
My experience
I decided to try creating something basic but ultimately useful. A charity I support needed a user-friendly but secure way to search through their existing membership database while on the go. On Day 1, I did a bit of basic research in the morning to familiarise myself with some basic frameworks. By evening, I had a working application locally hosted on my computer. By Day 2, I had implemented access control and also deployed my app for use. In terms of actual time spent on the computer, starting from scratch, was probably slightly more than a couple of hours to get a viable application going.
Crucially, I did not have to manually write a single line of code.
For instance, I wanted to add a filter to my existing search form. All I did was ask my AI coder agent (I used Cursor) to make the change for me. You can see screenshots of the app below and I'll include a link to a video out of my laptop so you can see the agent at work.
Some takeaways from my experience:
- The AI is clearly explaining what it is doing, and running the proposed changes by me.
- It's smart enough to go through all of my code and make changes across multiple files, not just the file I currently have opened.
- The ability to do this is not limited to a specific model. While Cursor uses Anthropic's Claude model for this, it would be easy to switch.
- This is not shown on the video, but the agent was able to seamlessly code across multiple frameworks to get the job done.
- Inputs are not limited to text! You can take a screenshot of a website and ask the model to replicate the elements and colours, and generate the relevant code.
Process: Watch the agent at work.
Agents are worth getting excited about
The experience for developers is an early indicator of how AI is going to disruptive. Seeing agents at work right in front of you makes it hard to deny their disruptive potential. But there are still many open questions on how this all pans out.
Does that mean developers are redundant? Absolutely not!
- Experienced developers can still work faster (and more robustly) that inexperienced ones (like myself) that are purely guided by the AI. They know when the AI is going in the wrong direction and can direct it more precisely.
- Further, there are many other aspects of developing that are more complex e.g. quality assurance, deployment, security, developing product roadmaps, etc. that are more complex.
- But it certainly means that the barriers to entry for more basic activities are lower! Frontend work, user interface design and basic data science are likely to be the easiest to build with AI, followed by backend work. Infrastructure management is likely going to be the most challenging - most companies don't make the intricacies of their infrastructure design easily discoverable online!
What happens to software? Satya Nadella has been creating waves with a statement that SaaS as we know it will end - since in his view, SaaS software is nothing more than a database with business logic on top and in the future, AI agents can take over all the operations in these businesses. Very worrying... if true!
- SaaS will continue to exist... as long as companies continue to outsource (and they will). But larger companies may adjust their pace of purchasing.
- AI will absolutely play a role in software. It would be very possible to see a world where customers will require AI agent access for all software (just like API access today). Agents can also be natively integrated into existing software platforms.
- The real risk is that certain categories of software is just getting easier to create (see above). A nice frontend to a database simply will not cut it anymore! I would be worried about software where there is some room for error (e.g. customer support), less worried about mission-critical, security-critical and highly-regulated use cases which would favor more deterministic approaches.
- But for any software company, we will have to ask difficult-to-answer questions on whether the company is future-proof.
2. Other surprising innovations over Christmas
At the same time, the AI companies were scrambling over each other to release new features and updates to their platform. OpenAI teased at a new model called o3, the inclusion of ChatGPT in iPhones, and better capabilities for voice and video generation. Google released Gemini 2.0 which they advertised to be "for the agentic era", featuring lower latency, better ability to understand pictures and speech, amongst other things.
But one surprise from left-field was the release of a large 671 billion parameter "open weights" model called Deepseek V3 which came not from a tech giant but from a Chinese quant fund! Not only was this model trained at a cost of around $6M (allegedly) but it was trained on H800s (i.e. the version of GPUs that Nvidia makes for China) over 2 months. Performance of the model beats or comes on par to OpenAI's GPT-4o. Even Western AI experts have acknowledged the quality of the accompanying tech paper and the engineering advances made by a comparatively under-resourced team. Link As a cherry on top - the company behind Deepseek is offering this model to the world at 1-2% of the cost of OpenAI, Anthropic and others.
I have no doubt that some enterprise users will think twice about relying on a Chinese company. But for many others, the cost-benefit analysis probably makes sense - especially if another third party can host the open source model. Some have dismissed Deepseek as ultimately "copying" GPT4o by training itself on GPT4's output. Perhaps, but this certainly raises the question on what moat any general model has. Technical edges seem to be short-lived - and the real moats (if any) will probably end up being infrastructure robustness, security and model specialization (including whether the model can be fine-tuned on specific data).
The other big takeaway: Chinese engineering innovation is still impressive, even when they are hampered by access to the best equipment.
Conclusion
A common musing at parties nowadays is what we should now teach our kids if AI is able to automate away a lot of what they could do. (This question is all the more salient to me since my eldest boy is going into primary school this year!)
I have my biases - they generally point towards (a) curiosity about fundamental structures; (b) a sense of purpose and (c) resilience.
Curiosity matters. A friend of mine gave me a wonderful quote from Godel, who speculated that:
"It would be a result of great interest to prove that the shortest decision procedure requires a long time to decide comparatively short propositions. More specifically, it may be possible to prove: For every decidable system and every decision procedure for it, there exists some proposition of length less than 200 whose shortest proof is longer than 10^20. Such a result would actually mean that computers cannot replace the human mind, which can give short proofs by giving a new idea."
Read another way - Godel believed that the human mind has a way of discovering new logical truths without arriving at it logically. I believe curiosity about how many things works fundamentally is the only way the human mind does this.
A sense of purpose - being by definition intrinsic - is something that an AI can never give to us, or to society at large. It has also been shown to be highly correlated to happiness.
And resilience - well, the future is going to be tough and tumultuous. These kids are going to need it.