#ai

TIL about Hinton’s Paradox where Geoffrey Hinton’s prediction in 2016 that we won’t have any more radiologists in 5 years isn’t just true but also that we have now more radiologists than ever. And of course, AI has integrated into every workflow of radiologists.

And this reminds me of another fantastic article that I read last year about how Cancer has a surprising amount of detail by @owl_posting. The world has fractals all around us. As we zoom in further and further in whatever we do, we get specialized and we collectively improve what we as a species can do and our expectations of ourselves get better. AI helps address the grunt work and a smart human uses AI to automate the grunt so that they can focus on the more important stuff to advance ahead.

Those are the people that survive.

The first casualty to agent driven development is Tailwind which is laying off 75% of its workforce since traffic to its docs have reduced, in spite of it being the most prominent CSS framework. This is not good and there are many more in the boat and we’ll see a few happening through 2026.

I also chanced upon this linkedin post by marcj who is switching entirely to closed source. And I don’t really blame him or any others who take this route.

Tailwind might have been bailed out by Google but is this sustainable? Sponsorships work but they don’t accurately capture market demand and is often just adequate to keep the ball rolling.

Gating agents to pay before accessing a specific framework would work and Stripe has support for this too but this is akin to getting a subscription to Netflix and then pay the producer of a movie a small fee before starting to stream.

A better alternative is to have model providers who train LLMs foot the bill. As much as I don’t like it, we’ll get to a point where LLMs will also prompt with ads in say comments or metadata that is spit out for humans to be aware of paid services that a given framework offers. At least, these will be scoped and not targetted or blanket ads like a search engine would surface.

We’re slowly getting into the territory of LLMs being a public good. LLMs are incredibly efficient in many every day usecases and we’ll get to a point where it would likely be hard to deal with life’s chores without a LLM. At which point, would LLMs be state-owned? Or is there a future where LLMs will be private owned but everyone would have access to one through subsidies?

Interesting times.

I love all the wonderful stuff that people share on socials about what they built over the holidays. AI assisted coding has certainly opened up the possibility that people who have a specific itch and had no idea how to solve it now have this magical tool to solve their usecase.

I’ve long enjoyed writing simple CLI tools with the UNIX philosophy in mind since I started writing code. These could be as simple as a function my shell rc or could be a long-winded utility script that gets invoked on the terminal.

But I’ll be honest - I suck with frontend. Not for lack of trying though. But LLM assisted coding has been a godsend for me personally when it comes to building tools with nice UX. I still default to writing CLI tools but there are many problems that are better served by a better UX. More importantly, these tools help me learn UX frameworks and patterns.

What’s yet to be seen is the impact this will have on the indie hacker economy. A lot of people make good money by writing a piece of software that solves a particular problem. When the cost of building software goes to 0, these are likely the first people to get affected. What stands between them and a user of their tool is sheer will and a few hours to build what they want.

My only hope is as people build more tools, they open source it as well. The world is better with an OSS renaissance and who knows, this wave of AI tools will usher us in it.

This tweet from Andrej Karpathy has been doing the rounds and I felt this quoted tweet from rahulgs was on-point based on my experience working with LLMs. The frontier is moving quickly for writing code at the least and the more harnesses and tools you provide for your engineer and the more context you provide your models, the better it is going to get over time.

Security is still an very important and sadly overlooked part at this point.

Of course, everyone is talking about this tweet by Andrej. I do feel this at times - there’s just a lot to keep up and honestly the best way I’m making progress is learning a thing or two every day and accepting that I can’t be knowing everything in one-go and trusting my future self to figure it out.

Having a bunch of pet-projects/problems to work on and putting these tools to use to work on these helps over time.