Alex Selimov
7bb4b03b67
All checks were successful
Build and Deploy Hugo / deploy (push) Successful in -17m5s
53 lines
3.3 KiB
Markdown
53 lines
3.3 KiB
Markdown
---
|
|
title: "Misc thoughts on AI"
|
|
date: 2025-08-01T07:45:26-04:00
|
|
topics: ['AI', 'Software development', 'LLMs']
|
|
|
|
---
|
|
|
|
## AI has become a very useful tool for me
|
|
|
|
If you know me personally, you know that I've been pretty slow on the LLM/GenAI uptake.
|
|
At this point, however, if you are a software developer that isn't integrating GenAI tools into your workflow I think you are probably going to be left behind.
|
|
I've been playing around with these tools a lot more lately, and I've come to a few conclusions.
|
|
|
|
|
|
1. Models that can be run locally (I've used a lot of different Ollama models) are mainly just ok. The best results I've had is with Gemma3:27b on an m4 macbook. Haven't tried qwen3 coder yet.
|
|
2. The amazing metrics you see referenced require 100b+ models that are not going to be feasible to run locally for most people.
|
|
3. Your application matters a lot. I find that these tools are extremely useful for java spring boot microservices but a lot less useful for scientific projects in Rust.
|
|
4. I really enjoy using the expensive LLM tools provided at work and enjoy using the free version of tools less.
|
|
5. Claude has given me the best performance so far.
|
|
|
|
I'm still debating which subscription to get for personal projects.
|
|
I know Claude is my favorite, but it's also much more expensive.
|
|
I also have a newborn right now and not much time for side projects, so I'm delaying the decision.
|
|
|
|
|
|
## AI has NOT become an irreplaceable tool
|
|
|
|
I am not a vibe coder and don't think I will ever give in to the vibes.
|
|
Even when I prompt claude code to make changes, I always know exactly what I want done and how I want it done.
|
|
Many times I'll make the changes myself because it's faster than trying to prompt Claude.
|
|
LLMs have definitely improved my efficiency, but I would be fine if all the GPUs in the world suddenly melted down.
|
|
I'm not a viber for a couple reasons:
|
|
|
|
1. I enjoy coding and problem solving.
|
|
2. I'm not convinced models are capable enough to be fully trusted.
|
|
3. I want to keep my own skillset because the future of AI is uncertain to me
|
|
|
|
## How does the economics of AI work?
|
|
|
|
From my very limited understanding, the fact that major AI companies are still requiring funding rounds indicates that AI is currently not profitable.
|
|
That makes sense to me since ChatGPT and Claude are available for free even though inference is certainly not free.
|
|
I think the significant portion of cost is due to model training.
|
|
It seems like model training costs won't be stopping anytime soon as companies race towards general and then "super" intelligence.
|
|
I'm not convinced either of those are a given.
|
|
As a result, prices will have to increase or ads will be introduced to recuperate investor money.
|
|
I'm also not sure how this applies to providers of open weight models.
|
|
Those prices can seem more reasonable, but are those subsidized as well? Finally, the current datacenter power consumption for model training seems entirely unsustainable.
|
|
If I recall correctly, Elon Musk is powering some of his datacenters with a constant flow of diesel fuel into generators because he can't get enough power from the power grid.
|
|
The scale of everything at the moment seems to border absurdity.
|
|
I think we are in for a bumpy ride for the next few years as the impact of GenAI technology and infrastructure works itself out.
|
|
|
|
|
|
Please let me know if I'm wrong about anything or with better predictions!
|