crazy month in AI

sonnet is not useful for many tasks, super slow, expensive
gpt-4.1 is useless
4o is weird, why even use it?
o1/3 all expensive
gemini pro is nice but still weirdly inhuman

meanwhile

Qwen3 is open source, can run on 100k machine, faster than sonnet and performs similarly.

just week ago you would pay much more for proprietary models, but now you can be cash positive if you buy hardware and run open source inference

proprietary models don’t have any moat anymore

Finaly got reduced bills for the solar, it was like ~$900/mo, now it is $250/mo. And i am returning more than a half of the electricity back to the grid and getting abysmal 1/10th of what it is worth. If i will add 2 batteries it would slash it to almost zero. Quite happy now, CA Gov wont take that much money from me. Right now it is ~5y to pay them off, not bad at all.

Tried to translate some russian tech-themed bands using Suno

miscommunication

if travelling - what car to use? Tesla or Porsche?

what board games do you guys like? answer in comments👇

New tiny tiny wheels

vibecoded board game in cursor this evening. it turns out sonnet can consistently produce satirical jokes about tech founders, gpt can create a good images for the cards but impossible to steer geometry. ultimately using cursor to generate cards using react turned out much easier way to create 100+ cards consistently.

superlocal acquired by foursquare!

https://x.com/alexkehr/status/1934653464661119217?s=46&t=A1zE0kfmZyf1QVe55CmQQA

let’s check

made better cards and simplified game, hope to test it soon

now my app suggests me what to do in the world of death stranding

i recently got very sick and i need your help and full support

please, what is a good tv shows to watch?

Couldn't help myself

It seems i can summarize what is the only valuable thing is left really that someone can provide - is the research.

Knowing that something works now costs much more, you cant really ask LLM to find you an insight or find a way to solve some particular problem, this is possible only with some fine tuning and very niche models and i doubt we can reach with generic model anytime soon. And why? How will 1t model would absorb all DNA variants? Specialized model can.

On small scale it is verifiable that multimodal tasks are still lacking in LLMs. We cant beat sota TTS models using LLM, we cant beat visual recognition either.

Meanwhile there are so many specialized knowledge exist that is not accessible by LLM.

On large scale, knowing how to cook apps is what really still valuable while low level monkey coding is not.

Happy Coder: A Mobile Client for Claude Code

Last week, we hosted a hackathon at my place with a simple goal: build a mobile client for Claude Code that we could use on the go. What started as a weekend experiment turned into something we're genuinely excited about. We've submitted Happy Coder to the App Store and are awaiting review. In the meantime, the project is fully open source—you can build and run it yourself today. We started with Kirill, but it quickly blown up to a 4 people.

Key Features
- End-to-end encryption - Your data stays private
- Native mobile (and soon desktop) experience - Built specifically for mobile
- Open source - Audit the code yourself
- Self-hostable - Run your own server or use ours

Contributors
- @bra1ndump - Turned our initial joke into a working prototype
- Thiago - Architected the packaged everything together and crafted beautiful UI components
- Karl Marx - Built extensive Tool Calling UI


Check out the code and documentation on GitHub: https://github.com/slopus/happy

---
Built with ❤️ during a weekend hackathon

Here is iPhone link!
https://apps.apple.com/us/app/happy-claude-code-client/id6748571505

And in another news, the company i was helping to build, Bee, just got acquired🥳.

https://x.com/JoannaStern/status/1947696733128995103