Default
Product process
PopTranslate
I finally finish this project, it works on my computer and my gf computer, I like this project.
Sad news is my review for this extension was rejected two times by google, I didn’t know the reason ans ask for further review.
Other product attempts
1) Text-to-voice in-browser
After know about transformer.js which aimed to bridge the gap between web and LLM, I tried this method to realize this project, but it failed again. I also meet some errors that I cannot solve in short time. In terms of the reason, transformer.js is not mature enough, more deep knowledge and techniqu in audio field are necessary to solve it. Maybe I will do that if I have free time in the future.
Consequently, I hold this proejct again.
2) Analysis 500 startups supported by YC Due to the lack of fresh idea, I analyse the 500 startups supported by YC in 2024, some conclusion as below
- 50%+ is LLM or AI related based on title, real value is higher than that.
- 64% is belong to B2B category. This is majority area of AI application.
- 10% is belong to healthcare categoty as the second biggest part.
- Education is smallest category, occupying 1.2%. There will be more opportunity if AI became more accurate.
- Projects that are great and suitable for me
- unsloth: LLM finetune accelerate
- gpt-pilot: Making AI deveop like a real developer, this will be the next stage of AI code
- transformer.js: bridge the gap between web and LLM, from huggingface
Request my first PR for this project and got merged, meanwhile, it’s the 1000th PR, just a coincidence.
The open source version of this project can accelerate most of open source models finetune stage 2x, pro version can provide more support, best performance is 30x. It worths more exploration.
Daily
- 9.2: PopTranslate
- 9.3: Publish PopTranslate
- 9.4: Reading
- 9.5: Reading
- 9.6: Try to reproduce a specific issue of unsloth
- 9.7: submit PR of unsloth. Try transformers.js
- 9.8: Rest
Personal life
Reading
- limu’s speech in shjt universion: recommendation for five stars
- AI tools help you grow fans
- Greppability is an underrated code metric: Useful and practical for developer, especially for how worked in big company
- Interest-base community have great growth: I have the same sense, interest is the final way to socializing
- Cost of LLM will continue to decrease – Andrew NG: Moore’s law will work in the latest field.
- Raise 1B with one html – ssi
- Photo has more power than text in terms of creation
- Next product of openAI will charge 2k/month?
- Architecture is the main problem when traing 400B+ model – limu
Exercise
three times this week, sad news.
But I start to running in the morning, just start in the monday. keep insisting.
Thought
1) Creating new product is a hard problem, I realize this more after I start to do it.
2) Creating entertainment content is a difficult task even TikTok is so popular all over the world. Hot content always need all necessary thing, fresh idea, capturing details, and so on. This is hard to created that by AI. In other words, AI is not enough smart to produce content which can attarct consumers. Consequently, AI is suitable for small business, this is the majority ustage of AI.
3) I’m tried, but I enjoy.
Next week
- Focuse on unsloth
- PR
- Reproduce
- Know the theory of acceleration and product a essay
- Reading: spend some time on papers
- Exercise: Running in the mornings
- Social account