This media is not supported in your browser
VIEW IN TELEGRAM
на русском значительно хуже войс
но движение рук это нечто 💪
но движение рук это нечто 💪
😁21🔥7❤1👍1👎1🤣1
my netflix these days:
https://www.deeplearning.ai/short-courses/preprocessing-unstructured-data-for-llm-applications/
https://www.deeplearning.ai/short-courses/advanced-retrieval-for-ai/
https://www.deeplearning.ai/short-courses/knowledge-graphs-rag/
https://www.deeplearning.ai/short-courses/building-multimodal-search-and-rag/
https://www.deeplearning.ai/short-courses/building-agentic-rag-with-llamaindex/
https://www.deeplearning.ai/short-courses/building-evaluating-advanced-rag/
https://www.deeplearning.ai/short-courses/preprocessing-unstructured-data-for-llm-applications/
https://www.deeplearning.ai/short-courses/advanced-retrieval-for-ai/
https://www.deeplearning.ai/short-courses/knowledge-graphs-rag/
https://www.deeplearning.ai/short-courses/building-multimodal-search-and-rag/
https://www.deeplearning.ai/short-courses/building-agentic-rag-with-llamaindex/
https://www.deeplearning.ai/short-courses/building-evaluating-advanced-rag/
🔥17❤8👍6😁1🤔1
Exposure to information isn't learning!
Learning is a change in actions. If you read/listen/watch anything, and your behavior doesn't change - you learned nothing.
If you don't change your behavior over years, and you are in the same place in your life, it means you are not learning.It means, you are a dummy.
Learning is a change in actions. If you read/listen/watch anything, and your behavior doesn't change - you learned nothing.
If you don't change your behavior over years, and you are in the same place in your life, it means you are not learning.
❤19👍6😢5🤔2👎1🔥1🥰1😁1
https://www.youtube.com/watch?v=EiPYgiu8-Hc
Makes sense. I have 0 motivation to hire juniors.
This old structure of "1 senior + few mid guys + few juniors" got replaced with "1 senior + few AI tools".
Makes sense. I have 0 motivation to hire juniors.
This old structure of "1 senior + few mid guys + few juniors" got replaced with "1 senior + few AI tools".
😐11👎9🔥3😱2
Made some wallpapers that motivate me. 🔥 If you like it, 😁 if you think it's cringe.
😁52🔥13❤4🐳3🥰1👏1
Remember this?
"Anthropic's CEO says that in 3 to 6 months, AI will be writing 90% of the code" https://www.businessinsider.com/anthropic-ceo-ai-90-percent-code-3-to-6-months-2025-3 What percentage of code do you create with AI tools?
"Anthropic's CEO says that in 3 to 6 months, AI will be writing 90% of the code" https://www.businessinsider.com/anthropic-ceo-ai-90-percent-code-3-to-6-months-2025-3 What percentage of code do you create with AI tools?
Anonymous Poll
21%
90%+
28%
50%+
37%
10%+
15%
0%
🔥4❤1
Im happily surprised by the results, I thought that much fewer people use AI.
For me, it’s 100%. I forgot when was the last time I wrote something on my own.
Last year, Cursor did this:
- wrote infra noscripts for setting up switches / routers / machines in an on-premise data center (which is crazy if you ask me. I thought this one was not going to work, but after a few hundred bucks in credits, it solved it. single noscript to set up all the hardware I had)
- built numerous full-stack apps (shoutout to the gemini 3 that now does amazing UIs for everything that i do)
- rewrote one app after it got a lot of traffic. The first implementation was "not optimal"... 20 minutes later - fully covered with tests, 20 minutes later - redis for queues and caching. same business logic, 100x performance + full documentation with system design-like graphs).
I specifically selected those 3 use cases because they show us that now AI can: build basic apps, scale them if needed and fully cover the infra part.
disclaimer: of course, without my expertise it would not work. on its own, it often hallucinates and ends up in a broken state. so, it doesn’t replace us, it empowers us to do 100x more.
I love the future🚀
For me, it’s 100%. I forgot when was the last time I wrote something on my own.
Last year, Cursor did this:
- wrote infra noscripts for setting up switches / routers / machines in an on-premise data center (which is crazy if you ask me. I thought this one was not going to work, but after a few hundred bucks in credits, it solved it. single noscript to set up all the hardware I had)
- built numerous full-stack apps (shoutout to the gemini 3 that now does amazing UIs for everything that i do)
- rewrote one app after it got a lot of traffic. The first implementation was "not optimal"... 20 minutes later - fully covered with tests, 20 minutes later - redis for queues and caching. same business logic, 100x performance + full documentation with system design-like graphs).
I specifically selected those 3 use cases because they show us that now AI can: build basic apps, scale them if needed and fully cover the infra part.
disclaimer: of course, without my expertise it would not work. on its own, it often hallucinates and ends up in a broken state. so, it doesn’t replace us, it empowers us to do 100x more.
I love the future
Please open Telegram to view this post
VIEW IN TELEGRAM
❤14👍8🔥4🤯2🥰1😁1
My Cursor workflow:
1. I first use my brain to understand what I need and how I would solve it.
2. I read the codebase and find the places where I need to make changes - this also gives me an understanding of the effort: whether the llm can one-shot it or if I need to split it into smaller steps.
3. I tag related files (the most important ones, not all) and describe the feature I want - I type this myself. I tried switching to voice, but when i type, my brain thinks more and i like it : )
4. I always run it first in Plan mode with Sonnet 4.5.
5. If I like the plan, I execute it with:
• Opus 4.5 when the code change is small and I understand it fits into the 200k context window (its fast and reliable, but worse on large changes)
• GPT-5.2 Codex if the problem is hard - it follows commands a bit better without deviations, but it’s super slow (slow, good in following plan)
• Sonnet 4.5 when it’s a big change across many projects (fast, reliable enough for day to day tasks)
• When I need to design UI, I go to Google AI Studio (gemini 3 pro). Usually 2-3 shots are enough to get a design I like. Then I export it as a zip, paste it into my codebase, and ask Sonnet 4.5 to integrate it into the project. Works like a charm.
If i dont like the plan - do 1-4 again.
6. Once the feature is ready, I accept all changes right away and test it. If everything works fine, I briefly read the PR and merge it. If there are errors, I prompt the model to fix them - but at this point, I study the implementation more carefully and make sure I understand the changes.
• I noticed that if it works on the first try, it usually means the plan was followed well, so I don’t need to pay much attention to it.
• But if there’s a bug, there’s a high chance something went wrong, and I need to make sure there were no deviations from the plan.
I usually spend around $300-400 per month on this ($200 sub + credits).
Would be happy to hear your workflows. These days everything changes super fast, and its nice to hear what other people are doing. That's the motivation for that post.
1. I first use my brain to understand what I need and how I would solve it.
2. I read the codebase and find the places where I need to make changes - this also gives me an understanding of the effort: whether the llm can one-shot it or if I need to split it into smaller steps.
3. I tag related files (the most important ones, not all) and describe the feature I want - I type this myself. I tried switching to voice, but when i type, my brain thinks more and i like it : )
4. I always run it first in Plan mode with Sonnet 4.5.
5. If I like the plan, I execute it with:
• Opus 4.5 when the code change is small and I understand it fits into the 200k context window (its fast and reliable, but worse on large changes)
• GPT-5.2 Codex if the problem is hard - it follows commands a bit better without deviations, but it’s super slow (slow, good in following plan)
• Sonnet 4.5 when it’s a big change across many projects (fast, reliable enough for day to day tasks)
• When I need to design UI, I go to Google AI Studio (gemini 3 pro). Usually 2-3 shots are enough to get a design I like. Then I export it as a zip, paste it into my codebase, and ask Sonnet 4.5 to integrate it into the project. Works like a charm.
If i dont like the plan - do 1-4 again.
6. Once the feature is ready, I accept all changes right away and test it. If everything works fine, I briefly read the PR and merge it. If there are errors, I prompt the model to fix them - but at this point, I study the implementation more carefully and make sure I understand the changes.
• I noticed that if it works on the first try, it usually means the plan was followed well, so I don’t need to pay much attention to it.
• But if there’s a bug, there’s a high chance something went wrong, and I need to make sure there were no deviations from the plan.
I usually spend around $300-400 per month on this ($200 sub + credits).
Would be happy to hear your workflows. These days everything changes super fast, and its nice to hear what other people are doing. That's the motivation for that post.
🔥16👍6❤5👀1