Leonid Bugaev
@leonidbugaev.bsky.social
By day Head of Engineering at Tyk.io, and Indie Maker at night. Building https://goreplay.org Helping OSS community to grow, and people learn to code: https://helpwanted.dev/ Living a nomad life with my beautiful family. Deep into AI.
Important change with AI, that it is ok to re-roll the dice to wipe all previous work, and re-do it with updated prompt, rather then trying to fix whats already done. The work itself becomes very cheap, the most expensive part here is capacity of human to review and think.
The recent agent-to-agent protocol more than just a technological solution, it is attempt to "fix" colapsing SaaS economy and API market, and moving to value based pricing.
The recent agent-to-agent protocol more than just a technological solution, it is attempt to "fix" colapsing SaaS economy and API market, and moving to value based pricing.
Now you can chat with documentation or code, via MCP, just by specifying github url - all search happens locally! More over you can create your own MCP servers, with pre-baked data! The easiest way to create MCP so far! All free and Open Source. github.com/buger/docs-mcp
GitHub - buger/docs-mcp
Contribute to buger/docs-mcp development by creating an account on GitHub.
github.com
Chatting with AI changed the "search" paradigm - simplifying it to Question -> Answer format. No wonder it become so addictive. MCP changes the game again, bringing it local and adding action to the flow: Question -> Answer -> Action Don’t underestimate its power.
Chatting with AI changed the "search" paradigm - simplifying it to Question -> Answer format. No wonder it become so addictive. MCP changes the game again, bringing it local and adding action to the flow: Question -> Answer -> Action Don’t underestimate its power.
Now you can chat with documentation or code, via MCP, just by specifying github url - all search happens locally! More over you can create your own MCP servers, with pre-baked data! The easiest way to create MCP so far! All free and Open Source. github.com/buger/docs-mcp
GitHub - buger/docs-mcp
Contribute to buger/docs-mcp development by creating an account on GitHub.
github.com
Important change with AI, that it is ok to re-roll the dice to wipe all previous work, and re-do it with updated prompt, rather then trying to fix whats already done. The work itself becomes very cheap, the most expensive part here is capacity of human to review and think.
Recently started spending more time working outside. Living in a sunny place, it’s almost impossible to use a laptop outdoors. I’ve been using these Xreal One glasses - like portable monitors, not VR. They look a bit odd on calls, but honestly, it’s working really well for me.
Recently started spending more time working outside. Living in a sunny place, it’s almost impossible to use a laptop outdoors. I’ve been using these Xreal One glasses - like portable monitors, not VR. They look a bit odd on calls, but honestly, it’s working really well for me.
It's been over two years since the new cursor was released. We now have juniors who've only programmed with AI, never experiencing coding without it. Would you hire these juniors today? Do you think they can become capable engineers?
Have become top 3 trending developer on Github, with @mitchellh right behind me 😱
Some future AI revolutions happen in some random discord discussions like this discord.com/channels/133...
Discord - Group Chat That’s All Fun & Games
Discord is great for playing games and chilling with friends, or even building a worldwide community. Customize your own space to talk, play, and hang out.
discord.com
Have become top 3 trending developer on Github, with @mitchellh right behind me 😱
Why artifact storage products are so expensive, like PackageCloud or CloudSmith. We pay a LOT for it now. Essentially it is a file storage with some layer on top of it... However configuring own deb/rpm repo also a pain. I wonder if it is a solved problem? What do you use?
Why artifact storage products are so expensive, like PackageCloud or CloudSmith. We pay a LOT for it now. Essentially it is a file storage with some layer on top of it... However configuring own deb/rpm repo also a pain. I wonder if it is a solved problem? What do you use?
It's always tempting to chase the hype—another flashy AI code editor, another trendy framework. But true innovation often hides in the quieter spaces, the "plumbing." These foundational tools quietly enable everything else to run smoothly and grow. 🧵
It's always tempting to chase the hype—another flashy AI code editor, another trendy framework. But true innovation often hides in the quieter spaces, the "plumbing." These foundational tools quietly enable everything else to run smoothly and grow. 🧵
A lot of problems can be solved by bruteforcing them. Imagine if we had Claude 3.5 10x faster and 10x cheaper, or 100x. Now imagine it is Grok 3. Our current models already good enough, for example see Manus. AGI is so close!
The more I get into the Rust and its ecosystem the more I think it can be one of the dominant languages in the AI era. Claude 3.7 knows it very well, built-in tooling is amazing and integrate very well with AI (like very detailed error messages). Type safe and super fast.
The more I get into the Rust and its ecosystem the more I think it can be one of the dominant languages in the AI era. Claude 3.7 knows it very well, built-in tooling is amazing and integrate very well with AI (like very detailed error messages). Type safe and super fast.
Its quite facinating that I actually stopped writing the code - I just think and review.
Rolling back to Claude Sonnet 3.5. The new 3.7 is too unpredictable, constanlty goes to loops in agentic mode. It defo better understand code (like knowing Rust much better), but worse at following instructions and history of conversation.
It’s fascinating that well-known engineering best practices—like domain-driven design, separation of concerns, unit testing, low cyclomatic complexity, detailed architecture docs, feature specs, and descriptive file names—are exactly what enable AI to scale in large codebases.
Rolling back to Claude Sonnet 3.5. The new 3.7 is too unpredictable, constanlty goes to loops in agentic mode. It defo better understand code (like knowing Rust much better), but worse at following instructions and history of conversation.
It’s fascinating that well-known engineering best practices—like domain-driven design, separation of concerns, unit testing, low cyclomatic complexity, detailed architecture docs, feature specs, and descriptive file names—are exactly what enable AI to scale in large codebases.
Just added internal Chat bot, which acts as internal R&D consultant, which helps you to shape R&D claims. It probably can help us shave $1M a year. Build boring stuff.
Reliable chat with a code is even more important then actually writing the code. It helps to architect before build (especially with AI), generate test cases, debug issues, and way more. And it can be used by everyone and not just engineers: QA, Docs team, Support, Sales end etc.