Pinned post
I always get asked "What's a good resource for a beginner to learn about Apache Kafka?".
Now I get to answer it.
(1/3)🧵
Now I get to answer it.
(1/3)🧵
Stanislav Kozlovski
Verified account This account is verified. Learn more@kozlovski
"The Kafka Guy" 🧠
Have worked on Apache Kafka for 7+ years, now I write about it. (& the general data space)
Low-frequency, highly-technical tweets. ✌️
Have worked on Apache Kafka for 7+ years, now I write about it. (& the general data space)
Low-frequency, highly-technical tweets. ✌️
536 Following 18K Followers 3.3K Posts
Latest posts 5
Anthropic's direct advice on what effort to run Fable 5.2 with.
`high` seems like the way to go!
`high` seems like the way to go!
is C# and the .NET ecosystem the most brutally underrated stack in the AI agent era?
• type checking at both run-time & build-time for correctness & safety
• world-class debugging/profiling capabilities - an agent can directly inspect & modify runtime state (via CSharpRepl)
• clean error traces (e.g compared to TypeScript slop)
• custom compiler rules (static analysis) for enforcing rules against your agent
• code generation reduces repetitive boilerplate that your agent would waste context space reading
• mature dependency management
• a unified first-party dev toolchain (the CLI that lets you do `dotnet new`, `dotnet build`, `dotnet format`, etc)
• first-class concurrency and multi-threading
• richly-documented FIRST-PARTY libraries (AI doesn't have to choose from N potentially malicious third-party libs)
• built-in env sandboxing (via Aspire) for parallel agent work on diff branches
• great ORM
• decent performance (better than Python or JS at least)
• type checking at both run-time & build-time for correctness & safety
• world-class debugging/profiling capabilities - an agent can directly inspect & modify runtime state (via CSharpRepl)
• clean error traces (e.g compared to TypeScript slop)
• custom compiler rules (static analysis) for enforcing rules against your agent
• code generation reduces repetitive boilerplate that your agent would waste context space reading
• mature dependency management
• a unified first-party dev toolchain (the CLI that lets you do `dotnet new`, `dotnet build`, `dotnet format`, etc)
• first-class concurrency and multi-threading
• richly-documented FIRST-PARTY libraries (AI doesn't have to choose from N potentially malicious third-party libs)
• built-in env sandboxing (via Aspire) for parallel agent work on diff branches
• great ORM
• decent performance (better than Python or JS at least)
Agent Scale is the 2026 version of Web Scale.
GitHub proves this - it has since added 4,000,000 CPU cores, 160 PB of high-speed storage and 50 Tb/s of network capacity this year alone.
Why?
It's commit rate exploded. Check this:
• 2024: commits were ~0.065B (billion) each month
• August 2026: 2.9B commits for that month alone
• September 2026: 7.4B commits 🤯
That's 1000% YoY growth, or 114x since Sep 2024...
How are they keeping this free?
Easy. Your data..
Users spend tokens on OpenAI, Claude, and others -> they generate variations of projects uploaded to GitHub -> Microsoft (GitHub's owner) trains LLMs on this -> Microsoft can end up with better models.
It's essentially model distillation at a higher level. Remember, GH has your commit descriptions too, allowing them to infer the user's intention and the ensuing edits to the codebase.
For whatever reason, they haven't taken full advantage of this yet.
But it's logical to conclude that Microsoft will come out with a good coding model some time in the future.
Perhaps after they fix their bureaucratic issues...
GitHub proves this - it has since added 4,000,000 CPU cores, 160 PB of high-speed storage and 50 Tb/s of network capacity this year alone.
Why?
It's commit rate exploded. Check this:
• 2024: commits were ~0.065B (billion) each month
• August 2026: 2.9B commits for that month alone
• September 2026: 7.4B commits 🤯
That's 1000% YoY growth, or 114x since Sep 2024...
How are they keeping this free?
Easy. Your data..
Users spend tokens on OpenAI, Claude, and others -> they generate variations of projects uploaded to GitHub -> Microsoft (GitHub's owner) trains LLMs on this -> Microsoft can end up with better models.
It's essentially model distillation at a higher level. Remember, GH has your commit descriptions too, allowing them to infer the user's intention and the ensuing edits to the codebase.
For whatever reason, they haven't taken full advantage of this yet.
But it's logical to conclude that Microsoft will come out with a good coding model some time in the future.
Perhaps after they fix their bureaucratic issues...
every good company has built some sort of "Software Factory" for themselves:
• Shopify built River - a Slack native agent that opens PRs to their monorepo. 1 of 8 merged PRs come from it
• Vercel built a webhook-to-queue pipeline that authors 25-35% of their weekly merged PRs
• Stripe built Minions to run tasks (in isolation) and open PRs. 1300 minion-produced PRs get merged each week
• Microsoft uses Copilot to turn issues into PRs. ~70% of such PRs get merged.
• Meta built a monitoring pipeline that cuts ~10hrs of investigation to ~30min.
Sources below 👇
• Shopify built River - a Slack native agent that opens PRs to their monorepo. 1 of 8 merged PRs come from it
• Vercel built a webhook-to-queue pipeline that authors 25-35% of their weekly merged PRs
• Stripe built Minions to run tasks (in isolation) and open PRs. 1300 minion-produced PRs get merged each week
• Microsoft uses Copilot to turn issues into PRs. ~70% of such PRs get merged.
• Meta built a monitoring pipeline that cuts ~10hrs of investigation to ~30min.
Sources below 👇