Mikhail Shilkov

Compute @poolsideai

[Mikhail Shilkov @MikhailShilkov]

@MeaganCojocar Any next level of details available? Where are those tokens saved, is it fewer turns, tool call efficiency, compaction, caching, faster successes or faster failures, etc? Are there trade offs or are Anthropic/OAI purely leaving tokens on the table?

[Mikhail Shilkov @MikhailShilkov]

We’ve released the first public Poolside models today: Laguna XS.2 and Laguna M.1! Also, the first models I helped train. An exciting day for us all!

[Poolside @poolsideai]

Today we’re releasing Laguna XS.2, Poolside’s first open-weight model.
It’s a 33B total / 3B active MoE model built for agentic coding and long-horizon tasks.
Trained fully in-house on our own stack. Runs on a single GPU. Released under Apache 2.0.
Links 👇
Weights: https://t.co/HSo8L2gM64
API: https://t.co/DMJtNFrace
Blog: https://t.co/BXEjQxtQoV

[Mikhail Shilkov @MikhailShilkov]

Getting "Noted on the malware reminder" on every file read with Opus 4.7. Did they have to bolt that reminder last minute to prevent some tail risk? Feels oddly raw.

[Mikhail Shilkov @MikhailShilkov]

Not sure I can trust these Claude Code /insights: "You are a power user who engages with Claude Code at an extraordinary scale—over 2,200 sessions and 22,000+ messages in just three weeks, with nearly 10,000 hours of interaction time."

Umm, no? You alright?

[Mikhail Shilkov @MikhailShilkov]

Today is my last day at Pulumi after 6+ years. Wrote about what kept me there that long. Thank you, Pulumians!

https://t.co/Lp4HKfxn8a

[Lance Martin @RLanceMartin]

https://t.co/iGYM1PJnxv

[Mikhail Shilkov @MikhailShilkov]

New post: how Claude Code skills actually work: folder layout, the Skill tool, skills in the system prompt, and what’s injected at invocation time. Reverse-engineered from live sessions.
https://t.co/90puFCmlwb

[Mikhail Shilkov @MikhailShilkov]

I poked at Claude Code’s web tools. Two paths: WebFetch (question → page answer) and WebSearch (discover links to fetch separately). Input/output schemas, implementation flow, and why it doesn’t do karaoke.
https://t.co/OxCP2Y7W9x

[Mikhail Shilkov @MikhailShilkov]

Compared Claude Code 1.x and 2.0 system prompts. Rigid "MUST/NEVER" rules became softer guidelines, some sections on code conventions removed, MultiEdit tool gone, SlashCommand tool added. Git safety got more specific. https://t.co/fMRCD1aLhM

[Christian Nunciato @cnunciato]

Dang. Windsurf with the @PulumiCorp MCP server is nice! ✨

[Mikhail Shilkov @MikhailShilkov]

I feel like I did something with AI
https://t.co/GjWxycv3VA

[Charlie Marsh @charliermarsh]

I think part of the virality of MCP is that it lets you "ship something" in AI. Building an MCP server is a concrete thing you can do in a landscape that can otherwise feel overwhelming. When you ship an MCP server, you get to say + feel like you "did something" with AI.

[Mikhail Shilkov @MikhailShilkov]

We launched a Pulumi MCP Server, the dev loop for your Pulumi projects has never been better https://t.co/akbg3OHOGE

[Mikhail Shilkov @MikhailShilkov]

Not yet sure if it’s a fool’s day joke or a next quarter deliverable

[Pulumi @PulumiCorp]

Introducing "vibe clouding", a new way to manage your cloud infrastructure 😎

Fully give into the vibes, embrace exponentials, and forget that the cloud even exists.

It even mostly works! ✨

[Mikhail Shilkov @MikhailShilkov]

Excited to share what my team at Pulumi has been cooking! Our new Components feature lets you author infrastructure in one language and consume in any other. Simple API for users, complex machinery behind the scenes. Full blog post below!

[Pulumi @PulumiCorp]

Introducing enhanced Pulumi Components: Create in one language, consume in any other—even YAML! Perfect for platform teams building reusable infrastructure that developers can easily adopt. Meet your engineers where they already are.

https://t.co/PsLJmHgWNY

[Mikhail Shilkov @MikhailShilkov]

And I closed an issue that I opened myself almost 7 years ago. My first ever issue in the Pulumi repo. So satisfying!

[J☁e Duffy @funcOfJoe]

A decision I made early in Pulumi was our best and worse decision: auto-naming.

Every resource gets a random suffix appended to its physical name. This ensures
• Guaranteed isolation from name collisions
• Zero-downtime replacements

Now thanks to @MikhailShilkov, you have more control over this. The easy things are easy and hard things possible. Check out his video to learn more:

https://t.co/7nbPmIYaFz

[J☁e Duffy @funcOfJoe]

A decision I made early in Pulumi was our best and worse decision: auto-naming.

Every resource gets a random suffix appended to its physical name. This ensures
• Guaranteed isolation from name collisions
• Zero-downtime replacements

Now thanks to @MikhailShilkov, you have more control over this. The easy things are easy and hard things possible. Check out his video to learn more:

https://t.co/7nbPmIYaFz

[Mikhail Shilkov @MikhailShilkov]

The next bottleneck to scaling LLM training may be the desire to spend 100s of billions of dollars, there are no other obvious limits

https://t.co/CJUhUMv0mH

[Mikhail Shilkov @MikhailShilkov]

Had to jump through the hoops but finally got Claude to search GitHub issues

[Mikhail Shilkov @MikhailShilkov]

What is state of the art for getting my personal/work context into Claude and ChatGPT? I looked at MCP but judging based on the quality of its GitHub server, it’s not really functional nor very actively improved.

[Mikhail Shilkov @MikhailShilkov]

I'll be in Seattle Oct 20-24! Want to catch up over coffee? I'd love to hear from you - drop me a line!