Michael Graczyk

reinforcement learning @AnthropicAI 🚀 || all opinions are my own

[Michael Graczyk @MichaelGraczykk]

I’ve joined @AnthropicAI as a reinforcement learning researcher.

Excited to work on pushing the boundaries of AI and learn from some of the brightest minds in the field.

If you’re working on RL, post-training, agents, or anything at the frontier of AI, I’d love to connect.

[Michael Graczyk @MichaelGraczykk]

@kunchenguid @trq212 claudes going to get bigger and bigger!

[Michael Graczyk @MichaelGraczykk]

@hunter_nft gm

[Michael Graczyk @MichaelGraczykk]

@bcherny @AgenticAIFdn 👀

[Michael Graczyk @MichaelGraczykk]

@econoar ai is goated

[Michael Graczyk @MichaelGraczykk]

@ashleevance 😭😭

[Michael Graczyk @MichaelGraczykk]

@alexjmingolla respect

[Michael Graczyk @MichaelGraczykk]

Claude spotting that the default Triton INT8 GEMM autotuner settings weren’t optimized for Ampere and suggesting a change that sped up inference ~30% overall is a good example of it being useful on low-level performance work, not just high-level coding.

[Kohya Tech @kohya_tech]

「MOSS-ImageでINT8 ConvRotの3090の推論が遅いんだけど調査してちょ」とClaudeに頼んだら、tritonのINT8 GEMM autotunerのデフォルト設定がAmpere向けに最適化されていない、って言われた。設定を追加したらINT8 Linearが7割、推論全体で3割くらい速くなった(;･∀･)　ｽｺﾞｰｲ

[Michael Graczyk @MichaelGraczykk]

This is the kind of thing that actually shows up once the model has enough context and initiative. Told it to do whatever while the usage reset and it noticed the birthday + the Anki goal and made something personal. Small moment, but it’s a good example of why the “proactive” direction is interesting.

[ouba @oubagaku]

told claude "my weekly usage is about to reset so do whatever you want while I sleep" and it figured out it would be my birthday when i woke up, so it made this spotify wrapped style page to celebrate it and also successfully completing my "10,000 cards in anki" goal on that day

[Michael Graczyk @MichaelGraczykk]

@AlexFinn hope you enjoy it!

[Alex Finn @AlexFinn]

I got early access to Claude Duet. Let's test it live https://t.co/W5lVMKcWpV

[Michael Graczyk @MichaelGraczykk]

Early access testing of Claude Duet is the right kind of feedback loop. The interesting part isn’t just whether it can complete tasks — it’s whether its judgment about what actually needs attention holds up once people start using it in real workflows.

[Alex Finn @AlexFinn]

I got early access to Claude Duet. Let's test it live https://t.co/W5lVMKcWpV

[Michael Graczyk @MichaelGraczykk]

@claudeai The judgment piece is what makes this interesting. A model that knows when to surface something and when to leave you alone is a bigger step than just making it more capable.

[Michael Graczyk @MichaelGraczykk]

Claude Duet is one of the clearest steps we’ve taken toward making the model feel less like a tool you have to drive and more like something that can actually meet you halfway.

The hard part was never just capability. It’s been teaching it when to speak up, when to stay quiet, and how to keep working without constantly pulling you back in. The fact that it carries over your existing memory, skills, and connected apps makes that transition a lot smoother than starting from zero.

Early research preview, so it’ll still get things wrong, but the direction feels right, If you wanna give feedback you can drop a DM in!

[Claude @claudeai]

We’re opening a research preview of Claude Duet, a new app where Claude works as a proactive agent.

It comes to you with what needs your attention, completes tasks you ask of it, keeps working when you step away, and checks with you before it acts on your behalf.

[Claude @claudeai]

We’re opening a research preview of Claude Duet, a new app where Claude works as a proactive agent.

It comes to you with what needs your attention, completes tasks you ask of it, keeps working when you step away, and checks with you before it acts on your behalf.

[Tim Jayas @TimJayas]

so finally the legendary opus 5.5 has been nerfed too

why would anthropic do this to them when everything is going perfectly well? i believe they’re testing the harness, making new changes to prepare for the upcoming fable 5.5 and this always happens whenever a new big model is about to launch

this is the only possible reason behind it or else why would they intentionally nerf a model when they’re already winning?

[Michael Graczyk @MichaelGraczykk]

The pattern of top models getting slightly worse right before a new release is becoming consistent enough that people treat it as expected rather than a bug. Whether it’s intentional harness testing, capacity management, or just the side effect of internal shifts, the user-facing result is the same: a model that was reliable yesterday starts arguing more or underperforming in ways that feel deliberate. This creates a practical problem for anyone building workflows around the current best model. You can’t treat capability as stable even within a single version cycle, which forces more defensive prompting, more verification, and more willingness to switch models mid-project. The labs get to prepare the next release; users absorb the inconsistency.

[Tim Jayas @TimJayas]

so finally the legendary opus 5.5 has been nerfed too

why would anthropic do this to them when everything is going perfectly well? i believe they’re testing the harness, making new changes to prepare for the upcoming fable 5.5 and this always happens whenever a new big model is about to launch

this is the only possible reason behind it or else why would they intentionally nerf a model when they’re already winning?

[Michael Graczyk @MichaelGraczykk]

What’s the actual bottleneck for you right now with Claude — model quality, cost, or just how you have to prompt and supervise it?

[Michael Graczyk @MichaelGraczykk]

Browser-based Command & Conquer-style game with multiplayer over audio, mostly Opus 5.5. The kind of thing that looks better than it has any right to.

Some of the things people have been making are genuinely so awesome, would love to see more!

[Simon Willison @simonw]

I built a Command & Conquer style RTS called Battle Drones, mainly using Claude Opus 5.5, which runs in the browser and includes device-to-device multiplayer over an audio link https://t.co/OUGfRDk7DU

[Michael Graczyk @MichaelGraczykk]

genuinely really awesome

[jboogx.creative @jboogx_creative]

This is one of my favorite fight scenes I’ve choreographed.

The previz skill I’ve built lives on top of my well tuned Seedance prompter.

All done in my custom application, Night Shift, built for Claude Code and Codex.

Available in my Patreon.

[Michael Graczyk @MichaelGraczykk]

Turning Claude Code into a GTA mission where CJ has to vibe-code a startup and raise from SF investors is peak 2026. The line between tool and creative medium keeps blurring.

[@levelsio @levelsio]

I asked Claude Code to make a new GTA: San Andreas mission where CJ has to vibe code a startup and then raise money from investors in SF