28 Comments
User's avatar
Alf Viktor Williamsen's avatar

The greatest unlock is using the Agent SDK inside sandboxes in the cloud.

Expand full comment
PDK's avatar

Claude code with Opus 4.5 feels like being on catnip the entire workday - I suspect this is how the normies felt about gpt4o

Expand full comment
Nathan Lambert's avatar

I avoided writing about my claude code psychosis and don't like referring to people as normies in a wide reaching publication, but yes I get your sentiment.

Expand full comment
Michael's avatar

My experience with their customer service has been so bad that I still recommend caution for consumer users. Got an account used for family stuff randomly banned, and the process was a complete waste of time and is not up to industry standards.

Expand full comment
João Vitor Gomes's avatar

I had this strange moment the other day: staring at the terminal, I flashed back to being a kid hacking away in cmd. Funny how the arc bends, you end up right back at the terminal :O

Expand full comment
JV's avatar

Didn't the default change from non-thinking to thinking recently? That might explain why suddenly people are finding it easier to do hands off stuff.

Expand full comment
Nathan Lambert's avatar

no idea!

Expand full comment
Alf Viktor Williamsen's avatar

--dangerously-skip-permissions

Expand full comment
Andrew VanLoo's avatar

YOLO

Expand full comment
phamvanty's avatar

14Rayint. endron. c. y enronjit🏥

Expand full comment
Michael's avatar

So you use claude code with desktop claude app?

Expand full comment
Nathan Lambert's avatar

ah I should've said, I use iterm2+Tmux so ONLY in the CLI and multiple Claude's at once :)

Expand full comment
Jean-Paul Paoli's avatar

Claude is making terminal sexy again

Expand full comment
Stuart Horner's avatar

I love Opus 4.5 but the cost is prohibitive. At this point I’m limited more by my wallet than my imagination/knowledge. I wish I had the funds to go full tilt with Gastown, there’s so much I could create.

Expand full comment
Josh Valdez's avatar

I feel like this is AGI now! I felt like the initial launch of Claude Code was a huge unlock. The models have gotten much better over the course of 2025 but when Claude Code first dropped, it blew my mind and made me feel like we lived in a whole new world: https://joshuavaldez.com/1-week-with-claude-code/

Expand full comment
Andrew VanLoo's avatar

@nathan

If you haven’t seen Gastown, check it out. It’s the next gen of agentic AI platform at scale.

Expand full comment
Andrew VanLoo's avatar

That’s the one. Mad hatter and all.

Expand full comment
Jean-Paul Paoli's avatar

Seems super interesting and solving a real issue …. Readme is not straightforward though if I may ;)

Expand full comment
Andrew VanLoo's avatar

Definitely agreed. ChatGPT and I spent hours unwinding it. We’ll see how we did today.

Expand full comment
Andrew VanLoo's avatar

How is Opus with advanced mathematics? My current project requires both coding expertise and advanced mathematics.

Codex has been the only model capable of doing the math successfully, and it has much larger usage limits.

Expand full comment
Nathan Lambert's avatar

excellent. It's been the best editor I could ask for for rlhfbook.com

(maybe a better way to say it is that they're comparable)

Expand full comment
Andrew VanLoo's avatar

For actually developing the math, describing it, or both?

I’ve been using Codex as my math phd side kick.

Expand full comment
Nathan Lambert's avatar

Ah, mostly math that's already known, so checking. For math research I'd use OpenAI or Gemini's Deep/UltraThink thing.

Expand full comment
Andrew VanLoo's avatar

Thanks!

Expand full comment
Alf Viktor Williamsen's avatar

Just tell it to use python tool. Send me an email and we can chat! I'm doing LLM + maths at scale. alf (@) 01.inc

Expand full comment
Andrew VanLoo's avatar

Sent

Expand full comment
Devansh's avatar

I have some insight re Claude Code you might find interesting. The big unlock comes from it's superior performance in iterating across tool use. It allows it to handle things like errors in terminals, sub agents, search etc much more naturally.

Deepseeks approach of RL across tool use is a good one to study here.

Expand full comment