Brian Shin

Working on something new

[Mackenzie Burnett @ciaomack]

excited to finally announce that @AmbrookAg has raised $29M in funding to date, including a $26.1M Series A led by Thrive Capital and @zoink.

thank you to everyone who has been part of our story so far. I can't wait for the story to come.

https://t.co/JjadSKWHMK

[Brian Shin @_brianshin]

@steipete @indragie @cleptric Hah, I’ve spent many hours debugging hairy React Native build errors on the iOS side, Mac can’t be worse… right?

[Brian Shin @_brianshin]

@indragie @cleptric @steipete Ah interesting, thanks! Perhaps I'll take a stab myself w/ Claude (haven't written much Swift)

[Brian Shin @_brianshin]

Update: career break is going well. (Just a week into the $100 Max Plan, such a steal.)

[Brian Shin @_brianshin]

@MaximeRivest Serverless LoRA vision LM support from a big (open) provider would help a ton! Didn’t find one in my recent search.

[Brian Shin @_brianshin]

@mitsuhiko I'm curious: is it because of latency with the multimodal inputs to the LLM or just Playwright itself being slow. If it's the former, perhaps a subagent with a smaller specialized model would help?

[Brian Shin @_brianshin]

Woah, ARC-AGI-3 is video games 🤔🎮

[ARC Prize @arcprize]

Interactive Reasoning Benchmarks are the next step in frontier evaluations

Hear @GregKamradt share why measuring human-like intelligence requires multi-turn environments

Including a sneak peak of ARC-AGI-3

Want to help us build interactive evaluations? We're hiring

[Brian Shin @_brianshin]

Sonnet 4: I'm in what appears to be a house interior (likely my home in Pallet Town based on the tile pattern). I can see my character (the sprite with a hat), a TV, and what looks like my mom (the other character sprite) in the room.

[Brian Shin @_brianshin]

Realizing that none of the LLMs really understand what's going on in this image.
Qwen 2.5: "it appears you are standing next to a Trainer in Pallet Town, and he is likely wandering around. Since he's standing in front of you, I think the best initial action would be to interact..

[Brian Shin @_brianshin]

@AnandButani Agreed. I think many folks (myself included) are discouraged from using models that aren’t readily hosted which seems like a mistake

[Brian Shin @_brianshin]

The sheer amount of engineering here to make this all "just work" out of the box is amazing. In other news, running your own LLM inference is surprisingly simple (at hobbyist scale.)

[Brian Shin @_brianshin]

Sounds like a great sabbatical project

[will brown @willccbb]

this is 50 cents worth of compute on https://t.co/6fpRViVIU6

[Brian Shin @_brianshin]

@anvisha @AmbrookAg Thank you Anvisha!

[Brian Shin @_brianshin]

Wrapping up my last day at @AmbrookAg after a great four years! My #1 learning was to how to have extreme agency with a scrappy team, from resolving outages to launching our first paid ads and AI products.

[Brian Shin @_brianshin]

No next plans as of yet — excited to take some time off to travel, hack on some AI side projects and move back to NYC. Would love to chat with folks doing interesting work here!

[Brian Shin @_brianshin]

@lateinteraction Noticed last weekend that the CoT module doesn’t take into account if a model has reasoning capabilities. Do y’all already have something planned here? I didn’t find a GitHub issue to this effect.

[Brian Shin @_brianshin]

While it’s not marketed as such, the [Vercel] AI SDK seems to be on the way to being a solid agent framework

[Lars Grammel @lgrammel]

AI SDK 5 concept: loop until condition

Often you want flexible control over when the steps loop stops.

[Brian Shin @_brianshin]

Kudos to the @CloudflareDev AutoRAG team - built a fully working RAG prototype in a few hours yesterday. Also handled the little things like creating a service account key inline!

[Brian Shin @_brianshin]

@wardpeet @calcsam @mastra Ah, great news! Looking forward to seeing that new version.

[Brian Shin @_brianshin]

Very excited to see more TypeScript based agent libraries, but @mastra seems too bloated to put in our monolith - anyone else have a slimmer alternative?