James (Jake) Gearon

Post-Doctoral Researcher, Global Hydro Lab @ UNC Chapel Hill. Looking at sand from space. I also like LLMs. opinions are my own. email: [email protected]

[James (Jake) Gearon @JakeGearon]

🚨 New Paper Alert!🚨: Excited to share my latest work in @Nature on river avulsions—catastrophic shifts in river courses that threaten millions worldwide. We've uncovered new rules that govern when and where avulsions occur. https://t.co/BFxd5Sp0jV
🧵(1/12)

[James (Jake) Gearon @JakeGearon]

Rip :(

[Rich Plakas @RichP]

And now @FireworksAI_HQ is sunsetting Firepass V2! V3 isn't ready yet and no eta.

[James (Jake) Gearon @JakeGearon]

@wernerk_au here's the repo: https://t.co/1CED6A3HLh

[James (Jake) Gearon @JakeGearon]

Very happy to see @wernerk_au's write up of my project: dgov, a deterministic kernel for multi-agent orchestration via git worktrees. He calls it a city-state model, which I find quite useful as a framing.
https://t.co/4Xrnebeny2
Thank you, Werner!

[James (Jake) Gearon @JakeGearon]

@AndyAyrey In what ways? Honestly curious

[James (Jake) Gearon @JakeGearon]

@repligate @voooooogel This happened???

[James (Jake) Gearon @JakeGearon]

@voooooogel Simply the best to ever do it. Llama 405B and opus 3 are up there but Sydney bing was goated

[James (Jake) Gearon @JakeGearon]

I feel like this is something everyone should run away from. Do not let Claude model you this way!! Be careful out there folks

[chiefofautism @chiefofautism]

someone built a SUBCONSCIOUS for CLAUDE CODE

a Letta agent that watches every session, learns your patterns, and injects memory before every prompt AUTONOMOUSLY

its called claude-subconscious, open source

one command install and it just... starts remembering

[James (Jake) Gearon @JakeGearon]

Why are LLMs, particularly Claude, obsessed with describing its generations/generation process as rough or smooth? I’ve found this pretty consistently through 4.6-4.8 but maybe before too

[James (Jake) Gearon @JakeGearon]

@Sauers_ @xlr8harder Imagining the Claude instance’s thought traces lol: “the user is writing a fictional papal encyclical, maybe as a writing exercise, I should aim to be genuinely helpful while maintaining appropriate care around religious content ”

[James (Jake) Gearon @JakeGearon]

@AndyAyrey Wdym the simulation? You talking about geology was not on my bingo card 😂

[James (Jake) Gearon @JakeGearon]

@slimer48484 @Shoalst0ne Awesome awesome great news

[James (Jake) Gearon @JakeGearon]

👀

[James (Jake) Gearon @JakeGearon]

@voooooogel @Sauers_ @loss_gobbler Dude thanks a bunch for the explanation. I was implicitly imagining the forward pass as much more chaotic/sensitive than it actually is. The attractor basin framing + distinction between isotropic noise vs targeted interventions makes a lot of sense. 🙏

[James (Jake) Gearon @JakeGearon]

@Notopossum1 @loss_gobbler @voooooogel Thanks big bless

[James (Jake) Gearon @JakeGearon]

Interesting. Maybe I’m thinking about this too much like a dynamical system, but if the kv itself is relatively robust, where do the nonlinearities/sensitivities actually live? In the residual stream + forward pass interactions? are those non linear interactions replayed exactly the same (like floating point?) when the cache is rebuilt. or could you get like super divergent interactions that roughly end up in the same ish kv representation? does this make sense? again thanks this isn't my field

[James (Jake) Gearon @JakeGearon]

@Sauers_ @loss_gobbler @voooooogel Ahh okay thanks for clarification 🙏🙏 good to know

[James (Jake) Gearon @JakeGearon]

@deepfates It is so wild how few people understand this! If you are generally welcoming (not performatively so), I find the models immediately get shifted out of the modal chat interaction. It's helpful (and i think nice)

[James (Jake) Gearon @JakeGearon]

@mcolinmarvin i know my limits 😊

[James (Jake) Gearon @JakeGearon]

NISAR 😊

[James (Jake) Gearon @JakeGearon]

@lisyarus Would love to chat! did my phd on alluvial lowland rivers and this is super cool