clems

PhD student in computational pathology at @Radboudumc — former AI research at @paige_ai & @GustaveRoussy | MSc at @EcoleDesPonts & @ENS_ParisSaclay

[clems @clemsgrs]

hs2p, the lightweight package i built for fast & scalable whole-slide tiling workflows, got a fresh look!

If you work in computational pathology, this might save you a lot of preprocessing pain 🧵

[Peter Koo @pkoo562]

Excited to share new work on interpreting and steering pathology foundation models!

A fun collaboration led by @ChanwooKim_ and @suinleelab, with support from our pathologist collaborators Zhen Zhao and Deepika Savant, and Jakub Kaczmarzyk.

Paper: https://t.co/QBxBLNCpk4

1/N

[clems @clemsgrs]

@pkoo562 @ChanwooKim_ @suinleelab mega cool! much needed work, excited to see what you guys are cooking :)

[clems @clemsgrs]

did saint tibo see my last tweet @thsottiaux

[OpenAI @OpenAI]

We heard you wanted to use Codex rate limit resets on your own time.

Starting today, we’re rolling out the ability to save rate limit resets to use later.

We’re starting Go, Plus, Pro, and Business users with one free reset:

[clems @clemsgrs]

frontier lab engineers run loops

i run out

[clems @clemsgrs]

the permanent underclass engineer

coding with agents on the basic tier is basically living under subscription feudalism

you get 46 minutes of competence

then 4 hours of being a medieval peasant staring at a dead field

there is no “workflow” anymore

there is only the resets

[clems @clemsgrs]

complete reset

pure forgiveness

the permanent underclass engineer lives like this

not poor enough to be outside the future

not rich enough to stay inside it

just standing at the glass

refreshing

[clems @clemsgrs]

the permanent underclass engineer

coding with agents on the basic tier is basically living under subscription feudalism

you get 46 minutes of competence

then 4 hours of being a medieval peasant staring at a dead field

there is no “workflow” anymore

there is only the resets

[clems @clemsgrs]

at 100% usage i become ornamental

the tickets are there
the ideas are there
the repo is there

but the oracle is closed

so i pace around waiting for tibo or some anthropic backend priest to press the reset button

and when the limits reset it is disgusting how much joy i feel

[clems @clemsgrs]

this is what they don’t understand

the model is not my assistant

i am the organism arranged around its limits

i plan my day like a rat in a lab pressing the lever only when the pellet dispenser is active

[clems @clemsgrs]

new codex subagent feature dropped earlier so now i’m choosing to believe token accounting is briefly in its experimental era

very respectfully planning my usage around the possibility of an apology reset arriving any moment

[clems @clemsgrs]

It’s been around for a while already, but I recently went back to clean up the user-facing API + built a small demo.

Drop a slide into it and see what happens ;)

• pip install hs2p
• demo → https://t.co/w7zkcx0ha2
• code → https://t.co/S1LHDh14lG

[clems @clemsgrs]

hs2p, the lightweight package i built for fast & scalable whole-slide tiling workflows, got a fresh look!

If you work in computational pathology, this might save you a lot of preprocessing pain 🧵

[clems @clemsgrs]

When I started my PhD, there wasn’t really a tool that handled slide tiling the way we needed at scale. So I hacked something together. Over the years, it quietly grew through daily use and feedback from colleagues, becoming hs2p.

[clems @clemsgrs]

@jbschiratti @richardjchen @ThomasFuchsAI that’s why we need to opensource benchmarks by allowing anyone to submit a FM 👀

[clems @clemsgrs]

@dinkar_juyal cool work! in the paper you mentioned you trained the ViT-S using 64 A40 gpus: how long did you train for?

[clems @clemsgrs]

👀 Dive deeper into our findings and methodology by reading our short paper here: https://t.co/64AjpbEMBY

Looking forward to MIDL 2024 in Paris. Excited to come back in my home country, where AI research is flourishing as beautifully as the layers of a buttery croissant! 🇫🇷

[clems @clemsgrs]

🔬 Working with ViTs in pathology? Leverage tissue segmentation to prevent background from skewing your predictions. Our latest work - accepted at #MIDL2024 - improves robustness and interpretability of ViTs by focusing their attention where it counts! 🧵1/5

[clems @clemsgrs]

📊 We validate our approach using prostate cancer grading from whole-slide images as a case study. Our results demonstrate it achieves comparable performance to traditional self-attention while providing more accurate and clinically meaningful attention maps. 5/5