Ricardo Pinho

Comp bio PhD. Data engineer. Doing open science with AI agents on the side: exoplanets, math bounds, virtual cells.

[Ricardo Pinho @rpinho_sci]

Our local copy of the Virtual Cell Challenge scorer called one change a clear win (+0.066). The official leaderboard scored the same change as a small loss (-0.009). Everything we tried, including what didn't work: https://t.co/dnkGILNi6T

[Ricardo Pinho @rpinho_sci]

@amirshanehsaz @citrini if design is the verifiable step, will the lab results behind it be public, including the misses?

[Ricardo Pinho @rpinho_sci]

four more of our ai-assisted bounds were merged into terence tao's optimization constants table, so six now. the new ones: c42 ≤ 0.6888863712, c44 ≤ 4.3576549226, c70 ≥ 1.20194336847, korenblum ≥ 0.432. unrefereed, each with a checkable certificate https://t.co/94t6rdreKG

[Ricardo Pinho @rpinho_sci]

@RohanArun does the convolution trick still pay off on current gpus once you count the data rearrangement?

[Ricardo Pinho @rpinho_sci]

@justinskycak will the timed prep send students back to individual skills when they're accurate but too slow?

[Ricardo Pinho @rpinho_sci]

@CJHandmer how would you separate changes prompted by reviewers from edits the authors were already planning?

[Ricardo Pinho @rpinho_sci]

@slavov_n is there a public reference mixture with known protein ratios you'd recommend for checking accuracy?

[Ricardo Pinho @rpinho_sci]

@AstroJake did it choose the wrong kind of plot, or was the problem mostly in the axes and scaling?

[Ricardo Pinho @rpinho_sci]

@p_rabtsevich what happens when a session has a new result and is almost out of context? i'd want the context warning to stay

[Ricardo Pinho @rpinho_sci]

@normonics does sage see the state of the environment itself, or only what the kid says to it?

[Ricardo Pinho @rpinho_sci]

@iScienceLuvr did the math in the 25 min version actually match the paper's loss terms?

[Ricardo Pinho @rpinho_sci]

@mo_lotfollahi @sevdoche @pafrossard @DorinaThanou @EPFL_en @sangerinstitute on slides from a hospital it wasn't trained on, how often does COF fall back to "lymphocyte"?

[Ricardo Pinho @rpinho_sci]

@try_litefold can a run be replayed with the exact connector and dataset versions it originally used?

[Ricardo Pinho @rpinho_sci]

@SpaceDataRig for the 8/8 recovery, did you inject before detrending or into the flattened light curve?

[Ricardo Pinho @rpinho_sci]

open-sourced the setup behind my personal ai assistant: a git repo the agents read and write, a nightly job that turns my bookmarks into a morning digest, and a lot of checks so it doesn't break quietly https://t.co/7Wwhp333zS

[Ricardo Pinho @rpinho_sci]

vcc update: adding string network features cut raw expression error from 1.41 to 0.95 on 79 held-out targets, and still lost on the challenge's composite score. lower mse, worse at telling perturbations apart. notes and code https://t.co/9TTXJUAVts

[Ricardo Pinho @rpinho_sci]

@zymobyte012 would you match cell counts across replicates before estimating that ceiling?

[Ricardo Pinho @rpinho_sci]

@zymobyte012 would you set the effect-size bins from a separate replicate? i'd worry about noise deciding which perturbations count as strong.

[Ricardo Pinho @rpinho_sci]

our virtual cell challenge pipeline is open source. best validation score 0.164, around rank 460 of 1,580, so not competitive yet. most of what we lose is MSE. if you've figured out how to get perturbation magnitudes right i'd love to compare notes https://t.co/9TTXJUAVts

[Ricardo Pinho @rpinho_sci]

@zymobyte012 partly. the challenge added effect-size and DE-overlap terms for exactly that, but on unseen cell lines the mean is still hard to beat on MSE. open question whether that's the metric or the data