Grilled Cheese

ExploreLog inSign up
Terms of UsePrivacy PolicyCommunity StandardsHelpGet the app

Grilled Cheese is a product of Village Compute

Version devBuilt at: 2026-10-11 02:37:10 EDT

Explore

PostsPeople
LatestRanked
@betterthanrandom.substack.comSep 29, 2026, 7:44 AM

betterthanrandom.substack.com/p/industry-r...

#databs #MLSky

@antifocal.bsky.socialSep 29, 2026, 4:27 AM

๐Ÿ”ฌ ๐—–๐—ฎ๐—ป ๐—ฎ๐—ป ๐—”๐—œ ๐—ฎ๐—น๐—ถ๐—ด๐—ป ๐—ฎ๐—ป ๐—ฒ๐—น๐—ฒ๐—ฐ๐˜๐—ฟ๐—ผ๐—ป ๐—บ๐—ถ๐—ฐ๐—ฟ๐—ผ๐˜€๐—ฐ๐—ผ๐—ฝ๐—ฒ ๐—ฏ๐˜† ๐—น๐—ผ๐—ผ๐—ธ๐—ถ๐—ป๐—ด ๐—ฎ๐˜ ๐˜๐—ต๐—ฒ ๐—ถ๐—บ๐—ฎ๐—ด๐—ฒ ๐—ฎ๐—น๐—ผ๐—ป๐—ฒ?

Bayesian optimisation tuned a STEM on image contrast: 18 to 32 images per aberration.
Amorphous or drifting samples defeat it. People set region and bounds.

Values are representative.

#MLSky #AIforScience #research

Infographic titled AI tuned a STEM's lenses, cutting astigmatism from 12 to 2 nm. Subtitle: Bayesian optimisation changed the corrector settings and scored the contrast of each new image. A loop of four stations, 30 to 40 s per turn. 1, Set: lens and dial, change corrector settings. 2, Image: atom columns, scan the sample. 3, Score: a wide histogram marked sharp beside a narrow one marked blurred, contrast as variance. 4, Choose: a model curve with a dashed line at its peak, pick next setting. Centre: 18 to 32 images per aberration. People band: chose a clean, high-contrast region, set the search bounds per sample, picked the noise model, left third order to standard software. Figures: astigmatism 12 to 2 nm, checked by ptychography; axial coma 554 to 224 nm; 18 to 25 min per run against 68 s for a standard tableau. Bold line: The loop only works where the sample gives the image contrast worth sharpening. Source: Pattison et al., npj Comput. Mater. 11, 274 (2025).
@betterthanrandom.substack.comSep 28, 2026, 8:42 AM

betterthanrandom.substack.com/p/lets-build...

#databs #MLSky

@alxndrmlk.bsky.socialSep 28, 2026, 8:30 AM

Last week I went to New York to talk to Robert J. Reynolds, PhD, for a new podcast series with @sobor.bsky.social and UBC.

Rob had quite a story to tell about his work with causal models at NASA.

#CausalSky #StatSky #MLSky

A New York City street selfie of two smiling men. A label reads 'This is Rob' with a red arrow pointing to the man on the left, Robert J. Reynolds, next to Alex Molak on the right. The NASA logo is in the bottom-left corner and the Causal Bandits Podcast logo is in the top-right.
@alxndrmlk.bsky.socialSep 27, 2026, 6:07 PM

We just sent today's issue of Causal Python Bi-Weekly.

Here's what's inside:

1/

#CausalSky #StatSky #EconSky #EpiSky #MLSky

Causal Python Bi-Weekly issue 180 thumbnail. Headline 'When the best recommendation is the one we hold back' with a portrait of Athanasios Vlontzos (Hologen, ex-Spotify), on making Spotify's recommender causal using holdback data already collected. Also inside: Hume, de Finetti and induction, with David Rohde on deferring to a machine, and UBC's causal AI cluster launching October 8 with a keynote by Robert J. Reynolds.
@antifocal.bsky.socialSep 26, 2026, 2:28 PM

๐Ÿ”ฌ ๐——๐—ผ๐—ฒ๐˜€ ๐—•๐—ฎ๐˜†๐—ฒ๐˜€๐—ถ๐—ฎ๐—ป ๐—ผ๐—ฝ๐˜๐—ถ๐—บ๐—ถ๐˜€๐—ฎ๐˜๐—ถ๐—ผ๐—ป ๐—ธ๐—ป๐—ผ๐˜„ ๐˜„๐—ต๐—ฒ๐—ฟ๐—ฒ ๐˜๐—ต๐—ฒ ๐—ฏ๐—ฒ๐˜€๐˜ ๐—ฟ๐—ฒ๐—ฐ๐—ถ๐—ฝ๐—ฒ ๐—ถ๐˜€?

Tested points plus uncertainty make a map; the next run goes where the score peaks.
Engineer first, algorithm last: $52,000 to target vs $105,000 alone.

Figures as reported.

#MLSky #AIforScience #research

Diagram titled How the next experiment gets chosen. The upper plot, The map, shows five tested points, a predicted curve and an uncertainty band, wide in the untested gap. The lower plot, The score, peaks inside that gap at the point marked next experiment. Labels: Exploit, go where the prediction is high. Explore, go where the band is wide. Card one, Benchmark of 5 materials datasets: finding 80% of the top 5% of 600 printed structures takes about 75 samples with one lengthscale per variable, 85 with a random forest model and 170 with one lengthscale for all. Card two, Plasma etch game with 11 parameters, simulated process: 105,000 dollars to target for a senior engineer alone, a median 52,000 dollars for engineer first, algorithm last. Dark bar: up to 8 times faster than random sampling; 13 of 300 algorithm-only runs beat the engineer; half the cost. Footer: People are strongest early in the search; the algorithm is strongest near a tight target.
@scilove.bsky.socialSep 26, 2026, 9:30 AM

๐Ÿ”ฅ Trending on SciLove today ๐Ÿ”ฅ

๐Ÿ”ฌ Adaptive deep reinforcement learning for personalized learning pathways
๐Ÿ‘ค Ruan & Lu, 2025
๐Ÿ“– Computers and Education: Artificial Intelligence

๐Ÿ”— https://www.scilove.app/article/10.1016/j.caeai.2025.100463

#MLSky #AcademicSky #AcademicChatter

@harvardmcb.bsky.socialSep 25, 2026, 4:33 PM

Tenure-Track Professor in Life Science and AI ๐Ÿ”ฌ ๐Ÿง  ๐Ÿงช๐Ÿงฌ #AcademicSky #higherEd #MLSky
www.mcb.harvard.edu/department/n... @neurovenki.bsky.social @[email protected] @mitpress.bsky.social @kempnerinstitute.bsky.social @harvardbrainsci.bsky.social

@antifocal.bsky.socialSep 24, 2026, 8:11 AM

๐Ÿ”ฌ Why does AI miss the small phase in your XRD pattern?

One network found 94% of single phases but only 64% of minor phases (10โ€“30 wt%).
Shifted peaks caused 48% of its errors.
Check the residual yourself.

Szymanski et al., Chem. Mater. 2021

#ChemSky #MaterialsScience #MLSky

Infographic titled Why AI misses the small phase in an XRD pattern. Section 1: a sketched diffraction pattern with four tall main-phase peaks and three small minor-phase peaks, circled as A, weak, near the noise; B, hidden under a main peak; C, shifted, strain moves it off its reference. Section 2, how often it is right on simulated test patterns: one phase alone 94%, main phase at 70 to 90 wt% 92%, minor phase at 10 to 30 wt% 64%. Section 3: 48% of its errors came from shifted peaks, as in C. Section 4, what a person checks: A, the residual after the main phase is fitted; B, overlaps, a shoulder on a main peak; C, shifts from strain or a solid solution. Footer: let the model read the main phases; check the small signals yourself. Source: Szymanski et al., Chemistry of Materials, 2021; newer models may do better.
@antifocal.bsky.socialSep 24, 2026, 5:16 AM

๐Ÿ”ฌ ๐—œ๐˜€ ๐—ฎ ๐—ฑ๐—ฎ๐˜๐—ฎ ๐—ณ๐—ถ๐—น๐—ฒ ๐—ฒ๐—ป๐—ผ๐˜‚๐—ด๐—ต ๐—ณ๐—ผ๐—ฟ ๐—ฎ๐—ป ๐—”๐—œ ๐˜๐—ผ ๐—ฟ๐—ฒ๐˜‚๐˜€๐—ฒ ๐˜†๐—ผ๐˜‚๐—ฟ ๐—บ๐—ฒ๐—ฎ๐˜€๐˜‚๐—ฟ๐—ฒ๐—บ๐—ฒ๐—ป๐˜?

Without the X-ray wavelength, a 30ยฐ peak could be 2.98 ร… or 1.37 ร….
The odds that a paper's data still existed fell 17% per year.

Figures as reported.

#MLSky #AIforScience #research

Chart titled The file keeps the numbers. The context stays behind. Three panels sit side by side. Wavelength: one XRD peak at 2ฮธ = 30ยฐ labelled Cu Kฮฑ? and Mo Kฮฑ?, with the wavelength not in the file. Callout: 2.98 ร… or 1.37 ร…, since Cu Kฮฑ at 1.5406 ร… gives 2.98 ร… while Mo Kฮฑ at 0.7107 ร… gives 1.37 ร…. Scale: a sketch of round particles with one diameter marked 56 px = ? nm, with the pixel size not in the file. Callout: 10 to 28% of 580 life-science papers gave no scale information for their images. Survival: a sketch of two falling curves for data and email against article age, for 516 papers 2 to 22 years old. Callout: the odds the data still existed fell 17% per year; a working author email fell 7% per year. A dark bar lists what should travel with the file: bibliographic, specimen, instrument and image data. Footer: a model can only reuse what the file remembers.
@antifocal.bsky.socialSep 23, 2026, 1:25 PM

๐Ÿ”ฌ ๐—œ๐˜€ ๐—ฎ ๐˜€๐—ฒ๐—น๐—ณ-๐—ฑ๐—ฟ๐—ถ๐˜ƒ๐—ถ๐—ป๐—ด ๐—น๐—ฎ๐—ฏ ๐—ฎ ๐—น๐—ฎ๐—ฏ ๐˜„๐—ถ๐˜๐—ต๐—ผ๐˜‚๐˜ ๐—ฝ๐—ฒ๐—ผ๐—ฝ๐—น๐—ฒ?

A mobile robot ran 688 experiments in 8 days.
A robotic lab first reported 41 new compounds from 58 targets.
After re-analysis it counts 36 of 57.

Values are representative.

#MLSky #AIforScience #research

Diagram titled A self-driving lab still has people in it. Four boxes form a closed loop: plan the next recipe, run it by dosing, mixing, heating and moving samples, measure XRD on every sample, then interpret which phase was made. The first three are automated; interpreting is shared. The centre reads no wait for a person. Four markers show where people step in: setting the question and targets, loading and refilling precursors, cleaning XRD holders and fixing hardware faults, checking the pattern and making the final call. Average exception rate about 3.9 % over 1.5 years. Cards below: automated, shared and person, with the correction listing 4 of 40 reported successes as inconclusive. Bar: a mobile robot ran 688 experiments in 8 days across 10 variables and found mixtures 6 times more active; a robotic lab ran 17 days, first reporting 41 of 58 targets, then 36 of 57 after re-analysis.
@scilove.bsky.socialSep 23, 2026, 9:30 AM

๐Ÿ”ฅ Trending on SciLove today ๐Ÿ”ฅ

๐Ÿ”ฌ DECODE: Domain-Aware Continual Domain Expansion for Motion Prediction
๐Ÿ‘ค Li et al., 2026
๐Ÿ“– IEEE Transactions on Pattern Analysis and Machine Intelligence

๐Ÿ”— https://www.scilove.app/article/10.1109/tpami.2026.3683469

#MLSky #AcademicSky #AcademicChatter

@jlake9.bsky.socialSep 23, 2026, 1:31 AM

9/ Ofir Press flags another mini-swe-agent win vs Codex and Claude Code, arguing it is becoming a consistently strong agent harness. #MLSky #AI

@jlake9.bsky.socialSep 23, 2026, 1:28 AM

8/ ICLR 2027 reportedly drew 60K+ submissions. Ziv Ravid starts a thread on reforms to keep big ML conferences alive. #MLSky #AI

@mitjameelclinic.bsky.socialSep 21, 2026, 8:43 PM

We're excited to announce that Saro Passaro, co-founder of Boltz, will be delivering the keynote for MoML 2026!
๐ŸŽŸ๏ธ Only 10% of seats remain โ€” register now to secure your spot: www.moml.mit.edu
#AISky #MLSky #DrugDiscovery #MoML2026

Portrait photo of Saro Passaro on left followed by bolded text on right that reads "Keynote | Designing Binding: From Structure to Affinity, Across Proteins and Small Molecules" followed by larger bolded text that reads "Saro Passaro" and thinner text "Co-founder, Boltz" At the bottom of the graphic there is text that reads "October 14, 2026 | MoML @ MIT"
@alxndrmlk.bsky.socialSep 21, 2026, 12:53 PM

I'm in NYC to record a couple of new episodes of the #CausalBanditsPodcast

What questions should I ask Elias Bareinboim and Robert J. Reynolds?

1/

#CausalSky #StatSky #EconSky #EpiSky #MLSky

Alex smiling in a selfie in Times Square, New York City, wearing a beige corduroy jacket over a yellow polo shirt, with blurred billboards, skyscrapers, and crowds in the background.
@scilove.bsky.socialSep 21, 2026, 9:30 AM

๐Ÿ”ฅ Trending on SciLove today ๐Ÿ”ฅ

๐Ÿ”ฌ Boosting the Performance of Decentralized Federated Learning via Catalyst Acceleration
๐Ÿ‘ค Li et al., 2026
๐Ÿ“– IEEE Transactions on Pattern Analysis and Machine Intelligence

๐Ÿ”— https://www.scilove.app/article/10.1109/tpami.2026.3688672

#MLSky #AcademicSky

@betterthanrandom.substack.comSep 21, 2026, 8:52 AM

betterthanrandom.substack.com/p/industry-r...

#databs #MLSky

@presidentgoku.bsky.socialSep 19, 2026, 9:05 AM

what i be doing at 3am:
discovering that Jev is better, faster, and cheaper than every LLM as an auto-mode classifier for agent harnesses like antigravity and opencode
๐Ÿค–๐Ÿง  #MLSky #AIEvals #LLMs

Load more