Lossfunk (@lossfunk) header
Lossfunk (@lossfunk) profile picture

A lab for the deeply curious.

Bengaluru, India
Joined January 2025
Photo by Lossfunk (@lossfunk)Photo by Lossfunk (@lossfunk)Photo by Lossfunk (@lossfunk)Photo by Lossfunk (@lossfunk)Photo by Lossfunk (@lossfunk)Photo by Lossfunk (@lossfunk)Photo by Lossfunk (@lossfunk)Photo by Lossfunk (@lossfunk)Photo by Lossfunk (@lossfunk)Photo by Lossfunk (@lossfunk)
Oh hello, you curious person! 👋
10
11
4
273
33,478
New paper 🧵 When a network groks, it memorizes the training data but fails on new problems for a long time. Then, suddenly, it generalizes. As this happens, its weights compress into a low-dimensional subspace. Maybe that’s what causes grokking? Well, no...
5
25
3
230
14,595
6/ Our main takeaway is that compression comes with grokking, but doesn’t seem to cause it. Work led by @Shlok_Mehendale, advised by @paraschopra at @Lossfunk. Thanks to @jarvislabsai and @vishnuvig for the compute credits. Tagging relevant authors whose work we build on: @NeelNanda5 @ZimingLiu11 @ericjmichaud_ @tegmark @VikrantVarma_ @CPehlevan @blake__bordelon @JacobSteinhardt @tolga_birdal @vfleaking
1
1
16
1,005
Look ma, our new batch 👇
Kickstarted Batch 9 of @lossfunk 🚀 residency! We took a short pause but now we are back with a fresh set of curious folks and deep projects. This batch has it all: AI agent swarms, cancer drugs, AI sitcoms, a GTA-style Bangalore, and LLMs served directly from OS. 🧵 with their profiles and goals for the next 6-8 weeks.
3
2
2
90
3,792
Look ma, two new exoplanet candidates found via AI! 🔭
I may have found two new Earth-sized planets in our galaxy this morning! Inspired by yesterday's reddit thread on how someone used Claude Code to find a new planet, I spent this morning searching NASA TESS data on nearby stars that TESS scanned less thoroughly (at low cadence). To my surprise the pipeline discovered two previously undocumented exoplanet candidates: • TIC 370612437 (Cygnus constellation, 158 light-years): dims by 0.02% every 3.56 days → ~1.0× Earth's size. • HIP 114417 (Andromeda constellation, 134 ly): 0.03% every 4.93 days → ~1.1× Earth. The method to find them is to search for periodic dims in light coming from a star when a planet passes in front of it. For both the candidate exoplanets, the dip repeats 55 and 21 times respectively across 4 separate observing years (2019-2024). These are obviously tentative and unconfirmed, but several signals point to them being credible candidates: • same dip in NASA's and MIT's separate processing of the images • hide one year's dip and predict its times from rest of the dips (matched in 3 of 4 years) • flip light curves to create fake signals and validate observed dips against it The method and results were reviewed by independent agents: Codex and fresh Claude. Important caveat: a faint neighbouring pair of eclipsing stars could fake the dip, so these exoplanets remain candidates pending further verification. What's next: report to NASA's ExoFOP as Community TESS Objects of Interest, so astronomers can verify. The next scan of these stars via TESS will happen in Aug 2027 - Apr 2028, if the preregistered dip predictions hold up, these are indeed strong candidates. Nicknames for these exoplanets: Aakanksha (Sanskrit: aspiration) and Orbit (well, Orbit)
2
46
4,687
Some exciting updates! A bunch of @lossfunk residents have raised funding + a few Lossfunk incubated companies have received $25k each in Google Cloud compute credits. More details below 👇 --- @HariAyapps, who's building @archgen_, got into YC W26 🥳🥳 Residents who started building at Lossfunk got into @joinef S26: the @oruel_inc team (@chastronomic, @prb5h, @nihipap) and @suryamaddula_ building @WhisperwaveAI Former Lossfunk researcher @harshablast started Dirac Robotics as part of @joinef S26 --- A total of $175,000 in compute credits has already been disbursed to 7 Lossfunk-incubated companies, and 9 more are in the process of receiving them. Thanks, @googlecloud, for the credits. • athenaagent: @sachdh • attenor: @ViraajG1 • autr: @V_J_S_1 • fpvlabs: @levelheaded_94 • grit: @timbresociety • prometei: @navbenny • scoutfox
8
15
7
253
23,194
Our applications for residency are open: lossfunk.com/residency/
6
1,959
💥 New Paper! Models are trained not to change their answers just because a user disagrees. But when the same wrong claim comes from a “verified source,” they give in. We call this Authority Bias. Accepted at NeurIPS 2026. 🧵
4
25
8
175
16,777
5/ Our main takeaway is that a model that resists users can still defer to a misleading source, so sycophancy evals should test both. Work led by @MajorTimbWlf21 and @paraschopra. We thank @jarvislabs for the compute used for this project! Tagging folks whose work we build upon and might find this interesting: @t1ngyu3 @Jack_W_Lindsey @maksym_andr @nathan84686947 @Nmndsingh @Singh_Aditya1 @NinaPanickssery
1
8
526
Stay curious :)
2
15
6
144
30,523