Karles Marín

Mastodon @karlesmarin@mathstodon.xyz

⚔ Compare ☆ Watch Own this channel?

🔄 Data last refreshed 4 hours ago

🤝

Make a deal with Karles Marín

🔒 Paid proposals held safely in escrow — released only when the work's approved.

Followers

Account age

0 yrs

🧰 Free analysis for Karles Marín

🕵️ Fake follower check 📊 Engagement rate 💰 What they charge

Known for

$Parts I–II of *Predicting How Transformers Attend* mapped the **distance** axis of attention — the decay exponent $\gamma$. Part III opens the orthogonal **depth** axis, and finds that one innocent question, "is the answer here yet?", quietly conflates four: information *present*, *linearly decodable*, *written*, *used*. We split residual computation into three separately measurable observables$ 2 views Parts I–II of *Predicting How Transformers Attend* mapped the **distance** axis of attention — the decay exponent $\gamma$. Part III opens the orthogonal **depth** axis, and finds that one innocent question, "is the answer here yet?", quietly conflates four: information *present*, *linearly decodable*, *written*, *used*. We split residual computation into three separately measurable observables

In 1935 two papers appeared.
Philip Hall asked when a family of sets admits a system of distinct representatives — when everyone can be married off.
Hassler Whitney asked what "independent" means once you strip the vectors away. He called the answer a matroid.
Same year. Neither cites the other. It took thirty years for anyone to notice they were about the same object.
Mathlib, in 2026,

1 views In 1935 two papers appeared. Philip Hall asked when a family of sets admits a system of distinct representatives — when everyone can be married off. Hassler Whitney asked what "independent" means once you strip the vectors away. He called the answer a matroid. Same year. Neither cites the other. It took thirty years for anyone to notice they were about the same object. Mathlib, in 2026,

📊 Post engagement

Avg engagement / post

1×

Engagement vs followers

Jun 2026

On Mastodon since

🔥 Top post: #introduction — I spent the last year measuring something oddly · 5 likes + reposts

📊 Activity & format

Posting cadence

6.2 / week

Posts several times a week.

Content mix

Mostly text

Recent: 8 text · 4 image · 0 video.

Recent posts

View on Mastodon ↗

🔥 Top post #introduction — I spent the last year measuring something oddly regular: in trained transformers, attention weight decays with token distance as a clean power law, $A(d)\propto d^{-\gamma}$, with $R^2>0.95$ across 40+ open models. The fun part: $\gamma$ behaves like a stat… ★ 5

A footnote I like. The target came from Theorem of the Day's list of theorems by women mathematicians. Hazel Perfect is the name that puts it there. In 1968 she applied Menger's theorem to introduce the gammoids — the class that co…

The Transversal Matroid Theorem (Edmonds–Fulkerson 1965; independently, and in an infinite version, Mirsky–Perfect 1967): the partial transversals of a finite family are exactly the independent sets of a matroid. The interesting part isn&#…

In 1935 two papers appeared. Philip Hall asked when a family of sets admits a system of distinct representatives — when everyone can be married off. Hassler Whitney asked what "independent" means once you strip the vectors away. … ★ 1 🎛️ And you can play with it: an interactive explorer computes the reduction $(a,b,c)\to(p,q,r)$ and the closed form $\pm\chi_p\chi_q\chi_r$ live for any representation — the support, the moments, all of it. No install, runs entirely in… 🧮 A Schur-function identity I can't prove — every 4-row partition with parts $\le 16$ checked (4845 of them), zero misses. New preprint (Part IV). For every partition $\lambda$ with at most four rows, \[s_\lambda(1,-1,t,t^{-1}) \;=\; 0… 🧮 New preprint, Part III: *A Centre-Charge Selection Rule for the Wilson-Line Potential* — the fundamental domain of gauge–Higgs unification is representation-dependent. If you compute a Hosotani potential on $T^2/\mathbb{Z}_2$, you prob… ★ 1 🧮 New preprint — Part II of a series: *Three Gates to a Quark Generation*. The question. In 6D gauge–Higgs unification, which representations of $SU(4)$ hold a Standard-Model quark generation as chiral zero modes on $T^2/\mathbb{Z}_2$?… ★ 2 A computer found one solution to a hard model-building problem. Then we proved you will *always* find one — and machine-checked the proof in Lean 4. 🧮 The setting: which fermions complete a quark block in a 6D $SU(4)$ gauge–Higgs model o… ★ 2 New Lean 4 / Mathlib formalization note — Part I of a new series on permutation groups. 🎼 One orbit–stabilizer counting identity, \[ \Big(\sum_{g}|gA\cap B|\Big)\,|\Omega| = |A|\,|B|\,|G|, \] runs underneath three different things: P. M. N… ★ 1 2/ Two open companions: 🧭 A bilingual (EN/ES) field guide that unpacks all of this from scratch — now with new interactive 3D widgets for the depth axis: the "operator, not content" band, the three clocks (decode < write < commit), and the… ★ 1

Parts I–II of *Predicting How Transformers Attend* mapped the **distance** axis of attention — the decay exponent $\gamma$. Part III opens the orthogonal **depth** axis, and finds that one innocent question, "is the answer here yet?", qu… ★ 2

🐘 Community & instance

Home server

mathstodon.xyz

Their home server on the fediverse — the instance a creator picks signals the community they belong to.

On Mastodon since

Jun 2026

An established account with real history on the platform.

💡 Facts

🗓️Joined Mastodon in 2026.

👁️Averages 2 views per post.

📤Posts about 6.2× per week.

🕵️ Fake follower check

Estimated

60/100

Fair Credibility score

82%

Real Real audience

Low Fake-follower risk

Medium Data confidence

Est. 82% real, active audience · Low fake-follower risk.
Strong engagement (~1× of followers engage each post) — an active, real audience.
Very new account (under 3 months) — higher throwaway/bot risk.

Heuristic estimate from engagement, follower ratios, account age & growth — a screening signal, not a guarantee.

About

Independent researcher measuring how transformers attend: attention decays as a power law $A(d)\propto d^{-\gamma}$, and $\gamma$ turns out to be a useful state variable — phases, horizons, what breaks at long context. I publish my refutations next to my results (they're better teachers). Also: an interactive field guide on transformers (ES/EN) and a browser-only diagnostics tool (EN/ES/FR/CN).

📸 Gallery

🔀 Audience overlap

Estimated

Estimated shared audience with similar creators — useful for avoiding overlap (or doubling down) when planning a campaign.

Vince Bliss 17%

Haloorodje.si Mojaoprema.si 15%

Steiner 14%

かみまち弓子@東村山市議会議員 13%

UCCIDI BIG SMOOKE !! 12%

Nino 10%

More like this

Find more →

Mastodon

Haloorodje.si Mojaoprema.si

@haloorodje@mastodon.social

2 Followers

Mastodon

かみまち弓子@東村山市議会議員

@kamimachiyumiko@mstdn.jp

2 Followers

Mastodon

UCCIDI BIG SMOOKE !!

@saladprod@techhub.social

2 Followers

Mastodon

Nino

@reinchek@techhub.social

2 Followers

Mastodon

Steiner

@steiner@techhub.social

2 Followers

Mastodon

Vince Bliss

@vincebliss@techhub.social

2 Followers

Auto-generated from Mastodon's public data — no affiliation with or endorsement by SocialDB. Is this you? Claim it · Remove

Karles Marín

Make a deal with Karles Marín

Known for

📊 Post engagement

📊 Activity & format

Recent posts

🐘 Community & instance

💡 Facts

🕵️ Fake follower check

About

📸 Gallery

🔀 Audience overlap

More like this

✉ Message Karles Marín

🎤 Event / appearance with Karles Marín