Maxine 💃🏼
LightNews LightNews
banner
maxine.science
Maxine 💃🏼
@maxine.science
480 followers 410 following 2K posts
🔬 Looking at the brain’s “dark matter”
🤯 Studying how minds change
👩🏼‍💻 Building science tools

🦋 ♾️ 👓
🌐 maxine.science
Posts Replies Media Videos
maxine.science
oh shap, here come dat boi
Photograph of Springer series mathematics textbook, Shun-Ichi Amari, Information Geometry and Its Applications
November 24, 2025 at 12:11 PM
maxine.science
welp if this mf is goin down might as well break out the better stuff
Photo of a pour-over coffee setup in front of assorted vegetables and spices like Old Bat and S&B curry. In the foreground a hand is holding a bottle of Caperdonich 18 year Speyside single malt Scotch
November 18, 2025 at 2:29 PM
maxine.science
it’s so fucking ready for prod.
Screenshot of git diff - ## + ** 100% Architecture Production-Ready Release - Feature Parity Achieved** - This project is organized as a Cargo workspace with the following crates: + A high-performance Rust implementation of Claude Code - the AI-powered coding assistant that lives in your terminal. This rewrite delivers **100x+ faster startup** and **16x smaller footprint** compared to the NPM version, with **zero runtime dependencies**.
November 15, 2025 at 1:48 AM
maxine.science
if vaporwave can solve this, i think we’re gonna be alright
Spotify screenshot of the album cover for 新しい日の誕生 by ２８１４
November 12, 2025 at 3:33 PM
maxine.science
Ok—nitpick not about the content, but in what way did this clearly-vibe-coded animation provide benefit in any form?

There is no axis in which this being an animation was not strictly worse for actually reading the results.

Call me old-fashioned, but results should be on-screen >20% of the time.
Screenshot of the linked website showing graphs animating for some reason, and caption: Video: pass@k curves of base models and their zero-RL-trained counterparts across multiple mathematical benchmarks. When kis small, RL-trained models outperform their base versions. However, as k increases to the tens or hundreds, base models consistently catch up with RL-trained models across all benchmarks and LLM families without exception. Eventually, base models surpass RL-trained models.
November 11, 2025 at 2:31 PM
maxine.science
and that’s why:
default Google slide reading AIs need shrinks, too. Maxine Levesque
November 11, 2025 at 2:53 AM
maxine.science
cf.:
Mamdani sign reading AGENTS for SOCIALLY-ACCEPTABLE SLAVERY
November 11, 2025 at 12:44 AM
maxine.science
ah, see—nat geo much more up my alley, as ever
National Geographic cover featuring a witch in a fabulous creepy blue dress over a fumey cauldron, reading THE HISTORY OF WITCHCRAFT From the greatest myths to the Salem witch trials
November 10, 2025 at 3:42 PM
maxine.science
“who he is”

uh, friends i hope i’m not the first one to tell you this but—
Photo of a Time special edition in the magazine rack of a supermarket bearing a portrait of Beardy White Jesus™️ reading JESUS Who he was. Who he is. His miracles, His mission, His love
November 10, 2025 at 3:17 PM
maxine.science
I know why they caged hand blinks
Photograph of walk signal at night with the red hand illuminated and the front grating hanging downward, fully exposing the hands. A Muni streetcar passes behind in the background.
November 10, 2025 at 2:48 AM
maxine.science
Bay Area edition
Zohran sign reading Waymo for Murdering small animals Zohran sign reading Crypto For poorly-executed financial crimes Zohran sign reading BCI for immersive pornography Zohran sign reading Agents For socially-acceptable slavery
November 9, 2025 at 2:36 PM
maxine.science
obviously the pulmonology clinic is in suite 420
Photograph of a directory in a medical building with white text on a backlit black background. UCSF Pulmonary and Sleep Disorders Center, 420
November 7, 2025 at 5:36 PM
maxine.science
So it begins.
Cap of a frame from “The Matrix Reloaded” showing a phalanx of Agent Smiths.
November 6, 2025 at 7:08 AM
maxine.science
Screenshot of the byline from the paper: Cameron Berg* AE Studio Diogo de Lucena AE Studio Judd Rosenblatt AE Studio Screenshot of the AE Studio website about page: Software is our labor of love. Who we are Our team consists of scrappy startup founders, academic PhDs, and formally trained technology consultants. We focus on combining all of our learnings and experiences to build the best products for our clients.
November 2, 2025 at 2:52 PM
maxine.science
the comedy is pretty great though

“Personal list, do not distribute”
Screenshot of clearly showing Maxine.science on two lists titled “transph*bes/terr*rists” and “transphobes and eugenicists”
November 1, 2025 at 7:09 PM
maxine.science
about the quality of “journalism” one would expect from business insider
Screenshot of article, reading: For the case of Baroness Marie Vetsera, a woman involved in a minor 18th-century royal scandal, Wikipedia had a substantial entry. But Grokipedia's was even longer, and formatted quite differently - and frankly, I thought it told a better narrative of the baroness's untimely demise. But it also had some bad quirks. For example, Grokipedia used a citation for a Facebook page about royal history that seemed to be viral clickbait written by AI. The last sentence is highlighted yellow.
November 1, 2025 at 2:48 PM
maxine.science
ok, i’m not saying they’re necessarily gonna kill us all.

… and yet
Image of the 1X NEO showing ear lights and beady eyes Image of Cyber an from Doctor Who
October 30, 2025 at 11:47 AM
maxine.science
The listing for POLICY:
Screenshot from the paper: Algorithm 2 Policy (Random) Require: Vocabulary V; visited set C; embedding matrix E € R/V|d Ensure: Next token ID and embedding 1: Sample a permutation L = (01,...,0(vl) uniformly from V 2: Define p(v; ) as the rank of v in I 3: u*= arg minueV\C p(v; T) 4: return v*, Ev*
October 28, 2025 at 2:50 PM
maxine.science
Like literally haha
Screenshot of Algorithm 1 from the paper at https://arxiv.org/abs/2510.15511: Algorithm 1 SIP-IT: Sequential Inverse Prompt via Iterative Updates Require: Observed layer-l states H(e) e RTxd; vocabulary V; tolerance e ≥ 0. Ensure: Recovered sequence § = ($1, ...,ST). 1: 5+ () 2: for t = 1 to T do 3: C+ Ø • tested candidates : for j = 1 to V do 5: v; + POLICY (V,C, S,l) 6: if h, E An,t(Vj; E) then 7: 8: § g田2j D new candidate token v; (see Alg. 2 and 3) • verify vj (see Def. D.2) • hit! break 9: else 10: C+ CU {vis 11: end if 12: end for 13: end for 14: return s
October 28, 2025 at 2:41 PM
maxine.science
When you get the ADHD under control
GitHub commit heat map showing four weeks of nothing followed by one week of deep green
October 24, 2025 at 8:34 PM
maxine.science
this one broke me
Screenshot Power Like most things in Omarchy, power monitoring and alerting are handled through a script, which is executed every 30 seconds via a systemd timer. That's your crash course on systemd timers right there, Omarchy users! This script queries upower and then uses awk to parse the battery percentage and state. It's almost comical how hacky the implementation is.
October 24, 2025 at 5:10 AM
maxine.science
it’s wild to say “does not require hierarchy” — like one can inspect the model and see structural hierarchy still 😂

hierarchy in the latent (fast) vs output (slow) passes; the arrangement of nested recursion loops wrt gradients, and wrt supervision / losses

like .. it uses the word “deep”, c’mon 🤣
Pseudocode of the TRM screenshotted from the paper def latent_recursion (x, y, Z, n=6) : for i in range(n): # latent reasoning z = net (x, y, Z) y = net(y, Z) # refine output answer return y, z def deep recursion (x, y, Z, n=6, T=3) : # recursing T-1 times to improve y and z (no gradients needed) with torch.no-grad(): for j in range (T-1): y, z= latent recursion (x, J, Z, n) # recursing once to improve y and z y, z = latent recursion (x, J, Z, n) return (y.detach), z.detach)), output head (y), Qhead (y) # Deep Supervision for xinput, y-true in train dataloader: y, z = y-init, z_init for step in range (N.supervision) : x = input-embedding(x-input) (y, z), y hat, q-hat = deep-recursion (x, y, Z) loss = softmax-cross-entropy(y-hat, y_true) loss += binary cross entropy (q hat, (y hat == y-true)) loss. backward () opt.step () opt.zero-grad() if q hat > 0: # early-stopping break Figure 3. Pseudocode of Tiny Recursion Models (TRMs).
October 23, 2025 at 2:05 PM
maxine.science
Oh yeah! This whole work was how I got turned in to Floating Points!

Literally next up on what I was listening to for cooldown haha. It’s got frickin’ Pharoah Sanders!
Screenshot of Spotify playing Movement 6 by floating points, pharoah sanders, and LSO. A macro shot of saxophone keys is visible as the dynamic wallpaper for the track
October 22, 2025 at 6:22 PM
maxine.science
claude code on prod after telling me it wont interfere with the structure
photograph of the east wing of the White House getting taken out by a backhoe
October 21, 2025 at 2:13 PM
maxine.science
meta
Screenshot of Bluesky with the OP skeet followed by the exact referenced “Ed Zitron low key sucks” skeet
October 21, 2025 at 2:16 AM