grand theft eigenvalue 🔆
LightNews LightNews
banner
akhilrao.bsky.social
grand theft eigenvalue 🔆
@akhilrao.bsky.social
6.7K followers 4.5K following 19K posts
personal account

he/him
Posts Replies Media Videos
akhilrao.bsky.social
sorry I meant Paul Samuelson's nephew
Jim Downey "Jeffrey Epstein? The New York financier?" meme from Conan, but the text says "Larry Summers? Ken Arrow's nephew?"
November 17, 2025 at 9:32 PM
akhilrao.bsky.social
ceci est stationnaire
A highly erratic time series line graph showing Flow on the y-axis against Time on the x-axis. The most notable feature is a single, very tall, sharp initial spike at the extreme left of the graph, representing a maximum flow value much greater than all subsequent points. After this initial peak, the flow drops significantly and settles into a dense, highly variable and noisy pattern of rapid, high-frequency fluctuations and spikes that continue consistently for the entire duration of the time series. It's quoting a post of a different time series. Looks similar but it's not actually!
November 16, 2025 at 9:44 PM
akhilrao.bsky.social
"yo"
Kendall Roy from Succession telling us he's twin track, he's dead but he's alive
November 16, 2025 at 4:00 PM
akhilrao.bsky.social
the other aspect of this, and I think it applies more broadly than elite college profs or students, is that people are extremely earnest. necessarily quite different from the vibes around folks successfully shaping or wielding power in more adversarial settings. it's easy to become an earnest mark.
It’s all a bit humiliating—or it is for me. Because I did take time for this reporter, despite my misgivings. I even went so far as to write my thoughts out for him, on the chance they might be clarifying or useful. “The first thing to say,” I told him, “is that Africana Studies at Bowdoin is less a singular pursuit than a suite of scholarly disciplines, condensed around a set of objects and questions.” And then, warming to the pedagogical project, I talked about anthropology, art and architecture, music, religion, the history of science, whole grand traditions of invention and resolve. I talk about James Weldon Johnson, and Gwendolyn Brooks, and Charles Chestnutt, and Nella Larsen, and Hortense Spillers, and a good deal else and I mean… would you listen to how pathetic that all sounds?
November 8, 2025 at 5:19 PM
akhilrao.bsky.social
any resemblance unintentional, etc
A five-panel "American Chopper Argument" meme showing two men, Paul Teutul Sr. (seated, grey goatee) and Paul Teutul Jr. (black goatee), arguing intensely in an office setting. Panel 1 (Sr.): Sr. is seated at a desk, leaning back with his arm outstretched, shouting: "INFINITIES ARE UNPHYSICAL, THE 2D REPRESENTATION IS INADMISSIBLE" Panel 2 (Jr.): Jr. is seated facing Sr., leaning forward, and shouting aggressively: "RETURN TIME SO LONG IT MIGHT AS WELL BE INFINITY" Panel 3 (Sr.): Sr. is shouting directly at the viewer with an intensely angry expression: "NOT GLOBALLY" Panel 4 (Jr. and Sr.): Jr. is standing, lifting a chair above his head as if to throw it. Sr. is shouting at him, and the text reads: "IT'S A VALID APPROXIMATION IN THE REGIONS OF INTEREST" Panel 5 (Sr.): Sr. is standing in a doorway, pointing emphatically and shouting: "YOU'RE KILLING THE INDUSTRY"
November 7, 2025 at 1:30 PM
akhilrao.bsky.social
ceci est stationnaire
A highly erratic time series line graph showing Flow on the y-axis against Time on the x-axis. The most notable feature is a single, very tall, sharp initial spike at the extreme left of the graph, representing a maximum flow value much greater than all subsequent points. After this initial peak, the flow drops significantly and settles into a dense, highly variable and noisy pattern of rapid, high-frequency fluctuations and spikes that continue consistently for the entire duration of the time series.
October 26, 2025 at 10:18 PM
akhilrao.bsky.social
hope springs eternal
A three-panel meme based on the movie Donnie Darko. Donnie Darko (Jake Gyllenhaal): "the bubble bursting might reduce GPU prices" Dr. Thurman (Katharine Ross): "relative prices?" Donnie Darko (looking dejected): (No text)
October 26, 2025 at 2:01 PM
akhilrao.bsky.social
seems like there might be more alpha in zzzzzimberman tho ...
The tool gives corresponding (or last) authors the highest weighting: 100% of the citations that their paper accumulates contributed to their Sh-index. First authors receive the second-highest weighting, of 90% of the total citations. Second authors get a 50% weighting. Other co-authors get 25% if there are six or fewer authors, or 10% if there are seven or more authors, according to a preprint published in September1 by the tool’s creators.
October 25, 2025 at 7:36 PM
akhilrao.bsky.social
sonnet 4.5 doesn't want to talk about this, it's dangerous
a normalized distribution function. P(lambda) equals 1 over pi times cos(lambda) over the square root of (sin i squared minus sin lambda squared)
October 21, 2025 at 7:22 PM
akhilrao.bsky.social
this is a good point. "machines are accountability sinks" cuts both ways. I expect this to be a more general result
Snippet from Tao's thread: This can lead to some redundant effort by multiple researchers to search in vain for the same non-existent literature on a problem, if the repeated failures to locate that literature are never reported; or to mistakenly believe a problem to be open, when in fact no serious literature review was previously attempted, and a solution was in the literature the entire time. But when applying an AI-powered literature review tool systematically to a large corpus of problems, it becomes more natural to report both positive and negative results ("Of the 36 problems reviewed by this tool, 24 (66%) returned additional results that we determined to be relevant, while 12 (33%) only produced references that were already known to us, or irrelevant"), which could help provide a more accurate impression of what the extant literature on a problem actually is. (4/4)
October 18, 2025 at 1:41 AM
akhilrao.bsky.social
gonna depict the virtuous LP as buff cheems and the hated VAR as sad cheems is what
buff cheems is local projections, sad cheems is VAR
October 17, 2025 at 12:42 PM
akhilrao.bsky.social
here, take this
Jim Downey looking surprised. "THE SOUTH SEA COMPANY? THE GOVERNMENT DEBT TRADERS?"
October 15, 2025 at 10:25 PM
akhilrao.bsky.social
concerning levels of graduate seminar particles
Gossip as market infrastructure - Smith requires market transparency. Villagers' gossip is epistemological labor to restore information symmetry when one party is opaque.
October 15, 2025 at 5:42 PM
akhilrao.bsky.social
lock in
a cute and slightly chonky little cat looking alert. text says "heavy breathing intensifies"
October 15, 2025 at 5:10 PM
akhilrao.bsky.social
inspired ty
fancy Pooh meme. regular Pooh says "Inputs that deliver publishable outputs", fancy Pooh says "preferred calibration"
October 14, 2025 at 5:34 PM
akhilrao.bsky.social
common mistake
Claude interface. Chat goes: View analysis Let me try a simpler approach: View analysis View analysis View analysis Ah, this is JavaScript not Python. Let me work with that:
October 13, 2025 at 5:49 PM
akhilrao.bsky.social
the subtitle is wrong: it should read "in earth orbit", not "active". oops.
line of R code that reads `filter(ExpandedStatus == "In Earth orbit") %>%`
October 7, 2025 at 12:55 AM
akhilrao.bsky.social
anyway all the data here is freely available online thanks to @planet4589.bsky.social. took me a few minutes to make this. kinda interesting

planet4589.org/space/gcat/i...
A scatter plot titled "Starlink Satellite Altitude by Launch Date" shows the mean orbital altitude (in kilometers) of active Starlink satellites against their launch date, ranging from early 2020 to mid-2025. There are 8229 satellites represented. The Y-axis, "Mean Altitude (km)", ranges from approximately 150 km to 550 km. The X-axis, "Launch Date", spans from 2020 to 2025. The data points are clustered at several distinct altitude bands. Early launches in 2020 show satellites at various altitudes between 300 km and 550 km. From late 2020 to mid-2021, a prominent cluster appears around 550 km. A denser band of satellites also emerges just below 550 km starting in late 2021 and continuing through 2022. Beginning in late 2023 and especially into 2024 and 2025, a significant shift is visible. While a cluster of satellites remains around 550 km, many newer launches are concentrated at lower altitudes, specifically around 400 km and then a very dense band appearing around 350 km. There are also many satellites launched in 2024 and 2025 that are at altitudes around 275 km. Some data points are scattered below 200 km, particularly for early 2020 launches and a few in late 2024 and 2025, indicating either deorbiting or initial deployment orbits. The data is from Jonathan McDowell's General Catalog of Artificial Space Objects (GCAT).
October 6, 2025 at 11:02 PM
akhilrao.bsky.social
preference drift in this type of cybernetic system, directed or otherwise, seems ever more interesting
Because these AI models are trained on such vast corpora outputs which are novel to them should mostly be actually novel, the question is how to get those outputs. My theory is that models will perform a kind of iterated tuning to extend their training distribution in a particular direction. Similar to how new music genres are made when a band cultivates excellence at the edge of the distribution of existing music, I think models will perform novelty search by going out to the edge of what is known and applying relatively objective standards to what they find there. People aren't really doing this right now because the best AI models are proprietary and expensive to tune.
October 2, 2025 at 12:35 PM
akhilrao.bsky.social
this may be right for AI, I think for H2H the issue is often about identifying that the generator is from an uninteresting or undesired class. a dense page of econ jargon may be slop to some readers but not others; slop can be positional that way
Ultimately I think what gives the impression of slop is when you can fully infer the generator. When you type in "fat cat rolling around on carpet 4k digital" the video you get back from that is entirely defined by the prompt.
October 2, 2025 at 12:33 PM
akhilrao.bsky.social
i was starting to worry
I hope you're doing well! We're currently looking to purchase surplus semiconductor manufacturing equipment and would love to help you free up valuable space and resources. Do you have any underutilized equipment in production or storage that you’d consider selling? We’d be happy to make the process easy and efficient for you.
October 1, 2025 at 7:34 PM
akhilrao.bsky.social
🥲
Visual Enhancements * Emojis for better section identification
September 27, 2025 at 4:23 PM
akhilrao.bsky.social
huh! super interesting, ty

vaguely recall, years ago, chatting casually with a CS grad student about identification... they loosely said something like "equivalent predictions mean it's the same model", which seems echoed here
Machine learning (ML) systems often exhibit unexpectedly poor behavior when they are deployed in real-world domains. We identify underspecification in ML pipelines as a key reason for these failures. An ML pipeline is the full procedure followed to train and validate a predictor. Such a pipeline is underspecified when it can return many distinct predictors with equivalently strong test performance. Underspecification is common in modern ML pipelines that primarily validate predictors on held-out data that follow the same distribution as the training data. Predictors returned by underspecified pipelines are often treated as equivalent based on their training domain performance, but we show here that such predictors can behave very differently in deployment domains.
September 23, 2025 at 5:04 PM
akhilrao.bsky.social
LLMs be like
Query successful A drawing of Karl Marx, depicted as a muscular superhero with a flowing white beard and hair. He's wearing red overalls and gray armor on his shoulders and arms, holding a large yellow hammer with a red circle on its head and a yellow, sickle-shaped boomerang. The caption reads, "I'M JOINING THE WAR ON TEXT CONTENT ON THE SIDE OF THE TEXT CONTENT."
September 22, 2025 at 8:18 PM
akhilrao.bsky.social
BCG was cooking
Accounting data is frequently misleading for cost analysis. The choice of treatment as expense versus capital can distort apparent cost change. Over time the experience curve has become recognized as essentially a pattern of cash flow. The average cost is by definition the total expenditure divided by the total output. The unit cost is the rate of change in that ratio. Projection of this relationship is frequently both simpler and more accurate for cost forecasting than even the most elaborate conventional accounting analysis.
September 19, 2025 at 5:19 PM