Vikram Saraph
@vikramsaraph.com
Software engineer, AI/ML researcher, and mathematician at Johns Hopkins APL. Former New Englander, current Marylander. Brown CS PhD and Notre Dame math alum. Nerd of sorts (computers, math, language, puzzles, games, books, music). Opinions are my own.
Almost every LLM that reaches a broad audience is quantized—compressed to run cheaply. USC's Emilio Ferrara argues this compression is treated as a safety nonevent, when it should be treated as a change to the deployed system. This has policy implications, he says.
The Model You Audit Is Not the Model You Ship
Almost every LLM that reaches a broad audience is quantized—with under-appreciated ramifications for AI safety, writes Emilio Ferrara.
techpolicy.press
Hmm, not sure I’m a fan of likes of posts by people you follow being more visible now in the UI. I know this data is all public by design anyways, but private likes on Twitter is something I think I prefer.
At The Anthem in DC, at intermission between Stars and Broken Social Scene
I am pumped to be seeing Metric perform live a week from now. Never been to a concert despite them being my favorite band.
Another absolutely bonkers math result from a frontier LLM (announced in a tweet, no less): www.reddit.com/r/math/comme...
From the math community on Reddit: The Jacobian Conjecture is False Per Anthropic (Link in Description)
Posted by overthinker020 - 724 votes and 288 comments
reddit.com
Nerd thing: so there's this tech/math/nerd culture that's hard to describe, but where there's a kind of enlightened playfulness, which if you like you can trace very far back, and which perhaps hits a kind of high watermark with Claude Shannon, routes through late 60s/70s hacker culture, and so on.
The Sheep Detectives is one of the weirdest movies I’ve ever watched
TIL of roaring bitmaps. Bitmaps that enable fast set operations: vikramoberoi.com/posts/a-prim...
A primer on Roaring bitmaps: what they are and how they work
A detailed primer on Roaring bitmaps explaining what they are, how they're different from traditional bitmaps, and how they work internally.
vikramoberoi.com
THE LAST TIME BOSTON HAD 2 CONSECUTIVE DAYS OF ABOVE 100 DEGREE TEMPS WAS AUGUST 12-13 1944
Our son seems to very conveniently pick the weekends when we *aren't* traveling (or otherwise have big plans) to get sick.
I am boosting this post to reach an audience that wouldn’t ordinarily see this. HOPE is a hacker conference with lots of cool stuff. While I definitely do not endorse all the content that’s been presented, it’s a venue that gives a voice or platform to people who might not otherwise have one.
Hackers On Planet Earth 🏳️🌈🏳️⚧️@hope.net · last mo.An important message from HOPE: www.2600.com/content/help...
I have finally learned to use the lab’s vinyl printer to make stickers and it’s awesome. Made 100+ of them today.
Seems like a good thing for privacy? Makes me wonder how much this is them feeling like they have to compete with Signal (article mentions same feature on Signal).
evacide@evacide.bsky.social · last mo.WhatsApp is rolling out usernames, so that you no longer have to give your phone number in order to do business in the many countries where WhatsApp is the default method of communication: www.bbc.com/news/article...
going to give this book a read: staffeng.com been a few years since I intentionally read a non-fiction book.
Staff Engineer: Leadership beyond the management track
staffeng.com
This is the first time I’ve ever used/owned AirPods and wow I didn’t know what kind of features I was missing out on
This has finally been fixed and I’m going to choose it’s because of me complaining on the Internet
Vikram Saraph@vikramsaraph.com · last yr.@nytimes.com still needs to fix its spelling of Bluesky (note the capitalization) in its sharing feature on the mobile app:
powerful essay. much to think about
The fall of the theorem economy
How AI could destroy mathematics and barely touch it
davidbessis.substack.com
About to try my hands DIY fixing our Toniebox with a soldering iron... we'll see how this goes
ML researchers know Python. Proof engineers know Lean. Never the two should meet.. Until now! Announcing Lean.py, effortless Lean to Python and Python to Lean bindings! - Write Lean tactics in Python - Access the Python ecosystem in Lean github.com/kiranandcode...
![/-- Expr-based equality check. Receives `Lean.Expr` trees (via the
`derive_python` Reflect registration) and delegates conversion entirely
to Python's `lean_to_sympy` module. -/
@[python "sympy_expr_eq_accept"]
def sympyExprEqAccept (lhs rhs : Lean.Expr) : IO Bool := do
init ()
let mod ← import_ "lean_to_sympy"
let checkFn ← mod.getAttr "sympy_eq_check"
let convFn ← mod.getAttr "expr_to_sympy"
let lhsSy ← convFn.call #[← Py.ofLeanObj lhs]
let rhsSy ← convFn.call #[← Py.ofLeanObj rhs]
(← checkFn.call #[lhsSy, rhsSy]).toBool](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:i32jjsch6xqcguzsf2lgbfyu/bafkreidjowsbwstojoc67hyrxgolgekqqduzv5afwd3wfrg4sr4lqvhpgm)
Kiran@kirancodes.me · 3mo agoPython and lean are the same language right? (this is lean code that is manipulating Python numpy objects)
![def runDemo : IO Unit := do
IO.println "demo_run:"
let a <- NDArray.ones .f64 [3, 4]
let b <- NDArray.ones .f64 [4, 2]
let c <- NDArray.matmul a b
let d <- NDArray.reshape c [6]
IO.println s!" {<- d.repr}"
#eval runDemo](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:i32jjsch6xqcguzsf2lgbfyu/bafkreieycwm352l4xweu5jzffnk54j255nqexcepx3mqnuhs37amap3kra)
One of the things the AI ethics people got right that the AI safety people mostly flubbed is however useful AI turns out to be, it will ruin a bunch of peoples’ lives if built and deployed carelessly
☀️ Starshine@starshine.bsky.social · last mo.My biggest worry about llms for like three years now has been the explosion in revenge porn and god I wish they didn’t work but they uh absolutely do
learned of spherical CNNs while reading about flow matching: research.google/blog/scalabl...
Scalable spherical CNNs for scientific applications
Posted by Carlos Esteves and Ameesh Makadia, Research Scientists, Google Research, Athena Team Typical deep learning models for computer vision, li...
research.google
I try to keep up with the kids but Saturday is the first time I’d heard of “unc”. And now I see unc as an answer to today’s NYT mini.
Contact is a totally new way to think about how nervous systems are organized. As part of this science, we're releasing all of our code and data #opensource (happening presently), and if you have a dataset that could benefit from a contactome, get in touch please :) github.com/aplbrain/
APL BRAIN
APL BRAIN has 118 repositories available. Follow their code on GitHub.
github.com
new post: how I develop recently using local models. the tooling is now good enough to do agentic workflows and everyone should give them a try! vickiboykis.com/2026/06/15/r...
Running local models is good now
Local agentic coding has gotten great over the past few months
vickiboykis.com