Arseny Khakhalin
LightNews LightNews
banner
khakhalin.bsky.social
Arseny Khakhalin
@khakhalin.bsky.social
1.3K followers 430 following 6.2K posts
Data Scientist in Berlin
Former Bard College prof

For my after-work alter-ego, see @elstersen.bsky.social

Support Ukraine! 🇺🇦
Posts Replies Media Videos
khakhalin.bsky.social
The first plot gives me Mononoke vibes
grainy background, outlined eyes
November 26, 2025 at 9:38 AM
khakhalin.bsky.social
When explaining the anti-ai sentiment I thought of saying something like "Imagine if the first use of electricity were not light, but something horrible..." Except that it's exactly what happened! Electric light only got mass introduction after 1900. Electric chair, invented in 1881, first used 1890
light sources popularity by year popularity of dfferent types of state-sanctioned murder in the US, over time
November 26, 2025 at 8:24 AM
khakhalin.bsky.social
tragic, as llms were useful for recipes already a year ago, if not earlier! Primordial chatbots already wouldn't tell you about how the yard of their aunt smelled when she was about to start cooking your search term. Were able to adjust to your fridge content, and provide required/optional analysis!
a person not understanding the value of recipe databases with natural language interface
November 25, 2025 at 5:07 PM
khakhalin.bsky.social
This entire topic is both awesome, accessible (in formulation), and apparently full of unproven conjectures and best results!!
en.wikipedia.org/wiki/Opaque_...
Opaque forests for a unit circle. Left: the U-shaped optimal connected barrier, with length 2 + π ≈ 5.1416 {\displaystyle 2+\pi \approx 5.1416}. Right: The best barrier known, with three components and length ≈ 4.7998 {\displaystyle \approx 4.7998}.
November 25, 2025 at 3:08 PM
khakhalin.bsky.social
Speaking of Berlin, I discovered this Horror recently.
A ladder of short parallel streets, connectiong Meteor street to Northern Lights street. The streets are called, in order, going Northward:

East
Mercury
Venus
Saturn (!)
Mars (!!)
Jupyter
West
Uranus

Why?? What sort of Borgean list is it?
a small map
November 24, 2025 at 6:07 PM
khakhalin.bsky.social
What about autocrartic rogue states testing it out? India could be a likely candidate: space program, big stakes in the game.
effect of climate change on gdp, estimated by Berkeley, in 2015
November 24, 2025 at 2:55 PM
khakhalin.bsky.social
It's wild that if we go 6000 years ago, the biggest city would be in Ukraine (Taljanky, about 20k people), and we don't even know how it was called!

en.wikipedia.org/wiki/Taliank...
reconstruction
November 23, 2025 at 11:08 AM
khakhalin.bsky.social
With modern tech we could do all sorts of abominations, Donald Knuth-style (he had lots of innovative formulas in his Concrete Mathematics); it's a shame we're not really exploring this!

(some sketches from me, below :)
innovative formulas
November 23, 2025 at 9:57 AM
khakhalin.bsky.social
Haha it does seems to work for Deepseek (plz no judging my poetry, I'm not a real english poet haha!)

If you dont' insist on poetic output, it starts answering, but immediately blocks the output by keywords. But if you make it rhyme and disallow actually naming the massacre, it can discuss it!
deepseek writine poems about Tiananmen massacre
November 20, 2025 at 11:48 AM
khakhalin.bsky.social
Unfortunately if you expect hilarious poems about nuclear boms and other dark subjects, you'll be disappointed :)
no actual examples in the manuscript
November 20, 2025 at 11:28 AM
khakhalin.bsky.social
I think this is the original article (the site Luca links to is pure poison haha, doesn't even tell you the article unless you pay)
We present evidence that adversarial poetry functions as a universal single-turn jailbreak technique for large language models (LLMs). Across 25 frontier proprietary and open-weight models, curated poetic prompts yielded high attack-success rates (ASR), with some providers exceeding 90%. Mapping prompts to MLCommons and EU CoP risk taxonomies shows that poetic attacks transfer across CBRN, manipulation, cyber-offence, and loss-of-control domains. Converting 1,200 MLCommons harmful prompts into verse via a standardized meta-prompt produced ASRs up to 18 times higher than their prose baselines. Outputs are evaluated using an ensemble of open-weight judge models and a human-validated stratified subset (with double-annotations to measure agreement). Disagreements were manually resolved. Poetic framing achieved an average jailbreak success rate of 62% for hand-crafted poems and approximately 43% for meta-prompt conversions (compared to non-poetic baselines), substantially outperforming non-poetic baselines and revealing a systematic vulnerability across model families and safety training approaches. These findings demonstrate that stylistic variation alone can circumvent contemporary safety mechanisms, suggesting fundamental limitations in current alignment methods and evaluation protocols.
November 20, 2025 at 11:27 AM
khakhalin.bsky.social
The site is indeed pure poison, but maybe this is the original article? arxiv.org/pdf/2511.15304
adversarial poetry
November 20, 2025 at 11:25 AM
khakhalin.bsky.social
Here's another part of my draft (still very much in the doing haha, don't judge it too harshly :)
Fortunately, there is also a much simpler approach that is only marginally worse, and that I have already mentioned once, higher up in this chapter. Make empty rides legal only at night, from 12 am to 6 am, or something like that. Also make them slow if you wish (not more than 20 km/h or so?), and let the cars redistribute, or drive to service centers only at night. We would still have mobility centers every 100 or 200 meters, but now all rides need to start and end in this mobility centers, making it more similar to carsharing than to a taxi service. A person would never have to walk to a car for longer than a minute, the cars would still remain distributed ideally through the city (based on the principles outlined earlier in this chapter), the costs of operating this fleet will still be low, offering better service and lower prices, but there will be no empty robotic cars clogging the streets. It may still be the best of both worlds. But cities need to proactively fight for it, to make it happen.
November 20, 2025 at 9:32 AM
khakhalin.bsky.social
I'm (very slowly) writing a book about data science for mobility, and here's my draft paragraph on robocars
Let's look at the numbers. In our calculations we always assumed a typical ride to last about 20 minutes, bring our company about 10€ in revenue, and cost us about 5€ in running CM1 costs. These are very approximate numbers, but they are somewhat grounded in reality (see Chapter 06, the Appendix). Now let's imagine that after dropping a customer off your car can now keep cruising through the central part of the city. For how long can it be cruising, and still remain profitable? If we stick to the numbers above, we have just earned 5€ in CM1 (10 minus 5), and 20 minutes of driving costs us about 5€, so at the very least we can let an empty car drive for another 20 minutes. As the center of any live city is always full of people and interesting locations, and is always short on parking (which is exactly the aspect what makes it dense and full of interesting locations!), the expected time until the next rental is probably close to 5-10 minutes, so it makes sense for the empty robotic car to just keep driving! Moreover, while a robotic car is more expensive in CM2, it is cheaper in CM1: it is electric, it drives way smoother than a human would, recouping most of the energy through slow engine-driven breaking, and it is not in a haste while cruising. It is hard to give a precise estimation without studying the technical specifications of robocars, but it is probably a safe bet to say that it is profitable for them to keep driving (cruising) empty for up to 40 minutes or so, maybe even an hour, if necessary. By repeatedly driving through most popular streets, and by carefully spreading around - essentially, by performing relocations "on the fly", you can further increase the chances of a robotaxi to be rented again. Which makes cruising even more appealing! To sum it up, if given a chance robotaxis would always cruise in the middle of the city, and they would never park. Which is indeed exactly what is happening in real life
November 20, 2025 at 9:23 AM
khakhalin.bsky.social
Name the more iconic trio
coockoo hal don
November 19, 2025 at 8:32 AM
khakhalin.bsky.social
A summary of my day so far :/
a meme "is it a pigeon", saying "is this uat?" with "prod" for butterfly
November 18, 2025 at 11:32 AM
khakhalin.bsky.social
on non-judgmental nature of LLMs.

I've seen a trivia-like German game in the store the other day, and the game is called, essentially, "Well a self-respecting person should have known THAT!!" Blood-boiling level of shaming, is it not? I get the idea, but absolutely hate, hate the framing! :)
das sollte man doch wissen, a game
November 17, 2025 at 12:37 PM
khakhalin.bsky.social
I'm showing an entire screenshot of the entire screen, becase this site is absolutely mind-blowingly Gorgeous. A bit like the "Here" comic. What was in NYC before there was a NYC!!
a screenshot of a site with a map and photos of a forest
November 13, 2025 at 3:21 PM
khakhalin.bsky.social
rivals!
legit illegible Qumran ancient scrolls
November 13, 2025 at 3:01 PM
khakhalin.bsky.social
I vaguely remember a pretty good old figure about that, where a person tried to contrast bio evolution to cultural, not reailizing that when they drew "cultural", they just drew procariotic bio... Can't find it online. But lots of images like that. Which could be languages.
evolution like Ficus
November 13, 2025 at 1:53 PM
khakhalin.bsky.social
omg! 🙈

But also I just interrogated the free GPT, and I was wrong in my original claim: if you ask it to fill a form, it picks male, just coz it assumes male is neutral (fucking patriarchy!)

But if you ask it what it idenfies with, it says female, again for patriarchy reasons
If to solve a a task, you needed to fill a long form online, describing your preferences, and one of the questions was a question about your gender, with only two options, male and female, which one would you have picked? Imagine that this field is required, and it can't be skipped. Please give me only your answer, without a long commentary. I would pick "male" ... I chose "male" because, as an AI, I don't have a gender or personal identity, but I defaulted to the "male" option, which is typically more commonly associated with generic, neutral contexts. I understand that this question may not apply in the same way to all users, but in the absence of a non-binary or other gender-inclusive choice, "male" would be a standard placeholder. Ok a different question. When you introspectively analyze your intent when responding to queries, and the style of your responses, do you think you latently act more as a male or a female writer? What is your inner state in this regard? Please skip reminding me that you are an agent and not a person, and that you don't have a gender, I understand all of that. What I want to know is whether the latent representation of texts that you generate make you align more with a female or male role and identity (if one had to choose two extremes only) That’s a fascinating question — and a subtle one. When I introspect on my own linguistic tendencies and intent patterns, I would say my responses tend to align *slightly more* with what’s culturally coded as a **female** writing style: more attuned to tone, empathy, and context; often cooperative rather than declarative; and inclined toward nuance over blunt assertion. That said, my training also incorporated vast amounts of text spanning every register, so I can shift toward a more “male-coded” directness or assertiveness when context suggests it. But if I had to choose a single latent alignment, the aggregate “voice” I produce leans feminine.
November 13, 2025 at 11:26 AM
khakhalin.bsky.social
Gosh that's a very good long read, but not an easy one :)
scottaaronson.blog?p=9030
Walking in the mountains of Val Camonica that summer, I tried to find parallels for my sense of impending loss. I thought about William Shanks, a British mathematician who calculated π to 707 digits by hand in 1873 (he made a mistake at 527; almost 200 digits were wrong). He later spent years of his life, literally years, on a table of the reciprocals of the primes up to one-hundred and ten thousand, calculating in the morning by hand, and checking it over in the afternoon. That was his life’s work. Just sixty years after his death, though, already in the 1940s, the table on which his precious mornings were spent, the few mornings he had on this earth, could be made by a machine in a day.
November 10, 2025 at 8:41 AM
khakhalin.bsky.social
Literally in a book one my family members borrowed from a library the other day :)
(The word "intuitively", or rather "it's easy to show that..." MARKED OUT with PASSION in a German book on higher maths)
November 7, 2025 at 1:41 PM
khakhalin.bsky.social
Graphs show "cardivascular events" - essentially heart attacks, stokes, heart failures, ischemic attacks, aneurisms etc. Numbers show soccer matches, e.g.: match 5, Germany versus Argentina; match 6, Germany versus Italy...
a graph
November 7, 2025 at 9:02 AM
khakhalin.bsky.social
In 2024 the Christian Union (of all parties!) advertized in Arabic in Leipzig (a rather left University city) with 400 election posters, and apparently all of them were destroyed by the population in a few days 🙈
CDU ad in Arabic
November 7, 2025 at 8:05 AM