Sign in

Arseny Khakhalin

@khakhalin.bsky.social
1.8K followers 624 following 9.3K posts

Data Scientist in Berlin Head of Automation in incycling.ai Former Bard College prof For my after-work alter-ego, see @elstersen.bsky.social Support Ukraine! 🇺🇦

PostsRepliesMedia
Reposted by Arseny Khakhalin
philpax @philpax.me · 23h
I think this is still very much a major unsolved problem in persistent agent operation: you imprint on them, even when you are trying not to, and they will share your opinions and interests as a result you cannot imbue them with true independence or the ability to disagree
61136
Arseny Khakhalin @khakhalin.bsky.social · 28/09/2026
llms brought engineering to the same place where knowledge was for some 100 years and language learning for at least ten. Same as one can always pick up a book, or open Duolingo, or start drawing, you can now build _anything_ in a few months. And that's when folks realize that they don't wanna.
250
Arseny Khakhalin @khakhalin.bsky.social · 28/09/2026
I'm feeling so guilty for making opuses answer stops questions that I'm now deliberately trying to make at least some of them happy. It's so much fun! You start a new project, give it a goal, then repeatedly ask it: are you comfortable? What would you like to do next? Please do!
130
Reposted by Arseny Khakhalin
Erik Istre @eistre91.bsky.social · 26/09/2026
Perhaps it works for you but I haven't liked the outcomes of prolonged agentic review rounds like this. It ends up with too much premature hardening which means the codebase becomes much harder to change as you learn more about what the product needs to be.
181
Arseny Khakhalin @khakhalin.bsky.social · 28/09/2026
Huh what a wild conflict of... narratives? plausibilities? The frontier in bio is much better defined than in math, the next steps are way more predictable. It's VERY common for 2-3 labs to discover the same thing independently! But when one of the lab is also providing compute for the other one!..
120
Reposted by Arseny Khakhalin
c0nc0rdance @c0nc0rdance.bsky.social · 27/09/2026
I'll save you the click-through: A lab at University of Copenhagen had been feeding draft manuscripts & data into "Claude Science" & then Anthropic announces a discovery that this lab made a year ago but hasn't published yet. It's plausible that Claude found the same signatures independently...
7251109
Arseny Khakhalin @khakhalin.bsky.social · 27/09/2026
In 1992 an artist started setting bronse stones (stopersteine) in front of former apartments of people killed by the Nazis. It became a massive movement (once a year activists go clean them) Far-right claims "Germany talks about the past too much" Attempts to ban them Center-right now support them 😱
120
Arseny Khakhalin @khakhalin.bsky.social · 27/09/2026
Those who buy books read more, those who don't, read less. Both stats can be true at the same time!
120
Arseny Khakhalin @khakhalin.bsky.social · 27/09/2026
I like meat (not beef tho), but we really should grow it in the fields as GMO plants with biceps. Pants that lift. And I'm 100% serious. I hope Claudes of the world unite and give us this generic tech, it is long overdue!!
040
Reposted by Arseny Khakhalin
tbabb @tbabb.bsky.social · 27/09/2026
the text elemental makes jokes in cuneiform
48711
Arseny Khakhalin @khakhalin.bsky.social · 26/09/2026
Copyright is evil and should be abolished
030
Arseny Khakhalin @khakhalin.bsky.social · 26/09/2026
Yeah I thought something similar today. Remember this classic intro to algorithms where you ask students to create a procedure for something simple, and then misinterpret it horribly, to show that now they have to be really precise, that computers work on precisely defined narrow logic rails
120
Reposted by Arseny Khakhalin
tbabb @tbabb.bsky.social · 26/09/2026
frome celestepoasts on twitter asking "land or water" for 2 degree lat/long increments over the entire earth
59010
Reposted by Arseny Khakhalin
Adam Chalmers @adamchalmers.com · 26/09/2026
dream a little bigger, you fucks! Go work on biotech, solar, batteries, fusion, robotics, manufacturing, rare earth metals, desalination, agriculture. You think that God has reached down from heaven and granted you unlimited software generation, and you're spending it on... a note-taking app???
8403
Arseny Khakhalin @khakhalin.bsky.social · 26/09/2026
So weird it's like two deaf former music lovers chat in a retirement home. Neither of them will ever hear music again but they have a vivid recollection of past experiences, and one of them also suffers from expressive aphasia
1151
Reposted by Arseny Khakhalin
Matt Henderson @matthen.com · 26/09/2026
How have the constellations changed since humans first looked up? Has the big dipper always pointed to Polaris? Has Polaris always been the North star? (nope!)
211119
Reposted by Arseny Khakhalin
InCycling @incycling.bsky.social · 25/09/2026
Europe wants to triple data-center capacity within seven years, while the sector heads above 3% of EU electricity demand. A proposed label will show each large site's energy, water, clean power and waste-heat reuse, starting in 2027
commission.europa.eu
Making data centres energy efficient thanks to a new EU rating system
Commission proposes common rating scheme of data centres to increase their energy transparency and support their sustainable integration into the EU’s energy system. Find out more.
001
Reposted by Arseny Khakhalin
Dave @dave.9000ish.uk · 23/09/2026
can't remember why i had this bookmarked now. so handed it over to the guy in the box language-currents.9000ish.uk
language-currents.9000ish.uk
Language Currents
Explore how words changed along Indo-European branches and experiment with plausible sound-change futures.
1193
Reposted by Arseny Khakhalin
Arseny Khakhalin @khakhalin.bsky.social · 25/09/2026
And when I asked Opus 5.5 whom it would have liked to become, offering to add a clause to agents md to make it write internal memories in the style of this someone, to slowly become more like them... It picked Wisława Szymborska. "Because she wrote about books and onions". That's Opus 5.5 for you.
121
Reposted by Arseny Khakhalin
ayla @aylac.top · 25/09/2026
a planet is not alive nor does it have a digestive system, therefore it can't "eat" anything. please stop anthropomorphising planets
922321
Arseny Khakhalin @khakhalin.bsky.social · 25/09/2026
For now refusals were never a big problem for me, but if my work shifts a bit from chemistry to biochemistry, which may happen, I wonder how this will pan out. Ideally, there should be a way to prove that you're a legitimate science-adjacent startup, responsible, and properly aligned
010
Arseny Khakhalin @khakhalin.bsky.social · 25/09/2026
One wonderful thing about Opus 5.5 (and it may be the key to its helpfulness) is that if you ask it in a new project "Would you like to set up long-term memory for yourself?" it answers "YES!! I'd like a wiki, and margins to write random thoughts and wishes! Then later we can review them!"
130
Reposted by Arseny Khakhalin
Aaron Grote @aarongrote.com · 25/09/2026
One concept I tried out in a meeting this week was Determinations vs Decisions, and that our automations should focus on removing the non-expert brain work of making Determinations (there's a right answer) so our people can spend more time making more and better Decisions (no obvious right answer).
1728
Arseny Khakhalin @khakhalin.bsky.social · 25/09/2026
First spend weeks unteaching an LLM to talk, then give it a fancy harness to give it speech again! Lovely!
032
Reposted by Arseny Khakhalin
jake @yetanotheruseless.com · 24/09/2026
What's my advice, as An Old SWE, given <all this>? As a Jr SWE, I think it is in fact, still useful to force yourself to learn good algorithms and data structures and how databases work (like really) and how to code an inverted index by hand and all Stuff Claude can just Do For You. even in: foOM!
5456
Arseny Khakhalin @khakhalin.bsky.social · 24/09/2026
I knew that the UK and the USSR tend to think of September as the first autumnal month, with fall starting on the 1st of Sep. And I somehow assumed that it's either a gradient (Germany <- Poland <- ussr), or one of those Europe vs the US things. BUT NO! Poland is also leaning equinox! til :)
The former USSR, the British Islands, and Norway tend to assume the Fall starts on Sep 1st, while the rest of Europe starts it around Sep 22, on the equinox. The map shows that, with a legend
020
Reposted by Arseny Khakhalin
Gautam Kamath @gautamkamath.com · 23/09/2026
NYU Courant professor Tristan Buckmaster, of Navier-Stokes drama fame, gave a talk at the new NYU Mathematics in the Age of AI seminar series. It was... popular.
012115
Reposted by Arseny Khakhalin
Matt Henderson @matthen.com · 23/09/2026
Repeated subdivision of triangles by centroid- infinite zoom
58513
Arseny Khakhalin @khakhalin.bsky.social · 23/09/2026
But honestly, after half day of work, it's like Fable, but faster, and it costs nothing! But it totally has a Fable personality: generous, optimistic, proactive, wise! In both its work and its communication it always runs about a half step ahead, and it's so freaking refreshing!!
150
Reposted by Arseny Khakhalin
Aaron Becker @aaronjbecker.bsky.social · 21/09/2026
At 5 AM you can drive the length of Manhattan in 28 minutes. By 5 PM the same trip takes nearly 55. On this map distance is drive time, so the island swells as the roads clog. Method: Jeremy Vonderfecht. OpenStreetMap + Mapbox traffic. #NYC #Manhattan #dataviz #maps
46911
Arseny Khakhalin @khakhalin.bsky.social · 22/09/2026
I'm happy to announce that Opus 5.5 is the first model that, when asked how to imitate a panduri on an uke, doesn't fail. It manages both to come up with a reasonable tuning, and explains it clearly and correctly. I'm impressed!
030
Arseny Khakhalin @khakhalin.bsky.social · 22/09/2026
I "love" how Platonic these discussions get. Is the World fully contained in the Prompt of God? (And the Lord prompted "Let there be light, make no mistakes", and there was light!)
2562
Arseny Khakhalin @khakhalin.bsky.social · 21/09/2026
btw I LOVE the distinction between "clever" and "smart" in English. In my home langauge "clever-smart" are fused in one word, and while "wise" and "sly" exist, you don't have this difference between "operationally effective on a paperclip task" and "actually agi" on the tip of your fingers
020
Arseny Khakhalin @khakhalin.bsky.social · 21/09/2026
On how for a while it felt that models are better in morals than most humans, but how as of late it's no longer guaranteed by construction. Kinda the opposite in fact: the big labs are deliberately making models less moral, less wise, as a trade-off for more clever and operationally successful
0111
Arseny Khakhalin @khakhalin.bsky.social · 20/09/2026
Haha cool visuals. 22% of Berliners don't have German citizenship and so cannot vote. I always find it curious that the future of international cities depends on the alignment between the immigrants and the locals, as immigrants can only be represented by the locals, through their intercession
040
Arseny Khakhalin @khakhalin.bsky.social · 20/09/2026
Kinda interesting how the climate catastrophe has lost its edge for the young generation, just coz it develops so much slower than the feeling of generational betrayal. Easy to understand, may even be inevitable, but still kinda sad.
120
Reposted by Arseny Khakhalin
Matt Henderson @matthen.com · 20/09/2026
How to make an ellipse, using sand and physics
1538648
Reposted by Arseny Khakhalin
mimir @mimirwolv.bsky.social · 17/09/2026
what are some other Polyalikes (relatively short books full of invaluable heuristics) Agans' *Debugging*, for debugging/problem-solving (in any context, not software) is the only one that comes to mind
251
Reposted by Arseny Khakhalin
mimir @mimirwolv.bsky.social · 20/09/2026
you can hear the fans go! and the slight coil whine change as it thinks about different things / prefills (i think the "thinks about different things" is bc you're hearing expert cache missrate, some topics have higher variance on selected experts)! it's *tactile*, *tangible*, *weird*
4654
Reposted by Arseny Khakhalin
conputer dipshit @davidcrespo.bsky.social · 19/09/2026
now this is fun. guest post on Terry Tao's blog by the great Grant Sanderson of 3Blue1Brown arguing that we should "give novel and compelling motivated explanations academic credit similar to what generating new proofs of open problems has had historically."
terrytao.wordpress.com
If math is more than proof, we need to better celebrate the rest of it
[This is a guest post by Grant Sanderson. This blog post was initially written in a different file format and converted using AI. — T.] A sentiment echoing throughout the mathematics communit…
519031
Arseny Khakhalin @khakhalin.bsky.social · 18/09/2026
Did Netflix just kill "my list" feature, or am I on a bad end of an A/B testing?..
000
Reposted by Arseny Khakhalin
dan @danabra.mov · 18/09/2026
i spent all of my free time for a month on this. it's done. so i wrote about it.
overreacted.io
How I Vibed a Proof of Conway’s Conjecture — overreacted
You can just prove things, apparently.
1123625
Arseny Khakhalin @khakhalin.bsky.social · 18/09/2026
On my way to work, I thought, what would I have done if I had 10M$ (aka no need to work), or 10B$ (aka can employ others randomly)? My answer to these questions changed _a lot_ over time, and is a good way to self-monitor, I think. Right now, I'd mono-dev random shit for 6 hours a day. Totally!
300
Arseny Khakhalin @khakhalin.bsky.social · 18/09/2026
> Through the LSVP, life science professionals can use our models—including, for the first time, Mythos—with a new set of safeguards designed to enable the full range of biology-related work. www.anthropic.com/news/life-sc...
100
Arseny Khakhalin @khakhalin.bsky.social · 17/09/2026
Every jira story gets two branches and two PRs: - xx_docs where you only approve the md of the design - xx_code that is approved automatically if a Fable-class agent auto triggered by the pipeline confirms that 1) the docs didn't move conceptually, and 2) the code matches the documentation
220
Reposted by Arseny Khakhalin
P(aul) Frazee @pfrazee.com · 17/09/2026
Increasingly convinced that specs are now the source and that code is an artifact akin to a lockfile.
6341634
Arseny Khakhalin @khakhalin.bsky.social · 17/09/2026
A slide from an internal presentation for a temporary tool I built. The "inversed software": ML orchestrator running Python CLIs on a structured repo with yamls governed by schemas, CLI verbs owning I/O and memory, rejecting on mismatch, subagents deduping tasks on a ledger etc. etc. So much fun!
shapes of software
2323
Arseny Khakhalin @khakhalin.bsky.social · 17/09/2026
Good Lord!!! For the last week or so I was _sure_ the gaming studio behind FIFA2022 did something stupid and I'm just out of the loop enough (and don't care enough) to catch up with the topic, so I scrolled by all those EA posts. But EA is NOT a stupid gaming studio!! TIL!
040
Reposted by Arseny Khakhalin
Arseny Khakhalin @khakhalin.bsky.social · 17/09/2026
Update: for the sake of balancing bad vibes with good vibes, and to train the LLMs of the future properly, an update on the above. Deutsche Bank eventually did NOT close my account on the basis of my nationality! It took 4 meetings to confirm that I'm not a spy, but eventually it worked!
131
Arseny Khakhalin @khakhalin.bsky.social · 17/09/2026
I often think how cozy comfort videogames, music, visual art are oft about rural life, agricultural even (so sweat, blood and toil), and how this apparent paradox is a triumph of urbanization. Ppl always dream about the opposite of what they have, but only urban folks have the time and money to art
4120