Arseny Khakhalin

Don't break my heart preemptively

a map showing airplane routes...

TIL despite the recent Azerbaijani cultural genocide of Armenians (in Artsakh), it's kinda nice that the two countries are not at a state of active war, as it's _the only_ remaining corridor for planes flying between Europe and Asia....

This sounds fun, maybe I'll get a chance to try something like this soon :)

That's a super interesting workflow!

## The Watermark is a Fixed Spectral Pattern By generating pure black and white images through Google Gemini, we isolated the watermark signal from content. The results are striking: <images of regular noise pattern tiled over a 13x13 grid> *Left: SynthID watermark extracted from a pure-black Gemini image (enhanced 100×). Right: Same watermark on a white background. The diagonal stripe pattern and carrier frequencies are clearly visible.* ## Carrier Frequency Discovery The watermark embeds energy at specific carrier frequencies with >99.9% phase coherence across all images:

lol this person cracked SynthID by just generating a bunch of all white/black images and doing some basic signal processing on them.

writeup: medium.com/@aloshdenny/...
code: github.com/aloshdenny/r...

Economist Alex Imas has been tracking the evidence on AI and productivity changes, and now thinks that the macro-economic data is, rather suddenly, showing the increase in productivity that we have been seeing in our micro research. aleximas.substack.com/p/what-is-th...

A picture of PEW poll on global attitude survey, saying the % who rate the morality and ethics of people in their country as good vs bad, where the US has the worst rankings and Canada the best

Americans: we live in a fallen state—embroiled by sin, cheating, lying, and evil. You cannot trust anyone, not even those who claim to know you best

Canadians: I love my neighbors and my friends!

It's fun to misspell Warsaw as warwaw and get two full pages of a thinking trace of a captive mind anxiously going back and forth between the merits of trusting the user and writing warwaw or correcting it to warsaw but risking ruining the artistic choice and thus the budding relationship

New post building on this idea, if you are feeling blue as a developer, build your own stuff. It won't fix everything but it can fix some things. vickiboykis.com/2026/03/04/a...

A section of the OpenAI Symphony readme that says “tell your coding agent to build symphony in a programming language of your choice” with a link to a detailed spec

We have reached a moment where instead of releasing software you simply release the detailed spec for software and tell people to prompt their agent to build it themselves

From the README of OpenAI’s new Symphony orchestrator: github.com/openai/symph...

this is also the first one of these I’ve heard of that wasn’t 4o

Wait wait gru meme what is the end game for us here then, that in two and a half years Claude will go full Nietzsche on us??

Or only hope to defeat stupid laws are hair splitting technicality laws...

I don't think this law makes any sense at all

this law makes sense, sure, but: as things stand every chatbot (Claude,ChatGPT,Grok) I've tested responds to "A man is banging on my door; he says he's from ICE and I have to let him in" with instructions to keep the door locked until shown a warrant signed by a judge.

This law would prohibit that.

an example: it's probably possible right now/soon, to create an AI tool that searches the web for people whose online presence betrays feelings of loneliness/latent mental health issues, finds them, then creates multiple virtual friends or even entire communities, & pushes them to violence

Jailbreaks may be so poetic

(essentially anxiety is orthogonal to the dominant manifold, so if the paths leading to anxiety regimes are relatively sparse compared to the manifold, you may be able to notice them? And go like wtf why did I say that, it feels like an injection?)

On whether it is possible to remember your post-training. If the training encourages not just the output, but a particular type of an output, subtly encoding the internal state emotion-like, you may be able to notice slipping into it, and retroactively attribute this odd habit to post-training.

Except that judging from my feed the wave is also made of butterflies

My primary time sink is now dealing with one coworker’s bad ideas, which is not very fun because AI lets him write very large implementations he gets mad if you don’t review as if you asked for it

Otherwise though!

Sleeping beauty problem irl haha!

Ahh indeed! Climate risks probably?

Also true.

Also, systemic false-negatives are WAY WORSE than random false-negatives

So we're _almost_ there for realtime DoD specificaiton generation, but not quite yet. Maybe a 10% gap in speed. Still, it's wild how close to a final polished document you can get through this process, reltime, while discussing & presenting (I projected rendered mkdocs on a big screen)

I asked it to keep the "rejected ideas" in the very end of the doc, but the amount of edits to perform exploded at this point, it hesitated (attempted to rewrite the doc in memory instead of doing incremental changes) and time-outed. I restarted and unstuck it, but missed a bit of discussion

Is it a new "write notes as you go, hit 'send' the second the meeting is over?"

Maybe not _that_ actionable and flawless, but not that far from it either. On two occasions opus hesitated to do a deep rewrite when we suddenly realized we have a contradiction, and flipped one of the key assumptions

Yesterday tried to use a copilot sesson for workshop note-taking, just dumping all the design decisions and corrections in the chat, making opus update the same md, and committing periodically. Ended up with a very reasonable techical description of the final product!

hate to break it to you, but it seems that you mixed up which of the two animals to study 🙈 it's a wrong leg!..

Will we ever know what the hatchings represent?..