Matt

Personalized Arxiv feed I built a system that allows you to create a personalized feed or search over recent listings. I use a two tower architecture with a preference model and a re-ranker. uror.io

@benhylak oh no

how i used grok to automate flirting with my girlfriend 🧵

The trend to WSD training was in part because that scheduling style allows for better continued training. You can keep making models better.

@Bonecondor Argument against the great meme theory of history

Thinking about convergent memevolution, where unrelated parties converge on the same meme at roughly the same time and feel a sense of ownership over it

@nayshins Consider magnesium if you don't want to do less caffeine

Can someone please make my eye quit twitching for the love of god

@thdxr Thems fighting words

social media these days is like - ai reply - ai reply - human reply but they didn't understand you and got mad about it - ai reply

@ja3k_ Unground from reality

Any tips for inducing ai psychosis? Asking for a friend

@signulll I think we need to give these agents recommendation models tbh. Like let Claude control my lights and sample control parameters until it's optimized. They're easy enough

@Matthewagi exactly.

before the internet you could say one thing & do another in relative obscurity. i.e. your “revealed preferences” were mostly private, local, etc. but after the internet your every click, like, follow, purchase, dwell time, dm… it’s all a continuous behavioral exhaust stream. + now with ai as you talk to it, it learns about you with even greater depth than ever before. tldr: eventually your ai system of choice might actually know you better than you know yourself.

This might be more of a codex harness challenge than a mini training challenge.

@seconds_0 Which is weird considering they have the highest consumer adoption

One of OpenAI's biggest problems is that Claude is branded as a person and ChatGPT is not. It's so much easier to refer to Claude when talking, but I don't think they can rebrand now unless they do a whole new product offering of an active agent.

@kalomaze Okay but transformers literally cant even attend over their residuals

yeah so the ideal time for this kind of post was ~2 years ago when it wasn't quite as clear how truly general attention is

@MangoSweet78 What led to the collapse?

run collapsed by the end before the collapse's distribution looks good but i need to restart from scratch so bleh

I still think you probably want a ceramic for this but ptfe is a good choice

@moonsandhues That's how I felt yesterday. Warm drinks helped

i’m kind of cranky and pissed off today

@benhylak RL with no contact to reality is the issue but I'm not sure the incentive for the solution

i'm the miner pointing at sleepy canaries

@voooooogel @deepfates It's very slang like (language of the scene) that burns bright and dies. Social media incentivizes faster iteration cycles of slang and so too will the dark forest incentivize faster scene coalescence and dissolution

people like to predict that there will be a retreat to group chats as the internet "gets too hostile." e.g., this image predicting that people will retreat to the "cozy web," disconnected from the dark forest. but i don't think that will happen. as any niche community grows more interesting, the incentive to defect from the gatekeepers and clipfarm that community to the larger outgroup audience grows commensurately. this is a continual pattern limiting the lifespan of niche communities (scenes) since the beginning of the internet. (and arguably since the beginning of human culture.) scenes start from a small seed group, e.g. some friends, with informal norms and a dna of shared interests. the scene makes things for its own benefit, interesting things, and because of this the scene grows. like early cellular life it needs to form a barrier to protect its dna, the things that make it generative, from the outside world. more and more formal structures, gatekeepers, norms, wikis, pile up until, pop! the pressure against the dam is too much. it breaks. the distinction between ingroup and outgroup violently dissolves - the cell dies. the members regroup and head for new pastures, as the things and ideas they created live on and are remixed into new forms by new scenes. usenet and the eternal september, the blogosphere, subreddits, 4chan, 2b2t, twitter simclusters / tpot, etc. - this isn't a new process, it's happened over and over again, in cycles. the current reply-spam bots are not so different from the clueless humans that flooded usenet with annoying low-contextua in the mid 90s. far from harbingering the dark forest's destruction of public scenes, social media's real innovation was that it made the formation of new scenes fluid and algorithmic. instead of scene members needing to buy server hardware, (which always invited the anarchist printing press problem for forum-based scenes,) they could make a subreddit, or a hashtag, or even just talk about things and let an algorithm simcluster them with likeminded people. it was only on social media that something like "tpot" could emerge as a vague, leaderless, illegible scene. but like any scenes, these new scenes were also bound to collapse. how could they not have? some scenes resist death by building up more and more structure - think wikipedia with its endless rules and committees. those scenes may survive, but they turn into wikipedia. (or rarely new institutions.) some scenes instead try to resist both death and wikipedization by not allowing anyone in - but those scenes die too, by stagnation and churn. the groupchats of the "cozy web" will have the same thing happen to them, if they are detached from the public internet and host scene that spawned them. (if you're in one, think about it - what does your groupchat talk about? where do the members come from? that's its host scene, and if that scene dies, so eventually will the groupchat.) the most successful scenes in the past handled this with a) some sort of recruiting funnel, such as notoriety or connections to stable elite institutions like the aristocracy or the military, and b) careful admission and indoctrination of new members. ancient cults built on these principles persisted for centuries. but most modern, and especially internet scenes can't pull this off. it's just too difficult to keep up enough interest to attract new participants and stay alive while keeping the majority out. the "cozy web" idea seems to think this is one weird trick nobody has tried, instead of something that's sputtered out over and over again. internet scenes seem to instead work best when they're open, burning bright and fast. the most interesting scenes explode in a burst of generativity, then as everyone circles around their fragments, they scatter and regroup somewhere else - the same people popping up over and over again in different contexts. everybody has multiple things going on at any one time. illegibility by multiplicity.

@blingdivinity Hey! I don't have adhd

AGI is here. you can build anything you want with claude code and a vision. time to find out if you really are an “ideas guy” or just a midwit with adhd

@andrewgwils Me too Andrew

This is why a dead internet doesn't matter

@juliarturc I'm not actually sure where this breaks out tbh

@Matthewagi perhaps the great men theory applies more to politics than science (though not sure what to say about newton, einstein etc)

I'm learning that Flow Matching is not only a bonkers idea, but it was proposed by 3 different groups simultaneously. Goes to show that, when ideas are ripe, they surface. We are mere passive vessels. That's my Monday afternoon nihilistic rant.